To Helena and Noah

CONTENTS

Introduction

PART I
THE EARLY YEARS
(PRE-1993)
1 Pain and Suffering
2 The Graphics Revolution
3 The Birth of Nvidia

PART II
NEAR-DEATH EXPERIENCES
(1993–2003)
4 All In
5 Ultra-Aggressive
6 Just Go Win
7 GeForce and the Innovator’s Dilemma
 PART III
NVIDIA RISING
(2002–2013)
8 The Era of the GPU
9 Tortured into Greatness
10 The Engineer’s Mind

PART IV
INTO THE FUTURE
(2013–PRESENT)
11 The Road to AI
12 The “Most Feared” Hedge Fund
13 Lighting the Future
14 The Big Bang

Conclusion: The Nvidia Way

Appendix: Jensen-isms
Acknowledgments
Notes
THE NVIDIA WAY
 INTRODUCTION

IN ANOTHER LIFE, JENSEN HUANG might have been a teacher. His

preferred medium is the whiteboard: at many of the meetings he attends at
Nvidia, where he has been the CEO since he cofounded the company in
1993, he will leap up, his favorite chisel-tip, dry-erase marker in hand, and
diagram a problem or sketch out an idea—even if someone else is speaking
or whiteboarding themselves at the same time. In fact, he alternates between
teacher and student, fostering a collaborative ethos among his employees to
develop their thinking and solve the issues they face. His sketches are so
precise that they could be turned into usable schematics for technical
documents; colleagues call him “Professor Jensen” for his ability to explain
complicated concepts on the whiteboard in a way just about anyone can
understand.
At Nvidia, the whiteboard is more than the primary form of
communication at meetings. It represents both possibility and ephemerality
—the belief that a successful idea, no matter how brilliant, must eventually
be erased, and a new one must take its place. Every conference room in the
company’s two headquarters buildings in Santa Clara, California, has a
whiteboard, signaling that each day and each meeting is a new opportunity,
and that innovation is a necessity, not an option. Whiteboarding also
requires active thinking and inevitably reveals how well (or not) any
employee, including an executive, knows the material. Employees must
demonstrate their thought process in real time, in front of an audience;
there’s no hiding behind neatly formatted slides or slick marketing videos.
 The whiteboard is perhaps the ultimate symbol of the unique culture at
Nvidia—the microchip designer that has grown from humble beginnings in
the 1990s, when it was one of dozens of computer graphics chip companies
and known mainly by hardcore gamers seeking the best performance for the
first-person shooter Quake and other games, into the premiere supplier of
advanced processors for our age of artificial intelligence (AI). The
architecture of the company’s processors is well suited for AI workloads
because of its ability to perform mathematical calculations simultaneously
—essential for training and running advanced large-language AI models.
Nvidia’s early recognition of the significance of AI and its forward-thinking
investments over the course of more than a decade—including enhancement
of hardware capabilities, development of AI software tools, and
optimization of networking performance—made the company’s technology
platform perfectly positioned to capitalize on and become the primary
beneficiary of the current AI wave. AI use cases are now plentiful.
Companies are harnessing Nvidia-powered AI servers to boost programmer
productivity by generating lower-level code that developers find tedious to
write, automating repetitive customer-service tasks, and empowering
designers to create and alter images on the basis of text prompts, enabling
the faster iteration of ideas.
Nvidia’s reinvention has paid off: on June 18, 2024, the company
surpassed Microsoft to become the world’s most valuable company, at a
market capitalization of $3.3 trillion. It reached that milestone on the back
of the immense demand for its AI chips; the company’s stock price had
tripled over the previous twelve months. To call Nvidia’s stock a
historically good investment is to understate things. Between its IPO in
early 1999 and the end of 2023, Nvidia investors enjoyed the highest
annualized compound return for any U.S. stock in history through a
compound annual growth rate (CAGR) of more than 33 percent.1 If an
investor had purchased $10,000 of Nvidia stock when it made its market
debut on January 22, 1999, that stock would have been worth $13.2 million
on December 31, 2023.
Nvidia’s culture starts with Jen-Hsun Huang, who is known simply as
“Jensen” to his friends, employees, suppliers, competitors, investors, and
admirers. (This is how I refer to him throughout the book.) He had already
found a modicum of fame before the AI boom, having been named to Time
magazine’s list of the one hundred most influential people in the world in
2021. But as Nvidia’s value reached $1 trillion, then $2 trillion and $3
trillion, his profile has grown commensurately. It is now common to see his
trademark leather jacket and shock of silver hair, always combed into a
simple side part, in articles and video clips, many of which described
Jensen as the “genius you’ve never heard of.”
To those of us who have covered the semiconductor industry, Jensen has
been a known quantity for some time. He has run Nvidia for its entire three-
decade history, the longest tenure of any current technology CEO. He has
driven the company not only to survive, but to surpass all of its competitors
in the unforgiving and volatile chip sector, and to surpass just about every
other company on Earth, as well. I have been following Nvidia in a
professional capacity for much of my career—first as an equity analyst and
now as a journalist—and I have seen how his guidance and strategic vision
have shaped the company over the years. Even so, my view has remained
that of the outside observer, dependent as much on interpretation as on
concrete fact. To learn the secrets of Nvidia’s success, I would have to talk
with many people, inside and outside the company. I would also have to
speak with Jensen himself: to become his student, much like his employees
do.

I GOT MY CHANCE JUST four days before Nvidia became the most
valuable company in the world. Nvidia knew I was writing a book, and, in
early June 2024, a representative offered to arrange a meeting with Jensen
to occur immediately after his commencement address to the California
Institute of Technology’s graduating class of 2024. I agreed, and a few
minutes before 10:00 a.m. on Friday, June 14, I found myself in front of a
stage, waiting for Jensen to appear. It was a perfect California day, clear
blue sky and warm sunshine. The students and their families took their seats
under a large white tent. David Thompson, the chair of Caltech’s Board of
Trustees, introduced Jensen and joked that the Nvidia CEO had drawn so
much attention earlier in the day, as the two made their way around campus,
that it felt like he had been walking alongside Elvis.
During his speech, Jensen suggested to the students that their graduation
from Caltech would mark one of the peaks of their lives. He mentioned that
he knew something about peaks, too. “We are both at the peaks of our
career,” he said. “For all of you who have been paying attention to Nvidia
and myself, you know what I mean. It’s just that in your case, you’ll have
many, many more peaks to go. I just hope that today is not my peak. Not the
peak.” He vowed to work as hard as ever to ensure that there would be even
more peaks ahead for Nvidia, implying that the new graduates should
follow his example.
After Jensen concluded, I was whisked over to the Keck Center for Space
Studies and led into a wood-paneled conference room with black-and-white
photos of pilots, astronauts, and presidents on the walls, where he was
waiting for me. We chatted a bit before I got to my prepared questions. I
explained that I was a PC gaming nerd who had been building my own
computers since the 1990s. I first encountered Nvidia while researching
graphics cards and invariably chose their products. And I mentioned that
earlier in my career, at a Wall Street fund, an investment in Nvidia became
my first big winner.
“Good for you,” Jensen deadpanned. “Nvidia is my first big winner too.”
We launched into a wide-ranging discussion of the company’s history.
Jensen knows that many of his former employees look back on Nvidia’s
beginnings with nostalgia. But he resists overly positive accounts of
Nvidia’s start-up period—and his own missteps.
“When we were younger, Tae, we sucked at a lot of things. Nvidia wasn’t
a great company on day one. We made it great over thirty-one years. It
didn’t come out great,” he said. “You didn’t build NV1 because you were
great. You didn’t build NV2 because you were great,” he said, referring to
the company’s first two chip designs, both of which were flops that nearly
killed Nvidia. “We survived ourselves. We were our own worst enemy.”
There were several more near-death experiences. But each time, amid the
stress and the pressure, the company learned from its mistakes. It retained a
core of die-hard employees, many of whom remain in the fold to this day.
Of course, there were also people who didn’t stay, requiring the company to
integrate new hires. “Every single time, somebody left, and every single
time we picked ourselves up. We healed the company into existence,” he
said.
He slipped into the third person. “If Jensen wasn’t even involved in the
first fifteen years of our company, I would really like that,” he laughed,
meaning he wasn’t proud of how the company was managed then or of his
own naïveté and lack of strategic thinking.
 I was put in the unusual position of defending Nvidia’s past to its
founder. I pointed out that the early decisions—about which I had learned a
great deal by this point in my research process—weren’t all bad ones.
While mistakes were made, some of the misfires were tied to factors that
were unpredictable or out of his or the company’s control. With hindsight,
many of them appeared to have been unavoidable.
“Yeah, that’s fine,” said Jensen. “I don’t love talking about our past.”

I FOUND THIS TO BE A pervasive attitude within Nvidia: that the

culture of the place discourages looking back, whether at errors or
successes, in favor of focusing on the future—the blank whiteboard of
opportunity. But you cannot understand what Nvidia is today without
understanding how it got to this point. This book is the first to tell the story
of Nvidia—the full story, not just Jensen Huang’s, although he is at the
center of everything. It explores the founding of Nvidia by Jensen, Curtis
Priem, and Chris Malachowsky, in a back booth at a Denny’s all the way
back in 1993—many lifetimes ago, to anyone working in technology.
Nvidia would never have come into existence without the contributions of
all three men. Jensen’s business acumen and hard-driving management style
were critical to Nvidia’s early success, but Priem’s chip architecture
prowess and Malachowsky’s manufacturing expertise were essential, too.
It is a thirty-year story, so far, and to tell it I interviewed more than one
hundred people. Many of them are current or former Nvidia employees and
know the company’s inner workings intimately—this group includes
Jensen, his two cofounders, and most of the early and current senior
management team. Others include the two original venture capitalists who
invested in Nvidia; technology industry CEOs; the partners who have
helped Nvidia manufacture and sell its chips; and those at other
semiconductor companies who competed with, and almost invariably lost
to, Nvidia.
Through these interviews, I began to understand what makes Nvidia
special. Its defining characteristic is not its technological prowess, which is
more a consequence than a root cause. It is not the financial resources and
the new opportunities that come from a high market valuation. It is not a
mystical ability to see the future. It is not luck. Rather, it is a unique
organizational design and work culture that I have come to call “the Nvidia
Way.” This culture combines unusual independence for each employee with
the highest possible standards; it encourages maximum speed while
demanding maximum quality; it allows Jensen to act as strategist and
enforcer with a direct line of sight to everyone and everything at the
company. Above all, it demands an almost superhuman level of effort and
mental resilience from everyone. It’s not just that working at Nvidia is
intense, though it certainly is; it’s that Jensen’s management style is unlike
anything else in corporate America.
Jensen runs the company in the way he does because he believes that
Nvidia’s worst enemy is not the competition, but itself—more specifically,
the complacency that grips any successful company, particularly one with a
long and impressive track record such as Nvidia. In my work as a journalist,
I have seen that companies tend to become dysfunctional as they succeed
and grow, largely because of internal politics, with employees focused not
on driving innovation or serving customers but on advancing their bosses’
careers. This jockeying distracts them from doing their best work and leaves
them constantly looking over their shoulder for the threat from the next
office over. It is something that Jensen has structured Nvidia to eliminate.
“Over the years, I realized what was happening, how people protect their
turf and they protect their ideas. I created a much flatter organization,”
Jensen said. His antidote to the backstabbing, to the gaming of metrics, and
to political infighting is public accountability and, if needed, public
embarrassment. “If we have leaders who are not fighting for other people to
be successful and [who are] depriving opportunities to others, I’ll just say it
out loud,” he said. “I’ve got no trouble calling people out. You do that once
or twice, nobody’s going to go near that again.”
Nvidia’s unique culture might sound strange or unusually grueling, even
for the tech industry. But among all the former Nvidia employees I talked
to, it was hard to find a dissenter. They all reported that the company was
largely free from the internal politics and indecisiveness typical in large
organizations. They mentioned how difficult it was to adjust to working at
other companies where direct, blunt communication is rare and there’s far
less urgency to get things done. And they described how Nvidia not only
empowered them, but also required them to fulfill their professional calling,
as a necessary condition of employment.
In a sense, that is the Nvidia Way in its purest form. It is the unwavering
belief that there is tremendous reward in doing your job the best you can. It
is the drive to persevere amid adversity. Or, as Jensen put it when looking
directly into my eyes: the secret to his company’s success is nothing more
than “sheer will.”

MORE PRECISELY, IT IS JENSEN’S personal will that has shaped

Nvidia. He has personally made the most consequential decisions in its
history. His ability to place the right major bets on emerging technologies
stems from his deep technical knowledge—a founder with an engineering
background. I have tried in this book to distill the Nvidia Way down to a set
of principles that anyone can learn from, if not use. But behind them all
lurks a question: Can you really separate Nvidia from its CEO?
As of this writing, Jensen is sixty-one years old. He has run the company
for thirty-one years—more than half his life. Nvidia is bigger, more
profitable, and more crucial to the global economy than ever before. Yet it
still relies on Jensen, as a business leader and a tone setter. Apple survived
the ouster of Steve Jobs in 1985 and his death in 2011; Amazon, Microsoft,
and Google all did well after Jeff Bezos, Bill Gates, and Larry Page and
Sergey Brin wanted to move on. Someday, Nvidia will have to face a
similar transition. It is not quite clear what the company might look like
post-Jensen—whether its culture will survive, whether it will maintain its
momentum.
After all, a whiteboard is only as useful as the person who holds the
marker. It can reflect genius, but it cannot create it.
 PART I

THE EARLY YEARS

(PRE-1993)
 CHAPTER 1

Pain and
Suffering

WHEN JENSEN HUANG WAS FOUR years old, his father visited New
York City and fell in love with America. From that point on, his parents had
one goal: to find a way to raise him and his older brother in the land of
opportunity.
It would not be an easy task. Jensen was born in Taiwan on February 17,
1963, to Taiwanese parents. They were not wealthy and moved around
based on the needs of his father’s work. They ended up in Thailand for an
extended period. Jensen’s mother taught her two boys English, each day
selecting ten words at random in the dictionary and asking them to spell the
words and memorize their definitions.1
After a wave of political unrest hit Thailand, Jensen’s parents decided to
send him and his brother to Tacoma, Washington, to live with their aunt and
uncle. Tacoma was once called the “City of Destiny,” because it sat at the
end of the Northern Pacific Railroad, but by the 1970s it was about as far
from the dynamism of New York City as it was possible to get: damp,
dreary, and smelling of sulfur thanks to the paper pulping and processing
plants that lay on the outskirts of town. Jensen’s aunt and uncle were recent
immigrants to the United States themselves and did the best they could to
help their nephews adjust to their new country while they waited for
Jensen’s parents to follow them across the Pacific.
 The two boys were difficult to handle. “We could never sit still,” Jensen
said. “We were eating all the candy in the cupboard, jumping off the roof,
climbing out the windows, tracking mud into the house, forgetting to close
the shower curtain, and flooding the bathroom floor.”2
Even though they had not yet made the move to the United States
themselves, his parents wanted to send their children to an American
boarding school so they could get a good education. They found one called
Oneida Baptist Institute, which was located in eastern Kentucky and
accepted international students. They could afford the tuition only by
selling nearly all of their possessions.
Jensen remembers the initial drive through the mountains of Kentucky,
past the single building that was the town of Oneida’s only gas station,
grocery store, and post office all at once. The boarding school had around
three hundred students, evenly split between boys and girls. But it was not a
prep school as Jensen’s family originally thought. Oneida Baptist Institute
was, instead, a reform school for troubled young people. It had been
founded in the 1890s to remove children from feuding families in the state
and thus keep them from killing each other.
As befit its original purpose, the school held its students to a strict
routine. Each morning, Jensen would cross the Red Bird River on a worn-
out swing bridge to attend his classes. He joined the swim team, played
soccer, and discovered new foods such as Jell-O, sausage, biscuits, and
gravy. He went to church twice a week and watched ABC’s Sunday Night
Movie on the weekends. Some evenings, he would play chess with the
school custodian. On others, he would help him fill the vending machines
and get a free soda in return. The occasional trip to town gave him an
opportunity to buy fudgesicles at the grocery store; otherwise, he would
content himself eating apples off of the tree outside his dorm room window.
Above all, there were the chores. Every student was required to work
every day. Already strong enough for extended manual labor, Jensen’s
brother was assigned to work at a nearby tobacco farm. For his part, Jensen
was on janitorial duty for his three-story dorm. “I had to clean the
bathrooms,” he said. “You can’t unsee that kind of stuff.”3
Jensen’s relative youth, and likely his different ethnicity as well, made
him a target of bullies. Even though the school was ostensibly designed to
reform its pupils, in practice oversight could be lax, and Jensen was beaten
up often in his first months on campus. Even his roommate was
intimidating: he was eight years older than Jensen, and his whole body was
covered in tattoos and scars from stab wounds. Eventually, Jensen learned
to overcome his fear. He befriended his roommate and taught him how to
read, and in exchange the other boy introduced Jensen to weightlifting.
Jensen took to it, and it gave him not only strength but confidence—the
ability and desire to stand up for himself.
Later in his life, Jensen’s executives would say he had developed his
tough, street-fighter mentality during his days in Kentucky. “Maybe this is a
bit of my early schooling, I will never start a fight, but I will never walk
away from one. So, if someone is going to pick on me, they’d better think
twice,” Jensen himself said.4
After a few years, Jensen’s parents moved from Thailand to Beaverton,
Oregon, a city on the outskirts of the Portland metropolitan area. They
withdrew the boys from the “boarding school” in Kentucky and enrolled
them in public school. Although Jensen was happy to be back with his
parents, he looked back on his time at Oneida Baptist as formative.
“I don’t get scared often. I don’t worry about going places I haven’t gone
before. I can tolerate a lot of discomfort.”5

ON THE FOURTH FLOOR OF the Elks Club building in downtown

Portland, inside an ornate ballroom with chandeliers and carved ceilings, a
man named Lou Bochenski had started a table-tennis club called the Paddle
Palace. It was open every day from 10:00 a.m. to 10:00 p.m. and had a
thriving juniors program for young enthusiasts. After school, Jensen would
often end up at the Paddle Palace, where he discovered both a talent and a
passion for the sport. He would also once again find himself doing janitorial
work, now to make some extra money—Bochenski paid him to scrub the
Paddle Palace’s floors.
This was not just charity on Bochenski’s part. His daughter, Judy
Hoarfrost, was a member of the “ping-pong diplomacy” team that visited
China in 1971. In fact, Hoarfrost and her eight teammates were the first
group of Americans to make a state-sponsored visit to China since the
Communist Revolution in 1949. Although they lost most of their matches,
their trip signaled a thawing in U.S.-Chinese relations—and helped increase
the profile of table tennis in the United States. Bochenski considered it his
duty to help discover promising young table-tennis players and develop
them into national-level talents.
Both Hoarfrost and Bochenski were impressed by Jensen’s skills and
work ethic,6 so much so that in 1978, Bochenski wrote a letter to Sports
Illustrated magazine that praised Jensen as the “most promising junior”
ever to emerge in the Pacific Northwest. He pointedly said that, unlike other
teenagers written up in the magazine whose families spent $10,000 annually
on traveling to tournaments, Jensen earned his own travel money.
“He is a straight-A student and very hungry to become a table-tennis
champion. He has played only three months, but I suggest you watch out for
him in another year,” Bochenski wrote.7 At the time, Jensen was only
fourteen.
At one point, he went to Las Vegas for a national table-tennis tournament.
But the lights and sounds of the city were too alluring for him. Instead of
resting up before his matches, he stayed up all night walking up and down
the Strip. He lost badly—and never forgot the sting of his own failure.
“When you’re thirteen or fourteen years old and you go to Las Vegas for
the first time, it’s hard to focus on the match,” he said three decades later.8
“To this day, I regret not being more focused on the tournament.”
When he was fifteen, he entered the U.S. Open Junior Doubles
tournament. This time, he knew better than to let himself get distracted and
came in third place overall.

JENSEN WAS ALWAYS A GOOD student. Learning how to interact

socially with other people, though, was more challenging.
“I was very introverted. I was incredibly shy,” he said. “The one
experience that pulled me out of my shell was waiting tables at Denny’s.”
When Jensen was fifteen, his brother helped him get a job at a Denny’s in
Portland. He would work at the twenty-four-hour diner for several summers
in high school and college. Jensen started where he always had, doing the
dirty work of washing dishes and cleaning bathrooms. “I did more
bathrooms than any CEO in the history of CEOs,” he would recall.9 He then
became a busboy and later waited tables.
He believed that Denny’s equipped him with a number of significant life
skills, including how to navigate chaos, work under time pressure,
communicate with customers, and handle mistakes (in this case, from the
kitchen). It also taught him to find satisfaction in the quality of his work, no
matter how minor the task, and to fulfill each according to the highest
possible standards. It didn’t matter if he was cleaning the same toilet for the
hundredth time or interacting with a new customer who had never been to a
Denny’s before and didn’t know what to order. He recalls pushing himself
to do the best he could, even if it meant chasing an absurd goal such as
being able to carry more cups of coffee at a time than anyone else on staff.
He learned to take pride in daily toil.
“I’m certain I was the best dishwasher, busboy, and waiter they’ve ever
had,” he said.
Except when it came to one common order. “I hated shakes because I
hated making them,” he said; it took a long time to prepare a single shake
and an even longer time to clean up afterward. He would try to nudge
customers to order Coke instead, and if they persisted, he would ask, “Are
you sure?”10 Already, he was learning another fact of working life: the
trade-off between having high standards and being efficient with one’s time.

JENSEN ATTENDED ALOHA HIGH SCHOOL in Beaverton, Oregon,

where he made friends in the math, computer, and science clubs. He spent
all his free moments programming BASIC on the Apple II and playing
games on teletype terminals, which looked like electric typewriters, and
which were connected to a larger mainframe computer.
He “fell in love” with video games, in particular the Star Trek mainframe
game, which was based on the classic Hasbro board game Battleship.11 He
also spent a good deal of time playing Atari and Konami games at the
arcade, including Asteroids, Centipede, and Galaxian.12 He didn’t have a
computer at home, so he had to go elsewhere to get his gaming fix. “We had
no money,” he said.13
The precocious Jensen had skipped a grade at his elementary school in
Thailand and again at Oneida Baptist Institute in Kentucky. He graduated
from Aloha High School at the age of sixteen and decided to attend Oregon
State University in Corvallis, both because of the low in-state tuition and
because his best friend, Dean Verheiden, was going there too. Together,
Jensen and Verheiden chose electrical engineering as their major and took
many of the same classes. Hoping to gain relevant work experience, Jensen
applied repeatedly for an internship at a local technology company called
Techtronic Industries but was rejected every time.
During his sophomore year, Jensen met Lori Mills, one of only three girls
in an electrical engineering class of two hundred fifty students. “I was the
youngest kid in the class. I was tiny. I was skinny. But I had a great pickup
line,” Jensen said, who by this point had grown out of his awkward phase
and had sharpened his social skills. “Do you want to see my homework?”14
The line worked. He and Mills began dating, and they got married shortly
after they both graduated in 1984. Jensen was invited to interview with
some of the largest semiconductor and chip makers in the country. He first
had his eye on Texas Instruments, whose offices spanned multiple zip
codes, but his interview went poorly, and he didn’t get an offer. He next
interviewed with two companies based in California. The first was
Advanced Micro Devices, or AMD, a company that Jensen had idolized
ever since he saw a poster of one of their microprocessors at Oregon State.
The second was LSI Logic, which made customizable microchips called
application-specific integrated circuits, or ASICs, for technical and
scientific uses.
He received offers from both companies and chose AMD, because he was
more familiar with its reputation. By day, he designed microchips; at night
and on the weekends, he took courses at Stanford so that he could get his
master’s in electrical engineering. On top of his work and his continuing
studies, he and Lori had a son, Spencer, and daughter, Madison. Because he
couldn’t take many classes at once, completing the master’s degree was a
long, arduous process; he finally finished after eight years. “I have a very
long-term horizon,” he said. “I can be impatient about certain things, but
infinitely patient about others. I plug away.”15
Between his job, his master’s, and his family, Jensen had achieved the
dream of so many immigrant parents, who make huge sacrifices to move to
the United States in order to give their children the chance for a better life.
“My father’s dream and my mom’s aspirations for our success are what
ultimately put us here,” Jensen said nearly thirty years later, when asked to
reflect on his past. “I owe them a great deal.”16
Yet Jensen’s well of ambition ran deeper still. The drive to do every job
to perfection and, at the same time, as efficiently as it could possibly be
done now led him to question his own work designing microprocessors.
Although he was good at designing microchips for AMD, he found it
tedious; at the time, it was still done manually, by hand.
One of his office mates had left for LSI and wanted Jensen to come with
him to the company. And Jensen, like most people in the chip-making
industry, had heard that LSI was now pioneering new software tools that
promised to make the process of chip design much faster and easier. The
idea intrigued him. Although he knew it would mean taking a risk, he felt a
need to work at a company that, to him, seemed to have a clear grasp on the
future of the chip industry. It was an early sign of his restless, forward-
thinking nature, which would lead him to pursue the cutting edge even if it
meant leaving safety and security behind.
He took the plunge and joined LSI. There, he was given a technical role
working with customers. He was assigned to a start-up called Sun
Microsystems, where he met two engineers, Curtis Priem and Chris
Malachowsky, who were working on a secret project that promised to
revolutionize how people used workstation computers—high-performance
computers built to perform specialized technical or scientific tasks, such as
three-dimensional modeling or industrial design.
Luck had clearly played a part in bringing Jensen to this new opportunity.
So had his own talent and skills. But as he saw it, the biggest single factor
that propelled him from scrubbing toilets to managing entire divisions of a
microchip company was his willingness, and ability, to put in more effort,
and tolerate more suffering, than anyone else.
“People with very high expectations have very low resilience.
Unfortunately, resilience matters in success,” he later said. “Greatness is not
intelligence. Greatness comes from character.”17 And character, in his view,
can only be the result of overcoming setbacks and adversity. To Jensen, the
struggle to persevere in the face of bad, and often overwhelming, odds is
simply what work is.
It is why, whenever someone asks him for advice on how to achieve
success, his answer has been consistent over the years: “I wish upon you
ample doses of pain and suffering.”
 CHAPTER 2

The Graphics
Revolution

AS A TEENAGER, CURTIS PRIEM taught himself how to program by

writing games in the computer lab of his high school in Fairview Park,
Ohio, just outside of Cleveland. The school had a Teletype Model 33 ASR
Coupler terminal, which connected to a mainframe computer located about
ten miles away and transmitted data over a phone line at a speed of about 10
characters per second. He wrote in BASIC, transferred his instructions onto
punched paper tape, and fed the tape into the Teletype’s tape reader in order
to run his programs remotely on the mainframe.
Priem’s most ambitious project was a billiards game. The program would
show a layout of balls on a billiards table by using text characters, and
players would take turns specifying the angle and speed at which they
would strike the cue ball. The mainframe would then calculate the collisions
and the resulting positions of the billiard balls. The program was massive;
its punched tape roll measured nearly nine inches in diameter and took
almost an hour to print every time Priem wrote a new version of the game.
When he submitted it as an entry in a local science fair, he won first prize.
Priem’s programming exploits brought him to the attention of Elmer
Kress, the chair of Fairview Park’s math department. Kress became Priem’s
mentor and allowed him to access the school’s lone mainframe terminal as
often as he wanted, once the other students had their chance to use it for
schoolwork. As Priem became more proficient at programming, he learned
how to digitize images by hand by use of a monochrome wheel, and he
wrote a program that could manipulate those digitized images on the
computer. Priem’s journey into computer graphics began with the simple act
of scaling and rotating a digitized photo of Kress.

WHEN CONSIDERING COLLEGES, PRIEM FOCUSED on three

schools: the Massachusetts Institute of Technology, Case Western Reserve
University, and Rensselaer Polytechnic Institute (RPI). Two factors led him
to favor the last: at RPI, professors, not teaching assistants, taught freshman
classes, and the school had recently announced that it would be acquiring an
advanced IBM 3033 mainframe computer, which would be made accessible
even to incoming freshmen. Although Priem was accepted at all three
schools, there was never any question where he would end up once he heard
about the new IBM.
At RPI, Priem immersed himself in computers. He hand-built his own
multibus computer, linking an Intel 8080 processor to two eight-inch floppy
disk drives and a monitor. And, of course, he spent plenty of time with the
university’s IBM 3033—a room-sized mainframe that was housed inside
RPI’s Voorhees Computer Center and which generated enough heat to warm
the entire building during the winter.
But Priem’s trajectory seemed to change during his sophomore year, after
his father lost his job. Without a steady income, his parents could no longer
afford to pay for his education. They asked RPI for assistance, but the
college didn’t offer any direct aid beyond a job at a campus engineering lab;
the wages Priem would earn there wouldn’t be nearly enough to cover
tuition. To fund his last two years at RPI, Priem enrolled in a work-
experience program sponsored by General Motors, which sought to fast-
track promising engineers into management positions. Each summer, Priem
and his GM Scholar cohort would work on a number of projects at various
assembly plants. During one stint, Priem programmed the machines that
produced compression-molded body panels for the Pontiac Fiero.
When Priem received his degree in electrical engineering in 1982,
General Motors offered him a full scholarship to continue his graduate
studies, provided that he take a job with the company afterward. RPI also
invited him to continue on as a graduate-level researcher in graphics.
 Priem had other ideas. Two years earlier, a pair of entrepreneurs in
California named Steve Jobs and Steve Wozniak had led their personal
computer start-up to a blockbuster IPO—and had earned themselves a more
than $100 million payday each in the process. Apple’s revenues had reached
nearly $300 million on the sales of the Apple II computer, making it the
fastest-growing company in history. The Apple II proved that there was a
huge market for personal computers that were smaller, cheaper, and better,
for both productivity and entertainment, than mainframes and
minicomputers. The emergence of the personal computer gave engineers
such as Priem opportunities not only to do what they loved, which was to
make cutting-edge graphics chips, but to do it in a setting that could earn
them a huge payday.
Priem decided to accept a job offer from Vermont Microsystems, a
hardware start-up that appeared to be on the cusp of its big break. It was
located in an old textile mill just outside of Burlington, about a three-hour
drive north of RPI’s campus. Vermont Microsystems made its own plug-in
boards, including graphics cards, for computer manufacturers. At a trade
show in Chicago, an IBM representative visited the company’s booth and
asked whether Vermont Microsystems could make a graphics card
specifically for the IBM PC. In true start-up fashion, the representatives at
the booth said absolutely. What they didn’t say was that they had precisely
one person on staff with the necessary knowledge and skills to make such a
card, and that person was the newly graduated, newly hired, twenty-three-
year-old Curtis Priem.
Overnight, Priem went from staff engineer to principal design architect
for the card that became the IBM Professional Graphics Controller, or PGC,
which was released in 1984. The PGC represented a significant upgrade
over the graphical capabilities of graphics cards of earlier IBM PCs. The
first PCs used a Monochrome Display Adapter (MDA) card, which could
render only green text on a black background that was eighty characters
wide and twenty-five characters tall. Subsequent models used Color
Graphics Adapter (CGA) cards, which gave PCs the ability to manipulate
individual picture elements (pixels) at a resolution of up to 640 × 200 and a
depth of up to sixteen colors. But engineers wanted more space to do their
work and grew tired of the limited range of purples, blues, and reds that the
cards could render.
 Priem’s PGC offered more colors and at higher resolutions than any other
IBM PC graphics card on the market: it could render up to 256 colors at
once at resolutions up to 640 × 480 pixels. The card could also run graphics
routines independently of the main central processing unit (CPU), leading
to faster rendering times. Priem had the card boot in CGA compatibility
mode and only activate its advanced features when needed.
Despite his initial excitement over the job and the significant
responsibilities swiftly granted to him, Vermont Microsystems ended up
being nothing like Apple. The company had difficulty hiring other qualified
engineers, partly because it refused to offer any employee stock options or
equity—a tool that many start-ups use to attract and retain workers and
keep them motivated while living under the inherent risk and pressure of a
company whose money could run out. No matter how hard Priem worked,
no matter the quality of the graphics cards he made, he would never become
as rich as Steve Jobs if he stayed where he was.
So he started looking west, to Silicon Valley. He booked a vacation to
northern California that was, in reality, a job search. Once he arrived, he
headed not to the beach but to a newsstand, where he bought a copy of the
San Jose Mercury News and turned straight to its Help Wanted section.
Amid the many openings at start-ups, one in particular caught his eye: a
hardware engineer role at a company called GenRad, which was at the time
one of the world’s leading manufacturers of testing equipment for circuit
boards and microprocessors. That meant the company had access to early
versions of the latest chips by most major manufacturers—a prospect that
would be hard for Priem to turn down.1 He interviewed at GenRad and
received an offer.
When he got back in Vermont, Priem submitted his resignation. He had
worked at Vermont Microsystems for only two years, and in that time he
managed to design one of its highest-profile products to date. He left on the
very day that the company shipped its first cards to IBM. As the launch
party started, Priem was ushered into his exit interview, after which he was
walked to the exit.
He didn’t know it, but GenRad was a company in crisis by the time he
joined. Despite its successful 1978 IPO and its command of nearly 30
percent of the electronics-testing market—a share that put it comfortably
ahead of rivals Teradyne and Hewlett-Packard (HP)2—a series of
management missteps saw the company’s very existence come into
question. Its executives had spent lavishly to break into the testing market
for semiconductors, only to have their operation fall flat. In order to build a
competitive moat around the business, executives began to insist that
manufacturers completely outsource their chip-testing functions to GenRad,
which caused friction with the company’s largest clients such as IBM and
Honeywell. And a failed merger with a company called LTX led to a crisis
of confidence in GenRad’s senior leadership, followed by an exodus of
talent that strengthened its rivals. Shortly after Priem arrived, GenRad
entered a nose dive from which the company would never fully recover.
After two years of corporate turbulence, Priem asked a tech industry
headhunter to find him a role somewhere else.
A man named Wayne Rosing offered Priem an interview at Sun
Microsystems. Sun was an early pioneer in high-end UNIX computer
workstations, which it sold for thousands or even tens of thousands of
dollars. It was founded by three Stanford graduate students—Scott
McNealy, Andy Bechtolsheim, and Vinod Khosla—in 1982.
Rosing was a former Apple employee who had run the engineering team
behind the Lisa desktop computer that was released in 1983, around the
same time that Priem was making PGC cards for IBM. Lisa was supposed
to change desktop computing forever: it would be the first mass-market
personal computer to feature a graphical user interface (GUI) instead of a
text-only command line and the first to feature five megabytes of hard-drive
storage in an era when most other computers had none at all. But a lack of
software that would make it competitive with similarly priced workstations
and its lofty price tag of nearly $10,000 doomed the Lisa even before it
launched. After disappointing sales, Apple hired a company to take its
remaining inventory of unsold units and bury them in a Utah landfill.
Rosing left Apple shortly thereafter.
During the development of Lisa, Rosing spent a considerable amount of
time assessing the capabilities of competing machines. The one graphics
card that he envied was Priem’s PGC. It was the sort of card that Rosing
wanted but could not make work for the Lisa, which shipped with basic
graphics that could support only a monochrome display at a resolution of
720 × 364 pixels, nowhere near the performance of PGC-powered IBM
machines. After joining Sun Microsystems, Rosing vowed to take
advantage of the increasing technical ability to render fast, beautiful, color
graphics. To do that, he needed someone who could design powerful
graphics chips. Hence his interest in Curtis Priem.
In Priem’s interview, when Rosing asked the young engineer whether he
could make a graphics card like the PGC at Sun, the reply was brief: “Yes.”

IT WAS EXACTLY THE OPPOSITE of what Sun’s executives wanted

Rosing to do. At the time, the company was focused on launching a new
line of computers called the SPARCstation series. These were UNIX-based
workstations designed for specific scientific and technical applications, in
particular computer-aided design (CAD) and computer-aided manufacturing
(CAM) programs that could be used to design complex physical objects,
from bridges to airplanes to mechanical parts. Sun believed that CAD and
CAM tools would make industrial design far faster, cheaper, and more
accurate than hand-drafting. And the company wanted the SPARCstation to
lead the way.
Bernie Lacroute, Sun’s vice president of engineering and Rosing’s
immediate boss, believed that the SPARCstation could dominate the market
on CPU power alone. He directed the SPARCstation team to focus on
improving the device’s main processor and to not do any work on its
graphics capabilities. He was satisfied with the graphics solution from the
previous generation of Sun workstations, in which most of the rendering
took place within the CPU.
Rosing disagreed vehemently. His experience with Apple’s Lisa taught
him the importance of speedy graphics. For the typical workstation user,
fast computation or extensive storage wouldn’t compensate for laggy
graphics. He thought the SPARCstation should have state-of-the-art
displays that could render a million pixels and hundreds of colors. To
achieve that, however, he would need to move graphics processing out of
the CPU and onto separate graphics-accelerator chips—like the PGC from
Vermont Microsystems. And he would have to do it behind his boss’s back.
So when Priem asked for clarification from Rosing, the answer was
almost completely open-ended.
“Curtis, do anything you want. Just fit it in the same size frame buffer as
the last workstation,” Rosing said. “As long as it fits in that area, you’ll
have a spot on the motherboard.”3
 This was as close to carte blanche as Priem—or any engineer—could
reasonably expect to get on any project. Priem could design and build
whatever he could dream up, so long as it could work within the data-
throughput constraints of the “frame buffer”—the memory that the
SPARCstation dedicated to graphics processing.
Priem realized he couldn’t tackle the project alone; he needed help. It
would come soon from another engineer, Chris Malachowsky, whom Sun
Microsystems had hired from Hewlett-Packard. The two men would share
an office and become known as the “closet graphics” team. They worked in
secret on the one thing that their boss’s boss did not want anyone working
on.

UNLIKE HIS OFFICE MATE, CHRIS MALACHOWSKY had come late

to the world of computers. Born in Allentown, Pennsylvania, in May 1959,
the son of an obstetrician and an occupational therapist turned homemaker,
he grew up in Ocean Township, New Jersey. As a teen, he loved carpentry
and considered becoming a cabinetmaker, but his parents pushed him
toward medicine. At that point, he had never considered electronics or
technology as a possible career path.
He graduated from high school at seventeen and enrolled at the
University of Florida, which was well known both for its medical school
and its school of construction management, and about as far away from cold
New Jersey winters as Malachowsky could get. In addition, the school’s
pre-med program had a unique philosophy: it wanted to give future doctors
a broad foundation of knowledge and thus made them take classes beyond
the life sciences. To fulfill his non-life-sciences requirement, Malachowsky
took a physics class and earned an A in the electrical section of the course.
He found that engineering came naturally to him.
He didn’t think too much of it until the lunch break between his MCAT
medical school entrance exam sessions. While Malachowsky laid down on
a picnic table and stared up at the Florida sun, he contemplated life as a
doctor and following in his father’s footsteps. Is that what he wanted to do
for the rest of his life? Be on call at all hours, working four-to-five-day
stretches with little to no sleep? He wondered, “Do I really want to know
what all the names on drug bottles mean?”
 “No,” he realized. “I like this engineering stuff. I’d rather be an
engineer.”
After finishing the MCAT exam, he headed back to his rental house,
stopping only to pick up a case of beer at a 7-Eleven on the way, and called
his parents as soon as he got home.
“Mom, Dad, I’ve got good news and bad news,” he told them. “The good
news is the test wasn’t that hard. The bad news is, I don’t want to be a
doctor anymore.”
He waited for a response, sure that his parents would be upset. But they
expressed relief.
“Good,” said his mother. “You never read directions anyway. We didn’t
think you’d be a good one. We thought you were doing it for your father.”
Malachowsky went on to major in electrical engineering and parlayed his
good grades into a job at Hewlett-Packard in California. He ended up
working in the manufacturing department, responsible for the production of
a new 16-bit minicomputer that HP was developing in its research and
development lab.
“It turned out to be great for me because it gave me a chance to learn
about how real computers were built,” he said.
While many people knew how to design a computer chip in principle,
few could design one that could be manufactured in large volumes, and at a
profit. When Malachowsky first arrived at HP, he saw that hands-on
experience in its manufacturing department could give him a practical
perspective on the industry that few others seemed to possess. On top of
that, HP had a reputation for molding young engineers into disciplined
veterans through its mentorship and training programs. Malachowsky knew
that his time at the company would prepare him for whatever opportunity
came next.
It was after his stint on the HP manufacturing floor that he was invited to
join the company’s research lab to develop new chips. He worked on the
HP-1000 Minicomputer product line and learned how to write embedded
control software for its communication peripherals. Later, he led the team
that would make the HP-1000’s CPU, which would be manufactured in the
same building where he started his career at HP.
While working every day on the most crucial element of the HP-1000, he
also pursued a master’s degree in computer science from nearby Santa
Clara University. Once both projects—the chip and the degree—were
finished, he and his wife, Melody, whom he married a year after college,
began to think about where they could start a family.
At first, they considered transferring to HP’s satellite office in Bristol,
England, but his wife didn’t like the idea of moving so far away. They then
considered the East Coast. Her family was in Northern Florida, while his
parents were in New Jersey. Halfway in between was North Carolina’s
Research Triangle, home to both world-class universities in Duke and UNC
and the offices of tech behemoths IBM and Digital Equipment Corporation,
or DEC.
Before making the transcontinental move, however, Malachowsky
decided to apply for jobs at other companies, solely for the purpose of
getting some practice interviewing. His first invitation came from the
nascent supercomputer division at Evans and Sutherland, a graphics
company otherwise known for making high-end flight simulators for
military training. He was rejected right away; his interviewers thought he
questioned the status quo too much and felt that he would be a poor fit at
the company. (Malachowsky believed their feedback didn’t bode well for
the company’s future. He was right. Evans and Sutherland’s first
supercomputer later failed to sell, and the looming end of the Cold War
meant that simulator demand from the military was already drying up.)
The second practice interview was at Sun Microsystems, where he had
applied for an unspecified position making graphics chips. Although
Malachowsky had no prior graphics experience, his curiosity got the better
of him, and he agreed to interview with the lead engineer, Curtis Priem.
What started out as a mere prep session ended up changing the course of
Malachowsky’s life—and the course of the entire tech industry.

“CURTIS WAS THE ONE WHO understood graphics,” Malachowsky

later recalled. “I turned out to be the build-it guy. Tell me what to do, what
needs to be done, and I’ll go figure out how to do it.”
In order to produce the high-quality graphics that Rosing wanted (but that
Rosing’s boss didn’t), Priem had designed a monstrosity of a graphics
accelerator. It would contain two dedicated ASICs: the frame buffer
controller, or FBC, which rendered high-resolution images at fast speeds;
and the transformation engine and cursor, or TEC, which could quickly
calculate the motion and orientation of objects as a user manipulated them.
Instead of relying on the CPU to perform all of these tasks, as earlier Sun
workstations had done, Priem’s accelerator would handle up to 80 percent
of the computational workload on its own—meaning the dedicated graphics
chips would do the limited set of functions that they did best, and the CPU
would be freed up to handle the myriad other tasks that it did better.
It was a good design, in theory, but now it was up to Malachowsky to
figure out how to make it a reality. Unlike HP, Sun did not make its own
chips. Instead, Malachowsky would rely on LSI Logic, headquartered in
nearby Santa Clara, which was then the global leader in fabricating custom
ASICs for hardware manufacturers. Malachowsky’s timing was fortuitous:
LSI had just introduced a new chip architecture called “sea-of-gates,”
through which they could fit more than ten thousand gate arrays onto a
single chip, a feat that no other manufacturer had been able to accomplish.
Although LSI’s own prototypes were impressive, Priem’s chip designs
would need to be larger still in order to produce sufficient processing power
for the SPARCstation. LSI’s executives recognized the potential to turn Sun
Microsystems into a big customer and agreed to take the contract—even
though, as Malachowsky later noted, they seemed nervous about their
ability to deliver.
To make sure that Priem and Malachowsky got the chip they had drawn
up, LSI assigned one of its rising stars to manage the Sun account—a
relatively new hire named Jensen Huang.
“This young kid had just joined them from AMD who had worked on
microprocessors,” Malachowsky said. “Curtis knew what he wanted, I
could design it, and Jensen helped us figure out how we were going to build
it.”
Together, the three of them worked out the manufacturing process that
would make Priem’s design ready for fabrication. As problems arose, each
man worked within his own area of expertise to solve them. But a small
team laboring on a high-pressure project could create tension.
“Curtis is so bright. He thinks so fast,” said Malachowsky. “He starts
with an idea and jumps to a solution and there are no breadcrumbs between
the two. I really felt that my biggest contribution was helping him articulate
[his ideas] for other people in a way that they could get behind. My
communication skills turned out to be equally important as my engineering
skills.”
 Sometimes, communication turned into outright conflict.
“Chris and I would have these knock-down, drag-out fights. Not
physically, but we’d be yelling, screaming at each other,” Priem recalled.
“He was trying to get something out of me about a decision on a chip. Then
when I told him what he wanted, I would just keep going because I couldn’t
settle down. Chris would then say, ‘No, no, we’re done. You gave me the
answer.’ ”
Priem would then storm out of the office, while the rest of his team—
which at this point consisted of two hardware engineers named Tom
Webber and Vitus Leung—would look at Malachowsky with alarm.
Eventually, someone would ask if the team would now be disbanded.
“We’re good,” Malachowsky would always reply.
Jensen saw more promise than peril in these explosive fights, too. He
called them examples of “honing the sword.” Just as a sword only becomes
sharper when it meets grinding resistance, the best ideas always seemed to
come from spirited debate and argument, even if the back-and-forth could
get uncomfortable. Already, he was learning to embrace conflict rather than
shy away from it—a lesson that would eventually come to define his
philosophy at Nvidia.
“We broke every tool that LSI Logic had in their standard portfolio,”
recalled Malachowsky. “Jensen was bright enough and savvy enough to say,
‘Look, I’ll fix these problems at the back end. You can ignore them. These
you’d better fix because I don’t know if I can handle those.’ ”
In 1989, the three men finalized the specifications for Sun’s new graphics
accelerator. The FBC would require 43,000 gates and 170,000 transistors in
order to do its job properly; the TEC, 25,000 gates and 212,000 transistors.
They would sit together on a single graphics accelerator, which was
packaged as the “GX graphics engine”—or just GX for short.
The “closet graphics” team got one more boost just as they were ready to
release the new chips. Bernie Lacroute, the executive who had shown such
antipathy toward graphics chips just a few years earlier, had recently asked
Wayne Rosing if he had followed his order not to put any effort toward
improving the SPARCstation’s graphics capabilities. Rosing replied in the
negative.
“Good,” said Lacroute.4
GX STARTED AS AN OPTIONAL add-on, for which Sun up-charged
customers $2,000. GX made everything on the display work faster: two-
dimensional geometry, three-dimensional wireframing, even the mundane
task of scrolling through lines of text was quicker and better with GX
accelerators than without them.
“For the first time, probably in history, the scrolling of the text in a
windowing system was faster than you could see,” said Priem. “It allowed
you to scroll up and down a large document without actually seeing the
FBC painting.”
But the best showcase for GX graphics was a game that Priem had been
working on in his spare time. Back at Vermont Microsystems, he had started
making a flight-simulator game featuring the A-10 Warthog. A squadron of
Warthogs was stationed at the nearby Vermont Air National Guard Base in
Burlington. After work, he would park his car at the end of the base’s
runway and watch the jets take off. His simulator program was meant to get
him even closer; it was supposed to let him fly the A-10 in its intended role
as a “tank buster” during an imagined Cold War conflict. But his personal
computer, an Atari 800, didn’t have enough graphical processing power to
render the complex physics of an A-10 in flight. He never finished the
game. In fact, no card then on the market could bring to life the game that
Priem had imagined.
Until the GX-enabled SPARCstation. For the first time, a realistic flight
simulator became possible. Priem bought a workstation for his own
personal use with his 60 percent employee discount, which shaved
thousands of dollars off of the price. After he spent sixty hours a week at his
day job, he would go home and get back to laboring on his new simulator
program that would take full advantage of the new GX chips. Finally, he
was able to realize his vision and complete the game, which he called
Aviator.
Aviator placed users in the cockpit not of the A-10 but of the high-
performance F/A-18 fighter jet and pitted them against other F/A-18s in an
aerial dogfight. The game fully modeled the F/A-18’s weapons, including
Sidewinder missiles, guns, and bombs. Priem rendered Aviator’s battlefields
realistically, purchasing satellite data to get elevations and land contours
right, and adding texture-mapped graphics. He even designed a hardware
device adapter to enable PC-compatible joysticks to work with Sun’s
workstations, so that players wouldn’t have to use the keyboard to control
their virtual aircraft.
Priem had a business partner for the game: Bruce Factor, who worked in
Sun’s marketing department and agreed to handle sales and marketing.
Factor quickly realized that Aviator could do more than just pass the time—
it could also help Sun move workstations. The game was a fantastic means
to demonstrate the GX’s graphical capabilities, running at a high resolution
(1,280 × 1,024 pixels) and at 256 colors, at a time when most other PC
games could only manage resolutions up to 320 × 200 pixels. Aviator also
allowed clients with multiple networked Sun workstations to play against
each other in real time, using Sun’s new “multicasting” protocol—a kind of
rudimentary local area network (LAN) that presaged the LAN party craze
of the 1990s and 2000s.
Priem and Factor gave out free copies of Aviator to every sales office at
Sun Microsystems. The company’s reps used it as a way to show off the
computer’s capabilities and would often buy more copies to present as gifts
to their workstation customers.
“I was getting every last bit of performance out of the hardware,” Priem
said. “Aviator became pretty serious. It was the best demo the Sun
Microsystems sales force used to show off the performance of a standard
workstation.”
Aviator was officially released to the public in 1991. It was demonstrated
at the annual conference of the Special Interest Group on Computer
Graphics and Interactive Techniques (SIGGRAPH). At the show, Priem and
Factor set up a network of eleven workstations so attendees could try
dogfighting against each other.
The process of developing Aviator taught Priem some important lessons
beyond game design. The game was hacked by a Sun employee within two
days of release, enabling people to play without paying for a copy of their
own. To prevent future hacks, Priem released a new version that could
disable itself if it detected any changes to the code and which would e-mail
him the details of the users trying to pirate the software. Later, Priem would
incorporate similar private-key encryption technology into his first Nvidia
chip design.
After a few years of torrid sales as an add-on option, the GX chips
became standard on every Sun workstation. Its success boosted the careers
of Priem and Malachowsky, who became graphic architects and were given
their own team, called the Low End Graphics Option group. Meanwhile,
LSI’s gamble on the chip had paid off handsomely. The company’s
revenues grew from $262 million in 1987 to $656 million in 1990, driven
partially by GX sales, even as it reduced the list price of each unit from
nearly $375 for the initial two-chip run to around $105 for the later one-
chip version. Jensen was promoted to director of LSI’s CoreWare division,
which made custom chips for third-party hardware vendors by using a
library of reusable intellectual property and designs.

IRONICALLY, GX’S SUCCESS HAD THE opposite effect on Sun

Microsystems. By the early 1990s, it had shifted away from the agile, start-
up-like environment that gave people such as Rosing, Priem, and
Malachowsky the independence to follow their instincts and display their
technical virtuosity. The culture was now becoming more bureaucratic,
more controlled, and thus slower. Project teams no longer competed to
come up with the most innovative ideas; they competed to create
PowerPoint presentations that would win over the most executives. In short,
Sun Microsystems had become political.
It was not an environment where Malachowsky or Priem wanted to be.
Priem, in particular, was bothered by a culture where “it was easier to
sabotage or get the other project killed than to come up with better
technology.” He just wanted to make good graphics chips and had no
interest in corporate infighting.
New chip design releases at Sun ground to a halt as cycle after cycle of
new proposals—many of which looked good on a slide but were either
technically or economically unviable—were approved one quarter, then
wound down the next.
“For two years, nothing exited the building,” Malachowsky said. “My
assessment was they had been so successful leading up to this time that they
were more concerned with protecting success than driving for it. It was
getting caught up in fear of failure. They stopped being very aggressive.”
Worse still, Sun actually tried to undo much of the progress that Priem
and Malachowsky had made with GX. During one pitch cycle, Priem’s team
proposed a new generation of graphics accelerators that would incorporate
cutting-edge video-memory technology from the Korean chip manufacturer
Samsung. But Priem lost out to a rival named Timothy Van Hook, who
believed that the best way to push the graphical envelope for Sun
workstations was to task the CPU with more high-end 3-D graphics
functions rather than rely on a dedicated graphics chip.5 Priem was
convinced that the idea wouldn’t work, from a technical perspective. But it
didn’t matter because Van Hook had one advantage that Priem didn’t: the
ear of one of Sun’s cofounders, Andy Bechtolsheim. Without an internal
champion of that stature, Priem knew that he and his group had no chance.
“Andy came and told me that our product line was a dead end,” Priem
said.
He soon realized that his days at Sun were numbered. Rumors swirled
that Sun leadership wanted to disband his team, fire him, and move
Malachowsky onto another chip project. Having worked side-by-side with
Priem for the past six years, Malachowsky was angry over the treatment of
his friend and one of the company’s most talented engineers.
“Chris knew every single struggle I went through, taking all the hits from
Sun management,” Priem said. “He respected me taking all the arrows in
the back. There were times I was so chastised by the VP of graphics that I’d
be out with HR walking around the buildings in the park crying. It was just
brutal.”
Bechtolsheim choosing Van Hook’s idea was the last straw for both men,
whose success with GX now meant little in what they saw as in increasingly
dysfunctional company.
“We realized our time was limited and neither of us wanted to work at
Sun,” Priem said. They already had a new project in mind: resurrecting the
next-generation accelerator chip that Sun leadership had passed on.
“Why don’t we just go build Samsung a demonstration chip?” Priem
asked Malachowsky. “We’ll just be consultants and show them the value of
this new memory device they are committing to build.”
Malachowsky thought it sounded like fun. They knew how to build chips,
and they knew that they had a plan for a good one. But this advantage could
just as easily become a liability too: in the high-stakes, multibillion-dollar
world of semiconductors, no company would think twice about stealing an
idea from a pair of engineers if it would give them even the slightest
competitive edge. Unless they had a partner with business savvy to match
their technical brilliance, they might as well not even bother.
Then another idea hit Malachowsky.
 “We knew a guy!” he recalled later. “We knew a guy who we were good
friends with who had moved into technology licensing and building systems
on a chip for other people. So, we reached out to Jensen.”
Malachowsky and Priem asked Jensen Huang for help writing a contract
to work with Samsung. The three started meeting to devise a business
strategy to deal with the Korean company. Then one day Jensen said, “Why
are we doing this for them?”6
 CHAPTER 3

The Birth of
Nvidia

CURTIS PRIEM AND CHRIS MALACHOWSKY’S idea for a graphics-

chip venture was perfectly timed. In 1992, two major developments—one
in hardware, one in software—accelerated the demand for better graphics
cards. The first was the computer industry’s adoption of the Peripheral
Component Interconnect (PCI) bus, a type of hardware connection that
transferred data among the expansion cards (such as graphics accelerators),
the motherboard, and the CPU at a much higher bandwidth than that
available from the prior Industry Standard Architecture (ISA) bus. The
process of designing higher-performance cards would be easier, and there
would be a far larger market for the resulting products.
The second development was Microsoft’s release of Windows 3.1, which
was intended to showcase the very latest in computer-graphics capabilities.
It introduced TrueType fonts, which rendered pixel-perfect text across all
Microsoft programs, and it supported high-quality video playback with its
new Audio Video Interleave (AVI) video-encoding format. Importantly, it
did not hide these developments under the hood. With garish screensavers,
user-customizable interfaces, and the constant nudges to use Windows
Media Player, the operating system wasn’t shy about showing off its
graphical prowess. In its first three months after its release on April 6, 1992,
Windows 3.1 sold nearly three million copies and proved that there was
strong demand for programs that could take advantage of the PC’s
improving graphics.
Priem and Malachowsky decided the PC market, rather than the
workstation market, represented the best opportunity for their start-up. They
were thinking, in part, of Priem’s flight simulator, which they planned to
make available to any gamer who had a personal computer—instead of
restricting it only to those who had access to Sun Microsystems hardware at
their place of employment. As they had at Sun, Priem and Malachowsky
would not manufacture the chips or circuit boards themselves, to keep costs
down. Instead, they would focus on designing the best chip possible and
would outsource production to semiconductor firms that already had the
expensive production infrastructure in place.
Still, Priem had little idea how they would stack up against the
competition. “I knew Chris and I were good, but I didn’t know if we were
good compared to the rest of the world,” he said.
Sun machines had always had a Windows-like graphics interface, and
PCs running Windows would soon need to support a similar multi-window
operating system environment, a feature Priem and Malachowsky had
already created. They knew their skill set would be valuable in the PC
market.
“You have to do all sorts of security protections and abstractions with ten
windows open,” said Malachowsky. “These were the kinds of things the PC
hadn’t had to deal with because they had a DOS environment that basically
owned the whole screen.”
In late 1992, Priem, Malachowsky, and Jensen met frequently at a
Denny’s at the corner of Capitol and Berryessa in East San Jose to figure
out how to turn their idea into a business plan.
“We’d show up, we’d order one bottomless cup of coffee. And then, you
know, work for four hours,” Malachowsky said.1
Priem remembers eating plenty of Denny’s pies and the Grand Slam
breakfasts—two buttermilk pancakes served with eggs, bacon, and
sausages. Jensen doesn’t recall his typical order, but he thinks it was likely
the Super Bird sandwich—turkey, melted Swiss cheese, tomato, and his
favorite addition, bacon.2
Jensen still needed to be convinced to leave his job. Between bites, he
peppered Curtis and Chris with questions about the size of the opportunity.
 “How big is the PC market?” Jensen asked.
“It’s big,” they replied, which was true—but obviously not detailed
enough to satisfy Jensen.
“Chris and I were just sitting there watching Jensen,” Priem said. He kept
working through his analysis of the PC market and the potential
competition. There was a place for their start-up, he believed, but he didn’t
want to leave his current job until he felt like the business model made
sense. He was grateful that Chris and Curtis had somehow decided he was
essential, even though he remembers thinking, “I love my job, you hate
your job. I’m doing well, you’re doing crap. For what reason do I leave
with you?”
He told them that he would join them if they could prove that the start-up
could eventually generate $50 million in annual sales.
Jensen reminisces fondly about the long conversations at Denny’s. “Chris
and Curtis were the two brightest engineers, computer scientists that I have
ever met,” he said.3 “Luck has a lot to do with success, and my luck was
having met them.”
Eventually, Jensen decided that $50 million in revenue was possible. He
was confident, as a gamer himself, that the gaming market was going to
grow considerably.
“We grew up in the video-game generation,” he said.4 “The
entertainment value of video games and computer games was very obvious
to me.”
The question then turned to who was going to make the first move. Priem
was ready for it to be him—the way things were going at Sun, he would
have to leave the company in a few months, anyway. But Jensen’s wife,
Lori, didn’t want him to leave LSI until Malachowsky also left Sun—and
Malachowsky’s wife, Melody, didn’t want him to leave Sun until Jensen
committed.
In December of 1992, Priem forced their hand. He submitted his letter of
resignation to Sun Microsystems, effective December 31. The following
day, alone in his house, he founded the new venture, “just by declaring that
this was started,” he later recalled.
Even this was a bit of an overstatement. Priem had no name for his
company. He had no funding. He had no employees. He didn’t even have
Malachowsky or Jensen aboard just yet. All he had was an idea—and some
leverage over his friends.
“I put pressure on both of them that we can’t let Curtis flail alone,” Priem
said, adding that he almost guilt-tripped them. “I think they joined together
and said because Curtis quit, they have to quit. Because they quit
simultaneously, they solved the problem with their wives, making sure we
were a team.”
Malachowsky agreed to stay at Sun Microsystems long enough to sign
off on his last project, a new upgrade to the GX lineup. Once his engineers
verified that the chip was 100 percent perfect, he was comfortable declaring
that his final day would be in early March of 1993.
“A good engineer doesn’t walk away from their responsibilities,” he said.
A good engineer doesn’t walk away without his tools, either. Before he
left, Malachowsky asked to take his Sun workstations with him to the new
start-up. Wayne Rosing, who was still his boss, agreed, and in his last days
on campus, Malachowsky made sure to upgrade as many components in his
devices as he could.
“They got upgraded by the maximum memory, maximum disk drives,
and maximum monitor size,” Priem said.
Jensen, too, wanted to leave LSI on good terms. He spent the first six
weeks of 1993 distributing his projects to other leaders within the company.
He officially joined Priem on February 17, which happened to be Jensen’s
thirtieth birthday.

ROSING THOUGHT THAT PRIEM, HIS protégé, was making a big

mistake. In January, while Priem was still “flailing alone,” Rosing invited
his now-former engineer to an off-site location where several Sun
employees were working on a secret project. After getting Priem to sign a
nondisclosure agreement, Rosing revealed that Sun was creating a new
general-purpose programming language that would eventually become
Java. Although the project was off to a promising start, Rosing believed that
it ran too slowly to be useful. He asked if Priem would be interested in
designing a new chip that could take some of the processing load off of the
CPU and accelerate the execution of the new language.
Priem was tempted, especially as he wasn’t yet sure whether Jensen and
Malachowsky were going to follow through on their promises to join him in
the new venture. “If I had said yes, it would have taken my career down a
totally different path.”
Although he seriously considered Rosing’s proposition, he had no real
interest in designing CPUs and was far too excited by the prospect of
designing his own graphics chips with his friends, even if it entailed huge
risk. He passed on Rosing’s offer.
Undeterred, Rosing tried again in February. This time, he didn’t try to
peel one of them away from the others. He tried to get all three at once. He
offered to license Sun’s entire portfolio of patents to their start-up, including
all of Priem’s and Malachowsky’s old GX chip designs. In exchange, they
would agree to make their new chips compatible with both Sun’s GX
graphics and IBM PCs.
After hearing Rosing’s pitch, the three men retreated to the parking lot of
Sun’s campus to debate the decision. Priem considered all the implications
of the proposal and declared it “interesting.” The partnership would give
them a large, brand-name customer right off the bat and would protect them
from any copyright-infringement claims from their former employer. But
the downside was that the agreement would force them to spend less time
and resources on the PC market, which was, in their view, where the real
opportunity lay. They weren’t even sure whether they could make a single
chip work for both the Sun and the PC platforms. They agreed to decline
Rosing’s proposal and go it alone.
During the parking-lot discussion, Priem revealed that he already had
basic specifications in mind for a new PC-based graphics accelerator. It
would have more colors and work with a larger frame buffer than that of the
GX chips he and Malachowsky had made at Sun. In many ways, it would
be an evolution of the GX chip they had worked on for six years. He
pointed out that Microsoft named their new operating system Windows NT,
with “NT” meaning “next technology.” That’s why, he said, he wanted to
call the chip the “GX Next Version,” or GXNV.
It sounded like “GX envy,” a phenomenon that was common among
Sun’s workstation competitors. Priem had heard stories about rivals, such as
Digital Equipment Corporation, who had lost customers to Sun’s sales
teams armed with GX graphics and copies of Aviator. The name indicated
their resolve to do it again—and to do it on their own terms, this time.
To emphasize the clean break with their past (and, likely, to prevent even
the faintest whiff of possible copyright infringement), Jensen told Priem to
“drop the GX.” Their new chip would be called the NV1.

THE THREE COFOUNDERS STARTED WORKING out of Priem’s

townhouse in the San Jose suburb of Fremont with little more than a vision
and Malachowsky’s Sun workstations. Priem cleared every room except his
bedroom and moved all of the furniture into his garage, setting up large
folding tables for all of their equipment. The first few weeks, there wasn’t
much to do. The three of them would get together each day and talk about
food.
“What did you do last night? What did you have for dinner?” Jensen
recalled them asking each other. The day’s big event would be the decision
on what to eat for lunch. “It sounds pathetic, but it’s true.”
After a time, they decided to make their first official hardware purchase,
ordering an IBM-compatible PC made by Gateway 2000, the mail-order
computer manufacturer that famously shipped its devices in black-and-
white cow-print boxes. Upon arrival, the machine utterly confused Priem
and Malachowsky, whose professional lives up to that point had been
focused on Sun Microsystems hardware and software.
“We were not PC people,” said Malachowsky. “It was funny. We were
going to take over the world, but we didn’t know anything about PCs.”
Fortunately, they would not be on their own for long. Once news of the
three cofounders’ new venture spread, several senior engineers at Sun
Microsystems quit and joined the fledgling start-up. Two crucial early hires
were Bruce McIntyre, a software programmer on the GX team, and David
Rosenthal, a chip architect who became the start-up’s chief scientist.
“I can’t believe how many amazing people joined us. We had a dozen
people working without a salary,” Priem said. “We didn’t pay them until I
think June, when we first got our funding.”
McIntyre and Priem took a Sun GX graphics chip and attached it to a
board that could plug into their Gateway. The hardware interface was easy;
the software integration was much harder. The Sun hardware processed
instructions in a manner that Microsoft’s operating system could not
understand. It took a full month of work to remap the GX’s graphics
registers to work with Windows 3.1, but eventually the team solved the
problem. Naturally, the first game they ported over to Windows was the
latest version of Priem’s Aviator, which they renamed Zone5.
 Now, the start-up had a staff. It had a viable demonstration product. It
only needed an official name, so that it could be legally incorporated. Priem
had already written down a list of potential options. One early leader was
“Primal Graphics,” which sounded cool and combined the first few letters
of two of the cofounders’ last names: PRIem and MALachowsky. Others
liked it, but the whole team felt that in order to be fair it had to include
Jensen’s name as well. Unfortunately, this made it impossible to come up
with a name that sounded remotely appealing. The other contenders
included Huaprimal, Prihuamal, and Malhuapri. The name-combination
idea was dropped.
Most of the other possibilities on Priem’s list incorporated “NV” as a
reference to their first planned chip design. These names included
iNVention, eNVironment, and iNVision—the kinds of everyday words that
other companies had already co-opted for their own brands, such as a toilet
paper company that had trademarked the name “Envision” for its
environmentally sustainable product line. Another name was too similar to
the brand of a computer-controlled toilet. “These names were all stinky,”
Priem said.
The last remaining option was “Invidia,” which Priem found by looking
up the Latin word for envy—in a sense, another callback to their work on
the GX, when he and Malachowsky believed that their rivals, both within
and beyond Sun, had envied their success.
“We dropped the ‘I’ and went with NVidia to honor the NV1 chip we
were developing,” said Priem, “and secretly hoped that someday Nvidia
would be something that would be envied.”
With a name in hand, Jensen sought out a lawyer and chose James
Gaither, who worked at the law firm of Cooley Godward. Gaither’s firm
was midsized, with fewer than fifty attorneys on staff. Even so, it had
carved out a niche for itself as the go-to firm for early-stage Silicon Valley
start-ups. During their first meeting, Gaither asked Jensen how much
money he had in his pocket. Jensen said $200.
“Hand it over,” said Gaither. He then told Jensen he now owned a large
equity stake in Nvidia.
Nvidia’s incorporation documents gave each of the cofounders equal
ownership. Jensen returned to the townhouse and asked his cofounders to
each invest $200 of their own to “buy” their shares of the company.
 “It was a good deal,” Jensen later observed, with typical dryness.
On April 5, 1993, Nvidia was officially born. That same day, Priem drove
to the Department of Motor Vehicles to order a vanity license plate:
NVIDIA.

THE FIRST TEST OF NVIDIA’S viability—the search for funding—

loomed. The world of venture capital (VC) was much smaller in 1993 than
it is today. Silicon Valley VC firms, most of which—then as now—were
headquartered on Sand Hill Road in Palo Alto, only accounted for about 20
percent of the nation’s total venture investments and competed with firms
based in Boston and New York. The entire VC industry was really a niche
in the economy, making just over a billion dollars in outlays per year (close
to $2 billion in today’s dollars).5 Today, Bay Area VC firms now dominate
the industry, investing more than half of the $170 billion in funding that
gets distributed every year.
Two things have remained constant about venture capital, however. The
first is that founders whose start-ups already produce revenue are far more
successful with their pitches than start-ups that have no products in the
market—and this was especially true in the early ’90s, when venture
interest in early-stage companies was at a ten-year low. The second is that,
as with many things in the business world, success depends as much on
who one knows as it does on how strong one’s business is. In Nvidia’s case,
the founders’ connections were extensive enough to make up for the
company’s nonexistent revenue stream.
Jensen’s decision to ease his way out of LSI Logic turned out to pay
immediate dividends during Nvidia’s fund-raising process. When he
submitted his resignation, his manager had immediately taken him to LSI’s
CEO, Wilfred Corrigan, a British engineer who pioneered several
semiconductor manufacturing processes and design principles that are still
in use today. Jensen’s manager wanted “Wilf,” as he was known throughout
the company, to talk the young engineer out of leaving LSI altogether. But
when Corrigan heard about Jensen’s vision for a new generation of graphics
chips, he asked him a question: “Can I invest?”6
Corrigan grilled Jensen on the start-up’s addressable market and strategic
positioning: “Who plays games?” “Give me an example of a gaming
company.” Jensen responded that if they built the technology, more game
companies would be founded. Existing companies in the space, such as S3
and Matrox, typically made 2-D accelerated-graphics cards, and games with
3-D graphics were only beginning to take off.
Still, Corrigan remained skeptical that Jensen’s business would be viable.
“You’ll be back soon,” Corrigan told him. “I’ll hold your desk.”
Nevertheless, Corrigan promised to introduce Jensen to Don Valentine at
Sequoia Capital. Valentine had invested in LSI Logic back in 1982, which
earned him a handsome payout when the company went public a year later.
He had hit it even bigger in other investments in tech companies such as
Atari, Cisco, and Apple. By the early ‘90s he was considered “the best
venture capitalist in the world.”7
Although Corrigan may have had doubts about Nvidia’s potential, he had
none about Jensen himself. When he called Valentine after his conversation
with the young, departing engineer, he didn’t pitch Jensen’s start-up idea; he
pitched Jensen.
“Hey Don,” he said, “we’ve got this kid who is going to leave LSI Logic.
He wants to start his own company. He’s really smart. He’s really good. You
guys should take a look at him.”8 Valentine agreed to meet with Jensen,
Priem, and Malachowsky and had a junior partner set an appointment for
the end of May. In the meantime, they would be free to pitch other potential
investors.
In mid-April, just weeks after Nvidia’s incorporation, the three
cofounders visited Apple’s headquarters to discuss the graphics needs for
the Macintosh line. Nothing came of the meeting.
Three weeks later, they visited the offices of Kleiner Perkins Caufield &
Byers, another venture capital firm that, like Sequoia, got its start in the
1970s and had made its own series of home-run investments. These
included America Online, Genentech, and Sun Microsystems—the last
being how the VC firm came to the Nvidia cofounders’ attention. At the
meeting, one of the Kleiner partners fixated on the topic of circuit boards,
insisting that Nvidia needed to bring board manufacturing in-house.
Nvidia’s plans were to design the graphics chip, have it manufactured by
someone else, and then sell the chip to a board partner, who would mount it
on a graphics card and sell that card to PC makers.
The partner’s insistence made no sense to Malachowsky. “Why would we
compete on pennies on a resistor?” he asked. “I mean, we have no special
expertise there. We’ll stick to what we’re good at, and if that’s not for you,
it’s not for you.”
Part of this was the typical, if not necessary, bravado of a start-up
founder, but part of it was Malachowsky’s practical nature shining through
once again. For all their ambition to take over the PC graphics market,
Nvidia had to focus its resources on the single best opportunity rather than
spread themselves thin chasing all possible ones. This was why they had
declined Wayne Rosing’s offer to make chips that could run on both Sun
workstations and IBM-compatible PCs. Now, it meant walking away from
the conversation with Kleiner Perkins, too.
The next meeting they took, with Sutter Hill Ventures, went more
smoothly. Once again, the cofounders’ prior connections meant that they
were not going in completely cold. Sutter Hill had also invested in LSI
Logic and had contacted Wilf Corrigan to ask about Jensen. Corrigan gave
the same enthusiastic endorsement he had given Don Valentine. But Sutter
Hill had already made some investments in graphics companies, and the
firm doubted whether a new start-up could really differentiate itself in a
market that they considered extremely competitive and highly
commoditized already. The only partner excited about Nvidia was Tench
Coxe, who had joined the firm a few years prior.
“It was a controversial deal,” Coxe recalled. “I was the young guy at
Sutter in a partnership of five guys.”
Coxe was impressed by the three cofounders. He already had Corrigan’s
endorsement of Jensen. In the meeting, he probed Priem’s and
Malachowsky’s expertise and was surprised by their depth of knowledge
about 3-D graphics and computer operating systems.
The positive meeting with Sutter Hill seemed to bode well for the big test
two days later: their pitch to Don Valentine at Sequoia. Although Nvidia
still didn’t have its own proprietary chip to show off yet, they could present
the Sun GX graphics card that they had hacked to work with their Gateway
2000 PC as a proof-of-concept. The chip was four years old at this point,
but it was still far more capable than any other Windows graphics card on
the market. To demonstrate that, they would play a twenty-minute session
of Zone5, electing to run the demo not on a standard monitor but through an
early virtual-reality headset made by another start-up. They believed that
the dazzling graphics alone would make their pitch successful.
 What the Nvidia team didn’t know was that Valentine hated product
demos. The Sequoia founder had sat through enough pitches to know that
entrepreneurs loved showing off their technology and would always present
well. He believed, however, that even more important than a flashy product
was a real understanding of the product’s potential market and competitive
position. The Nvidia cofounders were walking into a trap of their own
making.
The three cofounders were met at Sequoia’s offices on Sand Hill Road by
Mark Stevens, a newly promoted junior partner who had previously worked
at Intel and was now the firm’s semiconductor specialist. He led them to a
dark, wood-paneled conference room, where they set up the demo. After it
concluded, Valentine switched over to his preferred style of evaluating a
start-up: a rapid-fire series of questions designed not only to test the
founders’ expertise but also to see how they would perform under pressure.
Malachowsky later referred to it as Valentine “holding court.”
“What are you?” Valentine asked the three cofounders. “Are you a
gaming-console company? Are you a graphics company? Are you an audio
company? You have to be one.”
Priem froze for an instant. Then he blurted out an answer. “We’re all of
them.”
He went on a long and deeply technical explanation of how they could
integrate all the features Valentine asked about into their one proposed chip.
Although Priem didn’t say anything untrue about the NV1’s potential, his
flustered response was so dense that only an engineer could understand it.
To Priem, the plan was a sign of their ambition and expertise: they could
develop a single chip that could address multiple different markets at once,
expanding the potential of the chip without increasing its engineering
complexity all that much. To Valentine, it sounded like Priem was being
indecisive.
“Pick one,” he snapped. “Otherwise, you’re going to fail because you
don’t know who you are.”
Valentine then asked where Nvidia would be in ten years. Priem
responded, “We’re going to own I/O architecture.” It was another
engineer’s reply to a business question. What Priem meant was that he saw
future generations of Nvidia chips accelerating not just graphics, but other
computer-board operations such as sound, game ports, and networking.
Once again, however, his answer was impenetrable to everyone on the
Sequoia side. According to Malachowsky, it even confused his cofounders.
Stevens stepped in to bring the conversation down to a more practical
level. Who, he asked, did Nvidia expect to actually manufacture their chips?
The cofounders replied that they planned to use SGS-Thomson, a European
semiconductor firm that had only recently avoided bankruptcy through deep
cost cuts and by outsourcing production to Singapore and Malaysia. After
hearing this, Valentine and Stevens looked at each other and shook their
heads. They wanted Nvidia to work with the Taiwan Semiconductor
Manufacturing Company (TSMC), which had a better reputation.
Jensen attempted to steer the conversation back toward Valentine’s
preferred topics of market position and strategy, but by now even he was
flustered by the barrage of questions and the fact that the Nvidia team
seemed unable to produce a satisfying answer to any of them. The meeting
ended without a commitment from Sequoia.
“I did a horrible job with the pitch,” Jensen said, taking responsibility for
the entire performance. “I had a hard time explaining what I was building,
who I was building it for, and why I was going to be successful.”
After the meeting, Valentine and Stevens discussed what they had just
heard. They agreed that the three cofounders were bright and that the vision
to bring 3-D graphics to the PC platform had promise. Although they
themselves were not gamers, Sequoia had invested in Electronic Arts, the
publisher of software computer games that had recently gone public and
made Sequoia money. They were also invested in S3, the company that
primarily produced 2-D graphics-accelerator chips and which the Nvidia
cofounders thought they could beat, so they knew the market was viable.
Additionally, Valentine regretted passing on Silicon Graphics, which now
dominated the market for high-end graphics workstations.
Sequoia met with Nvidia’s cofounders two more times in mid-June. At
the last meeting, they decided to invest.
“Wilf says to give you money. Against my better judgment, based on
what you just told me, I’m going to give you money. But if you lose my
money, I will kill you,” Valentine told the Nvidia team.
Nvidia secured $2 million of Series A funding from Sequoia Capital and
Sutter Hill Ventures—$1 million apiece—at the end of the month.
 Nvidia now had enough money to fund the development of its first chip
and to start paying its employees. It was a humbling moment for Jensen,
Priem, and Malachowsky: they had succeeded on the strength of their
reputation, not their business plan or their demo. It was a lesson Jensen
would never forget. “Your reputation will precede you even if your business
plan writing skills are inadequate,” he said.

Chris Malachowsky and Jensen Huang in 1994. (Nvidia)

 PART II

NEAR-DEATH EXPERIENCES

(1993–2003)
 CHAPTER 4

All In

FINALLY, NVIDIA COULD STOP MERELY talking about its first chip
and start building it. The first order of business was moving the company
out of Priem’s townhouse and into a real office. With the money from Sutter
Hill and Sequoia, Nvidia could afford to rent a suite of offices inside a
single-story building that was located just off Arques Avenue in Sunnyvale.
The location was less than ideal—a nearby Wells Fargo bank would be
robbed several times during the company’s lease period—but it gave
Nvidia’s employees a sense of legitimacy.
For the first time, the company could also afford to pay its staff. Before
fund-raising, Nvidia had only a handful of employees, and they worked
without salaries, with the promise that at some point the money would flow.
Now, Nvidia went on a hiring spree, bringing twenty new people aboard to
fill both engineering and operations roles.
One such hire was Jeff Fisher, who was lured away from a graphics chip
maker called Weitek to run Nvidia’s sales department. During the interview
process he was impressed by each one of Nvidia’s cofounders.
“Great guys. All very different, but super smart,” he recalled. “Jensen is
an engineer at his core, but he could wear many hats. Curtis is an architect,
determined to solve the forward-backward compatibility unified
architecture. Chris could sling transistors like there’s no other.”
 Robert Csongor and Jensen Huang in front of Nvidia’s first office.
(ROBERT CSONGOR)

Robert Csongor, another of Nvidia’s earliest employees, was so excited

about his first day that he convinced Jensen to take a photo with him in
front of the Nvidia sign on the front door of the office.
“One day we’ll be big and famous,” insisted Csongor, “and this picture
will be cool.”
Before Nvidia increased its headcount, the three cofounders established a
chain of command. Priem and Malachowsky wanted to maintain the same
working relationship they had had at Sun: Priem would handle chip
architecture and products as the company’s chief technical officer, and
Malachowsky would run the engineering and implementation teams. They
simply assumed that Jensen Huang would make the business decisions.
“We basically deferred to Jensen on day one,” Priem said, telling him,
“you’re in charge of running the company—all the stuff Chris and I don’t
know how to do.”
Huang remembers Priem being even more direct: “Jensen, you’re the
CEO, right? Done.”1
With roles defined and project teams fully staffed, Priem launched into
the design for the NV1 chip. In the world of PC graphics, the constraints
were even harsher than on the Sun SPARCstation. Intel’s current generation
of CPUs, which powered most PCs, had difficulty performing the high-
precision “floating-point” math calculations that were helpful for graphics
rendering. Manufacturing capacity for chip designers was scarce and not
very advanced, which limited the number of transistors Nvidia could fit on
a single chip. And prices for semiconductor memory chips, which the
graphics accelerators needed to perform their increasingly complex
operations, were extremely high, at nearly $50 per megabyte, as a result of
rising demand for PCs.
Priem and his team planned to build a chip that could display graphics at
a resolution of 640 × 480 pixels, with high-quality textures and fast
rendering speeds. But they would have to invent their way around the PC’s
limitations. The biggest hurdle involved the cost of memory. If they used
standard chip-design methods on the NV1, the chip would need four
megabytes of onboard memory, at cost of $200. This alone was enough to
make any graphics card that used the chip unaffordable to most gamers,
who were used to much cheaper prices. Before this first era of powerful 3-D
PC chips, most 2-D-focused graphics chips cost under $10 and used limited
amounts of memory.
Priem attempted to solve the problem with a new software process for
handling textures, which was called forward texture mapping. The NV1
would render 3-D polygons by using quadrilateral shapes instead of the
traditional inverse texturing, which was based on triangles. The shift to
quadrilaterals would require less computational power and therefore lower
memory requirements. The only downside was a significant one: software
developers would have to completely rework their games in order for them
to take advantage of Priem’s forward texture mapping. If the NV1 tried to
run a game that was built around the older inverse texturing process, the
result would be slow rendering and poor graphical quality. However, Priem
was confident that in the fragmented world of PC video-game graphics,
where there was not yet a single dominant standard, Nvidia’s technically
efficient process would eventually win out.
As if inventing an entirely new texture-rendering process wasn’t enough,
Priem also wanted the NV1 to improve the audio capability of games. At
the time, the market leader in audio was the SoundBlaster sound card,
which to Priem’s ears produced unrealistic, tinny music. He added high-
quality wavetable synthesis to the NV1, which re-created digitized sound
from recordings of actual instruments, whereas the SoundBlaster’s audio
samples were completely synthetic.
 This alternative audio standard was another risky decision. Combining
graphics and audio in one card was an unusual move, as most computers
came with separate cards for each function. Yet Priem believed that this
meant there was a market inefficiency waiting to be corrected by a
technically superior multifunction card. The adoption of the new format
was not guaranteed: with such a strong incumbent in SoundBlaster, Priem
was banking on software makers switching from an inferior but widely
adopted standard in favor of a proprietary one that produced better audio
but that required more work to implement.
While Priem was working on the design, Jensen focused on convincing
Intel to support his new card. His contact at Intel was a young executive
named Pat Gelsinger, who was responsible for managing revisions to the
Peripheral Component Interconnect (PCI) expansion-slot standard for PCs
that all forthcoming graphics cards would use. Jensen wanted PCI to add
different types of throughput modes for the NV1 to take advantage of;
Gelsinger was resistant.
“I can remember vicious conversations that Jensen and I had on different
architectural points,” Gelsinger recalls.2
In the end, Jensen prevailed. Intel went with a more open standard, one
with better capabilities and that encouraged innovation. It was a victory not
only for Nvidia but also for the graphics industry as a whole—with an open
standard, peripheral card makers could dictate the pace of technological
improvements without having to wait for Intel to catch up. According to
Gelsinger, Nvidia owed its future success to “the open PCI platform that
enabled his graphics devices to really race ahead of everybody else.”
As the NV1’s design came into focus, Jensen and Malachowsky finalized
their partnership with the foundry that would be manufacturing all of their
chips, SGS-Thomson in Europe. Although Don Valentine and Mark Stevens
had been critical of SGS-Thomson’s suitability as a partner, Nvidia was
able to use the European chip-maker’s relative weakness as bargaining
leverage. Their agreement gave SGS-Thomson the exclusive license to
manufacture the NV1 chip for Nvidia and also to create a stripped-down
version of the NV1 that the foundry could resell as a mid-tier chip under its
own white-label brand. In exchange, the manufacturer would pay Nvidia
around $1 million per year to write regular software and driver updates for
all major Windows operating systems. SGS-Thomson essentially agreed to
fund Nvidia’s entire software division of around a dozen people in order to
secure the privilege of manufacturing the NV1 chip.3
In the fall of 1994, SGS and Nvidia presented the NV1 at COMDEX in
Las Vegas, one of the largest computer trade shows in the world. They
prepared three working prototypes installed in PCs. Right before the
convention opened, Priem and another engineer were still debugging the
software drivers and brought one prototype back to a hotel room to continue
to work on it. They decided to keep the other two machines at their booth. A
security guard strolling by recommended that the company hire someone to
guard their equipment overnight. The Nvidia team declined.
When they returned the next day to set up, they found that everything
was gone. The doors to the exhibition floor had been left unlocked, and
someone had walked in overnight and stolen their prototypes. Fortunately,
they still had one chip back in the hotel, and the NV1 made its official debut
as the conference opened.4
Amid the busyness of the trade show, the Nvidia team managed to secure
an introduction to representatives from the Japanese video-game and
console maker Sega.5 Impressed with the NV1 demonstration, Sega agreed
to begin working with Nvidia as it planned its next console. On December
11, 1994, Jensen and Curtis Priem flew to Tokyo to suggest a chip-
development deal to Sega management.6
It was the first step in what should have been a long and beneficial
relationship between the two companies. In May 1995, Sega and Nvidia
signed a five-year partnership, where Nvidia agreed to build its next-
generation chip, the NV2, exclusively for Sega’s next gaming console. In
return, Sega agreed to boost the NV1’s launch on PC by porting several
games originally developed for their current-generation console, the Sega
Saturn, and rewriting them to support the NV1’s forward texturing process.
Sega also purchased $5 million worth of Nvidia preferred stock.
When the commercial terms were settled, Curtis Priem took over as
Nvidia’s main point person with Sega, given the technology collaboration
the deal required. He would travel to Japan six times in 1995 to manage the
two companies’ joint projects. He oversaw the design specifications for the
NV2-based console, including how it would read game cartridges and
perform color compression. He also helped Sega understand the nuances of
porting their Saturn-based games to the PC.
 The NV1 had all the ingredients for a successful launch. It had a unique
marketing angle, as a single-chip multimedia accelerator with several new
texturing and rendering features. It had significant initial sales, including a
250,000-chip order from Nvidia’s primary board partner, a company called
Diamond Multimedia that packaged the chip in a $300 graphics card under
the “Edge 3D” brand. And it had a splashy launch partner in Sega, which
not only agreed to support the current chip but also had committed in
advance to Nvidia’s next one. The chip was officially announced in May
1995, and the entire company expected it to be a runaway success.

BUT NVIDIA HAD GRAVELY MISJUDGED the market. For one, over
the previous two years, memory prices had plummeted from $50 per
megabyte to $5 per megabyte, which meant that the NV1’s stinginess with
onboard memory was no longer much of a competitive advantage. As a
result, few game developers saw the need to rewrite their software to
support Nvidia’s new graphics standard. Sega’s PC ports, which included
Virtua Fighter and Daytona USA, ended up being some of the only titles
specifically designed to run on the NV1. Just about every other game ran
poorly on Nvidia’s new chip, which used an intermediate software wrapper
to perform inverse texturing and thus was prone to slow rendering.
It was a single game, the first-person shooter DOOM, that sealed the
NV1’s fate. At the time of the chip’s launch, DOOM was the most popular
game in the world: its kinetic visuals and gruesome, fast-paced combat
were unlike any other gaming experience ever produced. This was in large
part due to the technical wizardry of John Carmack, the game’s designer
and cofounder of its publisher, id Software. Carmack built the game using
the 2-D Video Graphics Array (VGA) standard and leveraged every
hardware-level trick he knew for maximum visual impact. Priem had been
sure that most game designers would switch to the NV1’s 3-D accelerated
graphics and leave VGA behind. So the NV1 chip only partially supported
VGA graphics and relied on a software emulator to supplement its VGA
capabilities—which resulted in slow performance for gamers playing
DOOM.
Even DOOM’s iconic soundtrack and sound design didn’t work properly
on the NV1. The chip’s proprietary audio format, which Priem had included
more as a flourish than a strict necessity, was not compatible with the
industry-standard SoundBlaster format made by sound-card maker Creative
Labs. However, most PC manufacturers required their peripherals to be
SoundBlaster compatible, and that wasn’t changing as quickly as Priem had
expected. To work around this, Priem wrote yet another emulator, this one
designed to produce sound rather than visuals. But it would break every
time Creative Labs updated its proprietary format and would stay broken
until Nvidia could follow up with a patch. NV1 users would have to endure
long periods where the sound on their games didn’t work properly.
It was a hard lesson in the value of backwards compatibility and the
dangers of innovating for innovation’s sake. Nvidia’s new card, which was
supposed to push the boundaries of the graphics industry, could not keep up
with the world’s most popular game. It was sunk by a lack of truly
compatible games and the ongoing support from most game makers for
inferior, though widely adopted, technical standards.
“We thought we had built great technology and a great product,”
Malachowsky said. “It turns out we only built great technology. It wasn’t a
great product.”
Sales were dismal, and most of the units sold during the holiday season
were returned. By the spring of 1996, Diamond Multimedia had returned
nearly all of the 250,000 chips it had ordered.
Jensen realized Nvidia had made several critical mistakes with the NV1,
from positioning to product strategy. They had overdesigned the card,
stuffing it with features no one cared about. Ultimately, the market simply
wanted the fastest graphics performance for the best games at a decent price
—and nothing else. Computer manufacturers also told Nvidia that
combining video and audio functionality onto one chip made it harder for
Nvidia to win a contract.
“The irony was that the thing that killed NV1 wasn’t the most important
thing, which was the graphics,” said Michael Hara, Nvidia’s director of
marketing at the time.7 “It was the audio. Games back then needed
SoundBlaster compatibility, and NV1 didn’t have it.”
“We really like your graphics technology, so if you guys ever want to get
rid of the audio, come back and see us,” Hara recalled being told several
times.
The NV1 could simply not stack up against other cards that were more
narrowly designed. Nvidia saw that it couldn’t again build things customers
would not pay extra for.
“We were diluted across too many different areas,” Jensen recalled.8 “We
learned it was better to do fewer things well than to do too many things
even though it looked good on a PowerPoint slide. Nobody goes to the store
to buy a Swiss Army knife. It’s something you get for Christmas.”9
Nvidia had spent nearly $15 million to develop the NV1. That money had
come from the initial investment from Sutter Hill and Sequoia, as well as
from SGS-Thomson and Sega.10 The company was counting on strong sales
of the NV1 to recoup most of its development costs, so that it could move
on to the next chip. The bad result, however, meant that Nvidia was now
facing a cash crisis. Jensen, Priem, and Malachowsky needed to secure
more money, and soon, or else their dream would come to an abrupt, and
self-inflicted, end.

DURING ONE OF NVIDIA’S VERY FIRST board meetings, director

Harvey Jones, a former CEO of a leading chip-design-software company
called Synopsys, asked Jensen about the NV1: “How would you position
this?”
At the time, Jensen didn’t realize that Jones was not merely asking about
the NV1’s feature set or product specifications. He was asking him to
consider how Nvidia would sell the new chip in a highly competitive
industry. He knew that products had to be presented in the clearest, most
precise terms in order to stand out.
“He asked me a simple question. I had no idea how simple it was. It was
impossible for me to answer because I didn’t understand it,” Jensen
remembered.11 “The answer is supremely deep. You’ll spend your whole
career answering that question.”
In the aftermath of the NV1’s failure, Jensen regretted not taking Jones’s
question a little more seriously. It frustrated him that he and the Nvidia
team had put in so much effort for so little reward, and he believed that it
came down to his own shortcomings as the leader of a new company.
“We were just bad at our jobs,” he said. “The first five years of our
company. We had really talented people, working super, super hard, but
building a company is a new skill.”
 Jensen vowed that he would absorb as much as he could about leading a
business to prevent himself, and his fledgling company, from ever making
the same mistakes again. In his search for answers to Jones’s question, he
gravitated to the book Positioning: The Battle for Your Mind by Al Ries and
Jack Trout. In it, Ries and Trout argue that positioning is not about the
product itself but rather about the mind of the customer, which is shaped by
prior knowledge and experience. People tend to reject and filter out
anything that doesn’t align with their existing worldview, which makes it
hard to change their minds with reason and logic. But emotions can change
quickly, and a skillful marketer can manipulate people to feel a certain way
about a product, if a company uses the right message. According to the two
authors, potential buyers didn’t want to be persuaded. They wanted to be
seduced.
But seduction requires a simple message, and Nvidia’s message with the
NV1 was far too complicated. It wasn’t superior to the competition in any
obvious way, and it was actually inferior under some circumstances.
“The customer’s always thinking of alternatives,” Jensen said. And in the
customer’s mind, the alternatives could do what the NV1 could not—they
could play DOOM. No amount of complaints about how the game used
older graphics standards, or didn’t to take advantage of the NV1’s
performance-boosting capabilities, would offset that one, easy-to-
understand, negative message. No matter how many times Nvidia pointed
to the NV1’s innovative audio and graphical capabilities, it could not
counteract what gamers saw with their own eyes and heard—or didn’t hear
—with their own ears.

THE NV1 DISASTER JEOPARDIZED THE company’s relationship

with Sega. The Japanese company had commissioned Nvidia to build the
NV2 for its next console after the Saturn and be a follow-up to the
successful, earlier Genesis console. The code name for the NV2 within
Nvidia was “Mutara,” after the location of the climactic space battle in Star
Trek II: The Wrath of Khan—during which the Genesis device fires,
collapsing the Mutara Nebula into a new, life-bearing planet. In the same
way, Nvidia now needed its NV2 chip to breathe new life into the
struggling company.
 From the start, things did not look promising. In spite of Priem’s direct
involvement and his several trips to Japan, Sega’s programmers
increasingly soured on Nvidia’s proprietary graphics-rendering technology.
In 1996, Sega informed Nvidia that the company would no longer be using
the NV2 in its next console. But Jensen had deftly worked into the initial
contract a clause for a $1 million payment from Sega if Nvidia was able to
produce a working prototype of a chip that could be installed onto a self-
contained motherboard that was about the same size as the older Sega
Genesis/Mega Drive motherboard.
Priem assigned a lone engineer, Wayne Kogachi, to build the NV2
prototype. It was an isolating and thankless job. Kogachi had only a single
chip and a motherboard to play with, and Priem had assigned the rest of the
engineering team to the company’s next chip, which was then called NV3.
The few interactions Kogachi had with his colleagues often involved
juvenile, late-night shenanigans, such as the time when the entire
engineering division started measuring and recording the circumference of
everyone’s head, a sort of jocular phrenology.
“Wayne had the largest circumference of anyone at Nvidia at the time,”
Priem remembered, with a laugh.
After about a year spent on the project, Kogachi was able to get an NV2
prototype working within Sega’s specifications. The milestone triggered the
$1 million payout, money that was a key lifeline during a time of crisis.
Still, it did not solve all of Nvidia’s woes. The majority of the $1 million
was immediately put into research and development on the NV3, and there
was not enough left over to pay for the many employees who had been
hired in anticipation of selling huge amounts of NV1s and NV2s, and who
had nothing to do now that both chip projects were essentially dead. To
conserve the company’s remaining cash, Huang elected to lay off the
majority of the staff: Nvidia went from more than one hundred employees
to forty.12
“We had a marketing team, we had a sales team, and all of a sudden we
had a road map that was no longer viable,” said Dwight Diercks, a software
engineer who survived the culling.
While Nvidia reeled from its missteps with the NV1 and NV2 and was
pivoting to focus on the NV3, a formidable new competitor had emerged in
the PC graphics market. Three alumni of Silicon Graphics, Scott Sellers,
Ross Smith, and Gary Tarolli, founded the company 3dfx in 1994, just one
year after Nvidia’s incorporation. In the 1990s, Silicon Graphics, or SGI,
was best known as a manufacturer of high-end graphics workstations used
for computer-generated movie effects, including the dinosaurs in Steven
Spielberg’s Jurassic Park. The founders of 3dfx intended to bring that same
level of performance to the PC market at a price gamers could afford. In the
fall of 1996, after two years of development, the company announced that it
was ready to launch its first graphics chip, branded as “Voodoo Graphics.”
3dfx decided to unveil Voodoo Graphics at a conference held by the tech-
focused investment bank Hambrecht & Quist in San Francisco. There, an
executive named Gordon Campbell planned a session that would
demonstrate how 3dfx’s chip could produce high-end, enterprise-grade
graphics on low-end, consumer-grade equipment. The centerpiece of his
demo was a 3-D cube, rendered with such precision that it could have
passed for something made by an SGI workstation.
“I was down in the basement in a little tiny room with a PC, a projector,
and our first chip on a card,” Campbell said.13
The 3dfx session was scheduled to take place at the same time as a
keynote session that featured Silicon Graphics’ CEO, Edward McCracken.
Initially, Campbell’s demonstration was sparsely attended, as most people
who might have been interested were listening to McCracken walk through
SGI’s corporate history. But partway through his presentation, McCracken’s
SGI workstation, which retailed for $85,000, crashed and brought the
keynote to a complete halt. As the crowd grew restless, word began to
trickle in that the downstairs session was more compelling—that a small
start-up had managed to engineer 3-D graphics on par with SGI’s machines,
but on a consumer PC card.
“People were kind of blown away,” said Campbell. “There were all going
you got to see this and dragging people in.”
The dueling sessions not only became part of 3dfx corporate lore but also
informed the marketing message for Voodoo Graphics, which went on sale
in October 1996. 3dfx pitched itself as the only start-up that could bring
SGI-level performance to personal computers, at a fraction of the cost.
These themes were reinforced everywhere in the launch materials for
Voodoo Graphics, such as a quote that Ross Smith, 3dfx’s head of
marketing, provided to a company called Orchid that featured the Voodoo
Graphics chip on their Righteous 3D graphics card.

Last year at Comdex [sic], Bill Gates played The Valley of Ra in the Orchid booth on a quarter
of a million dollar SGI Reality Engine-based Voodoo Graphics simulator. That same real-time
3D graphics performance is now available to PC consumers for $299 from Orchid. That’s
Righteous!14

Knowingly or not, 3dfx followed the exact principles laid down by Al

Ries and Jack Trout in Positioning. The company pitched its product as a
clear alternative to the other cards in the marketplace and appealed to its
customers’ emotions—the feeling of “beating the system” by getting
outsized performance at a great price—rather than try to convince them
with facts and performance statistics.
3dfx was offering more than just marketing puffery. In June of 1996, id
Software launched the title game in its new series of first-person shooters,
Quake. Just like DOOM had done three years prior with 2-D cards, the
original version of Quake pushed the capabilities of 3-D graphics cards to
their limits, in this case by rendering everything in real-time 3-D. In
January of 1997, id Software released an updated version of Quake, dubbed
GLQuake, which added support for 3-D graphics hardware acceleration—
the very feature that the Voodoo Graphics chip excelled at.
“Our sales just went crazy,” 3dfx’s chief engineer, Scott Sellers,
recalled.15
The company’s revenues exploded, from $4 million in the 1996 fiscal
year, to $44 million in 1997, to $203 million in fiscal 1998 after the release
of the upgraded Voodoo2 graphics card. The vast majority of this demand
came from Quake gamers; it was the killer app that motivated buyers to
upgrade their hardware so that they could get better graphics performance
and quality, all to make their gaming experience more immersive.
3dfx’s executives knew that Nvidia was under significant financial strain
and considered making an acquisition play for their fading rival. Even
though its first two chips had failed to gain traction, Nvidia still had some
of the best graphics engineers in Silicon Valley on its payroll. In the end,
however, 3dfx executives opted not to make a move. Its executives believed
that Nvidia’s bankruptcy was inevitable, and that it would be cheaper to
wait until Nvidia collapsed so that they could pick up its talent and assets
for a bargain.
“The mistake that we made at 3dfx is we should have killed them when
they were down,” said Ross Smith. “That was a huge tactical error on our
part, not buying them. We had them on the ropes.”
“We were very aware if the RIVA 128 chip had come back and there
were any bugs in it, they would be dead,” said Sellers, referring to the chip
that started as the NV3 and would later be sold as the RIVA series. “They
had no time. We were betting that if we just waited a little bit, they were
going to implode on their own.”
They did make one attempt to push Nvidia over the edge. Sellers had
previously worked with Dwight Diercks at another small start-up and knew
that the Nvidia engineer was exceptionally skilled at programming software
drivers for graphics chips, which are critical to a card’s success. Sellers
aggressively courted Diercks, trying to convince him to abandon the sinking
Nvidia and join the ascendant 3dfx.
“We were so close to getting him,” Sellers later said, with more than a
hint of regret in his voice.16
Diercks, for his part, seriously considered the opportunity.17 But he
remained at Nvidia for two reasons. One was curiosity: he wanted to see the
RIVA 128 through to production before he considered leaving. The other
was Jensen, who talked with Diercks and convinced him to stay. To this
day, Jensen claims that he “saved” Diercks; in turn, Diercks likes to joke
that if he had left, 3dfx would have bought Nvidia outright. Diercks is still
at Nvidia three decades later, overseeing the company’s software
engineering.

JUST AS QUAKE WAS PUSHING 3dfx to new heights, Jensen Huang

and Nvidia counted their dwindling reserve of cash and wondered whether
it was enough to see their next chip through to production. With $3 million
left in the bank, the company could afford to operate for nine more
months.18 To survive, it had to do more than just create an above-average or
merely good chip. It had to create the fastest graphics chip possible with the
manufacturing and memory technology available—something that could
win against 3dfx’s excellent Voodoo line.
 To beat such a formidable competitor, Nvidia would have to rethink its
entire approach to chip development. The NV1 had been designed to do
what Nvidia engineers wanted rather than what the market wanted. The
proprietary standards that Curtis Priem had included on the chip showcased
his technical acumen but ended up alienating manufacturers. In June 1996,
Microsoft made it even harder for new graphical standards to gain traction
when it released Direct3D, an application programming interface (API) for
graphical texturing that used the traditional inverse triangle approach.
Within a few months, game developers almost universally abandoned
small-scale proprietary graphics standards such as Nvidia’s in favor of one
of the two big and well-supported alternatives, Microsoft’s Direct3D and
OpenGL.
Jensen saw where the industry was headed and demanded that Nvidia’s
engineers follow the market rather than fight it.
“Guys, it’s time to quit polishing the turd,” he told his remaining
employees.19 “At this point it is clear we’re doing it the wrong way and
nobody’s going to support our architecture.”20
Malachowsky agreed with the new approach. “Instead of trying to be
smarter than the competition with a different technology like NV1, we just
had to out-engineer everybody else using the same basic tactics,” he said.
Jensen’s message inspired Priem to go big, literally, with the NV3. In
order to make a much faster chip, the team wanted to use a wide 128-bit
memory bus and design a graphics pipeline that would be able to generate
pixels at record-breaking speeds. Nvidia would have to produce a chip that
was physically larger than any that had ever been successfully
manufactured.
Priem cornered Jensen in a hallway in the Nvidia offices to ask for his
approval, despite the technical challenges involved.
“Let me think about it,” Jensen answered. He took the next two days to
map out the schedule, pricing, manufacturing plans, and business model for
the revised NV3. In the end, he not only approved the larger-sized chip; he
also wanted Priem and his engineers to add another 100,000 gates, or about
an extra 400,000 transistors—for a total of 3.5 million transistors on the
entire chip.21
“Jensen gave us the green light to fill up the chip with even more
functionality,” Priem said.
 “I wasn’t worried about my cost,” Jensen said years later, when asked to
explain his decision-making process. “I built a chip that physically was as
large as anyone could build at the time. We just wanted to make sure this is
the most powerful chip the world’s ever seen.”
As a nod to Nvidia’s ambitions—and, perhaps, as a way to signal a clean
break with its past design philosophy—the company decided to give the
NV3 an external brand that was different than its internal code name and
dubbed it the RIVA 128, which encapsulated the chip’s ultimate purpose:
RIVA stood for Real-time Interactive Video and Animation Accelerator, and
“128” was a nod to the 128-bit bus, which would be the largest ever
included on a single chip—another first for the consumer PC industry.
Given its financial position, Nvidia would have to make the RIVA 128 in
record time, and without the safety net of multiple quality-assurance runs.
Standard chip development usually spans two years, involving multiple
revisions to identify and fix bugs after a chip “tape-out,” when a finalized
chip design is sent for prototype manufacturing. The NV1, for example, had
three or four physical tape-outs. Nvidia could afford just one physical tape-
out for the NV3 before the company had to send it to production.
To shorten the timeline, Nvidia would have to shorten the testing cycle.
Jensen had heard about a small company, Ikos, that made refrigerator-sized
chip-emulation machines. The massive machines enabled engineers to run
games and tests on digital chip prototypes, bypassing the need to make an
actual chip for testing and fixing bugs, and thus saving time and resources.
An Ikos machine wasn’t cheap: a single one cost $1 million and thus would
cut Nvidia’s payroll runway from nine to six months—but Jensen realized it
would speed up the testing process by significantly more time. He argued
the point with the company’s other executives, who wanted as much time as
possible by raising more money. The CEO stood firm.
“We’re not going to find more money,” he said. Venture capitalists had
“ninety other companies to believe in. Why believe in us? We got to do it
this way.”
Jensen won the argument and bought an emulator from Ikos. As soon as
it arrived, Diercks and his software team began running tests on the digital
RIVA 128 in order to identify and fix issues with the chip. Diercks
remembered that the first conversation between his team and the hardware
engineers was a disaster.
 “Hey guys, we can emulate for the first time for our chip,” he said. “It’s
just booted DOS, very slow.”22
One of the hardware guys said, “Yeah, look at that. It’s already got an
error. The C colon was off by two pixels.”
Normally, chips don’t error out so quickly after boot up or on something
so basic. As a result, the hardware team assumed that the emulator wasn’t
working properly and that Jensen and Diercks had just wasted three months
of payroll cushion for no good reason.
“But it was actually the first bug we found in our hardware,” Diercks
said.
Working with the Ikos machine was an arduous process. The setup
consisted of two large boxes connected to a motherboard that was exposed
to the open air. Instead of a chip, wires were plugged into the socket to
transmit the data that, with a physical chip, would be sent from the chip to
the CPU. The software-emulated chip was much slower than a real
hardware chip.
“It took fifteen minutes to load Windows. I remember moving the mouse
just a tiny bit and having to wait for the screens to refresh frame by frame,”
tester Henry Levin said. “Trying to click a button was a nightmare because
if you moved it slightly, it would overshoot.”23
Levin drew a map on his desk so he would know where to place the
mouse to access specific parts of the screen without needing to wait for the
emulator to refresh every frame. The testers would run basic utilities, such
as drawing a triangle or circle. The running of benchmarks often meant
leaving the machine on overnight and returning the next morning to see
whether it had finished.
The emulator did not produce any bug reports automatically. Instead,
when a program froze, all Levin could do was take a screenshot and call
over one of the hardware engineers to figure out what happened or where
the corruption occurred. If it was a significant problem, the engineers would
go back to redesign a part of the chip.
One engineer recalled that the team attempted to run a longer benchmark
over a weekend. Somehow, an overnight cleaning crew gained access to the
testing lab and unplugged the emulator in order to plug in a vacuum cleaner.
The engineers returned to find their benchmark test completely ruined; they
would have to run it again from the start, which cost time. The cleaning
crew didn’t even need to be in there, as the lab did not have a carpet.
Unnecessary janitorial services were not the only challenge the team
faced. Nvidia didn’t have the time to start from a completely clean-sheet
design. So Priem, Malachowsky, and chip architect David Rosenthal
figured out a way to reuse parts of the NV1 but add support for multiple
new features, including inverse texturing, better math capabilities, and a
very wide memory bus. Even if Nvidia wanted to make a clean break from
its first chips, the DNA of those early designs would persist in the RIVA
128.
“We got it to work,” Malachowsky said.24
The company also now knew that its chips needed to have 100 percent
hardware support for the old VGA standard. The NV1 had tried to get by
with a solution that was half hardware, half software emulator, but this
approach caused significant issues in many DOS-based games, including
DOOM. Nvidia couldn’t afford to fall short on VGA once again.
But the company didn’t have anyone on staff with the expertise to design
a VGA core in-house. Incredibly, Jensen was able to source and license a
VGA core design from one of Nvidia’s competitors, a company called
Weitek.
“Jensen is the best deal cutter in the world. Hands down,” Priem said.
“Somehow Jensen is able to come up with these amazing business deals for
Nvidia that save the company again and again.”
Not only did Jensen sign a licensing agreement with Weitek—he was also
able to poach its VGA chip designer, Gopal Solanki, who became a project
manager and one of the CEO’s top lieutenants. A former Nvidia employee
said that they worked like “business soulmates.” Solanki was known for
being extremely tough and demanding—and for delivering. Jensen, for his
part, gave Solanki credit for saving the company.
“Gopal is really important,” Jensen said nearly thirty years later. “If not
for Gopal, we’d be out of business now.”25
“You always had a good feeling when Gopal was assigned to the next
generation of NV chips. You would know things would turn out okay,”
Priem agreed.
Nvidia unveiled the RIVA 128 in April at the 1997 Computer Game
Developers Conference. The conference was where hardware companies
showcased their latest products, in hopes of securing orders from PC
manufacturers and retailers. Nvidia’s timelines were so tight that it was
uncertain whether the chips would arrive in time for the conference—or
whether they would be in good enough shape to showcase. The samples
from the factory came only a few days before the event, and Nvidia
engineers worked feverishly to troubleshoot any software bugs that
remained. Their goal was to ensure the chips could run the Direct3D
graphics benchmark that hardware manufacturers would use to evaluate
their quality. Mere hours before the show started, the engineers managed to
get the chips stable enough to work without crashing at unpredictable
moments.
“Our dirty little secret was the RIVA 128 would only run that particular
test once and it did so tenuously,” said Eric Christenson, a regional sales
manager for Nvidia who attended the 1997 conference.26 “You had to treat
it with the utmost care and respect. It would just as likely lock up the
system mid-test if you looked at it the wrong way.”
Representatives from rival graphics-card manufacturers visited Nvidia’s
booth, mainly for an opportunity to heckle the company about its NV1
failure.
“Oh, you guys are still around?” one 3dfx employee said.
But enough people saw Nvidia’s benchmark tests—and were impressed
enough by the results—that RIVA 128 began to attract buzz. The industry
collectively understood that it might be something special. Near the day’s
end, 3dfx cofounder and head of engineering, Scott Sellers, approached the
booth and asked for a demonstration.
“In order to give the best experience, I’m going to power-down this
system to let you see a clean run,” said Christenson. “We’ll just re-boot the
system, open the app, and run the demo.”
Although he tried to sound as matter-of-fact as possible, Christenson was
taking a gamble in order to impress his rival. The chips were especially
prone to crashes after a reboot, so he couldn’t be sure that the device would
come back up. At the same time, if he ran a test without rebooting the
device, Sellers could claim that the benchmarks were less accurate than
they would have been otherwise.
Christenson held his breath while the device restarted. It came back
online without crashing. He ran the benchmark test. The results appeared on
the PC display. Sellers didn’t—couldn’t—believe them. They were not only
better than 3dfx’s benchmarks but also higher than anything Sellers had
ever seen from a consumer graphics card. Christenson assured him that the
test results were unassailable. Sellers realized the implications of the test.
First, the RIVA 128 could outperform 3dfx’s best cards; and second, the
company that 3dfx had left for dead was instead about to come roaring back
into the 3-D graphics market.
Walt Donovan, the chief architect at another 3-D graphics start-up, called
Rendition, also came over and looked at the RIVA 128’s test results. He
asked Nvidia’s relatively new chief scientist, David Kirk, a series of
questions about the chip and its performance. Listening to Kirk’s answers,
all Donovan could say was, “That’s amazing.” None of Donovan’s projects
would ever come close to the performance of the RIVA 128. In the span of a
single benchmarking test, his company had gone from competitive to failed.
Once he processed the situation, Donovan asked one more question.
“Can I have a job at Nvidia?” He was hired shortly thereafter.

WITH A WORKING PROTOTYPE CHIP that produced strong

performance numbers, Jensen now had leverage to raise more money.
“We did not want Sutter Hill’s or Sequoia’s money early,” Priem said. If
Jensen had gone back right after the NV1 or NV2 fiascos, when Nvidia had
no clear path forward, he would have faced a skeptical audience demanding
unfavorable investment terms—if they agreed to invest any additional
money at all. Now, however, the venture capital firms were highly
motivated to keep the company going just when it was on the verge of
possible success. Jensen asked for another investment round so that he
could buy chips from the fabs. Both funds agreed to invest again, with
Sutter Hill committing $1.8 million on August 8, 1997.27 (I have not been
able to determine how much Sequoia invested in this round, only that it did.
Sequoia did not offer the information after requests.)
In late summer, Jensen gathered the whole company in the office
cafeteria. He pulled a piece of paper from his pocket and read off some
dollar figures, down to cents. He folded the paper back into his pocket and
said, “That’s how much money we have in the bank.”
The room fell silent. The number was not very large at all—barely
enough to pay everyone’s salary for a few more weeks. One recent hire
remembered being brought to the verge of panic. “My God,” he thought,
“we’re almost out of cash.”
Jensen then pulled out another piece of paper from his pocket. He opened
it up and read, “One purchase order from STB Systems for 30,000 units of
RIVA 128.” It was the chip’s first major order. The cafeteria erupted in
cheers. Jensen had indulged in a bit of showmanship for dramatic effect.
The RIVA 128 was the company’s first big hit. The stellar reviews it
received upon release more than erased the bad memories from the NV1
launch.
“Whoever is a die-hard gamer will have to buy this card,” Tom’s
Hardware, a leading tech enthusiast website, said. It is “the fastest 3D chip
for PCs currently available.”
Within four months of the chip’s release, Nvidia had shipped more than a
million units and captured a fifth of the PC graphics market. PC Magazine
named the RIVA 128 an Editors’ Choice product, and PC Computing named
it the Product of the Year for 1997.28 Large PC manufacturers, including
Dell Computer, Gateway 2000, Micron Electronics, and NEC, all
incorporated the chip into their computers for the holiday season. The torrid
pace of sales allowed Nvidia to turn a $1.4 million profit in the fourth
quarter of 1997—the company’s first profitable quarter since its founding
four years earlier.
Jensen’s flair for the dramatic was apparent again at a company meeting
near the end of the year. In those days, he still preferred to wear sport coats
and jeans; he had yet to start donning his signature black leather jackets.
From his front coat pocket he produced a thick envelope that was stuffed
full of crisp new $1 bills. He walked around the room, giving each
employee a single bill from the envelope, as a symbol of the financial
lifeline that the RIVA orders had given them—and a reminder that their
situation was still too precarious for a lavish celebration.
Then, he went back to a woman named Kathleen Buffington, who
worked in operations and was responsible for packing graphics chips and
shipping them out to customers. He had already given her one dollar; now,
he gave her a second. He said to the whole company that she had worked so
hard to get all the chips out the door that she deserved a double bonus.
Jensen’s distribution of dollar bills was a much-needed moment of levity
and celebration for a company that had been operating on the brink of
failure for years. “The RIVA 128 was a miracle,” Jensen said. “When our
backs were against the wall, Curtis, Chris, Gopal, and David Kirk built it.
They made really good decisions.”29
 CHAPTER 5

Ultra-Aggressive

THE RIVA 128 DID MORE THAN ensure Nvidia’s survival. It also
served as a magnet for talent, drawing people from across the relatively
insular world of computer graphics to a small office park in Sunnyvale,
where they believed they would get a chance to work on something
extraordinary.
Caroline Landry was a chip designer for the Canadian company Matrox
Graphics when she first heard about Nvidia’s new chip. “I was in my late
twenties. I wasn’t completely up-to-date on all the industry trends, but I
knew Nvidia had released the first RIVA, which had taken the industry by
storm. It was way ahead of a product I was working on at Matrox that
wasn’t even close to taping out,” she said.1
Her boyfriend had recently found a job in the Bay Area, but Landry
wasn’t sure if she wanted to join him. Until, that is, a headhunter put her in
touch with Nvidia, and she flew down for a full day of interviews. She was
offered a job right away and accepted it just as quickly, solely on the
strength of Nvidia’s reputation. She was the company’s first female
engineer.
When she started, she had trouble adjusting to Nvidia’s intense culture.
She would often work until eleven at night during the week, and she would
put in full days nearly every weekend, too. She recalls that one time an
executive checked in on a late Friday afternoon to ask about her work
objectives for that weekend. “Canada was a good place to recruit from
because engineers there were paid much less than engineers in the U.S.,”
she said. “But quality of life is generally more important to Canadians.”
Landry mentioned to Jensen that some employees were griping about the
long work hours. His response was typically direct.
“People who train for the Olympics grumble about training early in the
morning, too.”
Jensen was sending a message: long hours were a necessary prerequisite
for excellence. To this day, he has not deviated from that view or altered
Nvidia’s expectation that employees adopt extreme work habits.
Landry also noticed Nvidia’s managers were quick to recognize special
talent. She joined Nvidia around the same time as Jonah Alben, a young
engineer barely out of school who was, as she put it, obviously “brilliant.”
Jensen saw Alben’s potential early, saying at a company meeting, “in 20
years I expect I’ll be working for Jonah.” Initially, Landry felt a bit jealous
about the attention her colleague was receiving but got over it. “At Nvidia,
you embrace your smart colleagues and don’t feel threatened. This is not
about your ego. This is about whether we make it or not. Be grateful that
you have people like that to work with,” she said. Alben would later rise
through the ranks to become the head of graphics processing unit (GPU)
engineering.
Jensen insisted that new hires should know exactly what they were
getting into the moment they walked in the door.2 He tasked Michael Hara,
Nvidia’s director of marketing, with giving a frank talk at every orientation
session. As Hara recalls, his speech was intended to encourage new arrivals
to not be afraid of speaking up and to offer fresh perspectives and new ideas
at every opportunity.
“We’re ultra-aggressive,” he told the new employees. “We don’t waste
time finding excuses for why things don’t work. We move on. If you came
here thinking you can just hide in the back, collect your paycheck, and go
home at five, you’re mistaken. If that’s what you think, you should resign
today.”
Hara remembers that the human resources employee who was handling
the new-hire class looked aghast. He continued his speech, undeterred.
“We don’t do things like anybody else. If you come here and say, ‘This is
how we did it before,’ we don’t care. We’re about doing things differently
and better. When we were just twenty-five people, Jensen taught us to come
here, take risks, do things outside the box, and make mistakes. I encourage
you to do all three. But don’t make the same mistake twice, because we will
fire you in a heartbeat.”
Hara meant it, too. John McSorley, Nvidia’s former head of human
resources, said that the company had a policy of hiring quickly—but also of
firing fast if a new employee wasn’t working out. Jensen’s primary
guidance to all of his hiring managers was simple: “Hire someone smarter
than yourself.” However, as Nvidia grew and began adding new employees
at a pace of more than one hundred a month, executives understood that
they would occasionally make the wrong decision. Better to correct those
mistakes as soon as possible than to let them fester and harm Nvidia’s
culture.
In Nvidia’s early years, even the longer-tenured employees could never
feel totally secure because the company had adopted an “up or out”
approach, with people either getting promoted on a regular cadence or
getting pushed out to make room for someone with greater potential. The
company handled personnel in the same uncompromising way that it
approached chip design.

SINCE NVIDIA’S FOUNDING, JENSEN HAS insisted that all Nvidia

employees work at the “Speed of Light.”3 He wants their work to be
constrained only by the laws of physics—not by internal politics or
financial concerns. Each project must be broken down into its component
tasks, and each task must have a target time-to-completion that assumes no
delays, queues, or downtime. This sets the theoretical maximum: the
“Speed of Light” that it is physically impossible to exceed.
“Speed of Light gets you into the market faster and makes it really, really
hard, if not impossible, for your competitors to do better,” a former Nvidia
executive said. “How fast can you do it, and why aren’t you doing it that
fast?”
It was not just a rhetorical question—Jensen used this measure to gauge
the performance of his employees. He would reprimand subordinates who
set goals that referred to what the company had already done before or what
the competitors were doing in that moment. As he saw it, he needed to
prevent the kind of internal rot that he observed at other companies, where
employees often manipulated their projects to provide steady and
sustainable growth that would advance their individual careers, when in
reality they were making only incremental improvements that actually hurt
the company in the long term. The “Speed of Light” notion ensured that
Nvidia would never tolerate such sandbagging.
“The theoretical limit of what you could do—that’s what Speed of Light
is. That’s the only thing we were allowed to measure against,” remembered
former executive Robert Csongor.
The RIVA 128 was a prime example of “Speed of Light” project
planning. Jensen had confronted two facts: most graphics chips take two
years to get from concept to market, and Nvidia had only nine months.
During the planning phase, Jensen had asked software engineer Dwight
Diercks, “What’s the main limiting factor in getting a graphics card to
market?”
Diercks responded that software drivers—the specialized programs that
enable the operating system and PC applications to interface with and use
the graphics hardware—were the primary obstacle, because they needed to
be completely ready by the time the chip was prepared for mass production.
In traditional production processes, the first step was to build a physical
prototype of the chip. Once it was complete, software engineers could begin
work on the drivers and fix any bugs they encountered. Then the chip’s
design was optimized at least one more time to accommodate the new
driver.
To save time, Jensen decreed that Nvidia would have to develop the
driver software for the RIVA 128 before the prototype chip was completed
—a reversal of the customary process. This would shave nearly a year off of
the production timeline, but it would require the company to find a way to
bypass the step of testing the software on physical chips. That was why
Nvidia invested $1 million in its Ikos emulator, even though every dollar
was precious: it would allow them to approach the “Speed of Light.”
(Later on, in 2018, Jensen considered replacing “Speed of Light” with a
metaphor suggesting something even faster than light—a physical
impossibility. He was frustrated with the increasing slowness across the
organization as it grew in size. He yelled at his executive staff that they had
to move faster than light, and then turned to Robert Csongor. “Rob, what’s
the Star Trek: Discovery propulsion system that allows them to
instantaneously travel somewhere?”
 “Well, Warp Drive is faster than the speed of light, but I think you’re
referring to the Mycelium Spore Drive,” Csongor responded.4
Both Jensen and Csongor were Star Trek nerds. Jensen shouted, “The
Spore Drive! We need to be like the Spore Drive.” Everyone started
laughing. They decided to stick with “Speed of Light” because it was an
easier concept to explain than the instantaneous “Mycelium Spore Drive.”)
Nvidia pushed against the limits of what was possible with the RIVA
128’s development process in other ways, too. The staff created a chip
larger than any that had ever been designed and then crammed it full of
even more transistors than they originally intended in order to improve
performance. They licensed VGA technology from their competitor so that
they would not have to build lower-priority components from scratch.
Jensen ruthlessly recruited the top engineers from rivals and even Nvidia’s
partners, including Weitek. All of this happened because Nvidia employees
did not let themselves think about what was likely to work or what they
could reasonably achieve. They only cared about what would be possible
with the maximum amount of effort and minimum amount of wasted time.
Much of what the company learned on the RIVA 128 became standard in
its future chip development. From that point on, Nvidia had software
drivers ready at the beginning of chip production: the drivers would already
have been tested across all the important applications and games and to
ensure compatibility with prior Nvidia chips. This approach became a
significant competitive advantage for Nvidia, whose rivals had to develop
separate drivers for different chip-architecture generations.5
Nvidia also decided to handle the maintenance of graphics drivers rather
than rely on PC makers and board partners to push out updates on their own
schedules. The company distributed new drivers each month. Jeff Fisher,
Nvidia’s former head of sales and current head of its PC graphics business,
explained that a frequent, centralized update process was the best means to
guarantee a consistently good user experience, ensuring that gamers always
had optimal performance for the latest software released by developers and
other companies. “Graphics drivers are perhaps the most challenging piece
of software in the PC after the operating system,” he said. “Every app
touches it, and every app release or update can potentially break it.”
WHEN GEOFF RIBAR WAS HIRED away in December of 1997 from
Advanced Micro Devices to serve as Nvidia’s CFO, he found that his new
boss had two impressive traits: Jensen was extremely persuasive and
extremely hardworking.6
“There may be people smarter than me,” Jensen once told his executive
staff, “but no one is ever going to work harder than me.”7
He was often in the office from 9:00 a.m. to near midnight, and his
engineers usually felt obligated to keep similar hours.
“I used to tell people at AMD, Intel, or anywhere else that if they wanted
to see how Nvidia was doing, they should visit the company’s parking lot
on weekends. It was always busy,” said Ribar.
Even for the marketing department, working sixty to eighty hours a
week, including every Saturday, was the norm. Andrew Logan, Nvidia’s
director of corporate marketing, remembers leaving the office to take his
wife to a 9:30 p.m. showing of the movie Titanic. On his way out, his
coworker shouted, “Oh, half day, Andy?”8
Tester Henry Levin recalls that whenever he found himself working late,
he was never the only one there. Even when he stayed to 10:00 p.m. or later,
Nvidia’s graphics architects would still be at the whiteboard, passionately
discussing chip optimization and rendering techniques. His contemporary,
Director of Materials Ian Siu, has imprinted on his memory the image of
colleagues spending the night at the office, even over the weekends, after
bringing sleeping bags to work. Employees would also bring their kids to
the office so that they could spend time with their families without leaving
their workplace.
“We worked our asses off all the time,” Siu said. He had fond memories
of the camaraderie in the office and of close relationships with his
coworkers.
Ribar rarely worked until midnight, but he would often arrive early in the
morning. He quickly learned that one big disadvantage to sitting near the
CEO at the office was that he was often the first person Jensen saw in the
morning. And Jensen was known to unload on the first person he
encountered, whoever it was.
“Jensen would often stew with his thoughts on something overnight
about products or marketing,” Ribar said. “It was almost never a finance
issue but it didn’t matter. If I saw him first, I would get the first blast from
him.”
As the day went on, no place in Nvidia headquarters was safe from a
drive-by grilling from Jensen. Kenneth Hurley, a technical marketing
engineer, was at a urinal when Jensen walked up to the one next to him.
“I’m not the kind of guy who likes to talk in the bathroom,” Hurley said.9
Jensen had other ideas. “Hey, what’s up?” he asked.
Hurley replied with a noncommittal “not much,” which earned him a
sidelong glance from the CEO. Hurley panicked, thinking, “I’m going to
get fired because he thinks I’m not doing anything. He’s probably
wondering what I’m doing at Nvidia.”
To save face, Hurley proceeded to list twenty things he was working on,
from convincing developers to buy Nvidia’s latest graphics card to teaching
those developers how to program new features on them.
“Okay,” Jensen replied, apparently satisfied with the engineer’s answer.

FEAR AND ANXIETY BECAME JENSEN’S favorite motivational

tools. At each monthly company meeting, he would say, “We’re thirty days
from going out of business.”
It was hyperbole, on one level. The tense, high-stakes RIVA 128 process
was not a complete outlier—as we’ll see—but it certainly wasn’t a regular
occurrence. Yet Jensen didn’t want to allow any complacency to creep in,
even in successful periods. And he wanted to confront new hires with the
kind of pressure they would face going forward. If they didn’t have what it
took, they needed to self-select out sooner rather than later.
But on another level, the line “We’re thirty days from going out of
business” was true. In the technology industry, a single bad decision or
product launch could be fatal. Nvidia had gotten lucky twice before, barely
surviving the disasters of the NV1 and the NV2 before succeeding—with
only months to spare—with the RIVA 128. That luck would not hold
forever. But a good corporate culture would fortify the company against the
dire consequences of most mistakes. And a mistake or downturn in the
market was inevitable.
Still, as Dwight Diercks said, “It always felt like we were at zero. And
the reason is that no matter how much money we had in the bank, Jensen
could explain why we were going to be at zero with three things happening.
He would say, ‘Let me tell you how. This could happen, this could happen,
this could happen, and all that money goes to zero.’ ”
Jeff Fisher pointed out that fear can be clarifying. Even today, although
Nvidia is no longer thirty days away from going out of business, the
company could easily be thirty days from starting down a path that will lead
to its destruction. “You’re always trying to look around corners and see
what we’re missing,” Fisher said.
That paranoia came to a head in late 1997. Intel had always been both an
important partner for Nvidia and a potential competitive threat. Nvidia’s
graphics chips all had to be compatible with Intel’s processors, because
Intel was the primary CPU maker for the PC market. But that fall, Intel
started telling industry partners it had its own graphics chip coming, which
threatened to take business away from Nvidia and other companies in the
space.
Just months after the RIVA 128 launched to great fanfare, Intel
announced its own chip, the i740. It was a direct challenge to Nvidia—its
new chip and its very existence. Unlike the RIVA 128, which had a four-
megabyte frame buffer, Intel’s i740 had an eight-megabyte buffer—twice
the size of the one on Nvidia’s chip—that the company was trying to set as
the new standard in the industry. Intel had the ear of every PC maker in the
world, as it supplied the vast majority of their CPUs. After Intel’s i740
announcement, “our sales pipeline started to dry up,” one Nvidia executive
said. If Intel could force adoption of the eight-megabyte buffer, the RIVA
128 would quickly be rendered obsolete.
“Make no mistake. Intel is out to get us and put us out of business,”
Jensen declared at an all-company meeting. “They have told their
employees, and they have internalized this. They are going to put us out of
business. Our job is to go kill them before they put us out of business. We
need to go kill Intel.”10
Caroline Landry and the rest of the Nvidia team worked even harder to
fight off the new competitor, a company that at the time was about 860
times Nvidia’s size in terms of revenue. She often worked past midnight,
staggering home and falling asleep for a few hours before waking back up
to do it again.
 “I’m super tired. I need to get up. It’s hard,” she told herself. “But we
need to kill Intel. Must kill Intel.”

CHRIS MALACHOWSKY SPEARHEADED the response to the Intel

threat. Throughout his career at Nvidia, he acted as an ultratalented utility
infielder. Jensen would assign him to manage a struggling part of the
business, whether operations, manufacturing, or engineering, and
Malachowsky would do whatever was necessary to fix the issue. Now, the
CEO needed him to go back to his roots as a chip architect and beat the
i740.
Even when engrossed in a time-sensitive project that required intense
concentration, Malachowsky found himself pulled into a mentorship role,
which he embraced. A new hire named Sanford Russell had just joined
Nvidia from Silicon Graphics but was having trouble getting up to speed on
Nvidia’s technology and culture. There was little formal orientation beyond
Hara’s confrontational pep talks, and few of the company’s processes were
written down.
One day, Russell noticed that Malachowsky would go home to have
dinner with his family and then come back late at night to work on the
RIVA 128ZX, an eight-megabyte version of the original chip intended to
compete with Intel. He realized that if he showed up at the lab at 10:00 p.m.
sharp and pulled up a stool across from Malachowsky, he could ask the
Nvidia cofounder whatever he wanted.11
Russell would ask him about a deep technical issue, Malachowsky would
talk through the topic for a few minutes and then quietly go back to work.
Russell sat there until Malachowsky asked him for the next question, at
roughly fifteen-minute intervals.
“I did this for weeks, patiently watching him, listening to him mutter,
‘Why isn’t this working?’ as he tried to bring the chip up. The whole
company was trying to get the chip working,” Russell said. “But Chris still
helped me climb the ladder of knowledge of all these chip things because he
built them. He was the guy who built the stuff and taught me while he was
trying to save the company.”
Russell was amazed that Malachowsky could hold the whole chip in his
head and work through it until he figured it out. At 2:00 a.m. one morning,
everything fell into place. Malachowsky shouted, “I got it! I got it! We’re
going to live!”
He had internalized Jensen’s paranoia and in the sprint to make the
original RIVA 128 had helped to future-proof it in one important way: he
had left himself some spare capacity in the chip’s silicon. Now, he was able
to use it to rework the chip so that it had an eight-megabyte frame buffer.
“It was a very elaborate change order, rewiring gates,” he recalled. “We
were able to make a feature change, on the fly, in metal.”
Once he hit on his solution, the company applied focused ion beam (FIB)
technology, which can modify chips at the microscale level. The FIB
instrument looks like an electron microscope but does not use electrons: it
uses ions to modify chip prototypes. The modified chips worked, saving
Nvidia’s RIVA series from instant obsolescence.
Malachowsky was able to inspire one of his newest employees while
doing so. In 2024, when Russell ran into Malachowsky at a conference, he
raised the long nights they spent together in the lab.
“You saved me, man,” Russell said, thanking him for getting his career at
Nvidia—which would last twenty-five years before he moved on—off to a
solid start.
Malachowsky demurred. “Nah. You were fine.”
“No,” said Russell with a chuckle. “No, I wasn’t.”

IN SOME INSTANCES, NVIDIA’S FOCUS on speed could lead to

lapses in quality—at least, relative to the high standards that Jensen set for
the company.
Andrew Logan, Nvidia’s director of corporate marketing, remembers
how one of Nvidia’s chips came in second place for a computer magazine
awards feature. In his prior job at S3, executives would be happy if their
products finished in the top three. Not at Nvidia.
“The first time we came in second place, Jensen sternly told me: Second
place is the first loser,” Logan said.12 “I never forgot it. I realized I’m
working for a boss who believes we have to win at everything. It was a lot
of pressure.”
By all measures, the original RIVA 128 was an excellent chip. It could
render high-resolution graphics at a much faster frame rate than that of its
competitors; even visually demanding games such as Quake ran at
maximum quality without any slowdown. It was also the largest chip ever
manufactured but could still be produced fast enough to meet initial
demand. Even so, the Nvidia team had to make some trade-offs to get the
chip out in time. On some types of images, the RIVA 128 would apply
dithering—a form of intentional noise that was designed to break up or
obscure obvious visual irregularities—to certain types of renders, such as
smoke or clouds.
Enough gamers had noticed the issue that a major PC magazine decided
to publish an exposé on Nvidia’s flagship graphics chip. It placed renders
from Nvidia’s RIVA series and from equivalent current-generation cards
from 3dfx and another competitor, Rendition, side-by-side-by-side in a
large, detailed image spread. Nvidia’s images were blurry and smudged,
and the magazine commented that they were the worst of the three—that
they “looked terrible.”
Upon seeing the article, Jensen called his several executives into his
office, where he had the issue spread open on a table. He demanded to
know why the RIVA 128’s output looked so bad. Chief Scientist David Kirk
responded that they had made some image-quality trade-offs in order to get
the chip done on time (and save the company). The reply only made Jensen
even more agitated. He demanded that Nvidia chips beat the competition
not on one metric but on all of them.
The shouting match got so loud that it caught the attention of Walt
Donovan. He was the chip architect who had seen the RIVA 128 demo at
the Computer Game Developers Conference and who had asked for a job at
Nvidia on the spot. He worked on the opposite end of Nvidia’s headquarters
from Jensen’s office, which shielded him from most of the CEO’s
broadsides. He also had a severe hearing deficiency, wearing hearing aids in
each ear. But this time, he couldn’t ignore the commotion and invited
himself into the argument.
Donovan assured Jensen that Nvidia’s next generation of chips, which
they were calling the RIVA TNT series, would not only solve the dithering
problem; it would also outpace the industry on every possible measure of
graphics quality. He pointed to the Rendition image, which the magazine
had rated the best of the three.
“That’s what RIVA TNT will look like,” he said.
It did little to placate Jensen, who at this point wanted to be left alone.
 “Get out!” he yelled.

JENSEN’S COMPETITIVENESS OFTEN MOTIVATED his employees

to do extraordinary things. But it could also reveal a petty side of the CEO.
Harry Levin, the chip tester who often spent late nights working on the
RIVA 128, once challenged Jensen to a game of ping-pong on one of the
communal tables inside Nvidia headquarters. He was well aware that Jensen
had been a nationally ranked ping-pong player during his teenage years.
And he was familiar with the CEO’s obsession with winning at business.
What Levin didn’t realize was that Jensen brought the same level of
intensity to any kind of competition, whether professional or recreational.
Levin considered himself a good recreational player, but never expected to
receive quite the beatdown he got at the hands of his superior.
“He just whooped me,” said Levin. “The game was up to twenty-one
points, and he let me score only once or twice. It was a very quick match.”
Jensen was so competitive that he challenged other employees even when
he was at a disadvantage. In high school, CFO Geoff Ribar had ranked
among the top fifty chess players in the country. His boss, however, could
not accept that someone else was better than him.
“Jensen knew about my chess skills. Being competitive, he was
convinced he was smarter than me and could beat me,” Ribar said. “There
was no way he could have been able to beat me, but he tried.”
Jensen attempted to close the gap between his and Ribar’s chess skills
through brute-force learning. He memorized chess openings and sequences
of moves so that he could control the board. Yet Ribar found his playing
style predictable. Whenever he saw a standard opening of the type that
Jensen had learned, Ribar would counter with an unorthodox move that
thwarted his boss’s strategy. Every time he lost, Jensen would swipe his arm
across the board, knocking over the pieces, and storm away. He would
sometimes later insist on a rematch on the ping-pong table. Ribar graciously
accepted, knowing Jensen was purposely shifting the competition onto
more favorable territory.
“He’s good at ping-pong,” recalled Ribar. “I’m okay, but he would just
kill me to get his revenge. It helped relieve his frustration from losing in
chess by beating me in ping-pong.”
JENSEN’S CHESS LOSSES WERE NOT the only thing fueling his
general frustration. Like other graphics-chip companies, Nvidia only
designed and prototyped its products—it did not actually manufacture them
at scale. Chip fabrication was instead outsourced to one of the small number
of dedicated chip-fabrication companies around the world. These
companies invested hundreds of millions of dollars in the clean rooms,
specialized equipment, and skilled personnel required to make tiny silicon
wafers into advanced computational devices.
Since Nvidia’s founding, it had partnered with SGS-Thomson, the
European chip conglomerate, to manufacture its chips. As Jensen and his
cofounders discovered during their initial meeting with Sequoia, SGS-
Thomson did not have the best reputation, and it had struggled to remain
competitive in the face of less expensive labor in East Asia.
Yet now that Nvidia was producing great chips and selling them in
massive quantities, SGS-Thomson’s weaknesses became far more difficult
to ignore. In late 1997, head of sales Jeff Fisher arranged for a team from
Gateway 2000 to tour the SGS-Thomson fabrication plant in Grenoble,
France. The RIVA 128 had been out for several months and was enjoying
robust demand from gamers. The trip was supposed to be a victory lap for
Fisher and Nvidia.
During the flight to France, Fisher learned that SGS-Thomson was
having yield problems with Nvidia’s flagship product. The fabricator
estimated that it would be able to fill only about half of Gateway 2000’s
allocation. As Fisher recalled, “We had to huddle with the SGS guys and
discuss how to message it to Gateway.”13
The disastrous factory tour was just the first warning sign of a full-blown
production crisis, which finally exploded over Thanksgiving. Fisher had
intended to enjoy a well-deserved break at his mother-in-law’s house in
northern Indiana. Instead, he spent nearly the entire holiday on the phone,
breaking the news to Dell and other computer makers that they would no
longer get as many of the winter’s hottest graphics cards as they had
ordered. In between calls with irate vendors, he talked to Jensen, to pass on
the CEO updates from SGS-Thomson.
“We had gotten all these customers signed up, customers that we always
dreamed of having, and now we were going to have to get in allocation
mode,” he said.
 Jensen always exhorted his employees to never make the same mistake
twice—and he now vowed never to settle for a fabrication partner that
couldn’t handle the levels of production that Nvidia needed. Fortunately, he
had another vendor in mind.
When Nvidia was founded in 1993, Jensen struggled to find chip-
manufacturing capacity. He had cold-called Taiwan Semiconductor
Manufacturing Company (TSMC)—the best-regarded manufacturer in the
world and the one Don Valentine at Sequoia had recommended that Nvidia
partner with from the start—multiple times without success. In 1996, he
tried a more personal approach. He addressed a letter to Morris Chang,
TSMC’s CEO, asking if the two men could discuss Nvidia’s chip needs.
This time, Chang called him, and the two men arranged for a visit in
Sunnyvale.14
 Comic strip about the Nvidia-TSMC partnership. (NVIDIA)

During the meeting, Jensen outlined Nvidia’s future plans, explaining

how Nvidia would need larger chip dies for its current generation and even
larger ones in the future. He managed to secure some production capacity
from TSMC to supplement SGS-Thomson’s capabilities, and the
relationship seemed to be going well. Chang periodically returned to
Sunnyvale to ensure that Nvidia had all the capacity it needed, jotting down
notes in a small black book. He even visited during his honeymoon in 1998.
“The biggest joy I get out of this job was to see my customers grow,
make money, and succeed,” said Chang, and this was especially true for a
fast-growing customer such as Nvidia.
The two CEOs and their companies had become so close in such a short
period of time—and the relationship between Nvidia and SGS-Thomson
had soured almost as fast—that in February of 1998, Nvidia made TSMC
its main supplier. The switch occurred just as Nvidia announced its newest
chip, the RIVA 128ZX, which debuted just eleven days after the debut of
Intel’s feared i740 and was positioned by Nvidia as a clear improvement
over the competing chip. It offered better performance than the i740 and an
eight-megabyte frame buffer, equivalent to the one on the i740, at a cost of
$32 per chip, only slightly higher than Intel’s $28 list price. It was supposed
to ensure Nvidia’s continued dominance of the PC market in spite of Intel’s
attempt to undercut the company.
Yet again, production issues appeared. In the summer of 1998,
manufacturing defects plagued TSMC’s production of the RIVA 128ZX.
The defects, caused by residue called titanium stringers, were scattered
randomly across different parts of the chip. As a result, it was impossible to
determine which chips were faulty and which were functional; it was clear
only that some large proportion of RIVA 128ZX chips was contaminated.
Chris Malachowsky came to the rescue once more.
“Why don’t we test every chip and run software on every part?” he asked
one day.
“You can’t possibly do that,” another Nvidia executive replied. “Why?”
asked Malachowsky.15
It was, on the face of it, an absurd suggestion. Nvidia would need to ship
hundreds of thousands of chips to company headquarters for manual testing
—it would have to convert some portion of its messy office and workshop
areas into a large chip-testing lab. It would be one of the greatest tests of
Jensen’s “Speed of Light” maxim.
The company converted one of its buildings into a large testing assembly
line with open computer cages, motherboards, and CPUs. “It was a massive
operation,” said Curtis Priem. “You would go home at 11:00 p.m. and walk
by the lab to see dozens of people just putting in chips.”16
The process was extremely finicky. Priem recalls that they had to redo
tests because defective chips would sometimes pass for reasons having
nothing to do with the chips themselves, such as if the power wasn’t
completely disconnected from the testing rig before the next run.
Initially, both Nvidia employees and management pitched in. But soon
enough, the strain of such high-volume precision testing began to burn out
the engineering team. To take the burden off of his people, Jensen hired
hundreds of low-skilled contract laborers, whom Nvidia employees called
“bluecoats” because of the color of their lab coats. Soon, bluecoats
outnumbered Nvidia engineers in the building. The extra manpower
allowed the company to test every chip before it was sent to the customer or
trashed.
There was a significant cultural and class divide between Nvidia’s
employees and the bluecoat testers. Caroline Landry observed the
increasing stratification between the less-educated, immigrant bluecoats on
one hand and the highly educated engineers on the other.
First, she noticed that no one wanted to sit with the bluecoats at lunch.
“Coming from Canada, we’re a little more egalitarian,” she said. Ignoring
the disapproving looks that followed her around the cafeteria, she “would
go sit with the bluecoats and get to know them. I would then get all these
remarks from other engineers, like ‘You had lunch with the bluecoats? How
come?’ It was weird. I don’t understand the mentality.”
The main divide was over food. Nvidia provided generous food perks:
breakfast, lunch, and dinner, as well as free snacks, ranging from candy bars
and chips to noodles. Seeing this, the bluecoats—who had not normally
gotten meal privileges at prior jobs—would come into the cafeteria and load
up on food, then empty the drinks and snacks closets after they were
restocked on Fridays.
“I came in one weekend and there were a bunch of people with grocery
bags full of stuff, carrying it out to their cars,” one Nvidia employee said.
“In their mind, it’s free. It’s not stealing. It’s there for the taking, so they
take it,” Landry said.
Nvidia employees complained so frequently that Jensen sent a company-
wide e-mail with the subject line: “Give a bluecoat your pork chop.” If the
testers wanted the main course from your lunch plate, you should hand it
over. Jensen thought Nvidia employees should show gratitude to the
bluecoats, as they were instrumental in helping the company navigate a
major crisis. Their assistance was far more valuable than the minor
inconvenience of running out of free employee snacks.

EVEN WITH HELP FROM THE BLUECOATS, Nvidia could not

overcome the production slowdown. Geoff Ribar had been hired as CFO
specifically to prepare the company for an initial public offering, which was
being underwritten by the investment bank Morgan Stanley. As Nvidia ran
out of chips to sell, however, it became significantly less attractive to
potential investors. Its quarterly revenue fell by half, from $28.3 million for
the quarter ending in April 1998 to $12.1 million in the quarter ending in
July 1998. Yet its expenses kept growing, which caused its net losses to
balloon from $1 million to $9.7 million quarter to quarter. Only six months
earlier, Nvidia had booked its first-ever profitable quarter. Now, it was
losing money at an alarming rate.
In a booming economy, Nvidia’s deteriorating balance sheet still might
have made it appealing to the right buyers. But the financial crisis that had
gripped East and Southeast Asia for nearly a year had also tempered
enthusiasm for risky IPOs. Morgan Stanley decided to pause the process.
The IPO would have given Nvidia a huge infusion of much-needed cash.
Instead, Ribar now calculated that at its current burn rate, the company was
“within weeks” of becoming insolvent.17 It was the RIVA 128 situation all
over again.
Jensen would have to rely on his persuasiveness and his talent to pull
Nvidia through the new crisis. He asked for bridge financing from Nvidia’s
three largest customers: Diamond Multimedia, STB Systems, and Creative
Labs. The companies believed in Nvidia’s technological prowess, as they
each bought millions of dollars’ worth of RIVA chips to use on their high-
end graphics cards. Jensen argued that bridge financing would give Nvidia
enough additional time and operating cash to recover from the temporary
setback. To sweeten the deal, he structured the loans as convertible notes
that, when the company introduced its IPO, could convert into equity at 90
percent of the eventual IPO price—which would give Nvidia’s potential
creditors a far higher potential upside than that of standard loan interest.
After two weeks of negotiations, in August of 1998 the three companies
agreed to loan Nvidia a total of $11 million. Jensen not only gauged their
confidence in Nvidia correctly; he had also managed to turn that confidence
into a closer relationship with his biggest customers.
Despite the financial lifeline, Ribar was ready to move on. The pressure
“drove my hair gray,” he later said. In October 1998, he was recruited by
Marvin Burkett, who had mentored Ribar when the two men worked at
AMD, to join the Japanese electronics company NEC and help turn around
its monitor division. He had lasted at Nvidia less than a year—not enough
time even for his first tranche of Nvidia stock to vest.

JENSEN’S RESPONSE TO NVIDIA’S near-death from a production

backlog was, paradoxically, to restructure the entire company in order to
ship new designs even faster. He began to call Michael Hara, Nvidia’s head
of marketing, into his office to brainstorm strategy. Jensen had observed
that no one company ever seemed to have a permanent lead in the industry.
Companies that led one year, such as S3, Tseng Labs, and Matrox, were
often displaced within one or two chip generations.
“Mike, I don’t get it,” he said. “If you look at the PC graphics industry,
why is it that one company can never hold a lead more than two years?”18
Now that Nvidia was one of the market leaders rather than a challenger
brand, Jensen became obsessed with the problem. He turned it into a joke:
“The only thing that lasts longer than our products is sushi,” he would often
say to Nvidia employees. Jensen saw that whichever company could solve
the problem would build a strong moat around its business.
Hara, who had worked at several of Nvidia’s competitors, explained the
market dynamics to Jensen. The whole industry moved according to the
rhythms of computer manufacturers, who refreshed their product launches
twice a year: in spring and fall. The fall cycle was the more critical one,
driven by August’s back-to-school season and leading into holiday
shopping. Computer makers felt compelled to put out updated devices every
six months that featured the latest and best-performing chips. They were
constantly shopping around for better chips to put in their PCs, readily
replacing existing vendors with new ones as faster, higher-quality
components became available.
 Chip makers, Nvidia included, took eighteen months to design and
launch a new chip and typically worked on only one at a time. Yet graphics
technology was advancing fast enough that chips were functionally obsolete
long before chip makers could come out with a new product.
“That doesn’t work. There’s got to be a way to solve this problem of the
design cycles,” Jensen said. The RIVA 128 had revealed that Nvidia could
design and launch a new chip in less than a year, although it had taken the
threat of imminent bankruptcy to galvanize the company into moving that
fast. How could Nvidia do what it had done to produce the RIVA 128,
though in a more repeatable and sustainable way?
After a few weeks, Jensen announced to his executive team that he had
figured out how to keep Nvidia ahead of the competition—forever. “We’re
going to fundamentally restructure the engineering department to line up
with the refresh cycles,” he said.
Nvidia would split the design team into three groups. The first would
design a new chip architecture, while the other two worked in parallel to the
first to develop faster derivatives based on the new chip. This would allow
the company to release a new chip every six months, in line with PC
makers’ buying cycles.
“We won’t lose our sockets because we can go back to the OEM [original
equipment manufacturer; the PC maker] and say, ‘Here’s my next chip that
uses the same software. It will have new features and will be faster,’ ” he
explained. Of course, the solution required more than reorganizing Nvidia’s
design teams. Many technical decisions the company had made earlier
would come into play as well.
Early on, Curtis Priem had invented a “virtualized objects” architecture
that would be incorporated into all of Nvidia’s chips. It became an even
bigger advantage for the company once Nvidia adopted the faster cadence
of chip releases. Priem’s design had a software-based “resource manager,”
essentially a miniature operating system that sat on top of the hardware
itself. The resource manager allowed Nvidia’s engineers to emulate certain
hardware features that normally needed to be physically printed onto chip
circuits. This involved a performance cost but accelerated the pace of
innovation, because Nvidia’s engineers could take more risks. If the new
feature wasn’t ready to work in the hardware, Nvidia could emulate it in
software. At the same time, the engineers could take hardware features out
when there was enough leftover computing power, saving chip die area.
 For most of Nvidia’s rivals, if a hardware feature on a chip wasn’t ready,
it would mean a schedule delay. Not, though, at Nvidia, thanks to Priem’s
innovation. “This was the most brilliant thing on the planet,” said Michael
Hara. “It was our secret sauce. If we missed a feature or a feature was
broken, we could put it in the resource manager and it would work.”19 Jeff
Fisher, Nvidia’s head of sales, agreed: “Priem’s architecture was critical in
enabling Nvidia to design and make new products faster.”20
Nvidia also began emphasizing backwards compatibility for its software
drivers, which it had first done with the RIVA 128. But it was a lesson that
predated Nvidia: Priem had learned it in his pre-Nvidia days at Sun
Microsystems. He had heard about a sales session for a new version of the
GX graphics chip, where the sales force had been told that the new chip was
compatible with old software drivers. If a customer installed the new GX
into an existing Sun workstation, it would just work. There was no need to
wait for new software to be installed before a customer could use the newly
purchased graphics hardware. The sales team stood up and gave the
presenter a standing ovation. When Priem was told about that reaction, he
made a mental note that the unified drivers feature solved a pain point for
salespeople—and thus for customers.
“We thought, okay, this must be important,” he said. “It turned out to be
very important to Nvidia.”21
Jensen saw emulation and backwards-compatible drivers not just as good
technical principles but also as competitive advantages. He believed that
embracing both would allow the company to implement his new accelerated
production schedule, which he called “Three Teams, Two Seasons.” He
believed Nvidia had a chance of always being ahead of the rest of the
industry. Jensen had long insisted that Nvidia chips would always be the
best in the market, and they almost always were: that was not going to
change. Now, the company would have three times as many chips in the
market, none of which was more than six months out of date. Even if a
competitor offered a slightly better product, PC makers would have no
motivation to switch away from Nvidia, knowing that a faster part would
arrive in six months and without the hassle of changing drivers.
Nvidia’s rapid iteration meant that “the competition will always be
shooting behind the duck,” as Jensen described it. Like a hunter who aims
at a moving target instead of ahead of it, other graphics-chip makers would
lag behind—there would be too many new chips coming out too fast.
Nvidia’s competitors would simply be overwhelmed.
“The number one feature of any product is the schedule,” Jensen later
said.22
By the end of 1999, Nvidia had reorganized its model for design and
production on the “Three Teams, Two Seasons” strategy. It had a
philosophy that demanded employees operate at the “Speed of Light,”
measuring performance against what was physically possible and not what
other companies were doing or what Nvidia had accomplished or failed to
accomplish in the past. And it had a corporate mantra—”We’re thirty days
from going out of business”—that served as a warning about complacency
and conveyed the expectation that everyone, from the CEO on down, had to
work as hard as they possibly could, even if it meant sacrificing their lives
outside of Nvidia.
 CHAPTER 6

Just Go Win

AS NVIDIA ACCELERATED ITS PRODUCTION schedules and

methods in order to dominate the graphics-chip market, its competitors
fought back. In September of 1998, 3dfx sued for patent infringement,
alleging that one of its rendering methods had been stolen by Nvidia. The
press release that announced the lawsuit included a link to a page on
Nvidia’s website about the technology in question. In response, Nvidia’s
marketing team modified the linked page, so that anyone who clicked
through from the press release saw a banner that read, “Welcome to
NVIDIA, the greatest 3D graphics company.”
Just a year earlier, 3dfx’s leaders had been so confident that Nvidia was
about to go bankrupt that they didn’t even bother making a play for their
struggling rival. Now, the situation had been almost turned on its head.
Under “Three Teams, Two Seasons,” Nvidia was preparing to release three
chips in the time it took 3dfx to launch just one. 3dfx’s most recent chip, the
Voodoo2, had come out in February 1998, and the company was only
partway through the development cycle for two next-generation chips,
which were code-named in its characteristically over-the-top style: Napalm,
scheduled for release in late 1999; and Rampage, scheduled for release in
2001. At its current pace, 3dfx’s premium releases would lag Nvidia’s by a
year or more.
And 3dfx’s leaders were not even confident about their slower release
schedules. The company’s engineers “wanted perfection in every product
that we shipped,” said Ross Smith, the marketing executive.1 “There was
feature creep in every chip, while Nvidia’s mentality was to ship whatever
was ready in the chip to meet the deadline and push features out to the next
chip.”
3dfx had become the victim of its own success in another way, too. Scott
Sellers, the cofounder who also served as the company’s head of
engineering, said that the Voodoo2’s brisk sales were making it difficult for
the company to manage its distribution channel and relationships with
graphics-card partners.
“We had some quality problems where some board manufacturers were
not following our design guidelines,” he said.2 “The poor quality was
starting to impact our customer satisfaction.”
The entire industry was well aware of Nvidia’s ability to turn challenges
into opportunities. 3dfx now sought to do the same thing. But its approach
contrasted sharply with Nvidia’s.
First, in an attempt to mimic Nvidia’s strategy of getting more chips out
into the market, 3dfx announced that it would add several new products to
its portfolio. These included the Voodoo Banshee and the Voodoo3, which
were designed as combined 2-D–3-D accelerators instead of the kind of
purely 3-D chips that 3dfx had produced to date. Unlike Nvidia’s product
road map, which created efficiencies by making multiple derivative
versions of a single chip for a focused area of the market, 3dfx had an
overly complicated lineup aimed at too many different customer segments
and didn’t plan to reuse a common core-chip design.
3dfx then decided to expand into an entirely new part of the graphics
industry. In December 1998, it bought the graphics-board manufacturer
STB Systems for $141 million. The move made sense on paper. STB was a
major board maker, and bringing it under the 3dfx umbrella would give the
company more control over its own board supply chain. It would also build
brand awareness directly with consumers, as now both the chips and the
boards could be sold under the 3dfx brand.
Even more important, from a strategic perspective, was 3dfx’s belief that
the acquisition would hurt Nvidia. STB had a close relationship with the
rival chip maker. It had given Nvidia the very first purchase order for the
RIVA 128, the order Jensen had dramatically unveiled at a company all-
hands meeting. Since the launch of that chip, it had become Nvidia’s
leading board partner, and it had loaned Nvidia money as part of the bridge
financing round three months earlier. With the purchase, 3dfx forcibly
ended that relationship. STB announced that from now on, its boards would
carry only 3dfx chips.
“We knew it was a bet-the-company strategy,” said Sellers. “We just felt
we could do it.”
But none of 3dfx’s strategic moves and product bets worked. It struggled
to make the 2-D side of its mid-tier chips, because it did not have as much
in-house expertise as it did for 3-D chips. When STB enacted its 3dfx-only
policy, other board makers retaliated by switching over to Nvidia chips,
which canceled out that supposed advantage. And Sellers’s assumption that
3dfx could effectively manage the new acquisition’s business was
completely wrong. 3dfx’s executives had no prior experience overseeing a
retail physical distribution channel or a complicated board-manufacturing
supply chain. Once under 3dfx, STB ended up pulling its new parent
company’s focus away from the core chip-design business.
Most of all, not a single one of these initiatives addressed 3dfx’s main
problem, which was that it was no longer producing high-performance
chips at the necessary speed. Perfectionism, management dysfunction, and
distracted leadership had combined to slow output to a crawl. The mid-tier
Voodoo3, which had been intended as a stopgap between 3-D chip
launches, was delayed until April 1999. Napalm and Rampage slipped even
further behind schedule.
“We should have really stuck to our knitting,” said Ross Smith. “If 3dfx
had brought out Napalm and Rampage on time, Nvidia would have never
had a chance.”
3dfx soon faced a complete operational meltdown. It failed to manage
STB’s inventories. Its mid-tier cards failed to sell. It simply ran out of cash.
The company’s creditors initiated bankruptcy proceedings near the end of
2000. On December 15, Nvidia bought 3dfx’s patents and other assets and
hired about one hundred of its employees. In October 2002, 3dfx filed for
bankruptcy.
When those former 3dfx engineers arrived at Nvidia, they expected to
find out that their victorious rival had some kind of unique process or
technology that allowed them to make new chips every six months. Dwight
Diercks remembers their shock when they found out that the explanation
was much simpler.
 “Oh my God, we got here and we thought there was going to be a secret
sauce,” one engineer said.3 “It turns out it’s just really hard work and
intense execution on schedules.” It was Nvidia’s culture, in other words,
that made the difference.

PERFECTING THE COMPANY’S OPERATIONS was only one piece

of the plan to future-proof Nvidia against institutional dysfunction. The
other was bringing the best talent to the company. Nvidia’s excellent
products attracted high-quality applicants. But often, Nvidia had to lure
talented people from its competitors. Rare was the occasion when it could
scoop up dozens of engineers from a rival, as it had with the demise of
3dfx. Instead, Jensen and his team learned the fine art of corporate
poaching.
In 1997, Jensen asked Michael Hara if he knew anyone good who might
want to join Nvidia. Hara floated the name of John Montrym, the chief
engineer of Silicon Graphics. Montrym was legendary in the industry for
his graphics subsystem work, such as RealityEngine and InfiniteReality. He
had some history with Nvidia’s cofounder Curtis Priem: the two had
worked together back at Vermont Microsystems.
Jensen invited Montrym to lunch at Nvidia’s office and went for a direct
approach. “John, you should think about coming to Nvidia because
ultimately I’m going to put SGI out of business,” he said, explaining how
SGI could not compete by selling thousands of workstations each year
when Nvidia’s access to a market of millions of PCs gave it much better
economies of scale.4 Montrym politely declined.
Chris Malachowsky and Nvidia’s chief scientist, David Kirk, tried next.
Over another lunch, they told Montrym, “All the work you did with SGI’s
RealityEngine and InfiniteReality, Nvidia is going to put on a single chip
for PC, and that will be the end of SGI. Where do you want to be working
when that happens?”5
Priem also tried to recruit Montrym. The two met at St. James Infirmary
Bar & Grill in Mountain View, California. Priem insisted that Silicon
Graphics would be a “dead end,” and that his old colleague should join him
at Nvidia.6 Once again, Montrym wasn’t convinced.
 Jensen then decided to try a different tactic—to persuade Montrym with
technology, not words. He told his development team to make a graphics
demo for their latest chip prototype in the form of a military-themed
immersive simulation, mimicking a move that SGI used when it wanted to
show off its own new technology. Then, he told Hara to call Montrym up
again and invite him over to Nvidia’s lab to see the demo in action.
“This time will be more fun,” Jensen assured Hara.
When Montrym arrived, Hara presented the new prototype. “Isn’t this
exactly what InfiniteReality does?” he asked.
The new pitch worked. Of course, Montrym knew that Jensen’s
assessment of the relative weakness of SGI was correct. His current
employer could afford to fund a new chip only every few years, due to its
smaller market. Nvidia, by contrast, was releasing a new design every six
months. Nvidia’s pace of innovation was far beyond what SGI could ever
achieve. In time, it would be so far ahead that SGI could no longer catch up.
But the demo was still powerful. It showed that Nvidia had so many
resources and so much talent at its disposal that in a matter of weeks it
could make a graphics engine that had taken Montrym much longer to
develop—and that they could do it just for the purpose of recruiting a single
person. Montrym resigned from SGI a week later.
Dwight Diercks said Montrym’s defection was “a watershed moment,
because so many engineers revered John, and they all wanted to come work
with John.” After Montrym joined Nvidia, every time the company posted a
job opening for software developers or chip engineers, résumés and
interview requests from Silicon Graphics employees would pour in.7
SGI was, understandably, unhappy about losing Montrym and feared
losing even more talent to Nvidia. In April 1998, SGI sued Nvidia for patent
infringement, alleging that the RIVA family of processors infringed on the
company’s high-speed texture-mapping technology.
While some Nvidia employees were initially worried about the lawsuit,
Andrew Logan, Nvidia’s director of corporate marketing, was excited.
“I got a call from the Wall Street Journal right now on my voicemail,” he
told his colleagues after the lawsuit was announced. “This is perfect. We’re
on the map!”
Jensen agreed. He walked around from office to office, shaking
everyone’s hand and saying, “Congratulations! We just got sued by the most
important graphics company in the world. We’re somebody.”
The lawsuit went nowhere: in order to succeed, SGI would have to prove
financial harm, yet the only evidence the company cited was Nvidia’s
internal sales projections and forecasts. Nvidia’s lawyers argued that these
forecasts were inherently volatile, because they were based on broad
assumptions about the market that were often wrong, and thus couldn’t be
relied upon as measures of anything real.
In July 1999, the two companies settled the lawsuit in an agreement that
seemed to benefit Nvidia most of all.
“We’d hire fifty of their employees and become a supplier to them of
their low-end graphics line. In the end, we gained a partner,” Diercks said.8
Once again, Nvidia won some of the best engineering talent in Silicon
Valley.

AS NVIDIA GREW, IT GAINED potential leverage over its supply-chain

partners, and it could have pressured companies into helping its own bottom
line. Yet Jensen’s philosophy of business relationships saw the company
maintain good terms with its most crucial suppliers.
Rick Tsai was TSMC’s executive vice president of operations when
Nvidia first started working with the chip manufacturer. Tsai, who later
became CEO of TSMC, was in charge of all manufacturing at the time and
served as Nvidia’s main point of contact. “I made wafers for Jensen,” Tsai
said. “His brilliance and charisma were obvious from the beginning.”9
When TSMC first began to work with Nvidia, the entire industry was
working on a smaller scale. Tsai recalled building his first eight-inch-wafer
fabrication plant for $395 million, an amount that would be enough to buy
only a single chip-making machine today.
Within just a few years, Nvidia’s success in graphics made it one of
TSMC’s top two or three customers. Tsai remembered Jensen would
negotiate hard over pricing and would repeatedly cite how the graphics
company had only a 38 percent gross margin. One particular dispute
prompted Tsai to travel to California and meet with Jensen at a restaurant
that wasn’t much better than Denny’s.
“We tried to resolve the dispute. I forgot the details,” Tsai said. “But it
really hit me. Jensen taught me his philosophy of doing business called
‘rough justice.’ ” Jensen explained that “rough” meant the relationship was
not flat but rather had ups and downs. Justice was the important part. “After
a certain period of time, let’s say a few years, it would net out to roughly
equal.”
To Tsai, this was a way of describing a win-win partnership, though one
that acknowledged there wouldn’t be a win-win every single time.
Sometimes one side would get the better of a specific deal or incident, and
the next time it would be the other side. As long as it was roughly 50-50
after a few years—not 60-40 or 40-60—it was a positive relationship. He
remembers thinking Jensen’s approach made a great deal of sense.

Jensen Huang at his desk in 1999. (NVIDIA)

“Those things struck me about Jensen as a person and also a

businessman,” Tsai said. “Of course, he’s not shy in calling me when our
wafers were not coming out in time for him. Not shy at all. But together
with him, we met and resolved many adversities. If you look at both
companies, you cannot find a better partnership over the last three decades.”

ON FRIDAY, JANUARY 22, 1999, Nvidia finally went public. With the
Asian financial crisis over and the company’s finances in solid shape, the
stock proved irresistible to investors. The company raised $42 million from
its stock sale, and its shares ended the first day of trading up 64 percent at
$19.69 per share. At that price, Nvidia was valued at $626 million.10
The tone at Nvidia headquarters was muted: instead of exuberance, there
was relief. After several quarters when the company had nearly run out of
cash, the proceeds from the IPO brought a sense of security, at least for a
little while. It was by far the largest single windfall the company had
experienced—far larger than the bridge financing or any of the venture
rounds.
“We’ve got some breathing room now,” former engineer Kenneth Hurley
said, recalling his feelings on IPO day. “We’ve raised some money. We’re
not going out of business.”11
Jensen was more defiant than enthusiastic. “We have had some setbacks,
but I’m told I’m the hardest CEO to kill,” he told a Wall Street Journal
reporter when asked for comment about the IPO.12
Still, Nvidia’s executive team allowed themselves a rare moment to enjoy
the accomplishment and to dream about what might come next. During an
off-site management meeting, they discussed what each of them would do if
the company’s stock price ever hit $100 a share. (The share price was $25 at
the time.) Head of marketing Dan Vivoli pledged to get a tattoo of the
Nvidia logo on his leg. Head of sales Jeff Fisher would get one on his butt
cheek. Chief Scientist David Kirk agreed to paint his nails green, and head
of human resources John McSorley signed himself up to get a nipple
pierced. Two of the three cofounders went even bigger: Chris Malachowsky
would get a mohawk, and Curtis Priem would shave his head and get a
tattoo of the Nvidia logo on his scalp. Jensen agreed to have his left ear
pierced.13 Vivoli recorded the pledges on a paper placemat and had it
framed for display. In that moment, no one thought they would have to
follow through on their pledges anytime soon. It was almost impossible to
believe that the stock price would quadruple in the near term.
 Curtis Priem’s Nvidia logo hair. (NVIDIA)

WITH THE MONEY FROM THE IPO, Nvidia pursued ever-larger

strategic partnerships. The company had hired Oliver Baltuch, a tech-
industry veteran, to manage significant relationships with big companies
such as Microsoft, Intel, and AMD. Baltuch was given the authority to
spend freely. It was a significant change from his previous roles, where he
had to adhere to tight budgets.
One of Baltuch’s younger colleagues, Keita Kitahama, was a recent
college graduate who had been hired to ensure Nvidia graphics cards
worked well with major monitor vendors. Kitahama was naturally shy and
didn’t know much about the business development process. One day, while
Baltuch was drinking his daily cup of tea, Kitahama approached him and
asked, “What’s the best way to do this?”
Baltuch replied, “You have the industry’s hottest commodity. Use it.” He
was referring to Nvidia’s latest graphics cards, called the GeForce. He
instructed Kitahama to talk to another product manager, Geoff Ballew, and
scour Nvidia’s headquarters for as many spare GeForce cards as he could
find. Then, he was to “call up every single monitor company and tell them
you want to visit them and give them a free GeForce card.”
To Kitahama’s surprise, the tactic worked. Monitor manufacturers not
only took his call—they responded eagerly to his offer. They wanted early
access to the company’s newest products, no matter how they got them.
Baltuch employed a similar strategy with Intel. At Intel’s annual forum
for developers, he showed up with a box of fifty Nvidia cards and visited
vendors at every single booth, offering to replace the existing cards in their
machines with Nvidia’s. He knew that Nvidia cards had a significant
advantage because they were far easier to swap in and out, thanks to their
backwards-compatible software drivers. This ensured developers could play
with the newest Nvidia cards with little fear of a clunky installation process,
frequent crashes, or poor performance.
Even the biggest tech companies could not resist the allure of free Nvidia
graphics cards. Intel was then building several thousand development
workstations each year and sending the computers out to software
developers around the world. At the time, there was competition among
about ten graphics-card manufacturers to get inside the Intel boxes. Nvidia
won the contract, both because it had a better product and because its
strategy of giving out free cards ensured that Intel already had hands-on
experience with Nvidia’s chips.
The same strategy was employed with Microsoft, which made the
DirectX APIs that developers used to display media and run games on
Windows. Like clockwork, Nvidia cards would appear at Microsoft
headquarters every time Microsoft updated its API. “We blanketed them
through every major release of DirectX,” said Baltuch. “You didn’t even
have to ask for the resources.”14
Jensen’s guidance was direct. “Just go win. The idea was who could run
faster to get all the land.”
Nvidia still wasn’t a large company, with only around 250 employees.
And it didn’t generate huge revenues; for the fiscal year 1999, it reported
$158 million in sales, a fraction of what other tech companies such as
Microsoft ($19.8 billion), Apple ($6.1 billion), and Amazon ($1.6 billion)
made. But it had spent years sharpening its focus on technical excellence
and product execution. Now, that focus was finally paying off in the form of
something intangible, yet all-important: industry influence.
The company had stayed away from the gaming-console market since
Sega canceled its contract for the NV2 chip. But a few years later, in 1999,
Microsoft hinted that it was developing its first console and that it would be
based on the DirectX API. Nvidia’s existing relationship with Microsoft
opened the door for an Nvidia chip to power the console. For several
months, the two companies attempted to work out an agreement.
Microsoft soon changed direction, though. In January 2000, the company
gave graphics start-up Gigapixel, led by founder and CEO George Haber, a
development contract to supply the graphics technology for Microsoft’s
Xbox console. Microsoft invested $10 million in Gigapixel and an
additional $15 million to develop the Xbox chip. Haber moved his thirty-
three employees into a Microsoft building in Palo Alto, California.15
On March 10, just two months after publicly announcing the Xbox, Bill
Gates was scheduled to appear at the Game Developers Conference (GDC)
to present the console’s specifications and Gigapixel as its graphics
supplier. He had invited Haber to attend his speech, a copy of which was
provided to Haber for review beforehand. In it, Gates was supposed to say
that the relationship between Gigapixel and Microsoft had as much
potential to change the industry as an older partnership that saw IBM
choose a little-known software start-up named Microsoft to supply the
operating system for the original IBM PC. He would tell the world that
Microsoft had chosen Gigapixel for one reason: it made the best graphics
technology in the world.16 It was the sort of publicity, and endorsement by a
technology legend, that every start-up dreams of.
That dream would not last. Even after the Gigapixel announcement,
Nvidia continued to make the case that it was the right partner for the Xbox.
Jensen and Chris Diskin, Nvidia’s senior director of sales and marketing
who managed the relationship with Microsoft, often had weekly meetings
with Microsoft during the negotiation process. Jensen and Diskin spent
hours on their pitch material, sometimes working until midnight and then
starting again at 8:00 a.m. the next day. Both men kept up the same intense
work rate that had seen the company through its deepest crises and during
the development of the original RIVA 128 and the race to get its derivative,
the RIVA 128ZX, out to compete with Intel’s i740 chip. This time, they did
not have the threat of bankruptcy hanging over their heads. Even so, they
pushed themselves just as hard to salvage an opportunity to break into a
lucrative new market.
It helped their case that Nvidia’s reputation was at a new high.
“We had plenty of advocates inside Microsoft,” said Diskin about the
Xbox negotiations. “Game developers came out and said, we want Nvidia
because it’s easier to develop on and less of a risk.”17
On Friday, March 3, just a week before Gates’s GDC speech, Microsoft
executives Rick Thompson and Bob McBreen called Diskin and said they
wanted to reopen the Xbox contract and work out an agreement. Two days
later, they flew down from Seattle to San Jose, California, and spent most of
that Sunday in a conference room at Nvidia headquarters. Jensen, Diskin,
Thompson, and McBreen agreed that Nvidia would replace Gigapixel as
Microsoft’s graphics-chip partner. The new console would instead use a
new, custom-designed chip from Nvidia, with Jensen and Diskin insisting
that Microsoft pay $200 million up front to cover the new chip’s research
and development, an amount that required personal approval from Bill
Gates himself. By ensuring a significant upfront payout and sign-off from
Microsoft’s CEO, they felt that they had locked down the Xbox job and
protected Nvidia against an outcome similar to the one that had just befallen
Gigapixel.
On Monday, Microsoft executives informed Haber they had decided to
go with Nvidia. He was stunned. The prior week, he had been talking to
Wall Street investment bankers about going public in a billion-dollar IPO
on the back of the Xbox deal, and even potentially buying 3dfx and other
graphics-chip companies; Nvidia executives were, after all, not the only
ones who daydreamed about market dominance. Now, he had nothing left
except a $15 million development fee that Microsoft had agreed to pay even
though the contract was now canceled.
To this day, Haber feels bitter about what happened with Microsoft and
the Xbox contract. “Today, I would be running a trillion-dollar company,
not Jensen, I presume,” Haber said.18
In the week of Bill Gates’s GDC announcement that Nvidia, in fact,
would be the graphics-chip supplier for the Xbox, Nvidia’s stock soared
above $100 per share. The executive team found itself having to follow
through on pledges that seemed like a joke less than a year earlier. But
follow through they did: Malachowsky got his mohawk; Fisher, Priem, and
Vivoli got their tattoos; Kirk’s nails turned green; and McSorley and Jensen
got their piercings.

IT WOULD BE ONE OF THE last happy moments at Nvidia for Priem.

In the late 1990s, the Nvidia cofounder had started to clash more frequently
with the company’s engineering staff. During the development of one
generation of chips, Priem had found a flaw in the chip architecture. He
fixed it and, without telling anyone, pulled down some chip documentation
from the communal file server and replaced it with an updated version. This
was how he was used to operating in Nvidia’s early years, and his
colleagues then had accepted it. Now, however, he was part of a much
larger organization and found that “the software team went haywire because
they were using the original documentation for some of the code in their
software.” When they learned that he had removed the documents entirely,
“they threw a fit.”19
Priem was adamant that the architecture had to be fixed. The engineers
asked Jensen to intervene. In a heated argument, Priem insisted that he
could do whatever he wanted because he had personally designed the
architecture for Nvidia’s chips.
“It was my architecture,” he kept repeating.
It was the wrong thing to say to Jensen, who wanted to build a more
communal culture. Nvidia as a whole was to be recognized for its
achievements—not individuals. After Jensen came back from important
business trips, Priem observed that he would always describe his own
actions using the plural “we” instead of the singular “I.” Priem was initially
skeptical, thinking, “What’s this ‘we’ stuff? I don’t know anything about
negotiating contracts with fabs. But Jensen was right. We all did it together.
We all share the credit.”
When it came to chip designs, Priem could get possessive. He tended to
talk about it as “his” work, “his” architecture. Jensen insisted that Priem do
the opposite: that he consider it the collective property of the entire
company, which it was. Jensen would reply, “No, it’s our architecture. You
didn’t do it. We did it. You don’t own those files.”
So after learning of Priem’s unilateral decision to fix the flaw in the chip
architecture, Jensen exercised his authority as CEO and overruled his
cofounder. He told Priem to roll back the changes, to restore the original
documentation files back to the server, and to never take action on chip
documentation again without first informing all of the people whose work
might be affected by the change. Using the old documents, the software
team was able to finish their code and eventually figured out how to fix the
flaw in the chip architecture the following year.
Later on, after Nvidia hired John Montrym and a new generation of 3-D
graphics engineers, Priem became more disruptive and began interfering
with product development. “I would get in the way of shipping,” Priem
recalled, because he always wanted a chip to be perfect and to defend it
against changes to the architecture he believed he had created.
Priem had started to realize his own deficiencies. He recalls being in an
executive meeting with a graphics specialist in anti-aliasing, a technique
used to smooth out jagged edges and soften the transitions between an
object’s line and the background. “Wow. I read his papers to mimic anti-
aliasing at Sun,” Priem thought, impressed by the presentation. “I give up.
How in the world does Curtis live up to all these specialists?”
Priem was now fighting with Jensen so often, and with such intensity,
that the company brought in a special workplace consultant to try to resolve
their differences. After many arguments, Jensen suggested Priem leave the
engineering group to work on defending Nvidia’s intellectual property and
patents. Priem accepted. “For architectures, my job was really done in the
first two years,” he said. “I worked on products for five years. I was pulled
off products and put on IP. That allowed the other 3-D experts we hired
from Silicon Graphics to come in and create better products than I could
have done.”
In 2003, a few years after his reassignment, Priem took an extended leave
of absence to deal with some issues he had with his then-wife. Jensen tried
to help Priem by using his connections to help Priem find the best marriage
counselors. But after three months, Jensen could no longer dodge employee
questions on the whereabouts of the company’s cofounder and chief
technical officer. He gave Priem a choice and an ultimatum: return to work
full-time, transition into a part-time consulting role for Nvidia, or resign.
Jensen even suggested a new mobile architecture project that Priem could
work on, so that he could oversee one last job before retiring. Priem decided
to leave Nvidia. “I was tired, beaten up, and demoralized. I needed to
resign,” he recalled. “I always wish I could have stayed.”
 Nvidia and the Holy Grail. (MICHAEL HARA)

Two decades later, Jensen still seemed pained over the circumstances of
Priem’s departure. When I explained to him that Curtis felt he didn’t have
the skills to keep up with other graphics engineers, Jensen offered a stern
reply: “Curtis is smart. He could’ve learned it.”

IN EARLY 2000, JENSEN AND Michael Hara, who by this point had
moved out of marketing and now ran the investor relations team, embarked
on a multi-city road show to meet with bankers, investors, and fund
managers to raise capital for Nvidia.
“We’re flying from city-to-city with the bankers,” recalled Hara. “They
kept asking Jensen, ‘What things do you watch? What do you find funny?’
”20
Jensen thought for a moment: “Monty Python and the Holy Grail.”
The 1975 comedy was the first feature film released by the British
comedy troupe Monty Python. Among its many memorable scenes was one
set during an outbreak of plague when two peasants pull a cart of dead
bodies through a grubby medieval village.
 “Bring out your dead!” one of them calls, banging a wooden spoon on a
metal triangle to announce himself.
A villager stops the cart to deposit an old man’s body, except the old man
isn’t dead.
“I’m not dead!” he protests.
“Here, he says he’s not dead,” says the cart master.
“Well . . . he will be soon. He’s very ill,” says the villager.
“I’m getting better,” says the old man.
The three of them banter for several more moments. The old man
continues to insist that he’s not dead, while the villager and cart master try
to persuade him that he really belongs on the cart. Eventually, the cart
master strikes the old man on the head, and the villager deposits the man’s
body.
“Ah, thanks very much!” he says, with apparently sincere gratitude.
Jensen felt that many of the questions from potential investors followed
the same grim logic: they thought Nvidia would die.
“Why would we invest in a graphics company?” they asked. “You would
be the fortieth one we’ve backed, and all the others have gone out of
business. Why would we do this?”
Investor pessimism became the theme of the road show. Investors
demanded more revenue from Nvidia. They assumed Intel would eventually
crush the entire graphics industry with a new chip. They expected Nvidia to
follow in the footsteps of so many competitors before it: Rendition, Tseng
Labs, S3, 3DLabs, Matrox, and others.
This attitude irritated Jensen, who believed Nvidia was nothing like any
other graphics company that had ever existed. The company’s pitch to
investors went like this: its chips were better than anyone else’s, it had a
strong, defensible position, and its business strategy allowed it to move
more quickly and innovate faster than any other chip maker.
But most of all, it had Jensen, who had learned how to manage the
company as an extension of himself. Everyone at the company shared his
singular focus on the mission. Everyone shared his work ethic. Everyone
worked as fast as humanly possible in order to keep Nvidia one step ahead
of the competition. And if anyone faltered or doubted, a sharp word from
Jensen quickly brought that individual into line.
 Some investors did believe in Jensen’s future vision for Nvidia and his
ability to keep the company trained on that vision. Morgan Stanley
ultimately raised $387 million for Nvidia when it made a secondary
offering of equity and convertibles in October 2000.
At the conclusion of the round, the Morgan Stanley team presented
Michael Hara with a full-color illustration of the road-show team, depicted
in a mashup of Monty Python and the Holy Grail. Nvidia’s deceased
competitors are shown as bodies on the plague cart. Investors who asked
irrelevant questions are miniature “knights who say Ni.” Jensen is the
valiant King Arthur, defeating a black knight in single combat.
“I’m better. I’m faster. You can’t beat me!” he says.
 CHAPTER 7

GeForce and the

Innovator’s
Dilemma

IN HIS FAMOUS BOOK The Innovator’S Dilemma, the late Harvard

Business School professor Clayton Christensen argued that a company’s
success often contains the seeds of the same company’s own failure, and
this was especially noticeable in the technology sector. He posited that
every industry was shaped not by random chance but by regular and
predictable cycles. First, start-ups would release a new, disruptive
innovation that was less capable than the market-leading offering from a
major company but positioned at the low end of a market. The successful
incumbent would ignore the less profitable niche, focusing instead on
launching products that sustained and added to its current, robust profit
streams. But new-use cases would eventually appear for the disruptive
innovation, and the start-ups that capitalized on it would usually be able to
iterate and innovate faster than the established corporation. Eventually, the
start-ups would have far more capable products, and by the time the
incumbent realized it was in trouble, it was too late. For instance,
Christensen wrote about how Control Data, the market leader for fourteen-
inch mainframe disk drives, failed to achieve even a 1 percent share of the
subsequent market for eight-inch minicomputer disk drives. Analogous
shifts happened to makers of 8-inch drives when the smaller 5.25-inch and
3.5-inch drives came out. Each time, the cycle began again, and a new wave
of incumbents fell to the start-ups.1
The Innovator’s Dilemma is one of Jensen’s favorite books, and he was
determined not to let such a fate befall Nvidia. He knew it would be hard
for a rival to surpass Nvidia’s high-quality chips, because competing at the
top of the market required massive investments in capital and engineering
talent. Influenced by Christensen, he decided that the threat was from a
low-cost player.
“I’ve seen that before,” he said. “We build Ferraris. All of our chips were
designed for the high end. The best performance, the best triangle rate, and
the best polygons. I don’t want to let someone come in and be the price
leaders, lock me out at the bottom, and climb their way to the top.”2
He studied the business strategies of other leading companies for
inspiration on how to fend off an attack from below. As he looked at Intel’s
product lineup, he noticed that its Pentium series of CPUs had a range of
clock speeds—a key measure of processor performance—but all of the
Pentium cores themselves shared the same chip design and theoretically had
identical features and capabilities.
“Intel just builds the same damn part. They are selling customers
different products based on speed binning,” he said, referring to a process in
which components that fail quality checks at high speeds can be repurposed
to work at lower speeds where they can function properly.
Jensen saw that Nvidia could stop throwing away parts that failed quality
tests as a matter of course. True, these parts were not suitable for the
company’s Ferrari-grade chips. But if they were otherwise functional at
lower speeds, Nvidia could repackage them into a less capable (and
therefore cheaper) version of the company’s mainline products. This would
increase the number of usable parts manufactured from every wafer of
silicon and improve the company’s yields—an industry measure of
production efficiency.
At an executive staff meeting, Jensen asked his operations manager,
“How much does it cost us to package, test, and assemble a part?”
The answer was $1.32, a small number in the expensive world of chip
manufacturing.
“That’s it?” Jensen replied, incredulous. This looked like a clear
opportunity to make something out of nothing. Rejected parts were
generating no revenue for Nvidia; they were thrown out. But by spending a
bit more to spruce up the rejected parts for use on less intensive chip lines,
Nvidia could create a whole new derivative product line that could turn a
profit, without the expensive and time-consuming process of research and
development. The line would serve as a defense against competitors for
whom the low-cost chip was the main product. With the new low-cost parts,
Nvidia could easily afford to price its chips down so much that its
competitors would be forced to sell at a loss. Nvidia might lose money on
its cheaper lines, but sales from its Ferraris would more than compensate.
More important, Nvidia would avoid the trap that had ended the company’s
one-time rival, 3dfx, which had begun spending so much time and money
developing new chips that it fell behind in the race to continuously
innovate.
The strategy was dubbed “ship the whole cow,” a reference to how
butchers find ways to use almost every part of a steer carcass, from nose to
tail—not just the prime cuts like the tenderloin and the ribs.
“It became a very powerful tool for us and allowed us to fine-tune our
offering,” Jeff Fisher said. “We could build a low-yield part at the high end
and create four or five different products in the stack. It helped push ASPs
(average selling prices) up.”3 It also allowed Nvidia to test the demand for
more expensive, top-end products among enthusiast gamers who were
willing to pay more for higher performance.
The rest of the graphics industry would soon follow suit, especially as
Nvidia’s rollout of “ship the whole cow” was responsible for nearly putting
another of its competitors, S3 Graphics, out of business.
“ ‘Ship the whole cow’ is something the graphics industry takes for
granted, but it was an important strategy that made a big difference,” Nvidia
board member Tench Coxe said.4 It was a testament to Jensen’s strategic
forethought and his restless desire to anticipate any threat to Nvidia’s
future. After all, now that Nvidia was a market leader rather than one start-
up among many, Jensen knew that he had a permanent target on his back.
“I don’t think people are trying to put me out of business,” he once said.
“I know they’re trying to.”5
JENSEN KNEW THAT TECHNICAL SPECIFICATIONS alone did not
sell a chip. Marketing and branding mattered almost as much. His
competitors all took different approaches to positioning their products in the
marketplace. Some went with outlandish, hypermasculine branding that
appealed to gamers’ self-conceptions: 3dfx’s Voodoo Banshee, the ATI
Rage Pro, S3 Savage, Righteous Graphics. Others went more technical or
industrial-sounding, such as the Matrox G200 or the Verite 2200. Nvidia
tended to split the difference, branding its chips with names that both
conveyed technical excellence and resonated emotionally, such as the RIVA
TNT—RIVA standing for (as we’ve seen) Real-time Interactive Video and
Animation Accelerator, and TNT for TwiN Texels, or the ability for the chip
to process two texture elements (texels) at once. But to the average
consumer, of course, it was a name “that had something to do with
explosions,” as one engineer put it.6
Yet in a crowded market, Nvidia decided to bend the rules in order to
stand out even more. In 1999, it launched the successor to the RIVA TNT2
series, which it called the GeForce 256. It was incontrovertibly true that the
GeForce 256 represented a significant advancement in traditional graphics
capabilities, which by this point was typical and expected of every new chip
generation Nvidia produced. It featured four graphics pipelines, allowing it
to process four pixels at the same time. It also integrated a hardware
transform and lighting (T&L) engine, which meant it could take on the
calculation tasks essential to the moving, rotation, and scaling of 3-D
objects. Those tasks were usually handled by the CPU; the GeForce 256
thus took even more computational burden off of the CPU and made the
entire computer run faster.
“By putting dedicated hardware to do that, you could suddenly process a
lot more geometry and make much more interesting pictures,” said former
chief scientist David Kirk.
The management team at Nvidia thought that this was far too technical to
pitch to customers. The typical in-house naming formula of acronym and
number wouldn’t work. Nvidia needed something bigger to market its new
product.
“We had to find a way to position this product as a much better graphics
processor for 3-D graphics than anything out there,” said Dan Vivoli. “The
chip is big. It has texture and lighting. We’ve got to charge a lot for it. We
need to come up with a way to make sure how great this thing is.” He
challenged his product-marketing team to come up with something brilliant.
Product manager Sanford Russell got to work on potential ideas. Russell
loved bouncing branding, naming, and positioning strategy off his
colleagues, including Jensen and Kirk.
“It was never, walk into a room with a PowerPoint and here’s the name.
It was constant discussion,” Russell said. “We were asking them about the
technology. What works. What doesn’t work.”7
Russell grabbed Michael Hara for a thirty-minute brainstorm session to
figure out how to market the GeForce 256 more effectively, and both
executives remember coming out of the room with the notion to call the
new chip the first entry in an entirely new product category altogether: a
graphics processing unit, or GPU, which would be to graphics rendering
what the computer’s main central processing unit (CPU) was for all other
computational tasks.
The technical specialists at Nvidia knew that their chips were special.
The average computer user, though, didn’t quite appreciate the complexity
or value of a graphics chip. Unlike the CPU, which sounded like the main
piece of equipment essential to any computer, graphics cards were just one
peripheral among many. A special designation for graphics chips that also
drew an explicit comparison to the CPU would, for the first time, make
them stand out as truly exceptional.
“My memory has it Mike Hara and I were in the room when we came up
with GPU,” said Russell. “It didn’t seem that important at the time. We
were working fourteen-hour days.”
He soon told Vivoli the GPU idea, and Vivoli liked it. “Sometimes Dan
would take a while to chew on an idea, but GPU rippled through pretty
quick.”
Within days, the marketing team committed to the GPU designation,
which would not only help Nvidia stand out from other graphics chips but
also make it easier to command a pricing premium. The world understood
that CPUs were supposed to cost hundreds of dollars. Nvidia chips were
sold at wholesale for less than $100 each, even though they were just as
complex as, and had more transistors than, CPUs. Once the company
started marketing all of its chips as GPUs, the pricing gap narrowed
considerably.
 Even so, the GPU moniker, when applied first to the GeForce 256, was
controversial among Nvidia engineers, who pointed out that a chip could
not truly be called a GPU unless it had several features that the GeForce
256 did not. The chip lacked a “state machine,” a technical term meaning a
dedicated processor that transitions from various states to execute and fetch
instructions, in the same manner that a CPU does for programming
commands. It was not programmable, meaning that graphics styles and
features could not easily be customized by third-party developers. Instead,
developers would have to rely on a fixed set of Nvidia-defined hardware
functions. Moreover, the GeForce 256 did not have its own programming
language.
But the marketing team argued that such features were already planned
for the next generation of graphics chips. And even without them, the
GeForce 256 delivered a step change in performance that would be obvious
to every gamer and computer enthusiast in the world. While not literally a
GPU, the GeForce 256 could still be a category-defining product. The chips
that would follow—the “true” GPUs that would be fully programmable by
outside developers—were coming soon enough.
So Nvidia’s marketing team pushed forward with the GPU name over the
dissent of Nvidia’s engineers. “We don’t need anyone’s approval,” Vivoli
told them. He didn’t think anyone outside of the industry would really care
about the technical definition. Besides, “we knew the next one was going to
be programmable. We decided to take a leap, make a stretch that this is a
GPU.”8
When Jensen announced the GeForce 256 in August 1999, he did not shy
away from hyperbole. “We are introducing the world’s first GPU,” he
declared in the press release. “The GPU is a major breakthrough for the
industry and will fundamentally transform the 3D medium. It will enable a
new generation of amazing interactive content that is alive, imaginative,
and captivating.”
It may have been the first time that the company engaged in significant
marketing embellishment for a major launch—and it worked. Vivoli made
the decision not to trademark “GPU” because he wanted other companies to
use the term, the idea being that Nvidia had pioneered an entirely new
category. Hyperbole would become reality: the GPU moniker became an
industry standard and helped Nvidia sell hundreds of millions of cards in
the ensuing decades.
 Vivoli had another idea for the GPU’s launch: to actively intimidate
rivals. An Nvidia marketer unfurled a banner advertising the GeForce 256
on a highway overpass that led straight to the headquarters of 3dfx (this was
before its eventual bankruptcy). The banner announced that the new Nvidia
GPU would change the world and crush the competition. The state police
quickly removed the banner, which was being displayed illegally, and
Nvidia received a formal reprimand. Still, it had served its purpose. “It was
Art of War. We wanted to demoralize them,” Vivoli said. Nvidia was
learning how to bend the world to its will.

MODERN GRAPHICS CHIPS ORGANIZE computation through what

is called a graphics pipeline, turning geometry data with object coordinates
into an image. The first stage of this process, called the geometry stage,
involves transforming object vertices, or points, in a virtual 3-D space
through scaling and rotation calculations. The second stage, rasterization,
determines the position of each object on the screen. The third stage, called
the fragment stage, calculates the color and textures. In the final stage, the
image is assembled.
Early graphics pipelines involved fixed-function stages each with a
handful of hardwired operations. Nvidia and its competitor graphics-card
makers each defined how its chips would handle all four stages in the
pipeline; third-party developers could not change how the chips rendered
anything, meaning that they could only create visual effects and artistic
styles from a menu of options set by the chip designers.9 Because every
programmer had to use the same handful of fixed-function operations, every
game on the market looked similar—none could stand out through visuals
alone.
David Kirk, Nvidia’s chief scientist, wanted to change all this by
inventing a true GPU. His idea was to introduce a new technology called
programmable shaders. These would open up the graphics pipeline to third-
party developers, giving them the ability to write their own rendering
functions and exert more control over how they presented their games
visually. The shaders would allow developers to make visuals in real time
that rivaled the best computer-generated graphics in movies. He argued that
developers would quickly adopt programmable shaders in their games, as
they knew far better than chip designers how to create cutting-edge visuals.
This, in turn, would push gamers to Nvidia cards, because they would be
the only cards on the market that could support the advanced new graphics.
The downside was that programmable shading, and therefore a true GPU,
could be enabled only by revising how Nvidia chips were designed. It
would be an expensive and time-consuming undertaking, even for an
established player.
Kirk knew that the technological upside would be clear to Jensen, who
would have the final say. He also knew that Jensen would fixate on cost:
how much Nvidia would have to invest to create the technology, whether
the market was ready for it, and how much more revenue it would bring in.
Although Jensen seemed enthusiastic initially, Kirk didn’t know yet
whether that was a good sign.
“One of the things that happens with Jensen is right before he’s about to
kill your project, he will sound optimistic when he’s talking to you about
it,” Kirk said.10
To ensure his project’s survival, he stoked Jensen’s ever-present fear of
being outflanked by the competition. He pointed out that Nvidia’s lead in
fixed-function graphics acceleration would inevitably erode; the fixed-
function operations of a traditional graphics chip would someday become
miniaturized enough for Intel to include in a section of their CPUs or in a
motherboard chip, obviating a separate graphics card entirely. He also said
that programmable shaders could one day potentially open up other markets
outside of gaming.
“Okay,” said Jensen after hearing Kirk’s thoughts. “I’ll buy that.”
In February 2001, Nvidia released the GeForce 3, whose programmable
shader technology and support for third-party development of its core
graphics functions made it the first true GPU. Kirk’s analysis was proved to
be correct. The GeForce 3 was a blockbuster success. By the third fiscal
quarter of 2001, Nvidia’s quarterly revenue reached $370 million—an 87
percent year-over-year increase. It was now generating sales at an
annualized run rate of $1 billion, achieving that milestone faster than any
other semiconductor company in U.S. history. The previous record holder,
Broadcom, had achieved it in thirty-six quarters; Nvidia beat that by nine
months. By the end of the year, the stock price had tripled over the prior
three quarters. The company was now twenty times more valuable than it
had been on IPO day, thanks to a combination of strategic vision, relentless
execution, and the paranoia that kept Jensen and his executive team on
guard for threats that could come from anywhere, at any time.

THE DRIVE TO CONTINUALLY DIVERSIFY Nvidia’s business led the

company straight to Apple. Historically, Nvidia had not sold much to Apple,
partly because Nvidia optimized its products for Intel-based CPUs, which
Apple did not use. But in the early 2000s, Nvidia won a small contract to
supply graphics chips for the consumer-oriented iMac G4. The computer
was the successor to the colorful all-in-one iMac G3 that marked Steve
Jobs’s triumphant return to Apple in 1998.
Chris Diskin, who had successfully won Microsoft’s Xbox business, was
put in charge of the overall sales relationship with Apple. He worked with
Dan Vivoli to figure out a strategy to get Nvidia’s GeForce chips into more
Apple computer products. The key break came thanks to an old, iconic
Pixar short film.
By this point, the centerpiece of Nvidia’s sales pitch to PC makers was
the graphics demo, when it would show off the advanced features and raw
computing power of its chips. In the past, the company had used third-party
games to wow their audiences. But as Nvidia’s cards grew more powerful,
older games no longer fully showcased the breadth and depth of a new
chip’s capabilities. Vivoli decided to put more time and resources into
creating better graphics demos for the sales team. He even hired a former
colleague from Silicon Graphics named Mark Daly for the specific purpose
of enhancing Nvidia’s demos.
Vivoli knew that graphics demos would have the most impact if Nvidia
had a thorough understanding of its audience. Earlier demos were targeted
at engineers, which is why they showed off the specific features and
capabilities of Nvidia’s new chips. Like the 3-D cube that Voodoo Graphics
impressed the Hambrecht & Quest conference with in 1996, such demos
were only impressive if you knew the computations that were occurring
“under the hood.” A nonengineer wouldn’t necessarily know what he or she
was looking at. So Vivoli changed the focus of demos away from a cold
demonstration of graphics performance—the visual equivalent of reading a
list of benchmark metrics—and gave them a sentimental side.
In a brainstorm meeting during the development of the GeForce 3, Daly
believed he had come up with the perfect means of showing off Nvidia’s
new chip. Pixar’s two-minute animated short Luxo Jr. had been a watershed
moment for computer animation. When this film about hopping table lamps
was first released in 1986, it demonstrated just what computer-generated
imagery (CGI) could do, even at that relatively early stage. But it had taken
a huge amount of computational power to produce. Each frame was built on
a Cray supercomputer and took three hours to render. At twenty-four frames
per second, the computer needed nearly seventy-five hours to generate just
a single second of film. Daly thought Nvidia should make a Luxo Jr. demo.
Vivoli gave him the green light. “That’s a great idea. Go make that
demo,” he said.
A few months later, Daly reported to Vivoli that the team was making
good progress, but there was the fact that Luxo Jr. was Pixar’s property. If
Nvidia used it for a public demonstration, the company ran the risk of
infringing on Pixar’s copyright.
Vivoli didn’t want anything to derail what was shaping up to be an
extraordinary showcase for the important GeForce 3 launch. He waved off
Daly’s concerns.
“That’s fine. Don’t worry about it. I’ll go figure it out,” Vivoli said. Both
he and David Kirk had contacts at Pixar, and they worked them to get
approval for the demo. Their request eventually reached Pixar’s chief
creative officer, John Lasseter, who had directed Toy Story, A Bug’s Life,
and would later direct Cars. Lasseter declined it. He wasn’t keen on Pixar’s
iconic character, which at the time was part of the company’s logo and
appeared prominently at the start of every Pixar movie, being used to sell
graphics chips.
In the meantime, Daly’s team had completed the demo, and it looked
every bit as impressive as he had envisioned. Vivoli thought, “What if we
showed the demo to Steve Jobs?” He believed a real-time rendered version
of Luxo Jr. would be effective, because it would touch on a landmark
moment in Jobs’s own career and the development of computers more
broadly. It would also show that the new chip was powerful enough to rival
the graphics capabilities of a supercomputer, yet precise enough to
faithfully reconstruct an important piece of art.
Vivoli and Diskin went to meet with Jobs at Apple’s headquarters.
During the first part of the demo, the Nvidia team showed off Luxo Jr.
using similar shots and angles as the original. It was suitably impressive.
Jobs said that it “looks good.”
 Then they ran the demo again, but this time Vivoli started clicking
around the demo, which changed the position or the angle of the camera.
The camera movements showed that, unlike a static video, Nvidia’s chips
could render the entire scene in real time. The user could change and watch
the scene from any angle with realistic lighting and shadow effects. Now
Jobs was blown away. It was impressive enough that Nvidia’s GPU could
render animations in real time, and with comparable visual fidelity, that had
taken Pixar’s supercomputers weeks to generate. But on top of that, it
offered real-time interactivity. Jobs decided that the Power Mac G4
computer would offer the GeForce 3 as a premium option.
Jobs also asked whether Apple could use the demo at the 2001 Macworld
in Tokyo. Vivoli told him about the copyright issue, to which Jobs replied
that he would check with the people at Pixar. Diskin and Vivoli later
laughed about this moment: Jobs was the CEO of both Apple and Pixar, so
he was effectively asking himself for permission.
Jobs ended the meeting after about twenty minutes to head to another. As
he prepared to leave, he had some parting advice for the Nvidia team.
“You guys have really got to do some work on mobile because ATI is
kicking your ass in laptops,” he said, referring to Nvidia’s main rival after
the demise of 3dfx.
Without hesitating, Diskin responded, “Actually, Steve, I think you’re
wrong.”
The room went quiet. Jobs fixed Diskin with his intense stare and said,
“Tell me why?” Diskin got the sense that not many people challenged Steve
Jobs, and it was clear he expected a good answer.
Diskin had one. He explained that Nvidia’s chips did indeed consume
more power—indeed, more power than most laptops could muster—
because they offered the higher performance that desktop users required.
But their performance and power consumption could easily be lowered to
meet laptop specs. In fact, Diskin argued, if Nvidia lowered the clock speed
of its chips down to match the speed—and therefore power requirements—
of ATI’s chips, Nvidia’s chips would actually offer better overall
performance. It wasn’t that ATI was kicking Nvidia’s ass in laptops, as Jobs
had thought. It was just that Nvidia had not needed to create a chip
specifically for a lower-power laptop model, when a throttled-down version
of its flagship line would do the job just fine.
 “We have more headroom,” Diskin said, summarizing his main
argument.
Jobs stared at him one more time. “Okay,” was all he said. The meeting
was over.
Thirty minutes later, Diskin got a call from Apple executive Phil Schiller.
“I don’t know what you told Steve, but we need your entire laptop team in
here tomorrow for a full day to review your silicon,” he said. Nvidia would
go from no presence in Apple laptops to a nearly 85 percent share of
Apple’s entire computer lineup in a matter of years. Diskin got the chance
to prove himself, and Nvidia’s chips, thanks not only to his demo but also to
his quick thinking and his willingness to challenge one of the most
intimidating figures in the tech industry.

NVIDIA WAS GOING FROM STRENGTH to strength. It had added one

hundred employees from its vanquished rival 3dfx, won an Xbox gaming-
console business deal that would go on to generate $1.8 billion in revenue
over its lifetime, and secured contracts to make chips for Apple’s Mac
computer lines. These achievements led to impressive financial growth and
a skyrocketing stock price. But all the new business required attention from
management and engineering, drawing focus away from Nvidia’s core GPU
products—which resulted in one of the worst product launches in the
company’s history.
In 2000, ATI Technologies acquired the small graphics firm ArtX for
$400 million. ArtX specialized in graphics chips for gaming consoles; its
founding engineers had worked on the Nintendo 64 console at Silicon
Graphics before starting out on their own, and the company had recently
secured a contract to develop the graphics chips for the N64’s successor, the
Nintendo GameCube. ATI’s purchase of ArtX gave it instant credibility in
the field of console games—and a group of engineers who immediately
started working on a chip called the R300. ATI would release the chip on a
dedicated graphics card, the Radeon 9700 PRO, which went on sale in
August of 2002.
Nvidia, in the meantime, was caught up in a legal dispute with Microsoft.
The tech giant had recently revised their vendor agreements around
information sharing and intellectual property rights for the Direct3D API.
The next major update to Direct3D, called Direct3D 9, would be released in
December of 2002, with significant improvements that would be essential
to the next generation of chips. There was a catch, though. Chip companies
could not get access to Direct3D 9 documentation and could therefore not
build around its new features until they signed the new agreement. Nvidia
felt that the new language was overly favorable to Microsoft and refused to
sign the new contract until it got better terms.
A business problem created an engineering challenge. Nvidia was
designing its next chip, which it called the NV30, without access to the
upcoming version of Direct3D’s technical specifications. “We ended up
developing the NV30 without a lot of guidance from Microsoft,” David
Kirk said. “We had to guess what they were going to do. We made some
errors.”
Confusion reigned, both from the lack of clear guidelines from Microsoft
and from a lack of coordination between Nvidia’s own teams. A former
employee remembers an incident when a group of hardware and software
engineers were standing in a cubicle and looking at the lackluster
performance data from the NV30 during its development phase. One
perplexed software engineer said that it was almost as if the hardware fog-
shader feature had been removed. The hardware architect replied, “Oh yeah,
we took that out. Nobody used that.”
The software team was stunned. Fog shaders were still widely used in
most games, as they allowed developers to save on graphics computation by
blurring object details the further away they were, as though in fog.
Nvidia’s hardware team did not consult anyone before they removed it, nor
did they seem to understand how important it was. All of a sudden, it was as
though Nvidia’s various teams were siloed—the very kind of organizational
structure the company had rejected from the start.
Another Nvidia employee recalls a similar meeting where a hardware
engineer was presenting a list of different features in the NV30. One of the
developer-relations employees noticed that the list was missing an
important feature called multi-sample anti-aliasing (MSAA), the technique
to smooth out jagged edges and soften the transitions between an object’s
line and the background. He asked, “What’s up with 4X MSAA? What’s
going on?”
The hardware engineer replied, “We don’t think it’s a big deal. It’s kind
of untested.”
 The developer-relations employee was aghast. “What are you talking
about? ATI is shipping this feature in a product already, and gamers love it.”
Once again, Nvidia’s own engineers seemed to be unaware of what the
market wanted. “NV30 was an architectural disaster. It was an architectural
tragedy,” Jensen later said.11 “The software team, the architecture team, and
the chip-design team hardly communicated with each other.”
Nvidia thus failed to ensure that the NV30 met all the benchmarks for the
biggest games of the season. New graphics chips were reviewed in the
press, and a common feature of reviews was the benchmarking process,
wherein independent reviewers test certain metrics, such as frame rate per
second, in specific, graphics-intensive games, and under different
resolutions. Standard benchmarks give gamers a series of quantitative
reference points so that they don’t have to rely on subjective analyses of a
graphics card’s quality (or on marketing by those cards’ makers). During the
NV30’s development, it became clear that the chip would not win many of
the benchmarks for the games that consumers then cared most about. For
the first time since the NV1, Nvidia was about to release a card that was not
at the very top of the market in terms of performance.
ATI, in contrast, had agreed to sign the contract with Microsoft so that it
could optimize the R300 with Direct3D 9 from the start. The chip and the
new card that housed it, the Radeon 9700 PRO, worked perfectly and was
fully compliant with Microsoft’s latest release of the API. It ran the latest 3-
D games, including Quake 3 and Unreal Tournament, at high resolutions
with little trouble. It was able to render pixels in more vibrant 24-bit
floating point color, an upgrade from the 16-bit color used in the industry’s
previous generation of chips. It had far better anti-aliasing capabilities than
its competitors, making polygons sharp and lines crisp. And it came out in
August 2002 in time for the fall back-to-school rush.
Nvidia’s NV30 did none of these things. It didn’t work well with
Direct3D 9, which meant that new games performed poorly under their
highest graphical settings. It had been optimized for 32-bit color, which
technically leapfrogged the 24-bit color system on the Radeon 9700 PRO,
but Direct3D 9 didn’t support 32-bit. Nvidia was forced to tell its graphics-
card partners to delay the launch of the company’s new products that
carried the NV30 by about five months. The delay at least allowed Nvidia
to try to make the chip more competitive against the Radeon 9700 PRO but
caused Nvidia to miss the critical fall launch window.
 Once they compared the NV30-based GeForce FX card with the Radeon
9700 PRO, Nvidia’s engineers decided to overhaul the design of their chip
to make it more competitive. They created software workarounds to
“translate” the new features of DirectX for the NV30. “We had to do
backflips to run DirectX 9 calls,” Dan Vivoli said. “A call would be made to
DirectX, and we’d have to turn it into something else our chip could run.”
These “calls,” or graphics instructions sent to the DirectX API, required
more processing power, which forced Nvidia to turn up the clock speed of
the NV30. The excessive heat that resulted led Nvidia to put a big dual-slot
fan over the chip, which made a loud, high-pitched noise whenever it
activated.
“It was a horrible experience for the gamers who were using the chip
because it was so loud,” Vivoli said. The fan noise became a constant
talking point among customers. The only engineering solution Nvidia could
come up with was to write an algorithm to change the timing of the fan’s
rotation, but that would take time, and it wasn’t that effective in the end.
To salvage a small shred of the company’s reputation, a member of the
marketing team suggested making a spoof video about the fan noise to
indicate that it was an intentional feature. “We just fell on our sword. We
put together a video where the GeForce FX was on the end of a leaf blower
blowing leaves. We showed people cooking off of it because it’s so hot,”
said Vivoli.
This at least partially placated the gaming community, which appreciated
Nvidia’s ability to poke fun at itself and admit a misstep. It also helped
undercut negative mentions of the card. Whenever a competitor tried to
point consumers to how loud the GeForce FX was, they found Nvidia’s
self-critical video.
Although the video was a PR win, it did little to help the chip’s fortunes
in the market. Compared with the R300, cards based on NV30 were more
expensive, ran hotter, ran games slower, and had an excessively loud fan.
Sales during the all-important holiday quarter fell 30 percent versus the
prior year, and the company’s stock price plummeted by 80 percent from its
peak just ten months before. Nvidia was reliving the NV1 nightmare all
over again. The company’s individual teams had lost touch with each other,
and the company as a whole had somehow lost touch with its core
consumer base.
 Jensen was furious at the poor planning and execution on the chip. He
called out its engineers at an all-company meeting.
“Let me tell you about the NV30. Is this the piece of shit you intended to
build?” he shouted.12 “The architects did a shitty job putting the product
together. How could you not see the leaf blower issue before it happened?
Someone should have raised their hand and said, ‘We have a design issue
here.’ ”
His criticism didn’t end after a single meeting. Later, he invited an
executive from Best Buy, which was then the largest electronics retailer in
the United States,13 to speak to Nvidia employees. The executive spent
much of the session talking about the NV30’s poor performance and
customer complaints about the loud fan. Jensen agreed: “He’s right. This is
crap.”
The only thing that saved Nvidia was that its competition did not press its
advantage very hard. ATI had decided to peg the price of its R300-based
graphics cards at $399, the same price as NV30-based cards. If ATI had cut
the price of the R300 aggressively enough, the company could have
destroyed demand for the inferior NV30-based cards and likely bankrupted
Nvidia. Dwight Diercks said ATI had plenty of profit margin to work with,
because its chip had an enormous cost advantage over the poorly designed
and bloated NV30. “If Jensen had been running ATI,” said Diercks, “he
would have put Nvidia out of business.”

JENSEN REFLECTED ON THE NV30’S failures. Ultimately, it was his

responsibility to ensure that Nvidia’s teams collaborated effectively, no
matter how large the company got. He now saw that it may have been
asking too much to integrate the former 3dfx engineers into Nvidia’s culture
all at once. “NV30 was the first chip that we built as one company, after we
had brought in the 3dfx employees,” he concluded many years later.14
“Organizationally, the harmony wasn’t very good.”
The Innovator’s Dilemma taught Jensen, as it has generations of business
leaders, how to protect his company from competition. It helped him
understand the threat from low-cost competitors, which is why he launched
lines of low-tier and mid-tier Nvidia chips made from parts that were not
quite good enough for the top-of-the-line chips. It convinced him to
diversify Nvidia’s partner portfolio beyond just consumer desktop PCs, into
gaming consoles, Mac, and laptops. And it compelled him to made big
strategic investments, such as adding programmability to Nvidia chips to
make true GPUs.
But one of Christensen’s subtler messages escaped Jensen, at least for the
first decade of Nvidia’s existence. It was not enough to look at external
measures of success: revenue, profitability, the price of the stock, or the
pace of product launches. A truly sustainable business spent just as much
effort looking inward in order to keep its internal culture aligned. As Nvidia
established itself as the dominant player in the graphics industry, the
company’s executives got distracted by its partners, its investors, and its
finances. It failed to see the growing problem within its own walls—
complacency. And was almost destroyed because of it.
But Jensen is known for following his own rule to not make the same
mistake twice. He summoned the same vigilance that he had trained on
external threats and turned it toward internal ones. He resolved the
contractual disagreement with Microsoft, ensuring that his architects would
never have to work in the dark again when it came to Direct3D. He made
sure his staff constantly talked to game developers, so the features most
important to them, and to gamers, would always be incorporated into
Nvidia’s chips. He required his teams to ensure the upcoming GPUs would
be optimized for the most popular games, with the goal of dominating the
benchmark tests in reviews. Most important, he pushed his teams to work
with “intellectual honesty”—to always question their own assumptions and
embrace their own missteps, so that the company could solve them before
they snowballed into a disaster like the NV30.
Nvidia barely survived its first ten years. It had achieved many things:
technical wizardry, a blockbuster IPO, and relative longevity in an industry
where most competitors last only a few years. It had been humbled by its
failures: its near-bankruptcy due to the NV1 and NV2; the production
problems that halted the successful RIVA 128 in its tracks; and, most
recently, the NV30 debacle, which signaled a deep organizational problem
that the company would need to confront. It had become a large public
company, with the same challenges and the same tendency to entropy as
any other large public company. Jensen would have to evolve into a
different kind of leader for Nvidia to succeed in the coming decade.
 PART III

NVIDIA RISING

(2002–2013)
 CHAPTER 8

The Era of the

GPU

ONE OF THE EARLIEST REFERENCES to the technology that would

eventually turn Nvidia into a trillion-dollar company was in a PhD thesis
about clouds. Mark Harris, a computer science researcher at the University
of North Carolina at Chapel Hill, wanted to find a way to use computers to
better simulate complex natural phenomena, such as the movement of fluids
or the thermodynamics of atmospheric clouds.
In 2002, Harris observed that an increasing number of computer
scientists were using GPUs, such as Nvidia’s GeForce 3, for nongraphics
applications. Researchers who ran their simulations on computers with
GPUs reported significant speed improvements over computers that relied
on CPU power only. But to run these simulations required computers to
learn how to reframe nongraphics computations in the terms of graphical
functions that a GPU could perform. In other words: the researchers had
hacked GPUs.
To do so, they utilized the GeForce 3’s programmable shader technology,
originally designed to paint colors for pixels, to perform matrix
multiplication. This function combines two matrices (basically, tables of
numbers) to create a new matrix through a series of mathematical
calculations. When the matrices are small, it’s easy enough to perform
matrix multiplication by use of normal computational methods. As matrices
get larger, the computational complexity required to multiply them together
increases cubically—but so does their ability to explain real-world
problems in fields as diverse as physics, chemistry, and engineering.
“Really the modern GPU, we kind of stumbled onto,” said Nvidia
scientist David Kirk.1 “We built a super powerful and super flexible giant
computation engine to do graphics because graphics is hard. Researchers
saw all that floating-point horsepower and the ability to program it a little
bit by hiding computation in some graphics algorithm.”
Using GPUs for nongraphics purposes, however, required a very specific
skill set. Researchers had to rely on programming languages designed
exclusively for graphics shading, including OpenGL and Nvidia’s Cg (C for
graphics), which was introduced in 2002 to run on the GeForce 3.
Sufficiently dedicated programmers such as Harris learned how to
“translate” their real-world problems into functions that these languages
could execute, and they soon figured out how to use GPUs to make
progress in understanding protein folding, determining stock-options
pricing, and assembling diagnostic images from MRI scans.
The academic world initially referred to the use of GPUs for these kinds
of scientific purposes with cumbersome terminology, such as “application
of graphics hardware to nongraphics applications” or “exploiting special-
purpose hardware for alternative purposes.” Harris decided to coin
something simpler: “general-purpose computing on GPUs,” or “GPGPU.”
He created a website to promote the term. One year later, he registered the
URL GPGPU.org; on the site, he wrote about the budding trend and traded
advice with others on the best ways to use GPU programming languages.
GPGPU.org quickly became a popular destination for the community of
researchers who wanted to harness the power of Nvidia’s new devices.
Harris’s intense interest in GPUs earned him a job at Nvidia. After he
received his PhD from UNC, he moved across the country to Silicon Valley
to join the very company whose cards he had learned how to hack. He was
surprised to discover that the term he had invented, “GPGPU,” was widely
used by employees. “People at Nvidia saw potential and were using this
silly acronym that I made up,” Harris said.
Though he didn’t know it, Nvidia had brought him on to help the
company make GPGPU much easier. Jensen was quick to grasp that
GPGPU had the potential to open up the market for GPUs far beyond mere
computer graphics. “Probably the most important influences and early
indication that we ought to continue was in medical imaging,” he said.2 Yet
the fact that all GPGPU work had to be run through Cg—a language that
was proprietary to Nvidia and optimized only for graphical functions—had
become a barrier to wider adoption. In order to generate more demand,
Nvidia would have to make its cards easier to program.
Harris learned there was a chip team within Nvidia working on a secret
project code-named the NV50. Most chip designs were only one or two
generations removed from the current architecture. The NV50 was Nvidia’s
most forward-looking chip under development: it would not be released for
several years. It would have its own dedicated compute mode, so that its
GPU would be easier to access for nongraphics applications. Instead of Cg,
it would utilize extensions to the C programming language, a widely used
general-purpose language. And it would enable parallel compute threads
with access to addressable memory—in essence, allowing the GPU to
perform all the functions of a secondary CPU that might be needed in
scientific, technical, or industrial computing.
Nvidia called this programming model for chips the Compute Unified
Device Architecture, or CUDA. CUDA made it possible for not only
graphics programming specialists but also scientists and engineers to
leverage the GPU’s computing power. It helped them manage the intricate
web of technical instructions that were necessary to execute parallel
computations on the GPU’s hundreds, and later thousands, of computing
cores. Jensen believed it would expand Nvidia’s reach into every corner of
the tech industry. New software, rather than new hardware, would transform
the company.

TWO OF THE MOST IMPORTANT figures in the early development of

CUDA were Ian Buck and John Nickolls. Nickolls was the hardware expert.
He joined Nvidia in 2003, becoming the hardware architect for the
company’s early GPU computing effort. He collaborated closely with the
chip team to ensure that important features were incorporated into GPUs,
such as larger memory caches and different methods for performing
floating-point math. Nickolls understood that better performance was
necessary if Nvidia hoped to drive adoption of GPU computing. (Nickolls
tragically never saw the full success of his work on CUDA. He is
considered by many inside Nvidia as an unsung hero for the company. The
executive passed away in August 2011 after a battle with cancer. “Without
John Nickolls, there’d be no CUDA. He was the most influential of the
technologists in our company that ultimately willed CUDA into existence,”
Jensen said.3 “He worked on CUDA until his death. He was the one that
explained CUDA to me.”)
Buck worked on software. He had previously interned at Nvidia but left
to pursue a PhD at Stanford. During his studies, Buck developed the
BrookGPU programming environment, which provided a language and
compiler for GPU-based computing. His work attracted the attention of the
Department of Defense’s research arm, the Defense Advanced Research
Projects Agency (DARPA)—and also that of his former employer, Nvidia,
which licensed some of the technology Buck had worked on. In 2004, he
was hired by Nvidia.4
The early CUDA team was small and tight-knit. Buck’s software group
consisted of three engineers: Nicholas Wilt and Nolan Goodnight, who
worked on the CUDA driver API and implementation, and Norbert Juffa,
who wrote the CUDA standard math library. Others focused on hardware
compilers, which convert human-readable code into machine-usable code
that can be executed by a computer processor. They included Richard
Johnson, who designed the Parallel Thread Execution Language (PTX)
specification, which served as a virtual hardware compiler target for
CUDA; Mike Murphy, who built an Open64 (x86-64 architecture) compiler
for CUDA to PTX; and Vinod Grover, who joined in late 2007 and worked
on compiler drivers.
It was imperative that both groups work in close harmony. “Any
computer architecture has a software side and a hardware side. CUDA is not
just a piece of software,” said Andy Keane, a former general manager for
Nvidia’s data-center business. “It’s a representation of the machine. It’s a
way you access the machine, so they have to be designed together.”5
The original plan was to launch CUDA solely on Nvidia’s Quadro GPUs,
which were meant for high-end scientific and technical workstations. But
this carried some risk. All new technology presents a chicken-or-egg
problem. Without developers creating applications that took advantage of
the new chips, there would be no reason for users to adopt it. Without a
large installed base of users, developers wouldn’t want to create software
for the new platform. Historically, when a company drives adoption on both
fronts, as Arm Holdings did with its ARM chip architecture for mobile
phones and as Intel did with its x86 processor for personal computers, the
result is usually market dominance for decades. Those companies that fail
to drive adoption, such as PowerPC (with its RISC processors) and Digital
Equipment (with its Alpha architecture), have found themselves relegated to
the dustbin of computer history in the span of a few years.
First impressions matter. If Nvidia initially released CUDA solely for
high-end workstations and didn’t provide sufficient software support, it
might lead developers to pigeonhole it as a tool for a narrow range of
technical professions only. “You can’t just throw technology over the wall
and expect [people] to adopt it,” said marketing executive Lee Hirsch. “You
can’t simply say, ‘Here’s our new GPU, go crazy.’ ”
Instead, Nvidia would have to do two things: make CUDA available to
everyone, and make it applicable to everything. Jensen insisted that they
launch CUDA across Nvidia’s entire lineup, including its GeForce line of
gaming GPUs, so that it would be widely available for a relatively
affordable price. This would ensure that CUDA was synonymous with
GPUs, or at least with Nvidia GPUs. Jensen understood the importance not
just of launching new technology but also of saturating the market with it.
The more people who had CUDA in their hands, the faster the technology
would establish itself as a standard.
“We should just push this everywhere and make it a foundational
technology,” Jensen told the CUDA team.
The move was extremely expensive. Nvidia introduced the NV50, which
had been officially rebranded as the G80 for use in its GeForce line of
graphics cards, alongside CUDA in November 2006. It would be the
company’s first GPU chip with a computing function. The G80 boasted 128
CUDA cores, which are extra hardware circuits used to support CUDA
functionality. The GPU was able to run up to thousands of computing
threads concurrently across those cores by using a hardware multithreading
feature. By comparison, Intel’s main Core 2 CPU at the time only had up to
four computing cores.
Nvidia invested immense amounts of time and money into making the
G80. It took four years to develop the GPU computing chip in comparison
with the one-year gap between generations of GeForce chips. It cost an
astronomical $475 million,6 or around a third of Nvidia’s total research and
development budget for those four years.
That was just for one version of CUDA-enabled GPUs. The company
invested so much in converting its GPUs for CUDA compatibility that its
gross margin, a measure of its profitability, fell from 45.6 percent in the
2008 fiscal year (covering January 2007 to January 2008) to 35.4 percent in
the 2010 fiscal year. As Nvidia increased spending on CUDA, the global
financial crisis destroyed consumer demand for high-end electronics as well
as corporate demand for GPU-powered workstations. The combined
pressures caused Nvidia’s stock price to fall by more than 80 percent
between October 2007 and November 2008.
“CUDA added a ton of cost into our chips,” Jensen acknowledged.7 “We
had very few customers for CUDA but we made every chip CUDA
compatible. You can go back in history and look at our gross margins. It
started out poor and it got worse.”8
Nevertheless, he believed so strongly in CUDA’s market potential that he
remained committed to the course he had chosen, even as his investors
demanded a strategic course correction. “I believed in CUDA,” he said.
“We were convinced that accelerated computing would solve problems that
normal computers couldn’t. We had to make that sacrifice. I had a deep
belief in its potential.”
Yet upon its release, the G80 failed to gain traction, despite rave reviews
from tech publications such as WIRED and Ars Technica.9 A year after its
launch, about fifty financial analysts arrived at Nvidia’s headquarters, now
in Santa Clara, to listen to Jensen and the company’s investor relations
teams present their argument for why Wall Street should continue to believe
in Nvidia when all signs pointed to its being on the wrong track.
All morning, management provided details on its plans to expand high-
performance GPU computing to new markets such as industrial and medical
research applications. The company estimated that the GPU computing
market would rise to more than $6 billion in a few years, even though it was
near-zero at the time. In particular, Nvidia foresaw a demand for enterprise
data centers powered primarily by GPUs and had brought in Andy Keane,
who had extensive experience in hardware business development and
product marketing at various start-ups, to run a new division dedicated to
them. After the morning presentations, it was clear that the analyst group
was skeptical about CUDA and saw it mainly in terms of the negative
impact it was having on Nvidia’s profit margins.
Lunch was held under a tent in the parking lot: a buffet of sandwiches,
bottled water, and soda. An analyst from Hudson Square Research named
Daniel Ernst got some food and sat at an empty table. He was soon joined
by other analysts, and eventually by Jensen as well. The other analysts
began peppering the CEO with short-term financial questions; they wanted
to know the exact impact of CUDA on the company’s profit margins, given
that Nvidia was about to transition to a new manufacturing technology for
its next generation of chips. It was all material Jensen had covered earlier in
the day, and he dutifully reiterated the company’s official guidance, which
called for an eventual rise in long-term margins after the short-term impact
of research and development. This failed to satisfy the analysts, who
remained focused on the next few months, not the next few years.
Ernst felt Jensen was growing frustrated and would soon leave the table,
so he decided to ask him something different. “Jensen, I have a two-year-
old daughter at home. I bought a new Sony A100 DSLR camera and
regularly download photos to my Mac to do some light editing in
Photoshop. But whenever I do this, my Mac slows down as soon as I open
one of these high-resolution images. It’s even worse on my Think-Pad. Can
a GPU solve this problem?”
Jensen’s eyes lit up. “Don’t write about this because it’s not out yet, but
Adobe is a partner of ours. Adobe Photoshop with CUDA can instruct the
CPU to off-load the task to the GPU, and make it much faster,” he said.
“That’s exactly what I’m talking about with the coming ‘Era of the GPU.’ ”
Ernst, at least, was impressed. He saw that CUDA was no fad, but could
be central to Nvidia’s future. He was irritated by the questions from other
analysts about the company’s financial profile. He was happy Nvidia was
willing to sacrifice short-term margin to capture more of CUDA’s gigantic
upside. The “Era of the GPU” would create so many opportunities that
Jensen saw it as his mission to prepare Nvidia to take advantage of it—even
if no one could know exactly what those opportunities would be.
Everything else, including corporate financial concerns, was entirely
secondary.
It would not be easy to match Jensen’s vision with market reality. Nvidia
had the product and production problems solved, but now Jensen asked his
team to figure out ways to create a market for CUDA—to “solve the entire
problem,” as he put it. This would require a systematic analysis of the needs
of every industry, from entertainment to health care to energy, and to not
only analyze potential demand but also figure out how to unleash it via
special, GPU-centered applications in each field. If developers didn’t yet
know what to do with CUDA, Nvidia would teach them.

FOR SEVERAL YEARS, NVIDIA’S CHIEF scientist, David Kirk, had

been receiving requests from top universities around the country, asking for
support from the chip maker.
Nvidia saw a prime opportunity to both help universities and drive
adoption of its GPUs. After making several ad hoc donations, Kirk
formalized a program with Caltech, the University of Utah, Stanford, the
University of North Carolina at Chapel Hill, Brown, and Cornell. Nvidia
would provide graphics cards and financial donations to the schools, and in
exchange the schools would use Nvidia hardware in graphics programming
classes. “It was not entirely selfless,” Kirk said. “We wanted them to use
our hardware instead of AMD’s hardware for their teaching.”10
The program solved an ongoing issue Nvidia had with its university
donation program. Whenever Nvidia made cash donations, universities
would charge an overhead or administration fee, thus reducing the
donation’s impact on actual research. By switching to a more hardware-
based donation model, Nvidia was able to ensure that students and not
administrators would reap the most benefits from the company’s assistance.
Earlier on, Nvidia had established an internship program, where some of
the most talented students from the partner schools and others would gain
work experience at the corporate office and potentially be evaluated for
future employment. This was how CUDA engineer Ian Buck got his first
exposure to Nvidia, for example.
Kirk hoped to leverage these same relationships to promote CUDA after
its release. He and his colleague David Luebke started a new program they
called the CUDA Center of Excellence, offering schools CUDA-capable
machines if they committed to teaching a class on the subject. He visited
universities, telling students, professors, and department heads they needed
to change how they taught computer science because parallel computing
was going to become far more important. He gave more than one hundred
talks over the course of a year, traveling around the world, sometimes doing
multiple talks per day. There were no takers.
“Nobody knew how to program in CUDA, and nobody was devoting
effort to it,” Kirk said. “Nobody wanted to hear it. I was literally hitting a
wall.”
Eventually, he found himself pitching Richard Blahut, the head of the
Electrical and Computer Engineering Department at the University of
Illinois at Urbana-Champaign. Blahut told Kirk it was a really good idea but
said that if Kirk was serious about it, he should teach the class himself.
Kirk’s initial response was no. At the time, he was living in the
mountains in Colorado and didn’t have any interest in teaching, let alone in
Illinois. But Blahut pressed him, adding the school would match him with
one of its top professors, Wen-mei Hwu, who routinely won teaching
awards. “You two can teach the class together, and it will be guaranteed to
be a success as Hwu can show you how to teach the material,” he said. Kirk
agreed.
In 2007, Kirk flew from Colorado to Illinois every other week to give his
lectures. At the end of the semester, the students carried out CUDA research
programming projects and published their work. Other researchers around
the country began to request lectures and teaching materials from Kirk and
Hwu, so they recorded their classes and made their videos and notes freely
available online.
The following year, Nvidia named the University of Illinois at Urbana-
Champaign the first CUDA Center of Excellence and provided the school
with more than $1 million as well as thirty-two Quadro Plex Model IV
systems, each with sixty-four GPUs—the most advanced machines Nvidia
made.
“David Kirk and Wen-mei Hwu were the evangelists,” said Bill Dally,
Kirk’s eventual successor as Nvidia’s chief scientist. They “taught teacher’s
courses around the country to basically spread the religion of GPU
computing, and it really took off.”
Other schools heard about Kirk’s class and began to explore how they
might also start to teach parallel computing themselves. But because Kirk’s
was the first course of its kind, there was no common syllabus or set of
standards, no textbook to use. So Kirk and Hwu wrote one. Their first
edition of Programming Massively Parallel Processors, which was
published in 2010, sold tens of thousands of copies, was translated into
several languages, and was eventually used by hundreds of schools. It was a
major inflection point in attracting attention, and talent, to CUDA.
Having built an academic training pipeline for CUDA, Nvidia now
moved to drive adoption among nonacademic researchers. In 2010, outside
of academic computer science and electrical engineering departments,
almost no one used GPUs for scientific research. But gaming provided a
glimpse of what was possible. PC games—in particular, first-person
shooters—could increasingly produce realistic simulations of physics.
When they used GPU processing in its traditional, graphics-acceleration
role, these games could calculate the path of a bullet, from the moment it
was fired from a gun to the effect of wind on its trajectory to the spalling it
produced when it hit a concrete wall. All of these applications relied on
various permutations of matrix multiplication—the same math used to solve
complex scientific problems.
Nvidia’s director of business development for the life-sciences industry,
Mark Berger, was responsible for expanding the use of GPUs in chemistry,
biology, and materials science. He followed much the same playbook that
Oliver Baltuch had used when trying to raise Nvidia’s profile among its
prospective partners in the tech industry, as we saw in chapter 6.
First, he gave away GPUs to researchers and informed them about
Nvidia’s substantial investments in creating basic software libraries and
tools for CUDA. Although the company might not have been familiar with
the esoteric computational problems scientific users might perform, it
recognized that those users would rather spend their time designing
experiments than building the foundational math libraries they all needed.
As a result, the developer tools Berger provided alongside the cards
themselves made adoption of CUDA much faster—and helped him
establish strong relationships with scientists.
“I got to be Santa Claus and send a ton of GPU boards out to all my
developers, and everybody loves Santa Claus,” he said.11
Second, he started conducting annual two-day technology summits where
Nvidia employees could interact with and learn from scientists themselves.
Dozens of researchers from the life-sciences industry—chemical engineers,
biologists, pharmacologists, as well as the software developers who
supported their work—arrived in Santa Clara from across the United States
but also from Europe, Japan, and Mexico. On the first day, Nvidia’s
engineers would tell them about future improvements to CUDA, including
software and hardware advances. The scientists and developers would then
give their feedback.
“Our engineers aren’t clairvoyant,” Berger said. “They don’t know where
the hockey puck is going to be. At one point, I had well over a dozen
features that were in CUDA or the hardware because of the input my
developers gave.”
The scientists and researchers appreciated Nvidia’s transparency and
willingness to listen, in turn. “They saw us as resources,” said Ross Walker,
a biochemistry professor at the University of California, San Diego. “We
could go tell them, ‘we need this feature,’ and they would change the design
of the chip or add it to CUDA. There was no way on Earth Intel would have
ever done anything like that.”
Jensen himself loved attending the summits and sitting down with real-
world users of CUDA to get their insights. During one of the first annual
meetings, he gave a keynote speech in which he recalled his early days in
the industry. When he started in chip design, he had to design the silicon,
get it back from the factory, and put it under a microscope to see where the
defects were. “He had a real affinity to my guys . . . who were simulating
what was going to happen on a molecular level,” Berger said.
Jensen then pivoted to explain how simulations had changed the chip
industry. He was among the first generation of engineers who were able to
do a large amount of virtual debugging of chips before they went out to be
manufactured. This was, he argued, the same revolution that CUDA
promised to bring to the sciences. Instead of the expensive and manual
process of designing and testing new drugs by hand in the lab, they could
do it virtually with software. CUDA-powered GPUs could make their
research cheaper, faster, and far less prone to human error.
It was new territory for Nvidia. Since his very first meeting with Curtis
Priem and Chris Malachowsky at the Denny’s in East San Jose, Jensen had
always focused on the importance of clearly defining market opportunities
and developing new business strategies. Even in 1993, he had to convince
himself there was a $50 million annual revenue opportunity in PC graphics
if he was going to leave behind steady employment and cofound Nvidia. To
survive after the failure of the NV1 and NV2, he had to recalibrate Nvidia’s
strategy to go after the very top of the market. Here again, the opportunity
was clear: although PC graphics was a crowded field, almost no one made
truly excellent chips; that would be Nvidia’s niche. To avoid the endless
cycle of corporate obsolescence, where the top-selling company one year
was often leapfrogged in the next year, he had to push his teams to ship
three chips per design cycle, instead of just one. And to diversify the
company’s lines of revenue, so that weak demand in one area would not
doom the entire business, he aggressively pushed into new market
segments: into console graphics, even when Microsoft originally signed
another graphics-chip partner for its Xbox; into Apple’s Macintosh series,
despite the fact that Nvidia had little experience with the Mac architecture;
and even into the professional workstations that it had originally shunned,
with its Quadro line that was optimized for computer-aided design.
Now, however, Jensen had overseen the invention of an entirely new
computing technology in the GPU and had to build a market for it from
scratch. He realized that the opportunity could be astronomically huge—
that it could unlock so much potential not in gaming but in business,
science, and medicine. To realize that potential and make his market, he
would have to develop an entirely new skill set—and teach the company,
his investors, and himself the value of patience and persistence in an
industry that always expected the next great thing on a very short timeline.

PROFESSOR ROSS WALKER CREATED one of the new-use cases

for GPUs in the form of a biotechnology program called Assisted Model
Building with Energy Refinement, or AMBER. The program simulates
proteins in biological systems and has become one of the most popular
applications used by academics and pharmaceutical companies to research
new drugs. It was originally designed for high-powered computers, and
therefore its reach was limited to only the best-funded research groups in
the world. But Walker saw that it could run on little more than a few
consumer-grade GPUs working in concert. This has made it one of the
most-used tools in the biosciences. The software has more than one
thousand university and commercial licenses and is credited in more than
fifteen hundred academic publications per year. And it owes its success to
its compatibility with Nvidia’s CUDA architecture.
Walker received an undergraduate degree in chemistry and a PhD in
computational chemistry at Imperial College London. He then worked at
the Scripps Research Institute in San Diego as a postdoctoral fellow and as
a research scientist investigating computational simulation software focused
on enzyme reactions. At a bar one night, he met some employees of the San
Diego Supercomputer Center and introduced himself.
“We know you,” they said. “Your name is written on our whiteboard as
the person who uses all our computing power.”
Walker was offered the lead position in biosciences at the center, which
was based at the University of California, San Diego. He accepted. Yet even
though he continued his work on AMBER and was appointed a professor,
he grew increasingly disillusioned with the academic process—especially
when it came to the allocation of precious computing resources.
He sat on a committee that reviewed proposals and awarded time on the
center’s computers to the winning research teams. At any given session,
there were typically fifty proposals to read, and most of the members of the
committee would spend only a few minutes discussing each one. It was
demoralizing. “I know people spend three months of their lives, their blood,
sweat, and tears writing these, and we’re spending five minutes deciding
their fate,” he said. Most proposals were rejected: funding rates hovered in
the low single digits.
Even worse, computing capacity on the supercomputers tended to go
toward those people and groups who were already successful. Famous
scientists such as Klaus Schulten, who developed computer models that
could simulate protein and virus structures down to the atomic level, and
Greg Voth, who developed “multiscale theory” algorithms that could
simulate the behavior of complex biomolecular systems, were given
precedence, as Walker saw things.
“But the reason they could do the famous papers was because they were
getting the time on supercomputers,” Walker said. “Other people who had
great ideas never got the time and couldn’t make their impact. It wasn’t how
good your science was, but it was whether you could get access to
computing time.” It was a catch-22: the only way to get supercomputing
resources was to already have gotten supercomputing resources.
Walker remembers once rejecting an application from Schulten to get
emergency priority time on the supercomputer to work on molecular
dynamics simulations of the H1N1 virus, popularly known as “swine flu,”
during the 2009 outbreak. He knew any exploratory research would take
years to lead to a drug and thus wouldn’t change the outcome of the
pandemic. But Walker’s decision was overruled, and he believed it was
because Schulten was able to pull some political strings.
To Walker, it was just another example of how limited resources, along
with the politics and bureaucracy of the academic world, had created a
bottleneck that restricted progress for the entire field. He was dispirited; he
wanted computing power to be available on merit, but there was no way to
change the prevailing dynamic, with everything running through a small
number of extremely powerful, but extremely expensive, supercomputers.
He saw the need for a new kind of technology that could make computing
power more accessible. “That was my driving force,” Walker said.
He initially looked into commissioning custom-designed application-
specific integrated circuits, also known as ASICs, that were optimized
specifically for AMBER. But though less expensive than supercomputers,
they still cost tens of thousands of dollars apiece—and researchers would
have to spend even more to build special computers around them. Even if
he could find a designer and fabricator, most researchers could not afford to
buy the chips. Those who could probably had easy access to a
supercomputer anyway.
Walker next examined gaming consoles and decided his best candidate
was the Sony PlayStation series. But he ran into a wall here, too. Although
PlayStations were cheap enough, Sony made it difficult to hack into the
console’s firmware and software. There was no way for Walker to use them
for nongaming purposes.
Still, thinking about the PlayStation gave him an idea. Even though he
never could successfully hack the console, his investigation into its graphics
capabilities convinced him that retail-grade graphics chips were powerful
enough to run AMBER. All he needed was an open platform that he could
actually program. Then he realized the workstations inside his lab—the
ones that his colleagues used to create 3-D visualizations—all had high-end
GPUs comparable to the ones in the PlayStation. Although the workstations
cost tens of thousands of dollars each, they were a step closer to the
consumer-grade hardware that he wanted AMBER to run on. Maybe they
could work as a proof of concept.
He first experimented with using the Brook programming language
created by Ian Buck. He ran his first tests on graphics cards made for the
Radeon series by AMD, Nvidia’s primary competitor. But these cards had
immature software and were not easily programmable. He then started
talking to Nvidia about using their CUDA architecture to run his molecular
dynamics models.
It was a perfect match. Walker found CUDA a much easier programming
environment to work with, while Nvidia saw an opportunity to extend its
reach into the scientific computing world. The company gave Walker
technical resources to help redesign AMBER so that it could not only work
on CUDA but also use its computing capabilities to the fullest. “We made
the decision from day one to move everything onto the GPU, so the CPU
would become irrelevant,” Walker said.
In 2009, Walker released the first GPU-enabled version of AMBER. It
ran up to fifty times faster than the previous version.
Walker had broken the grip of the academic bureaucrats and realized his
dream of democratizing computing power. CUDA made it possible for
scientists to carry out important experiments on affordable hardware instead
of relying on the expensive and scarce supercomputing resources of a few
elite universities. For the first time, the tens of thousands of postdoc
graduates who used AMBER could do substantial science-computation
experiments on their own hardware, at their own pace, and without having
to compete with the luminaries of their fields—a competition that they
would inevitably lose. Students could outfit a PC with a few Nvidia gaming
GeForce cards and have an enormously powerful machine at a reasonable
price. “You could buy a $100 CPU and four $500 GeForce cards and have a
workstation that was as powerful as a full rack of servers. It was a game
changer.”
In its 2010 annual report, Nvidia mentioned the success of AMBER at the
very top of its discussion of “high-performance computing” products. It
appeared above other major announcements, including of its partnerships
with Hewlett-Packard, of the launch of a new GPU-powered “seismic
software suite” for oil prospecting, and of the use of GPUs by “the
investment banking division of a leading European financial institution.” To
further cement the relationship between Walker and Nvidia, the company
appointed him in November 2010 to the CUDA Fellows program, which
recognized research and academic leaders for their “exceptional work” in
using CUDA within their disciplines and raising awareness of the platform.
Just as Jensen had predicted, GPUs were making advanced computing far
more accessible and cheaper, which in turn made a program such as
AMBER far more accessible. And the widespread adoption of AMBER
transformed how the entire field of molecular dynamics conducted research.

THERE WAS ONE ISSUE where Walker and Nvidia diverged, however.
Walker was used to dealing with academic institutions, whose main priority
was advancing scientific knowledge. Nvidia was a business, with revenue
targets to hit and investors to please. And the company’s executives had not
expected Walker to get AMBER to work so cheaply. Nvidia’s high-
performance computing division started to recommend that scientists use
the company’s higher-end Tesla general-purpose graphics cards, which
retailed for about $2,000—four times more than the GeForce cards that
Walker tended to use. The company claimed that it based this
recommendation on the GeForce line’s lack of error-correction features,
which could leave AMBER outputs vulnerable to the accumulation of small
but harmful mathematical errors. The self-detection and self-correction
features on the Tesla line were not available on the less expensive GeForce
cards.
Walker disagreed. He ran a series of tests that proved the GeForce line’s
lack of error correction did not result in any issues with AMBER’s output.
He then set out to prove pretty much the opposite: that the error-correction
features on the Tesla line were superfluous, at least for AMBER. He
arranged for some of his contacts who worked at the Los Alamos National
Lab—one of the Department of Energy’s foremost research facilities and
where the atomic bomb was developed—to run the same tests with Tesla
cards, just to see how many errors the cards actually needed to correct.
There was no difference between the performance of the cheaper GeForce
cards and the more expensive Tesla cards. Clearly, as he saw it, Nvidia was
treating AMBER as an opportunity to up-sell cards as much as an
opportunity to advance molecular-simulation technology.
“Nvidia’s argument was you can’t trust the results. I have the data to
prove you can,” Walker said. “We ran these simulations for two weeks and
did not see a single ECC error. This is like the worst environment, top of the
mountain, next to a nuclear lab. Radiation is as high as you can get within
the U.S. Still, no errors.”
The conflict between Walker and Nvidia escalated. First, Nvidia changed
the level of math precision in its gaming cards, which had an imperceptible
effect on PC games—but a potentially catastrophic one for research tools
that relied on the cards to do advanced calculations. In response, Walker
and AMBER developers figured out a way to work around the precision
change, so that they could continue to run their simulations on GeForce
without any accuracy issues. Then Nvidia began to enforce purchase
controls for GeForce cards on its suppliers, making it difficult for people
like Walker to buy large quantities of them in a single order. Walker
criticized this move on the global mailing list for AMBER users, calling it
“a very worrying trend that could hurt us all and have serious repercussions
on all of our scientific productivities and the field in general.”
He became increasingly frustrated with Nvidia’s attempt to wring more
money out of him, when he had done so much to make CUDA more than
just a niche product for well-resourced developers and academics. The
architecture wouldn’t have been as successful if Nvidia had restricted its
use to cards that cost thousands of dollars; it would have been almost as
expensive to use CUDA as it would have been to design a custom ASIC.
“A key to Nvidia’s success was allowing CUDA to run on GeForce cards,
making it so that poor scientists could do work equivalent to those with
multimillion-dollar computers,” he told me, years later. “Once it reached a
critical mass, they slowly tightened the thumbscrews on GeForce and made
it harder to do.”
Walker later joined GlaxoSmithKline, the pharmaceutical and
biotechnology company, as head of scientific computing. The first thing he
did was build a data center cluster by using thousands of GeForce gaming
cards that cost only about $800 apiece.
This caught the attention of Nvidia’s vice president of health care,
Kimberly Powell, who called Walker and said, “You’re at GSK now. You
need to be buying our enterprise products.”
“No,” countered Walker. “I should be doing what’s best for my employer.
That’s my job.”

JENSEN MADE NO APOLOGIES for Nvidia’s aggressive approach to

chip sales. In fact, he insisted that salespeople take the same stance with all
clients, regardless of size.
Derik Moore was known as one of the best salespeople in the industry
when Nvidia poached him from ATI. He remembered getting a call from an
Nvidia executive who told him, “You have been kicking my ass for over a
year, so we’re actually wondering if you would like to come to work for
Nvidia?”12
Moore managed enterprise sales to large computer companies such as
Hewlett-Packard, which bought large quantities of GPUs for its PCs and
laptop lines. Nvidia wanted him to bring his book of business with him—
and it was willing to pay generously for it. At ATI, he made about $125,000
per year, which in 2004 was well above the mean salary for a sales rep.
During the recruitment process, Nvidia offered him almost double that.
He soon learned why. While still working at ATI, he once drove past
Nvidia headquarters around 7:00 p.m. and saw that the office was almost
entirely full. His manager, who was riding with him, remarked, “Oh, they
must be having an evening meeting.”
Now that he was on the inside, he realized that “evening meetings” were
the norm rather than the exception. He started to regularly work on
weekends, something he never did at ATI. He recalls being forced to take
part in a conference call on Christmas Eve to discuss a sales shortfall and
what the company could do to recover the business. No personal time or
day was truly his own. Still, he saw that the commitment expected of him
was expected of everyone else, too, all the way up to Jensen, which made
the sacrifice easier. “There was a sense of dedication and hard work that
resonated through the organization,” he said. “The work ethic was
contagious.”
Hard work was not always enough to shield him from Jensen’s critiques.
Within a few years of arriving at Nvidia, Moore’s work with HP’s server
division had increased annual sales to HP from $16 million to $250 million.
One day, two senior executives from HP’s server group came to Nvidia’s
headquarters. Because they were high ranking, Jensen asked whether he
could join the meeting. Moore was delighted to have him there.
The server business was higher stakes than general enterprise sales, as
the cards that Nvidia sold for such purposes were often used for mission-
critical enterprise applications and thus needed to be more reliable. The
customers were also more litigious. The HP executives asked whether
Nvidia would give HP unlimited indemnification against lawsuits if
something went wrong—basically asking Nvidia to assume all the legal risk
if faulty GPUs caused a failure of HP’s servers. This was a surprise to
Moore, who didn’t know the HP executives would start a legal negotiation.
He was glad to have Jensen present at the meeting to reply to the
unexpected question.
Jensen pointed out the problem with unlimited liability. The graphics
chip is a small part of the server, so Nvidia could not indemnify the full
value of the server. Doing so would entail a massive and unreasonable
financial risk. Instead, he proposed that Nvidia tie indemnity to something
more concrete: the amount of annual business the server group did with his
company. If HP spent $10 million per year on cards, Nvidia would
indemnify them for up to $10 million in the event of component failures.
The protection would grow as the business grew. The HP executives
accepted the deal on the spot, and Moore walked out of the meeting pleased
with the outcome.
Afterward, he went to Jensen. “Thank you for coming to the meeting. I
really appreciated it,” Moore said.
Jensen, however, saw the meeting differently. “It ended well, but Derik,
let me tell you your failure here.”
The remark shocked Moore. “It scared the crap out of me,” he recalled.
“The failure here was you did not tell us what the company was going to
ask in advance,” Jensen said. “Nobody likes surprises. Don’t ever let that
happen again.”
Jensen referred to his sellers as Nvidia’s Green Berets. He needed them
to be self-sufficient and aggressive. Moore had failed to match Jensen’s
expectation for the role—that each seller become the “CEO of your
accounts.” When they met with their customers, they needed to know more
about those customers’ businesses than the customers themselves do. They
had to anticipate how much customers were willing to pay for Nvidia’s
superior products. Jensen, for his part, would provide them with whatever
resources were necessary: the “reinforcements” behind the elite vanguard.
One such group of “reinforcements” was Nvidia’s developer-technology
engineers, who acted as consultants and implementation experts for Nvidia
products. They would sometimes visit customers to fix problems that came
up or figure out how to get a particular program to work better on Nvidia
GPUs. These engineers would ensure that as many partners as possible
knew how to utilize Nvidia’s cards for maximum effect.
 All this came at a premium for customers. The company never
discounted its chips, not even to match its competitors’ pricing, unless it got
something in return—a sticker on a partner’s computer, a logo on a boot-up
splash screen.
“We don’t sell on cost. We don’t believe our products are commodities,”
Moore’s sales leader told him. “We believe that we bring exceptional value
to the customer, and we extract value for our brand.”

JENSEN DOESN’T LIKE DESCRIBING the strategy around CUDA as

the building of a “moat.” He prefers to focus on Nvidia’s customers; he
talks about how the company has worked to create a strong, self-reinforcing
“network” that helps CUDA users. Indeed, CUDA is an incredible success
story. Today, there are more than 5 million CUDA developers, 600 AI
models, 300 software libraries, and more than 3,700 CUDA GPU-
accelerated applications. There are about 500 million CUDA-capable
Nvidia GPUs in the market. The platform is also backwards compatible,
meaning developers can be confident that any investment in writing
software will be usable on future chips. “All the invention of technologies
that you build on top of Nvidia accrue,” Jensen said. “If you were early
there and you were mindful about helping the ecosystem succeed with you,
you end up having this network of networks and all these developers and all
these customers who built around you.”13
Nvidia invested heavily in deep learning from the outset, dedicating
substantial resources to creating CUDA-enabled frameworks and tools. This
proactive approach paid off when artificial intelligence exploded in the
early 2020s, because Nvidia was already the preferred choice of AI
developers everywhere. Developers want to build AI applications as quickly
as possible with minimal technical risk, and Nvidia’s platform is far more
likely to have fewer technical problems, because the user community has
already, over more than a decade, fixed bugs or figured out optimizations.
Other AI chip vendors never really had a chance.
“If you have AI applications built on top of CUDA and Nvidia GPUs, it’s
a huge undertaking to move over to Cerebras, AMD, or whatever,” said Leo
Tam, head of engineering at Amicus.ai and a former research scientist at
Nvidia. “There’s more than just putting your programs onto different chips.
It’s not simple. As a user, I can tell you they never work perfectly. It’s not
worth it. I’m working on 99 problems for my start-up already. I don’t need
another problem.”
Nvidia had seen the opportunity early on and had seized it. Amir Salek,
former Nvidia director of hardware engineering, noted that Nvidia was very
quick to integrate important AI software libraries into CUDA, so that
developers could easily use the latest innovations in the field without
wasting time building or integrating their own software tools.
“If you wanted to write a new AI model or algorithm, CUDA gave you
access to library components that are highly optimized and ready to use as
opposed to you going all the way down to the nitty gritty details of like
moving bits from here to here,” Salek said.14
For these reasons, and others, it is hard to describe Nvidia’s action as
anything other than the construction of a competitive moat. Nvidia made a
general-purpose GPU that represented the first major leap forward in
computational acceleration since the invention of the CPU. The GPU’s
programmable layer, CUDA, was not only easy to use but also opened up a
wide range of functions across scientific, technical, and industrial sectors.
As more people learned CUDA, the demand for GPUs increased. By the
early 2010s, the market for general-purpose GPUs that had once looked
moribund appeared to be on the ascent.
Jensen’s strategic brilliance ensured that competitors would have
difficulty breaking into a market that Nvidia had created and which was
effectively based on its proprietary hardware and software.
Nvidia’s current position—among chip designers and in the national and
global economies—seems unassailable. As Amir Salek put it: “The moat is
CUDA.”
 CHAPTER 9

Tortured into
Greatness

THE COMPANY THAT CREATED CUDA and opened the way to the
era of general-purpose computing on GPUs had much in common with the
company founded in a Denny’s booth in 1993. It still prized technical skill
and maximum effort above all else. It still made strategic decisions for the
long term rather than try to juice its stock price over the short term. It still
operated with the necessary paranoia of a leading business in a volatile
industry, always trying to correct course before it started down the slope
toward irrelevance and obsolescence. And its CEO still managed the
company directly, involving himself deeply in product decisions, sales
negotiations, investor relations, and more.
What had changed, however, was Jensen’s relationship with his
employees. Nvidia in 2010 was no longer the start-up with a few dozen
people on the payroll, where he could get as much face time as he wanted
with each one of them, regardless of their level or their job function. Now, it
had 5,700 employees, and although many of them worked at the
headquarters in Santa Clara, it had satellite offices across North America,
Europe, and Asia.1 Jensen had learned that corporate culture tended to
atrophy as more people from more locations joined the company, and an
atrophying culture could hurt product quality—as the company had learned
with the “leaf blower” fan on the NV30-based GeForce FX 5800 Ultra. He
had always tried to give direct feedback to his employees as much as
possible when the company was small, in order to consistently reinforce his
principles and ensure everyone had a clear idea what was expected of them.
But in the new, larger Nvidia, he found that it was difficult to reach all the
employees on a consistent basis.
Jensen decided to offer Nvidia employees more direct criticism in larger
meetings, so that more people could learn from a single mistake.
“I do it right there. I give you feedback in front of everybody,” he said.
“Feedback is learning. For what reason are you the only person who should
learn this? You created the conditions because of some mistake you made or
silliness that you brought upon yourself. We should all learn from that
opportunity.”
Jensen displayed his trademark directness and impatience in all settings.
He would often chew people out for fifteen minutes straight, regardless of
the venue. “He does it all the time. It’s not even at company[-wide]
meetings. It’s during smaller meetings or alignment meetings,” a former
Nvidia executive said. “He can’t let it go. He just has to make it punitive a
little bit.”
One well-known example occurred when Nvidia was making one of its
first plays in the mobile phone and tablet market, with the Tegra 3 chip. At a
company all-hands meeting in 2011, Jensen asked the cameraman to
repeatedly zoom in on the project manager of Tegra 3, a man named Mike
Rayfield, as Jensen gave him feedback. While everyone in the audience got
a good look at Rayfield’s face, Jensen launched into him.
“Mike,” he said, “you need to get Tegra done. You got to tape Tegra out.
Guys, this is an example of how not to run a business.”
“It was the most embarrassing, humiliating thing I’ve ever seen,” another
former Nvidia employee said. When asked about the incident, Rayfield later
said in an e-mail, “That was not the only ass kicking I got [from Jensen],”
with a smiley face at the end of his comment. Less than a year after the
Tegra chip came out—nearly eight months behind schedule—he left Nvidia.
He wasn’t pushed out; he resigned voluntarily.
Jensen’s at-times harsh approach was a deliberate choice. He knew that
people would inevitably fail, especially in a high-pressure industry. He
wanted to offer employees more opportunities to prove themselves,
believing that they, in every case, are often just one or two epiphanies away
from solving their problems themselves.
“I don’t like giving up on people,” he said. “I’d rather torture them into
greatness.”
The method is not intended as a means to show off how much smarter he
is than his employees. Instead, he sees it as a guard against complacency.
Jensen’s time, and the time of his employees, is best spent trying to solve
the next problem. Praise is a distraction. And the deadliest sin of all is
looking back at your past accomplishments as if they will protect you from
future threats.
Former sales and marketing executive Dan Vivoli remembers receiving a
phone call from Jensen when he was driving to the office a day after Nvidia
put on a marketing event for the GeForce 256. Vivoli was proud of the work
his team had done.
“How’d the launch go?” Jensen said. Vivoli went on for five minutes
about every part of the event he believed was a success. “Uh-huh, uh-huh,
uh-huh,” Jensen said. Vivoli stopped talking and Jensen asked, “What could
you have done better?”
“That’s all he said. There was no ‘attaboy.’ There was no ‘great job.’
There was none of that. It doesn’t matter how well you think you did,”
Vivoli said. “It’s okay to be proud, but the most important thing is trying to
improve.”2
Jensen does not appear to be any less reproachful toward himself. A sales
executive named Anthony Medeiros recalled one meeting where Jensen
revealed a habit, if not an active practice, of self-criticism.
“I’ll never forget this. We had done fantastic. We just blew the doors off
the quarter. Then, during our quarterly review meeting, Jensen stood up in
front of us.”3
The first words out of Jensen’s mouth were, “I look in the mirror every
morning and say, ‘you suck.’ ”
Medeiros was struck by how someone so manifestly successful could still
think in such terms. But it was, for better or for worse, the same approach
that Jensen wanted everyone else at Nvidia to take toward themselves and
their own work. Do your job. Don’t be too proud of the past. Focus on the
future.
JENSEN’S PREFERENCE FOR THE DIRECT approach also shaped
Nvidia’s corporate structure as the company grew. Early on, Nvidia nearly
drove itself out of business because of a lack of internal alignment. The
strategy for a chip did not match what the market wanted, as with the NV1.
Or an excellent chip was hamstrung by poor execution on the
manufacturing side, as with the RIVA 128. Or a dispute with a key partner
created a cascade of technical problems that eventually doomed an entire
chip line—the story of the NV30. In all three cases, Jensen placed the
failure not on external factors but squarely on Nvidia and its inability to get
out of its own way. “When we were a small company,” he said, “we were
plenty bureaucratic and plenty political.”4
Over time, Jensen thought about how he would create an ideal
organization from scratch. He realized he would choose a much flatter
structure, so that employees could act with more independence. He also saw
that a flat structure would weed out lower performers who were
unaccustomed to thinking for themselves and to acting without being told
what to do. “I wanted to create a company that naturally attracts amazing
people,” he said.5
Jensen believed that the traditional corporate pyramid, with an executive
suite at the top, multiple layers of middle management in the middle, and a
foundation of line workers at the bottom, was antithetical to fostering
excellence. Instead of a pyramid, he would remake Nvidia into something
that looked more like a computer stack, or a short cylinder.
“The first layer is the senior people. You would think that they need the
least amount of management,” Jensen said. “They know what they’re doing.
They’re experts in their field.” He did not want to spend time on career
coaching—because the majority of them had already reached the pinnacle
of their careers. As a result, he rarely held one-on-one meetings with his
direct reports, at least when it came to such open-ended topics. Instead, he
focused on providing them collectively with information from across the
organization, as well as with his own strategic guidance. This would ensure
that every part of the business was aligned and allow Jensen to manage
more executives in a manner that actually added value.
Nvidia’s current structure stands in contrast to that of most American
companies, whose CEOs have only a handful of direct reports. In the 2010s,
Jensen had forty executives on his leadership team, or the “e-staff,” each
reporting to him. Today the number is more than sixty.6 He has steadfastly
refused to change his management philosophy, even when, for example,
new board members joined Nvidia and recommended that he hire a chief
operating officer to reduce his administrative burden.
“No, thanks,” he would always reply. “This is a great way to make sure
everybody knows what’s going on,” he would add, referring to his direct
communication with much of the rest of the company.7
The large number of executives in e-staff meetings has fostered a culture
of transparency and knowledge sharing. Because there aren’t many levels
between the e-staff and the most junior person at the company, everyone in
the organization can provide assistance on problems and prepare in advance
for potential issues.
Oliver Baltuch, a former marketing executive, was impressed by how
responsive his Nvidia colleagues were compared with those at his prior
jobs. “The biggest difference was you only had to ask somebody once to get
something done. It just got done,” he said. “You never had to ask a second
time.”8
Andy Keane, a former general manager of Nvidia’s data-center business,
remembered Jensen explaining the traditional structure of the company’s
main competitors on a whiteboard, a structure he dubbed “the upside-down
V’s.” This was how most companies were built. “You become a manager
and you build your upside-down V. You defend it. Then you become a vice
president and you get more upside-down V’s of people under you,” Jensen
said.
Keane said that at other companies, talking to executives one or two
levels above your direct manager was frowned upon. “Nobody likes it. It’s
just insane, right?” he said. “Nvidia was never like that.” Keane himself
talked to his direct manager once or twice a month but talked with Jensen
two to three times a week. “Jensen created a company that he could manage
directly,” he said. “There is a vast culture difference between Nvidia and
other companies.”9
Keane was also surprised by the sheer openness he found at Nvidia. He
joined at the general-manager level and was allowed to attend every board
meeting and off-site board event. When a typical CEO would have eight or
nine people in a room for big executive meetings, Jensen would have a
packed house. “Everyone could hear what he was telling the executive
staff,” Keane said. “It kept everybody in sync.”
When there is important information to share or an impending change in
the direction of the business, Jensen says he tells everybody at Nvidia at the
same time and asks for feedback. “It turns out that by having a lot of direct
reports, not having one-on-ones, [we] made the company flat, information
travels quickly, employees are empowered,” Jensen said. “That algorithm
was well conceived.”

MANY LARGE CORPORATIONS ARE DIVIDED into business units

managed by competing executives. These units are locked into long-range
strategic plans and must fight among themselves for resources. As a result,
most organizations tend to move slowly. There is indecision. Big projects
get stuck awaiting approvals from multiple stakeholders and hierarchies
across the company. Any decision-maker can unilaterally slow things down
by playing internal politics. When things go poorly, companies must shed
workers to meet budgetary targets, even if those workers are top
performers. All of this contributes to short-term thinking and the hoarding
of information at the corporate level. Instead of shaping a company into a
single, cohesive team, the usual corporate structure creates the kind of toxic
environment that drives good people away.
As Jensen put it, “you want a company that’s as large as necessary to do
the job well, but to be as small as possible,” and not bogged down by
overmanagement and processes.
To get there, he decided that rather than rely on a permanent class of
professional managers, whose only job was to be in charge of things, he
would create a far more fluid system that would orient Nvidia around its
business goals. And even as he took the long view, he would get rid of the
practice of long-term strategic planning, which would force the company to
stick to a particular path even if there were reasons to deviate from it.
“Strategy is not words. Strategy is action,” he said. “We don’t do a
periodic planning system. The reason for that is because the world is a
living, breathing thing. We just plan continuously. There’s no five-year
plan.”
He began to tell his employees that their ultimate boss was the mission
itself. The idea was to make decisions for the good of the customer, not to
boost the career of the executive above them. “The concept of the mission
is the boss makes a lot of sense because ultimately we’re here to realize a
particular mission, not in service of some organization,” Jensen said.10 “It
got people thinking about the work and not the organization. The work, not
the hierarchy.”
Under the “mission is the boss” philosophy, Jensen would start every
new project by designating a leader, or a “Pilot in Command” (PIC), who
would report directly to Jensen. He found that this created far more
accountability—and a far greater incentive to do a job well—than did the
standard divisional structure.
“We always have a PIC for every project. Whenever Jensen talks about
any project or any deliverables, he always wants the name. Nobody can
hide behind, ‘such and such a team is working on that,’ ” former finance
executive Simona Jankowski said.11 “Everything has to have a name
attached to it because you have to know who’s the PIC, who’s accountable.”
In exchange for that level of accountability, PICs were granted the weight
of Jensen’s authority and received priority support throughout the
organization. After Jensen organized Nvidia’s employees into groups
centralized by function—sales, engineering, operations, and so on—they
were treated as a general pool of talent and not divided by business units or
divisions. This allowed the people with the right skills to be assigned to
projects on an ad hoc basis. It also helped mitigate some of the ever-present
job insecurity that plagues corporate America.
“Nvidia doesn’t constantly fire people and rehire them,” said Jay Puri,
head of global field operations.12 “We take people that we have and we are
able to redirect them into a new mission.” Managers at Nvidia were trained
not to get territorial or feel like they “own” their people and instead got
used to them moving around between task groups. This practice removed
one of the main sources of friction at large companies.
“Managers don’t feel like they get power by having large teams,” Puri
continued. “You get power at Nvidia by doing amazing work.”
Jensen found that the changes made Nvidia much faster and much more
efficient. Decisions could be made quickly, as employees were empowered
to contribute to every decision, regardless of rank. Arguments were decided
on the basis of quality of information, data, and merit—not on a leader’s
need to get promoted or earn a bonus, or that leader’s ability to pressure
others into going along with him or her.
Most of all, the flat structure freed Jensen to spend his precious time
explaining the reasoning behind his decisions at meetings instead of
adjudicating turf wars. Not only did he see flatness as key to Nvidia’s
strategic alignment, keeping everyone focused on the mission; he also saw
it as an opportunity to develop his junior employees by showing them how
a senior leader should think through a problem. “Let me reason through
this. Let me explain why I did that,” Jensen said. “How do we compare and
contrast these ideas? That process of management is really empowering.”
Of course, employees’ constant exposure to Jensen and his decision-
making process included his public dressings-down of executives and PICs.
He justified those potentially painful moments as an efficiency gain for the
company: offering private one-on-one feedback behind closed doors would
slow him and the company down by requiring the scheduling of separate
meetings, but it also deprived junior employees of a learning opportunity.
“I don’t take people aside,” he said. “We’re not optimizing for not
embarrassing somebody. We’re optimizing for the company learning from
our mistakes. If a leader can’t handle the slight embarrassment, they can
come talk to me. But it’s never happened.”13

NOT EVERYTHING CAN BE COMMUNICATED in meetings. With

such a large and distributed organization, Jensen needed to somehow keep
tabs on what was going on inside Nvidia in order to make sure everyone
had the right priorities. At other companies, an executive would rely on a
formal status update from those under him or her. But Nvidia management
believed that formal status reports tended to consist of information that had
been sanitized so thoroughly that it was useless. Anything smacking of
controversy—current problems, expected roadblocks, personnel issues—
would be removed in favor of presenting a cheerful picture of harmony to
those in charge.
So Jensen asked employees at every level of the organization to send an
e-mail to their immediate team and to executives that detailed the top five
things they were working on and what they had recently observed in their
markets, including customer pain points, competitor activities, technology
developments, and the potential for project delays. “The ideal top five e-
mail is five bullet points where the first word is an action word. It has to be
something like finalize, build, or secure,” said early employee Robert
Csongor.14
To make it easier for himself to filter these e-mails, Jensen had each
department tag them by topic in the subject line: cloud service provider,
OEM, health care, or retail. That way, if he wanted to get all of his recent e-
mails on, say, hyperscaler accounts, he could easily find them through a
keyword search.
The “Top 5” e-mails became a crucial feedback channel for Jensen. They
enabled him to get ahead of changes in the market that were obvious to
junior employees but not yet to him or his e-staff. “I’m looking to detect the
weak signals,” he would tell his employees when asked why he liked the
Top 5 process. “It’s easy to pick up the strong signals, but I want to
intercept them when they are weak.” To his e-staff, he was a little more
pointed.
“Don’t take this the wrong way, but you may not have the brainpower or
the wherewithal to detect something I think is pretty significant.”15
Every day, he would read about a hundred Top 5 e-mails to get a snapshot
of what was happening within the company. On Sundays, he would dedicate
an even longer session to Top 5’s, usually accompanied by a glass of his
favorite single-malt Highland Park scotch whiskey. It was the thing he did
for fun: “I drink a scotch, and I do e-mails.”
Top 5 e-mails became a source of new market insights. When Jensen got
interested in a new market, he used the e-mails to shape his strategic
thinking in near real-time. For example, after reading several Top 5 e-mails
from employees that discussed machine-learning trends, Jensen decided that
the company wasn’t moving fast enough to take advantage of that market.
“I keep seeing this. I don’t think we have enough invested in this technique
called RAPIDS,” former executive Michael Douglas recalls. Jensen
promptly told his staff to add more software engineers to the development
of a RAPIDS CUDA library, which became an important resource for
accelerating data-science and machine-learning workloads on GPUs.
Driven by Jensen, Nvidia’s e-mail culture was and remains unrelenting.
“One thing I learned pretty quickly is if you got an e-mail from him, you
acted on it,” Douglas said.16 “Nothing stays. Nothing festers. You answer
and move on it,” former head of human resources John McSorley said.17
Jensen would often respond to e-mails within minutes of receiving them
and wanted a response from an employee within twenty-four hours at most.
The responses had to be thoughtful and backed by hard data. Those that fell
short of his high standards would get a typically sarcastic response: “Oh, is
that right?”
Because of Jensen’s lightning-quick reactions, employees learned to time
their Top 5 e-mails strategically. “You always have to be concerned if you
send it on Friday night because Jensen will respond to you late Friday,” a
former employee said.18 “It would wreck your weekend.” As a result, most
employees sent their Top 5 late on Sunday night, about the time that Jensen
would be settling in at his home office with scotch in hand. They could then
start working on his directives at the beginning of the work week.
Former life-sciences alliance manager Mark Berger unwittingly triggered
all of Jensen’s pet peeves around Top 5 e-mails when he sent one of his first
ones, which attempted to forecast GPU sales in his market. Jensen thought
Nvidia hadn’t made sufficient progress in the life sciences and now
perceived a lack of rigor in Berger’s analysis. The chief executive asked
him whether he had bothered to talk to research professor Ross Walker, who
had built a lab of scientists at the San Diego Supercomputer Center at UC
San Diego.
Berger admitted he hadn’t consulted with Walker, believing that the
academic would not know the specifics of how GPUs were used in research
labs. Jensen unleashed a tirade and challenged Berger to figure out a way to
gather more information.
The experience rattled Berger—and made him a better employee. “The
one thing with Jensen is you don’t bullshit him,” he recalled years later.
“You bullshit him and your credibility is dead. The appropriate answer is, ‘I
don’t know, Jensen, but I’ll find out.’ ”19
Sufficiently chastened, Berger got in touch with Walker right away. The
two men designed a survey for other life-sciences academics who did work
with GPUs. The survey took thirty minutes to complete, but Berger
incentivized scientists to complete it by offering them an entry in a raffle
for a gaming GPU. Berger and Walker received comprehensive responses
from three hundred fifty scientists on what software they had installed, the
size of their modeling projects, what features they wanted from Nvidia, and
their backgrounds. It was a treasure trove of data, and when Berger
presented it in a follow-up meeting, Jensen was finally satisfied that he had
done his due diligence on his market.

JENSEN HAS ALWAYS BEEN TRYING to achieve the closest thing

possible to the Vulcan mind meld from Star Trek—the complete fusion of
his employees’ minds with his own. As we saw in the Introduction, perhaps
his favorite tool for showing the rest of the company his thought process is
the whiteboard.
Jensen’s preference for whiteboarding runs counter to the way that the
rest of corporate America talks to itself—through PowerPoint presentations
where a speaker goes through a series of slides presenting information that
is usually accepted by the audience at face value. He has always hated how
static such meetings are, with little opportunity to work together or discuss
topics in depth.
At the whiteboard, Jensen will sketch out how to organize a particular
market, how to accelerate growth of a particular product, and the software
or hardware technical stacks involved in a particular case. His
whiteboarding creates a specific kind of meeting, one dedicated to solving
problems, not reviewing things that have already been done. “When Jensen
gets into a meeting, he wants to prioritize what the important issues are,
then starts with the top one and works toward solving that problem,” said
Jay Puri.20
Unlike Top 5 e-mails, whiteboarding has been a common practice at
Nvidia from the start. The company designed its two current main
headquarters buildings, Endeavor and Voyager, built in 2017 and 2022,
respectively, to encourage collaboration. Each building has a fully open
working space and wall-to-wall whiteboards in dozens of conference
rooms. Employees at every level are expected to use those whiteboards as
much as possible.
Every quarter, for example, Jensen convenes a meeting with a few
hundred of Nvidia’s leaders in a big conference room. Each general
manager had to go up in front of the room and discuss his or her business.
The general managers are expected to use the whiteboard to talk through
their business story, explain what they do, and face challenges to their
underlying assumptions. Jensen sits in the first row, next to some other
senior executives, and asks detailed questions of the person at the
whiteboard—questions that often required further whiteboarding.
“It wasn’t really a business review, but something forward-looking,”
recalled Andy Keane. Jensen looked at quarterly results as the final
scorecard of decisions made and implemented months or years ago. He
wanted everyone to constantly reflect on how they could have made better
decisions then, and how they would use those lessons to make better ones
now and in the future—especially when it came to allocating resources and
deciding on strategy. Even when the numbers were good, he wanted people
to remain aggressive. “It was always about how to do better. There was this
constant push, push, push,” Keane said.
The whiteboarding process helped executives distill what was essential.
They all started with an empty board; they had to forget the past and focus
on what was important now. “Every meeting was all about the whiteboard,”
former Nvidia executive David Ragones said.21 “It’s a give and take. While
you are whiteboarding, he’d jump up to another whiteboard and write his
thoughts there. He wanted to see your understanding and how you’re
thinking through the issues and then develop his own thinking.”
At the conclusion of a meeting, Jensen would summarize the new ideas
the group had developed on the whiteboard. That way, he could ensure that
there would be no misunderstanding on direction or responsibilities.
His subordinates found that he expected them to be ready to whiteboard,
even when they were traveling. Whenever Michael Douglas went on a
business trip with Jensen, he would make sure there was a big whiteboard at
each of their destinations—even if he had to arrange to rent or buy one on-
location. “If five people are forced to carry that whiteboard in, that’s the
right size,” Douglas said. “He needs all that whiteboard real estate.”22
Besides good scotch, one of Jensen’s few indulgences is his preferred
brand of whiteboard marker. He insists on twelve-millimeter-wide, chisel-
tip markers that are sold only in Taiwan. He wants employees sitting in the
back to be able to see his writing and diagrams. Nvidia employees must
keep a ready stock of these markers handy at all times.
Jensen is nonchalant about the prevalence of whiteboarding culture at
Nvidia, almost as if it was a fallback option. “We have to use a whiteboard
because I don’t have a projector. I don’t have a TV and I don’t like slides,
so we just talk and draw,” he said with a shrug.23
 But there is more to it than that suggests. Whiteboarding forces people to
be both rigorous and transparent. It requires them to start from scratch every
time they step up to the board, and therefore to lay out their thinking as
thoroughly and clearly as possible. It becomes immediately apparent when
someone hasn’t thought something through or bases their logic on faulty
assumptions, unlike with a slide deck, where you can hide incomplete
thoughts in pretty formatting and misleading text. At the whiteboard, there
is no place to hide. And when you finish, no matter how brilliant your
thoughts are, you must always wipe them away and start anew.

NVIDIA BECAME A MATURE COMPANY not as a result of the size of

its revenues, the refinement of its internal structure, or the collective brain
power of its employees. Rather, it became mature when Jensen learned how
to consistently turn the organization away from internal political
dysfunction and disorder. Through mechanisms such as direct public
feedback, the Top 5 e-mail, and the requirement to present ideas on a
whiteboard rather than as a static PowerPoint, Nvidia equips its workforce
with powerful weapons in the constant struggle for accuracy and rigor and
against groupthink and inertia. It is these operational principles that have
allowed Nvidia to move quickly to take advantage of new opportunities.
If Nvidia had not evolved from its early, more conventional form, it
would not have invented the GPU or designed CUDA; it probably wouldn’t
have survived into a second decade, even with Jensen in charge. But the
organizational dynamic he eventually created—one that represents the exact
opposite of the “best practices” in most of the rest of corporate America—
has made it possible for the company to withstand, and thrive amid, the
pressures of an eternally unforgiving market.
 CHAPTER 10

The Engineer’s
Mind

EARLY IN MY WORKING LIFE, I switched careers, leaving consulting

to join a small technology fund as a stock analyst. I recall the first time I
attended a major Wall Street investment conference, where I looked forward
to the breakout Q&A sessions with CEOs that followed their main
presentations. At a session with the late Gerald Levin, the CEO of the newly
merged AOL Time Warner, I asked a basic, somewhat skeptical strategy
question about how the conglomerate planned to use AOL’s technology and
platform. Levin’s response stunned me. Instead of providing a cogent
answer, he launched into a lecture on the power and capabilities of AOL
Instant Messenger, speaking in such a jumble of buzzwords that I struggled
to make sense of it.
As a technology enthusiast who had built several computers and spent
considerable time on the then-nascent Internet, it was clear to me that Levin
had little understanding of how AOL’s products actually worked. I found
myself wondering how a business executive with such limited technical
knowledge could find himself running one of the largest media and
technology companies in the world.
Yet as I learned soon enough, Levin was not an outlier. Activist investor
Carl Icahn has a theory that much of corporate America mismanages the
succession process in choosing new CEOs. He calls it anti-Darwinian—the
very antithesis of the ruthless process of natural selection that allows only
the best equipped of a species to survive and reproduce.1
Icahn observed that competent executives often get sidelined in favor of
more likeable but less capable ones because of behavioral incentives inside
companies. The personalities who ascend the corporate ranks resemble
college fraternity presidents. They become friendly with the board of
directors and are not threatening to the current CEO. They’re not prodigies,
but they’re affable, always available for a drink when you are feeling down.
As Icahn put it, these figures (they are mostly men) are “not the smartest,
not the brightest, not the best, but likeable and sort of reliable.”
CEOs want to survive. Naturally, then, they prefer not to oversee a direct
subordinate who is brighter and could potentially replace them. They tend
to opt for someone slightly less astute than they themselves are. But when
the CEO eventually departs, the glad-handing executive who is now on
good terms with the board of directors frequently gets elevated,
perpetuating the “survival of the un-fittest” as the new CEO starts a similar
cycle.
Over the past few decades, I’ve seen several examples of the congenial,
nontechnical executive with a business background becoming the CEO of
major technology companies. As with Gerald Levin and AOL Time Warner,
the outcomes have been mediocre or worse.
Microsoft’s Steve Ballmer is the classic example. Ballmer started his
career as a marketing manager at Procter & Gamble, then began studies
toward an MBA from Stanford before joining Microsoft in 1980. He was
the first business manager hired by Bill Gates; he held positions in
operations, sales, and upper management but had little hands-on experience
with technology.
He had a poor reputation in the tech sector. Walt Mossberg, a former
columnist for the Wall Street Journal, once recounted an interaction with
Steve Jobs at Apple.2 Mossberg was sitting down to interview Jobs when
the Apple CEO asked him about his recent trip to Microsoft. Jobs seemed
particularly interested to know whether Ballmer remained firmly in control
of the software giant. When Mossberg confirmed that he was, Jobs paused,
then pumped his arms and yelped, “Yes!” Mossberg elaborated that while
Jobs held Gates in high regard, he had little respect for Ballmer.
 Jobs was right. Under Ballmer, Microsoft missed the shift to mobile
computing and also made a series of terrible acquisitions, including
aQuantive and Nokia. Microsoft’s stock price fell more than 30 percent
during Ballmer’s fourteen years as CEO.
Apple had previously faced its own challenges under a chief executive
with more of a business than technical background. Jobs was famously
ousted by Apple’s board of directors in 1985, who replaced him with John
Sculley, a former marketing specialist at PepsiCo. Sculley met with some
initial success, including with his strategy of selling incrementally better
computers at higher and higher prices. He then made several misguided
technology-product decisions, such as introducing the Newton personal
digital assistant and selecting PowerPC processors for the Mac in the early
1990s. The stagnation in technical innovation brought Apple to the brink of
bankruptcy later in the decade.
While Ballmer and Sculley could sell different versions of Windows or
expensive PowerBook laptops better than anyone, they couldn’t predict
where technology was going next. Apple wasn’t able to upgrade its
operating system to modern standards until it acquired Jobs’s NeXT
Computer, whose technology became the foundation for Mac OS X.
Intel offers another example. Bob Swan joined the chip maker as its chief
financial officer in 2016 and rose to the position of CEO two years later.
Swan had a primarily financial background; he had previously held CFO
roles at eBay and Electronic Data Systems, the company founded by former
IBM salesman H. Ross Perot. Under Swan’s leadership, Intel suffered from
repeated delays in moving to more advanced chip-manufacturing
technologies and its next generations of processors, falling behind its main
CPU competitor, Advanced Micro Devices. Worse, it seemed Swan was
mainly focused on executing a substantial multibillion-dollar stock buyback
program and issuing billions in dividends to lift the company’s stock price,
which siphoned money away from R&D investments. Intel floundered so
badly that it lost significant market share across its businesses and yielded
pole position in the technology of CPUs to AMD, which was then led by
Lisa Su, who, in contrast to Swan, had a strong engineering pedigree.
Swan also proved a poor manager and allocator of Intel’s resources. Like
Nvidia, in the late 2010s Intel invested heavily in AI. In 2016, the company
acquired deep-learning start-up Nervana Systems for $408 million to
develop AI chips. The next year, Intel hired Raja Koduri, the former head of
AMD’s graphics-chip unit, to lead its GPU efforts. As CEO, Swan further
expanded Intel’s AI portfolio by acquiring Israelbased Habana Labs for $2
billion in 2019. But Intel had no coherent strategy; it pursued multiple,
independent, AI-related chip projects that divided both resources and
attention.
This was largely the result of Swan’s unfamiliarity with the technical
aspects of the business he ran. He lacked the knowledge to make informed
decisions about where the company should focus its time and to know who
should really be in charge of making those decisions. Instead, he was too
easily influenced by whoever could put together the best presentation—
even if, according to one former Intel executive, that presentation had no
basis in reality.
Under Swan, Intel made a string of poor product decisions. On the AI
front, it shut down Nervana Systems, even though the start-up had a
promising product that was nearly ready. Instead, the company restarted its
AI efforts with Habana, effectively negating the prior several years of
development time.
Nvidia’s head of GPU engineering, Jonah Alben, commented on Intel’s
AI plans after the company acquired Habana. “Intel’s AI strategy is like
throwing darts. They don’t know what to do but feel like they need to buy
something, so they are buying everything,” he said.3
In 2021, Swan resigned as Intel’s CEO and was replaced by Pat
Gelsinger, who came with an impressive background in engineering. One of
his first decisions was to halt the stock buybacks.

NVIDIA WAS ABLE TO AVOID similar pitfalls because it had a technical

CEO in Jensen. “When you meet Jensen Huang, even with dozens of other
graphics companies, you realize this is a guy you want to do business with,”
said Tench Coxe, one of Nvidia’s early investors, who continues to serve on
its board today. “What made him great is he is an engineer and he is a
computer scientist.”4
Former product manager Ali Simnad recalled working on a Wi-Fi
product that was never released, in part because of Jensen’s intense
diligence.
 “Jensen was very scary,” he said.5 “You would go to a meeting and he
would know more about the product than you do.” During the product
meeting, Jensen made clear that he understood all the technical details of
the various Wi-Fi standards. The product wasn’t crucial to Nvidia’s strategy,
but Jensen still made the time to master the technology and specifications.
“He knew everything. In every meeting we attended, he was probably the
most prepared person.”
Jensen is known for being active on numerous internal Nvidia topical e-
mail discussion groups to keep up with trends and expand his knowledge.
On the “deep learning” list, where engineers discuss the latest technology
developments in AI, Jensen has a habit of forwarding articles of interest.
“You very much knew what Jensen was thinking about,” former Nvidia
senior research scientist Leo Tam said.6
Former marketing executive Kevin Krewell recalls meeting Jensen on the
street outside the NeurIPS conference in Barcelona, Spain, in 2016.
NeurIPS is an academic conference held in December, where machine-
learning and neuroscience experts present their latest findings. It’s not like a
SIGGRAPH or GDC, which are known to portions of the general public—
NeurIPS is more hardcore.
Krewell knew Jensen wasn’t scheduled to speak and asked him what he
was doing at the conference. Jensen replied, “I’m here to learn.”7
Nvidia’s CEO had not assigned someone to attend and take notes on his
behalf. He had shown up himself so he could absorb the recent
developments in artificial intelligence. He wanted to be deeply involved in
the space, attending sessions and talking with presenters, students, and
professors. Later on, he began hiring many of the people he met at the
conference.
Jensen has said many times that he could not do his job effectively
without in-depth familiarity with the technology itself. “It’s essential we
understand the underpinnings of the technology so you have an intuition for
how the industry is going to change,” he once remarked.8 “Our ability to
extrapolate and see down the road is really vital because technology is
changing fast, but it still takes us several years to build a great solution.”
Only with domain expertise can he decide which projects to support,
estimate how long they will take, and then allocate resources properly to
generate the best long-term returns.9
 There can be a downside to being so in the weeds: it can cause decision
paralysis. A good leader must make decisions, even when lacking total
precision. This was a lesson Jensen learned early on in an engineering class
taught by Professor Donald Amort at Oregon State. In his lessons, Amort
always used round numbers.
“I hated that,” said Jensen. “We were working with exponents and
numbers from the real world accurate to three decimal points.”10 Yet Amort
rejected such precision if it slowed him down too much; he would round up
0.68 to 0.7, for example. He was teaching his students to not lose sight of
the big picture. “It used to drive me crazy. But I learned over the years that
false accuracy is pointless.”
Jensen applied the round-number rule at Nvidia. His employees call it
“CEO math,” half-jokingly and half-affectionately. It allows him to do big-
picture strategic thinking without getting bogged down. He can determine
the size of a new market and its potential to drive profit for Nvidia quickly,
and then spend more mental energy on the more complex and intuitive tasks
of analyzing the competitive landscape and developing an entry strategy. As
Tench Coxe noted, “It’s easy to make a spreadsheet tell you whatever you
want to see, but Jensen getting comfortable using CEO math was a great
growth for him.”11
Jensen’s approach to math—direct, clipped, and oriented toward the big
picture—is also how he communicates with Nvidia’s employees more
generally. Because everything at Nvidia is under his purview, he has to be
efficient with his outbound messages. According to former sales executive
Jeff Fisher, “His e-mails are short and sweet. Sometimes too short.”12
“Like a haiku,” agrees Bryan Catanzaro.13
The comparison is an apt one. The three-line Japanese poems can often
be impenetrable or ambiguous—and it can be a challenge for new Nvidia
employees to get used to the brevity of Jensen’s e-mail communications.
Even veterans can often spend hours debating what a particular e-mail from
the CEO means, and when they can’t decide among themselves, they will
check back with him to get clarification.
But that is, on one level, what he wants. Most senior leaders at Nvidia
agree that Jensen relies on his people to exercise their good judgment on
interpreting his directions. He doesn’t want to control every decision; in
fact, being overly prescriptive can stifle the very independence and bias
toward action that he seeks to cultivate. Rather, he wants to make sure that
they have done their diligence and considered all possible effects of their
decisions. Catanzaro emphasized that Jensen’s approach is not merely about
his personal preferences.
“We’re all busy,” he says. “We all have way more e-mail than we can
read. The message is you should have empathy for the people you’re
presenting your work [to]. Don’t just dump everything on them. Give it to
them in a way that piques their interest, so if they want to, they can ask for
more details. Jensen is trying to help us be a more effective company and be
careful how we use each other’s attention. If you want to be impactful in a
large organization, don’t waste other people’s time.”

THE PUREST EXPRESSION OF Jensen’s engineering background is

his seemingly limitless capacity for work. In business, as he sees it, work
ethic may well be more important than intelligence. “It doesn’t matter how
smart you are because there is always someone smarter than you,” he said.
And in a global world, “your competition doesn’t go to sleep.”14
Neither does Jensen. Although he has changed and matured as a leader—
in, for instance, his strategic vision, his understanding of graphics and
accelerated computing, and his ability to run an organization—the one
constant in his three-decade tenure as CEO has been his commitment to
long hours and maximum effort.
An operations executive claimed that Nvidia isn’t a 24/7 company, but a
25/8 one. “I’m not kidding. I wake up at 4:30 a.m., and I’m on the phone
until 10:00 p.m.,” she said. “It’s my choice. It’s not for everybody.”
Another product manager noted that many employees don’t want to
embrace the grind and end up leaving after a few years. He himself tended
to arrive at the office before 9:00 a.m. and rarely left before 7:00 p.m. Once
home, he had to log on from 10:00 p.m. to 11:30 p.m. every night to talk to
partners in Taiwan. “On weekends, if you could not reply to an e-mail
within two hours, you had to let the team know the reason why you
wouldn’t be able to respond,” he said. When he reviewed his calendar, he
discovered that he had been either traveling for work or at the office for
almost half of his weekends over the past year.
Nvidia’s extreme work culture stems from the chief executive himself,
who lives and breathes his job and looks down on anyone who isn’t as
committed. “I don’t actually know anybody who is incredibly successful
who just approaches business like, ‘This is just business. This is what I do
from 8 to 5, and I’m going home, and at 5:01, I’m shutting it down,’ ”
Jensen has said.15 “I’ve never known anybody who is incredibly successful
like that. You have to allow yourself to be obsessed with your work.”
Employees dread whenever Jensen goes on a rare vacation because he
tends to sit in his hotel and write more e-mails, giving them even more
work than usual. During Nvidia’s early days, Michael Hara and Dan Vivoli
tried to stage an intervention. They called Jensen: “Dude, what are you
doing? You are on vacation.”
Jensen replied, “I’m sitting here on the balcony watching my kids
playing in the sand and writing e-mails.”
“Go out and play with your kids!” his subordinates insisted.
“No, no, no,” Jensen refused. “This is when I can get a lot of work done.”
When he goes to the movies, Jensen says he never remembers the film
because he spends the entire time thinking about work. “I work every day.
There’s not a day that goes by I don’t work. If I’m not working, I’m
thinking about working,” Jensen said. “Working is relaxing for me.”16
He lacks sympathy for anyone who works less than he does, and he does
not believe that he has missed out on anything in life by giving himself so
completely to Nvidia. When 60 Minutes interviewed Jensen in 2024 and
asked about employees who said working for him was demanding, that he
was a perfectionist and not easy to work for, he simply agreed.
“It should be like that. If you want to do extraordinary things, it shouldn’t
be easy.”
In all my years covering business, as a consultant, an analyst, and now as
a business writer, I have never met anyone quite like Jensen. In the field of
graphics, he is a pioneer. In the harsh technology market, he is a survivor.
And he has been a CEO for more than thirty years—making him, as of this
writing, the fourth-longest currently-serving CEO in the S&P 500, after
only Warren Buffett of Berkshire Hathaway, Stephen Schwarzman of
Blackstone, and Leonard Schleifer of Regeneron. Within the tech sector, his
tenure at Nvidia has been longer than Jeff Bezos’s twenty-seven years at
Amazon, Bill Gates’s twenty-five years at Microsoft, and Steve Jobs’s
fourteen-year second term at Apple—and none of them are still in charge.
He is closing in on the overall record in tech set by Larry Ellison, who
cofounded Oracle and spent thirty-seven years as its CEO before stepping
back to the CTO position in 2014.
What separates Jensen from almost all of his competitors is easy to
understand yet hard to implement. He challenges the division of the
executive world between those CEO-founders who are technically oriented
but naïve in the world of business and those who are business-minded
operators but who have no technical acumen. He shows it is possible that
one person can serve both roles; in fact, in the highly technical
semiconductor industry, his ambidextrousness may be a key to success. This
is also why he has an almost symbiotic relationship with Nvidia. In many
ways, he is Nvidia, and the company is Jensen, stretched to the proportions
of a multinational corporation with tens of thousands of employees and
billions of dollars in revenue.
Of course, this reality raises a question that most likely won’t be
answered for some time: What happens when he and the company part
ways, as they inevitably will?

THE STAKES COULD NOT BE higher. Jensen always reminds Nvidia

employees that the company is just one bad decision away from taking the
road to obsolescence. The history of Intel, Nvidia’s sometimes-partner and
sometimes-rival, illustrates this risk all too clearly.
In 1981, IBM introduced the IBM PC, revolutionizing the world of
computing. The computer manufacturer made two critical choices for the
PC that would define the industry. The first was choosing an Intel 8088 chip
as the PC’s processor. The second was deciding on MSDOS, from a small
software start-up called Microsoft, as the PC’s operating system. IBM made
an important strategic error, though. At the time, the company was so
confident in its size and distribution might that it didn’t secure exclusivity
on Intel’s and Microsoft’s offerings. Soon, “PC-compatible” clones with
identical hardware but lower prices flooded the market. PC makers such as
Dell and HP priced IBM out of the very product category it had created, and
IBM sold its PC division to Lenovo in 2005.
But one consequence of IBM’s error was the close association of
Microsoft and Intel. For the past four decades, the two companies have
dominated the computer industry. The business partnership was eventually
dubbed “WinTel,” a neologism combining Windows, the name of the
operating system Microsoft would later develop, and Intel.
WinTel is an instance of what analysts call “lock-in.” Corporations built
more of their business processes around custom applications that ran on
Microsoft Windows PCs and servers powered by Intel x86 processors. Once
this happened, it was too difficult to switch to another operating or
computing system, such as Apple’s Mac ecosystem. Enterprises can’t just
take millions of lines of code written for Windows and put them on another
chip architecture. Rewriting software dependent on specialized Windows
libraries and utilities would have been an enormous task that CIOs deemed
too complicated and not worth the technical risks.
However, the fortunes of Microsoft and Intel diverged as each one
reacted to disruptive new technology. After Satya Nadella took over as CEO
of Microsoft in 2014, the company pivoted and bet aggressively on the rise
of cloud subscription software and cloud computing, earning itself a strong
number-two position behind Amazon Web Services in the latter category.
Intel, by contrast, missed a pair of generational opportunities: the arrival
of smartphone processors and the rise of AI software. In 2006, Steve Jobs
asked Intel CEO Paul Otellini whether the chip maker would be willing to
supply processors for the upcoming iPhone. In a fateful decision that would
prevent Intel from participating in the future of the smartphone chip market,
Otellini declined. “There was a chip that they were interested in, that they
[Apple] wanted to pay a certain price for and not a nickel more, and that
price was below our forecasted cost. I couldn’t see it,” he said in a 2013
interview with The Atlantic. “The world would have been a lot different if
we’d done it.”17
Also in 2006, Intel sold its XScale unit, which was developing power-
efficient ARM-based processors for mobile devices, to Marvell Technology
for $600 million. This left the company without important expertise just
before the smartphone market came to be dominated by such processors.
(Arm Holdings, which returned to public markets in 2023, licenses its
power-efficient chip-architecture designs, well suited for mobile devices, to
semiconductor companies and hardware makers, including Apple and
Qualcomm.)
Compounding matters, Intel made a series of missteps in its core
business. It was slow to purchase and introduce new chip-manufacturing
equipment from the Netherlands-based company ASML, which uses the
advanced chip-manufacturing technology called extreme ultraviolet (EUV)
lithography, and it underinvested in production techniques that are based on
EUV lithography. As a result, Intel fell behind TSMC in its ability to
produce more advanced chips at high volume. In 2020, when Intel
announced another round of delays in the transition to seven-nanometer
manufacturing, many customers abandoned it for competitors such as
Advanced Micro Devices, which designs semiconductors and pays TSMC
to make them. And in the same year, Apple started to replace Intel as a Mac
processor supplier with its internally designed chips that are based on the
ARM-chip architecture that powers the iPhone and which are now used
across its entire Mac lineup.
As for GPUs, current Intel CEO Pat Gelsinger laments that the company
failed to break into the category with its own in-house product that would
have competed with Nvidia’s.
“I had a project called Larrabee that, when I was pushed out of Intel, got
killed shortly thereafter,” he said. “The world would be different today had
that not occurred.”18
Gelsinger had been an executive champion of the project and headed the
enterprise-computing division at Intel before he left in 2009 for the data-
storage company EMC. The Larrabee GPU was canceled in 2010, and Intel
did not restart its GPU efforts until 2018.
While Intel made error after error, Nvidia was intensely focused on
inaugurating the era of the GPU. Under Jensen’s leadership, the company
invested so much in CUDA that it became a foundational ecosystem for AI
developers. Nvidia also made smart acquisitions, including the high-speed
networking leader Mellanox, to fill out the company’s data-center-
computing product offering. Nvidia took these decisions in the face of
demands from Wall Street to reduce costs and increase profits—exactly the
kind of strategy that Intel adopted when it declined to pursue ARM
architecture and GPUs. It was an instance of the innovator’s dilemma: Intel,
as the incumbent, failed to capitalize on new technology, allowing the more
agile Nvidia to undercut its entire business model.
So far, every major computing era has seen technology favor the big
players who can develop a market-leading platform—a “winner take most”
dynamic. WinTel’s dominance in PCs is a model for Nvidia’s leadership in
AI hardware and software. In an August 2023 report, Jefferies analyst Mark
Lipacis estimated that WinTel generated an incredible 80 percent of the
operating profit of the PC industry era.19 With the rise of the internet,
Google captured 90 percent of the search market.20 And Apple has been
able to generate nearly 80 percent of the profits of the smartphone industry
era.
This history may suggest that most of the spoils of the AI era will accrue
to Nvidia. The combination of CUDA and Nvidia’s GPUs, which are the
only chips that can run the platform, is comparable to the “lock-in” power
Microsoft’s Windows operating system and Intel’s x86 processors achieved
during the PC boom. Just as corporations built on top of Windows and its
libraries, AI model makers and enterprises are building on top of CUDA
software libraries.
Of course, Nvidia could falter and miss new computing waves, just as
IBM and Intel did. If it hopes to continue to remain relevant, it will have to
remain vigilant. Gelsinger commended Jensen for never giving up on his
vision for accelerated computing. “I have a lot of respect for Jensen because
he stayed true to his mission,” he said. But this is not just a question of
strategic vision. Nvidia continues to operate like a technology company, not
an investment vehicle. It does not focus on margins and profit-taking at the
expense of developing new innovations, even when those innovations can
drag on Nvidia’s bottom line.
“We can only continue to be relevant if we invest,” Jensen once said. “In
my business, if you don’t invest, you’ll be out of business soon.”
He believes, in other words, that in the highly technical chip industry,
innovative engineering matters far more than financial metrics. That belief
is perhaps the single thing that most differentiates Jensen from his peers.
 PART IV

INTO THE FUTURE

(2013–PRESENT)
 CHAPTER 11

The Road to AI

BY 2005, NVIDIA’S CHIEF SCIENTIST, David Kirk, was considering a

change. He had joined Nvidia in early 1997 during the development of the
RIVA 128 chip, which had saved the company. Since then, he had overseen
the launch of several chip architectures and witnessed Nvidia oscillate
between near-death experiences and market-defining successes. He needed
a break from the long hours and the stresses of the job, but this required
finding a worthy successor. Kirk knew of no one in the industry who could
meet his—and Jensen’s—high standards for the role of chief scientist at
Nvidia. But Kirk did have an eye on an academic who had an impressive
pedigree. The question was, how would the company lure him away from
his current post?
Professor Bill Dally had nothing left to prove in the field of computer
science. He was a living legend: after earning his bachelor’s degree in
electrical engineering in 1980 from Virginia Tech, he went to work at Bell
Labs on some of the earliest microprocessors ever invented. In 1981, while
working at Bell, he got his master’s in electrical engineering from Stanford,
and then in 1983 he enrolled in the computer science PhD program at
Caltech.1 Dally wrote his dissertation—Richard Feynman, the Nobel Prize–
winning theoretical physicist and pioneer in quantum mechanics, sat on his
committee—on the topic of concurrent data structures, a technique for
structuring information on a computer so that it can be used by multiple
computing threads simultaneously. Today, this is known as parallel
computing, and Nvidia relies on the technique for its entire line of advanced
processors.
After receiving his PhD, Dally taught at MIT, where he worked on both
cutting-edge supercomputers and cheaper machines that used off-the-shelf
parts. After eleven years in Cambridge, he returned to Stanford to chair its
computer science department and eventually was named to one of the
university’s coveted endowed professorships, becoming the Willard R. and
Inez Kerr Bell Professor of Engineering.
Kirk took note of Dally’s work in the early 2000s and invited him to
consult on the Tesla chip architecture that eventually powered the GeForce
8 series. This was Nvidia’s fifth-generation “true” GPU after the first,
programmable GeForce 3, but one of its first to really take advantage of
parallel computing. It was the first move in what would become a six-year
courtship.
“It was a long, slow hire. Once we hooked him, we just reeled him in
slowly,” Kirk said. “Bill was another essential piece because he’s like a
master of parallel computing. That’s what he’s been doing his whole
career . . . he had a vision of how parallel computing should work.”2
In 2008, Dally took a sabbatical to consider his next move. The next year,
Kirk finally won, convincing him to make the switch to industry. Dally
resigned from his post at Stanford and joined Nvidia full-time, aiming to
bring his theoretical work to commercial applications.
Kirk hired Dally not only to succeed him as chief scientist, an important
position with many duties across the company. He also knew that Dally
could accelerate Nvidia’s development of GPU technology.
For the first fifty years of computing history, the most important chip
inside the computer was the central processing unit, or CPU. The CPU is a
generalist, capable of performing a wide variety of tasks. It moves from
task to task with great speed and can dedicate significant processing power
to each operation. Nevertheless, it can handle only a few operations
simultaneously because of its limited number of cores, which process only a
few computation threads at once.
The GPU, in contrast, is optimized for volume over complexity. It
contains hundreds or thousands of tiny processing cores, enabling it to
break down tasks into numerous simpler operations executed in parallel.
While a GPU is less versatile than a CPU, it can vastly outperform a CPU
in processing speed for many applications.3 The secret to a GPU’s success
is parallel computing—the field Bill Dally had pioneered.
At Nvision 08—a conference held in San Jose aimed not at industry
insiders but at graphics enthusiasts—Jamie Hyneman and Adam Savage of
the TV show Mythbusters put on a presentation at the request of Nvidia.
They said that Nvidia had asked them to come up with a practical
demonstration of the differences between a CPU and a GPU—“kind of a
science lesson,” as Savage put it, “about how a GPU works.”4 They brought
on stage two machines that were designed to perform the same task—paint
a picture—in two different ways. The first machine was called Leonardo, a
remote-controlled robot that consisted of a paintball gun on a swiveling arm
that was mounted on a pair of tank-like treads. Hyneman piloted the robot
across the stage to a point in front of a blank canvas, where it began to
shoot paintballs according to a preprogrammed algorithm. Over the course
of thirty seconds, Leonardo produced a clearly legible smiley-face image in
a single color, blue. This, Savage explained, was how CPUs might perform
a task: “as a series of discrete actions performed sequentially, one after the
other.”
The second machine, Leonardo 2, was more like a GPU. It was a hulking
rack of eleven hundred identical tubes, each of which was loaded with a
single paintball. The tubes were connected to one of two giant compressed-
air tanks, which would launch the entire complement of paintballs
simultaneously. Whereas Leonardo took nearly half a minute to paint its
simple smiley face, Leonardo 2 took less than a tenth of a second to splatter
an entire canvas with a full-color image that was a recognizable
approximation of the Mona Lisa. “Kind of like a parallel processor,” said
Hyneman in his trademark deadpan.
Rendering computer graphics is a computationally intensive task, but it is
far less complex than, say, recalculating every math formula in a million-
cell spreadsheet. As a result, the most efficient way to make a computer
better at rendering graphics is to give it access to many more specialized
cores that can process more software threads in parallel, all optimized for
the small set of tasks related to graphics processing. To be more proficient
at what it was designed to do, a GPU doesn’t need more flexibility or more
brute-force power; it simply needs more throughput.
 Over time, the distinction between CPU and GPU has blurred, especially
as the kind of matrix math that GPUs can perform has been found to be
applicable in fields as diverse as computer vision, physics simulation, and
artificial intelligence. The GPU has become more of a general-purpose chip.

SOON AFTER HE STARTED AT NVIDIA, Dally began to redeploy the

company’s research teams to work on parallel computing. One of the first
big projects he had a hand in involved internet cat photos.
One of Dally’s former colleagues at Stanford, the computer science
professor Andrew Ng, was collaborating with Google Brain—one of
Alphabet’s AI research labs that would later merge into Google Deep-Mind
—to find better ways to conduct deep learning through neural networks.
Unlike early neural networks, which required humans to “teach” the
networks what they were looking at, deep-learning neural networks were
entirely self-directed. Ng’s team, for example, fed their deep-learning net a
random sampling of 10 million still images taken from YouTube and let it
decide which patterns occurred frequently enough for the net to
“remember” them. The model was exposed to so many videos of cats that it
independently developed a composite image of a cat’s face without human
intervention. From then on, it could reliably identify cats in images that
were not a part of its training set.5
To computer science veterans such as Dally, this was a tipping point.
“There are really three things that are needed to make deep learning work,”
he said.6 “Number one, the core algorithms have been around since the
’80s. There have been improvements like transformers, but by and large,
around for decades. Number two, datasets. You need a lot of data . . .
labeled datasets were interesting things [that] started to emerge in the early
2000s. And then Fei-Fei Li put together the ImageNet dataset. And those
were a huge public service because having that big dataset [and] making it
public made it available for a lot of people to do very interesting things.”
Ng’s work had shown the power of applying well-known and well-
understood algorithms to sufficiently huge datasets. Although his deep-
learning model’s ability to recognize cats captured the headlines, it was able
to do far more. With more than 1 billion parameters, the Google Brain
neural network could identify tens of thousands of different shapes, objects,
and even faces.7 Ng had needed Google, which gave him access to a rich
dataset for deep learning, a dataset that just happened to be one of the
largest content libraries in the world: YouTube, which Google has owned
since 2006. Even his home institution of Stanford, with its large research
budget, could not provide him that kind of training material. (Google was
not acting out of altruism: in exchange for access to its data, it retained the
rights to commercialize anything that Ng developed using that data.)
But the third thing “needed to make deep learning work,” according to
Dally, was the hardware, and this was proving more difficult to solve. Ng
had used one of Google’s data centers, building his own deep-learning
server by chaining together more than two thousand CPUs, with sixteen
thousand computing cores between them.8 Ng’s feat was impressive, to be
sure. But he now faced the same challenge that Ross Walker had faced at
the San Diego Supercomputing Center: as exciting as his proof-of-concept
work might have been, it still put the promise of deep learning far beyond
the reach of most organizations. Even well-funded research groups would
not be able to purchase thousands of expensive CPUs, let alone rent space
at a data center that could store, power, and cool such a massive computing
system. In order to truly unlock the potential of deep learning, the hardware
would have to become much more affordable.
After leaving Stanford to join Nvidia, Dally kept in touch with Ng. They
got together over breakfast one morning, and Ng revealed his work with
Google Brain. He described successfully demonstrating how deep-learning
theory could be applied to a real-world problem: the automated recognition
of objects in photos, without human tagging or intervention. Ng detailed his
approach of combining the extensive dataset of YouTube clips with the raw
power of tens of thousands of traditional processors.
Dally was impressed. “That’s really interesting,” he said. Then he made
an observation that would change the trajectory of artificial intelligence. “I
bet GPUs would be much better at doing that.”9
He assigned his Nvidia colleague Bryan Catanzaro, who had a PhD in
electrical engineering and computer science from the University of
California, Berkeley, to help Ng’s team use GPUs for deep learning. Dally
and Catanzaro were confident that the computational tasks involved could
be broken down into smaller, less complex operations that a GPU could
execute more efficiently. They developed a series of tests that conclusively
proved their case—in theory. The challenge, in practice, was that deep-
learning models were too large to run on a single GPU, which could handle
only models that had 250 million parameters, a fraction of the size of Ng’s
Google Brain model. While it was possible to get up to four GPUs on a
single server, “chaining” multiple GPU servers together to increase their
collective processing power had not been attempted before.10
Using Nvidia’s CUDA language, Catanzaro’s team wrote a new
optimized routine to enable the distribution of computation across many
GPUs and manage the communication among them. The optimizations
allowed Ng and Catanzaro to consolidate the work once performed by two
thousand CPUs across a mere twelve Nvidia GPUs.11
Catanzaro had demonstrated that with some skilled software work, GPUs
could provide “the spark that ignited the AI revolution,” according to
Dally.12 “If you think of the fuel as building the algorithms and the air as
being the datasets, once you have the GPUs, it makes it possible to apply
those to each other. Without that, it just wasn’t feasible.”
Catanzaro’s CUDA optimizations also brought him into direct contact
with Jensen for the first time. “All of a sudden, he got really interested in
the work that I was doing. He was e-mailing me to ask me questions about
what I was trying to do, what deep learning was, how did it work,”
Catanzaro recalled. “Also, of course, what the role of GPUs could
potentially be in bringing that to pass.”13
Jensen wanted to sell more GPUs, of course. But in order to do that, he
would need to find the “killer app” that drove GPU adoption. Deep learning
had the potential to be just that—but only if someone could show its use
beyond identifying house pets.

IN THE SAME PERIOD WHEN Catanzaro was working to help Ng

develop his deep-learning neural network project, a University of Toronto
research team showed that such networks could outperform the best human-
created software in solving the most challenging computer-vision problems.
The milestone had its roots back in 2007, when a newly minted professor
of computer science at Princeton named Fei-Fei Li (whom Dally referenced
in a quotation above) began working on a new project. At the time, the field
of computer vision was intent on developing the best models and
algorithms, because the assumption was that whoever designed the best
algorithm would necessarily get the most accurate results. Li turned that
assumption on its head, proposing that whoever trained on the best data
would get the best results, even if they hadn’t designed the most refined
algorithm.14 To give her fellow researchers a head start on the monumental
task of collecting the necessary data, she began compiling a catalog of
images, each of which was manually tagged based on its content. After two
years of work, the database had grown to more than 3 million images with
one thousand different and mutually exclusive categories, which ranged
from the specific (magpie, barometer, power drill) to the broad
(honeycomb, television, church). She christened her database ImageNet and
announced it to the academic world in the form of a research paper. At first,
no one read the paper or paid much attention to the other ways she tried to
draw attention to her research. So she contacted the University of Oxford,
which maintained a database similar to hers and sponsored an annual
competition in Europe for computer-vision researchers. She asked whether
Oxford might be willing to cosponsor something similar in the United
States, using ImageNet. The university agreed to do so, and in 2010 the first
ImageNet Large Scale Visual Recognition Challenge took place.15
The rules were simple: the competing models would get fed random
images from ImageNet and would have to correctly assign them to
categories. For the first two contests in 2010 and 2011, the results were not
very good. During the inaugural competition, one model misclassified
almost every single image, and no team scored better than 75 percent
correct.16 In the second year, the teams did better on average—the worst
performer got around half of the images correct—but once again, no one
correctly categorized more than 75 percent of the images.
In the third contest, which occurred in 2012, University of Toronto
professor Gary Hinton and two of his students, Ilya Sutskever and Alex
Krizhevsky, put forward an entry they called AlexNet. Unlike the rest of the
field, which had started developing algorithms and models before
optimizing them for use on ImageNet, the AlexNet team took the opposite
approach. They used Nvidia GPUs to support a small-scale deep-learning
neural network that was fed ImageNet content and which then “learned”
how to build relationships between images and their associated tags. The
team did not set out to write the best computer-vision algorithm possible; in
fact, they did not write a single line of computer-vision code themselves.
Instead, they wrote the best deep-learning model they could—and trusted it
to figure out the computer-vision problem on its own.
“GPUs starting around the Fermi generation, they were powerful enough
that, you know, you could do an interesting-sized neural network and an
interesting-sized amount of data in a reasonable amount of time,” said
Dally, referring to the chip architecture that powered the GeForce 500
series, which was first released in 2010. “So AlexNet was trained in two
weeks.”17
The results were astounding. Once again, the 75 percent barrier held for
most competitors. But AlexNet categorized almost 85 percent of the images
correctly—and it had done so on its own, through the power of deep
learning. AlexNet’s win gave Nvidia a huge PR boost, as Hinton and his
students had needed only a pair of off-the-shelf, consumer-grade GPUs that
cost a few hundred dollars apiece. AlexNet forever associated the company
with what is still considered one of the most important events in the history
of artificial intelligence.
“When Alex Krizhevsky and Illya Sutskever published their ImageNet
paper, [it] really took the world by storm,” said Catanzaro. “One of the
things that people often forget is that it’s primarily a systems paper. That
paper is not about a fancy new mathematical concept of how to think about
artificial intelligence. Instead, what they did was they used accelerated
computing to dramatically expand the dataset and the model that they were
applying to this particular problem. And that ended up yielding some great
results.”18
Alex Krizhevsky and Illya Sutskever’s work stoked Jensen’s interest in
artificial intelligence. He started talking frequently with Bill Dally and was
focused on how much of an opportunity deep learning, and specifically
GPU-powered deep learning, would be for Nvidia. There was considerable
debate within the executive team on the topic. Several of Jensen’s key
lieutenants were against investing more in deep learning, in the belief that it
was just a passing fad. But the CEO overruled them.
“Deep learning is going to be really big,” he said at an executive team
meeting in 2013. “We should go all in on it.”

THOUGH HE DIDN’T QUITE REALIZE IT, Jensen had spent the first
twenty years of Nvidia’s history preparing the company for this moment.
He had staffed Nvidia with the best talent he could find, including by
poaching from rivals and partners alike. He had created a culture that prized
technical brilliance, maximum effort, and, above all, a total commitment to
the company. He had built a company in the image of his own focused yet
far-ranging mind. Now, he would pull every lever at his disposal to navigate
Nvidia to the very center of the tech industry—as the company whose
hardware could bring about the AI-powered future.
The first step was to assign significantly more personnel and funding to
AI. Catanzaro estimated that there had been only a handful of people
working on AI-related projects. But as Jensen began to grasp the size of the
opportunity before Nvidia, he used the “one team” philosophy to quickly
reallocate resources.
“It was definitely not a single day when the entire company changed
forever,” remembers Catanzaro. “It was a period of several months where
Jensen was increasingly interested and started asking increasingly deep
questions and then started encouraging the company to swarm to machine
learning.”19
After the “swarming,” Nvidia released a torrent of new features designed
specifically for the AI market. Jensen had already taken the big and
expensive decision to make the company’s entire hardware lineup
compatible with CUDA, so that researchers and engineers could program
Nvidia GPUs for their specific needs. Now, he asked Dally to come up with
AI-focused improvements.
Jensen announced the change in strategic focus in a company all-hands
meeting. “We need to consider this work as our highest priority,” he said.20
He explained that Nvidia had to get the right people working on AI. If they
were currently assigned to something else, they would change focus and
work on AI, because it was going to be more important than anything else
they could possibly be doing.21
Catanzaro turned his GPU-optimization work into a software library
Nvidia called CUDA Deep Neural Network, or cuDNN. This became the
company’s first AI-optimized library. It would evolve into a must-have for
AI developers. It worked with all leading AI frameworks and allowed users
to automatically employ the most efficient algorithms for whatever GPU
task they needed. “Jensen was excited,” said Catanzaro. “He wanted to get
that productized and shipped as soon as possible.”
 Another promising path was to modify the level of precision of the math
that Nvidia’s GPUs could perform. At the time, the company’s GPUs
supported 32-bit (single float, or FP32) or 64-bit (double float, or FP64)
mathematical precision; either math type was a requirement for many
scientific and technical fields. But deep-learning models didn’t need to be
that precise. Models only required the GPUs to perform 16-bit floating-
point calculations, because the networks were resistant to calculation errors
during training. In other words, Nvidia GPUs did math that was overly
precise—and therefore much slower—for deep-learning models. To make
the GPUs run faster and allow these models to run more efficiently, in 2016
Dally implemented FP16 support on all Nvidia GPUs.
But the real task was making bespoke hardware circuits that were
optimized for AI. When Nvidia pivoted to AI, its architects were already
working on the next generation of GPUs, called Volta. The new line was
several years into development; making even a small change to the chip
design at that point would be costly and difficult. But Dally realized, with
some prodding from Jensen, that if the company did not try to make AI-
optimized chips now, it might not get another opportunity for years.
“The entire team—the GPU group, Jensen, and myself—agreed to
incorporate significantly more support” for AI despite how late in the
development process they were, said Dally. That “support” included the
development of an entirely new type of tiny processor, called the Tensor
Core, which was integrated into Volta. In machine learning, a tensor is a
type of data container that encodes multiple dimensions of information,
especially for complex content types such as images and videos. Because of
their richness, tensor-based calculations require large amounts of processing
power. And the most interesting forms of deep learning—image
recognition, language generation, and autonomous driving—required the
use of ever-larger and ever-more-complex tensors.
In much the same way that traditional GPUs marked an improvement
over CPU-based computation because of their ability to handle a smaller
subset of tasks more efficiently, Tensor Cores were an improvement over
traditional GPUs because they were optimized to run an even more
specialized subset of tasks at even higher efficiencies. In Dally’s words,
they were “matrix multiple engines,” made for deep learning, and deep
learning alone. A Volta-based GPU with Tensor Cores could train a deep-
learning model three times faster than the same GPU with standard CUDA
cores.22
All of these innovations and changes came with an operational cost.
Dally and his team made the last tweaks to the Volta line mere months
ahead of its schedule for tape-out, the last step before the locked design
entered production. It was almost unheard of for a chip maker to do such a
thing voluntarily instead of in response to a major defect found at the last
minute.
“That was a decision of how much chip area we are going to spend
because we think this evolving AI market is going to be a big market,”
recalled Dally. “It turned out to be a good call. I think that it was a real
strength of Nvidia that we could do that.”23
In a sense, Nvidia was doing what it had always done: spotting a big
opportunity and racing to get its products to market before anyone else
realized the potential was even there. Jensen came to understand early on in
the AI race that it wasn’t just about who could make the fastest chip for
deep learning. It was just as much about how everything—the hardware and
software infrastructure—worked together.
“Having an architecture and an attention mechanism that allowed for
scaling of these models really was also a kickstart in the industry,” recalled
Jensen in 2023.24
Dally agreed with Jensen’s assessment. “What is more important is
building up the whole ecosystem of software early on,” he said. Nvidia
wanted to produce “all sorts of software to just make it really easy for
people to do deep learning efficiently on GPUs” because presenting a ready
framework and a library of support software made it all but inevitable that
third-party developers, researchers, and engineers would turn to Nvidia first
when they thought of AI.
Just as CUDA had made Nvidia’s name in the insular world of academic
AI researchers, its next generation of hardware would arrive just in time for
those same pioneers to try their luck in the commercial market. Soon, the
center of gravity in AI would shift away from Stanford, Toronto, and
Caltech and move to start-ups and well-established tech companies alike.
Geoffrey Hinton and Fei-Fei Li would end up at Google. Andrew Ng
worked as the chief scientist at Baidu—originally the largest search engine
in China and now a tech conglomerate. And Ilya Sutskever, Hinton’s
student and one of the three researchers who made the AlexNet
breakthrough, would cofound a deep-learning start-up called OpenAI that
would bring the AI revolution to public consciousness.
The one thing all of them had in common was that in their academic
lives, they had used Nvidia GPUs to do their groundbreaking research. And
Nvidia would continue to be their preferred choice as they transformed AI
from an obscure academic field into a global obsession that created a huge
appetite for new chips, AI servers, and data centers.
The work of Bill Dally and Bryan Catanzaro allowed Jensen to pick up
an early signal of the potential of the new technology. Within a decade,
Jensen was sure that AI would create “the largest TAM [total addressable
market] expansion of software and hardware that we’ve seen in several
decades.”25 He refashioned Nvidia around AI in a matter of years, moving
with the “Speed of Light” intensity. In fact, only by taking extreme
measures—bucking the industry-wide tendencies toward static
organizations, long development timelines, and parsimonious R&D
spending—was Jensen able to prepare Nvidia to take advantage of the AI
earthquake when it finally occurred. And even then, no one—not even
Jensen—knew just how violently the ground was about to shift underneath
the entire tech industry.
 CHAPTER 12

The “Most
Feared” Hedge
Fund

THOUGH FEW KNOW IT, THE HISTORIES of Nvidia and Starboard

Value, perhaps the most famous activist hedge fund in the world, are
intertwined.
Jeff Smith, the founder of Starboard, grew up in the Long Island town of
Great Neck. In 1994, he received an economics degree from the University
of Pennsylvania’s Wharton School and started his career in investment
banking. Later, he joined a small hedge fund called Ramius Capital, which
merged into Cowen Group.1 In 2011, Smith and two of his partners spun
out Starboard Value as an independent fund, one that would be “focused on
unlocking value in underperforming companies for the benefit of all
shareholders.”2
According to a 2014 Fortune magazine article, Smith quickly earned a
reputation as the “most feared man” in corporate America for his aggressive
activist investing.3 By then the fund had more than $3 billion in assets
under management, generating strong returns of 15.5 percent per year. It
had replaced more than eighty directors on thirty different corporate boards;
the boards it shaped included those of the biotech company SurModics and
of the hair salon company Regis. It had experienced a rare loss in a proxy
fight to add directors to AOL’s board in 2012 but continued to set its sights
on ever-larger targets.
In late 2013, Starboard Value made its highest-profile move to date: the
fund announced that it had accumulated a 5.6 percent stake in the largest
owner of full-service chain restaurants in the country, Darden Restaurants,
the owner and operator of Olive Garden, Red Lobster, LongHorn
Steakhouse, and other national chains. Darden’s sales had been falling for
years, and the company had decided to divest from Red Lobster entirely,
citing the increasing costs of seafood.4 Smith disagreed with the decision;
he blamed Darden’s struggles on mismanagement and argued that Darden
divesting itself of Red Lobster would actually destroy value for
shareholders, not create it. Starboard believed that Darden already had
everything it needed to survive, except good leadership.
In September 2014, Starboard released its proposal to turn Darden around
in the form of a PowerPoint presentation that ran to almost three hundred
slides. The presentation deck generated significant attention in the national
media; business journalists noted the plan’s especially pointed tone
(“Darden has been mismanaged for years and . . . is in desperate need of a
turnaround”), while others made light of some of its cost-saving
suggestions, such as asking waiters to be less generous in handing out
unlimited breadsticks.5 But Starboard’s plan was both comprehensive and
logical—even the breadsticks suggestion, which was intended to increase
the number of touch points between staff and guests. And Starboard
suggested that it genuinely cared about Darden’s brands for more than just
financial reasons: “Olive Garden has a special place in our hearts,” one
slide said.6 The hedge fund’s combination of sentimentality and rigor won
over Darden’s shareholders; Starboard secured a proxy vote victory and
replaced the company’s entire twelve-person board. Darden’s CEO soon
resigned, and the company implemented the Starboard-approved
turnaround. Smith’s victory cemented his reputation for being both
thorough and tough.
A year before Starboard’s widely chronicled victory over Darden, Smith
had made a less publicized move on Nvidia.
Early in 2013, Nvidia’s shareholders were getting restless. The stock
price had been roughly flat for four years, and the financial performance
was mixed. In its latest quarter ending in January, sales were up 7 percent
year-over-year, but earnings were down 2 percent.
Nvidia had a strong balance sheet of about $3 billion in net cash, which
was a significant asset when the overall market value of the company was
$8 billion total. However, its growth rate was only in the single digits,
which resulted in a price-to-earnings (P/E) multiple of just 14 times
earnings. After backing out Nvidia’s cash on hand, Starboard believed that
the company was severely undervalued, and its core assets had far more
room to grow. The fund pounced: according to Securities and Exchange
Commission 13F filings, the hedge fund accumulated a stake of 4.4 million
shares in Nvidia, worth about $62 million, during the quarter ending in June
of 2013.
Some executives at Nvidia weren’t excited about having Starboard as an
investor. One senior Nvidia executive said the company’s board was very
worried that the activist fund would force a reorganization of the company,
install its own board, and make Nvidia cut back on its investments in
CUDA—the kind of drastic reshaping that it would attempt with Darden the
following year. Another Nvidia executive said Starboard wanted a board
seat, but the board had pushed back.
Still, the relationship never became too antagonistic. “I don’t think it ever
got to what I would call a crisis stage. You know DEFCON 1?” one Nvidia
executive said, referring to the alert system used by the U.S. military for
nuclear war. DEFCON 5 indicates peace, while DEFCON 1 means nuclear
war is imminent. “It got to DEFCON 3.”
The Starboard team met several times with Jensen and other Nvidia
leaders to discuss strategy. Looking back on the investment years later,
Smith said that Starboard primarily advocated for an aggressive stock
buyback program and a de-emphasis on non-GPU projects such as phone
processors.7 Starboard refrained from applying additional pressure after the
meetings. The hedge fund eventually got its wish on the buybacks. In
November 2013, Nvidia made two announcements: a commitment to buy
back $1 billion of stock by fiscal 2015 and the authorization of an
additional $1 billion stock buyback. The stock price rallied about 20 percent
in the ensuing few months, and Starboard sold its position in Nvidia by
March the following year.
 Far from a contentious relationship, Nvidia and Starboard seemed to
work well together in this brief period.
“We were incredibly impressed with Jensen,” said Smith.
For his part, Jensen recalls the meetings with Starboard but doesn’t
particularly remember what was discussed. Before he knew it, Starboard
was no longer an investor. But that wasn’t the end of Starboard’s influence
on the chip industry, and on Nvidia.

A COMPANY CALLED MELLANOX was founded in 1999 by several

Israeli technology executives, led by Eyal Waldman, who became its CEO.
Mellanox provided high-speed networking products for data centers and
supercomputers under the “InfiniBand” standard and soon became an
industry leader. It had impressive revenue growth, going from $500 million
in 2012 to $858 million in 2016. However, its high research and
development spend left it with very thin profit margins.
In January 2017, Starboard bought an 11 percent stake in Mellanox. It
sent a letter criticizing Waldman and his team for their disappointing
performance over the prior five years. Mellanox’s share price had fallen
even though the semiconductor industry index had risen in value by 470
percent. Its operating margins were half of the average of its peer
companies. “Mellanox has been one of the worst performing semiconductor
companies for an extended period of time,” read Starboard’s letter. “The
time for fringe changes and marginal improvements has long passed.”8
After a long series of discussions with the board, Starboard and Mellanox
reached a compromise in June 2018. Mellanox would appoint three
Starboard-approved members to its board and give the hedge fund
additional future rights if Mellanox didn’t meet certain undisclosed
financial targets. Even with those concessions in hand, Starboard retained
the option of waging a proxy fight to replace Waldman. Alternatively,
Mellanox could choose to sell itself to a company that could generate better
returns on its assets than it could as an independent company. The
groundwork was laid for what would be one of the most consequential
transactions in the history of the chip industry.
In September 2018, Mellanox received a nonbinding purchase offer from
an outside company at $102 per share—a premium of almost a third over its
current stock price of $76.90. Mellanox was now fully in play. It solicited
an investment bank to seek other bidders and eventually expanded its list of
potential buyers to seven in total.
Jensen wasn’t thinking about acquiring Mellanox when it became
available, according to another Nvidia executive. But he quickly saw the
strategic importance of the asset, decided Nvidia had to win the auction,
and joined the hunt in October.
Eventually, the list was narrowed down to three serious bidders: Nvidia,
Intel, and Xilinx, which made chips primarily for industrial uses. The three
potential buyers got into a multi-month bidding war, with Intel and Xilinx
topping out around a bid of $122.50 a share. Nvidia went just a little bit
higher, at $125 per share. It won the bidding war on March 7, 2019, for an
all-cash offer of $6.9 billion.
Days later, Nvidia and Mellanox made the deal public and held a
conference call with analysts and investors.
“Let me tell you why this makes sense for Nvidia and why I’m excited
about it,” Jensen said. He talked about how the demand for high-
performance computing would rise—how workloads including AI,
scientific computing, and data analytics required enormous performance
increases, which could only be attained through accelerated computing with
GPUs and better networking. He explained how AI applications would
eventually require tens of thousands of servers connected to one another
and working together in concert, and the market-leading networking
technology from Mellanox would be critical to make that possible.
“Emerging AI and data-analytics workloads demand data-center-scale
optimization,” he said. Jensen was predicting that computing would move
beyond one device—that the entire data center would become the computer.

JENSEN’S VISION CAME TRUE JUST a few years later. In May 2024,
Nvidia disclosed that the portion of the company that was formerly
Mellanox had generated $3.2 billion in quarterly revenue, up more than
seven times from the final quarter in early 2020 in which Mellanox reported
as a public company. After just four years, the former Mellanox business,
which had cost Nvidia a one-time fee of $6.9 billion, was generating more
than $12 billion in annualized revenue and growing at triple-digit rates.
“Mellanox was frankly a wonderful thing thrown in our lap by activists,”
a senior Nvidia executive said. “If you talk to AI start-ups today,
InfiniBand, Mellanox’s networking technology, is incredibly important to
scale the computing power and make everything work.”
Brian Venturo, cofounder and CTO of CoreWeave, a leading GPU cloud-
computing provider and a customer of Nvidia’s, argues that InfiniBand
technology still has the best solution to minimize latency, control network
congestion, and to make workloads perform efficiently.
Mellanox was a happy accident for Nvidia in some respects. Jensen
wasn’t on top of it from the start. But once Nvidia identified and understood
the opportunity, it made the decision to pursue Mellanox aggressively. It
was a great deal, though the outcome depended on Nvidia’s ability to
execute once the new business became part of the company. In those ways,
Mellanox was a typical Nvidia achievement: the company pounced when
others didn’t, and Mellanox helped power Nvidia’s rise to dominance in the
AI space.
“It’s absolutely going to go down in history as one of the best
acquisitions ever,” Nvidia’s head of global field operations, Jay Puri, said.
“Jensen realized that data-center-scale computing requires really good high-
performance networking, and Mellanox was the best in the world at that.”9
After seeing Nvidia achieve all that is has over the past decade, Jeff
Smith of Starboard Value had one summarizing thought, too.
“We never should have exited the position.”
 CHAPTER 13

Lighting the
Future

LIGHT IS A FIENDISHLY COMPLEX Natural Phenomenon. Sometimes

it behaves like a particle; sometimes it behaves like a wave. Sometimes it
bounces off objects, sometimes it scatters through them, and sometimes it is
absorbed completely by them. Unlike, say, the movement of an object
through space or the deformation of an object on impact with another, light
is not governed by a single set of physical principles. Yet we are exposed to
it from the moment we open our eyes; we know intuitively how it “works”
in real life.
Light may thus be the most important visual element in computer
graphics and also the hardest to reproduce. Without good lighting, images
become flat, harsh, or unreal. With good lighting, images can approximate
the work of the Old Masters—conveying emotion and drama even in simple
compositions. It can take a human artist or photographer a lifetime to
control light in his or her work. For years, it looked like computers might
never reach the same level of skill.
Most early computer graphics failed at creating convincing lighting
because the computations were just too difficult for even the most advanced
processors. The best rendering algorithms could only model the physics of
light in simple ways, resulting in flat textures, fuzzy shadows, and unnatural
surface reflections. Even after two decades of steady improvement in most
other areas of graphics—and even after the invention of the GPU, which
made graphics rendering better and more efficient in almost every way—
light remained intractable.
Then came David Luebke. In 1998, Luebke received a PhD in computer
science from the University of North Carolina at Chapel Hill and wanted to
pursue computer graphics as an academic career. He spent eight years as an
assistant professor at the University of Virginia but found himself
increasingly frustrated by the slow pace of his work. Each time his team
invented a new graphics technique for particle rendering or texture mapping
onto objects, it would be obsolete by the time the peer-review process for
the resulting paper had been completed more than six months later. The
reason for the near-immediate obsolescence of Luebke’s work was Nvidia,
which was constantly releasing new GPU features that were superior to
those his team was inventing in the lab. “I was very loose in the socket and
thinking about leaving academia entirely,” he said.1
Then, out of the blue, he got a call from David Kirk, Nvidia’s chief
scientist, who was familiar with Luebke’s work. “We’re starting a long-term
research group at Nvidia,” he said. “Would you be interested?”
Luebke didn’t hold a grudge against Nvidia for continually outpacing his
own work. On the contrary, he realized that he wanted to join the leading
organization in computer graphics—especially if it meant that he would
help define where it went next.
In 2006, he became the first hire at a new division called Nvidia
Research. In his first weeks on the job, Luebke had lunch with Steve
Molnar, a system architect at Nvidia and a long-time friend, and asked him
what he thought a research group at Nvidia should do. For instance, should
it be organized around pursuing patents? Molnar thought about it for a
while and said, “I don’t see Nvidia as some kind of IP fortress. Our strength
is just outrunning the other guy.”
It was a fair observation. Nvidia had remained at the very forefront of
innovation primarily through its operational excellence and strategic
discipline. It had fast release cycles and a clear sense of what its priorities
were—and funding speculative research without a clear commercial end
goal had not been one of them. Nvidia Research almost seemed to be at
odds with the company’s core competencies.
 Yet Kirk had championed the new division precisely because he saw that
the most complex problems in computer graphics would require sustained
research over time, even if commercialization could take a lot more time.
Within a few weeks of starting, Luebke had three new coworkers. At their
first team lunch with Kirk, they asked where they might start. Kirk was
noncommittal: he told them it was up to them to figure out what their job
would be. He did offer some basic guidelines, at least. They should work on
something important to the company. They should create significant impact
with their projects. And they should focus on innovations that would not
occur in the regular course of Nvidia’s business—inventions that would not
be possible without dedicated, long-term work, of the kind that the rest of
the company was not set up to do.

RAY TRACING—A TECHNIQUE THAT SIMULATES the behavior of

light rays as they bounce off or pass through objects in a virtual scene—was
just such a project. In theory, ray tracing would mean far more lifelike
illumination effects than what was currently available on the market. In
practice, it proved to be so demanding that computing hardware could not
handle it.
The conventional wisdom at the time was that CPUs were better than
GPUs at ray tracing because they could perform a wider and more diverse
set of calculations. Intel’s internal research group pushed this notion hard;
they argued that, due to the complexity of light’s behavior in the real world,
only a CPU could accurately model it.
Within six months of the founding of Nvidia Research, the team had
conducted experiments that seemed to indicate not only that GPUs had
become powerful enough to handle ray-tracing calculations, but also that
they could do it faster than the current generation of CPUs. Excited about
the potential to both solve and commercialize a longstanding problem in
computer graphics, Luebke scheduled Nvidia Research’s first meeting with
Jensen.
Normally when Jensen attends a presentation, the speakers get only a few
minutes of uninterrupted time before it becomes a back-and-forth
discussion. In this case, however, Jensen listened for the entire hour-long
presentation. “I think he was being very patient with us to let us have our
say,” Luebke said.
 After Luebke finished, Jensen offered him several pieces of feedback.
Ray tracing had obvious potential in the gaming market. But the CEO
suggested that Luebke and his team shouldn’t ignore other fields. For one,
ray tracing could be useful to promote Nvidia’s Quadro workstation
graphics cards, which were low-volume sellers but accounted for nearly 80
percent of the company’s profit at the time, because of their high price
points. Impressing the professional and technical market might end up
being better for the company.
With Jensen convinced that ray tracing was worth pursuing, Luebke next
went to an Nvidia GPU engineering team design session. Luebke’s team
had several ideas for how to realize the computational capabilities needed
for ray tracing, including making changes to the processors that lay at the
heart of the GPUs themselves. Accustomed to free-wheeling academic
discussions, Luebke and his team assumed that the engineers would be open
to something similar. “We showed up at a Fermi chip architecture meeting,”
he said, referring to a generation of chips that was then under development.
“We just wanted the ability for a bunch of threads to be running next to
each other on the same CUDA core.”
Like Jensen, the Fermi GPU architects accommodated their new
colleagues and their unorthodox, noncorporate behavior. “This is pretty low
cost. I think we can do this,” said Jonah Alben, the head of GPU
engineering—but there was a catch. “You guys need to understand, we need
to make these decisions on data.”
The Nvidia Research team received the message and learned an
important lesson. It was acceptable to think out loud, but to make
significant decisions, the GPU hardware team required evidence to justify
committing time and resources. “You can’t just say it’s obvious, this is a
good idea,” Luebke said.
For the next year, the researchers devoted themselves to providing that
evidence. They worked on proof-of-concept technology and created
algorithms to demonstrate that GPUs could be used cost-effectively for ray
tracing. It was engrossing, exciting work, and not just for the researchers
themselves. Brian Catanzaro, who was an intern at the time, remembered
that Jensen attended a ray-tracing research team meeting in 2008. He asked
no questions. He didn’t bring a computer. He was there just to listen to the
team talk about ray tracing for an hour.
 David Kirk was so convinced by the team’s results that he pushed Nvidia
management to move quickly to get Luebke’s ideas into production. The
first step was acquiring start-ups that had specific expertise in ray tracing.
Nvidia pursued and bought two: Mental Images, based in Berlin, and
RayScale, in Utah. Luebke and Kirk flew out to Utah to show the RayScale
cofounders Pete Shirley and Steve Parker that ray tracing could work much
better on GPUs than the CPUs they were using.
Right after RayScale joined Nvidia, its employees worked alongside the
Nvidia Research team to create a demo for the 2008 SIGGRAPH
conference. This was the same conference where, in 1991, Curtis Priem had
debuted his Aviator flight simulator to the world and with it showed what
was possible with computer graphics. Nvidia attended the conference
frequently, and now, after nearly two decades, it was ready to present the
next evolution in computer graphics. The team presented a GPU-powered
demo of a sleek, shiny sports car driving through a city filled with the kinds
of effects that only ray tracing could produce: reflections off of curved
surfaces, sharp shadows, distorted reflections, and motion blur.
“It was a pivotal moment for the company. That was the start of
something big,” remembered Luebke. “The story that GPUs couldn’t do ray
tracing was completely put to bed with the demo.”
There were several Intel employees in attendance for the demo.
Afterward, they came up to the Nvidia Research team and asked whether it
really was running on a GPU. When Luebke confirmed that it was, he saw
them start tapping furiously away on their Blackberries. Intel’s research
teams never produced another paper about ray tracing on CPUs.
The following year, at SIGGRAPH 2009, Nvidia launched OptiX, a
CUDA-based, fully programmable ray-tracing engine for Quadro cards,
which would accelerate ray tracing for photorealistic rendering, industrial
design, and radiation research. To support the release, Steve Parker and the
former RayScale employees were spun out of the research arm and joined
Nvidia’s main business.
“We always viewed Nvidia Research as an incubator. If something
succeeds, we sort of push it out of the nest and it becomes a product,”
Luebke said.
In just three years, Nvidia Research had transformed from a group that
pursued speculative computing projects into a reliable source of new
business opportunities for the company. Still, there was a long way to go to
make ray tracing accessible to the masses. The demo that Nvidia had shown
at SIGGRAPH 2008 was still beyond the capabilities of consumer-grade
graphics cards. While OptiX enabled engineers to render ray-traced scenes
faster, unless it was a very simple scene the ray tracing couldn’t be done in
real time, as it was too computationally intensive. The company decided to
set aside any thoughts on making progress in ray-tracing applications in
gaming.
Years later, in 2013, David Kirk approached Luebke again. “We need to
revisit ray tracing,” he said. “What would it take to make it the center of
graphics?” He thought it was time for real-time ray tracing in games.
Luebke was so excited by the prospect that he sent an e-mail to all Nvidia
employees on June 10, 2013, which became known as the ray-tracing
moonshot e-mail. “For some time, we’ve been planning this new initiative
around ray tracing,” he wrote. “What could we do if ray tracing were a
hundred times more efficient and what would it take to do it a hundred
times more efficiently?”
Luebke wasn’t exaggerating the scale of the problem. Only such an
efficiency gain would make real-time ray tracing possible for cheaper
consumer graphics cards. Getting to that point would require new
algorithms and the creation of new specialized hardware circuits. It would
also require new perspectives on just what was possible with GPU
technology.
A key contribution came from an Nvidia team located in Helsinki, whom
employees in Santa Clara came to call “the Finns.” Timo Aila, who joined
Nvidia through an acquisition in 2006, was the first employee in Helsinki.
Over time, Aila and his colleagues became a kind of internal strike team,
assigned to the toughest research questions facing Nvidia. Now, they took
on the challenge of research into a new specialized ray-tracing processor
core inside GPUs. They were supported by early Nvidia employee and chip
architect Erik Lindholm, who flew out to Finland.
“The Finns are this crack research team where just everything they touch
turns to gold,” Luebke said.
After Nvidia Research presented to the GPU architecture team and
gained its support, engineers in the United States were assigned to work
with the Finns on the ray-tracing cores in March 2014. In 2015, the Finns
traveled to Nvidia headquarters to hash out the remaining issues. By 2016,
the project was nearly completed, and Nvidia Research fully turned it over
to the company’s engineering team. Although the ray-tracing technology
was too late to make the launch of the Pascal architecture, which was rolled
out later that year, Nvidia prepared to release dedicated ray-tracing cores
with the next architecture, which would be called Turing.
“My job throughout all this is protecting this effort, making sure they got
the care, feeding, and attention they needed,” Luebke said, referring to the
Finns.
Jensen would introduce Turing, with its dedicated ray-tracing cores, as
part of his keynote address for SIGGRAPH 2018, exactly ten years after
Nvidia Research’s demo proved in a stroke that ray tracing belonged on
GPUs rather than CPUs. Most of his speech was devoted to introducing the
Turing architecture, as well as the improved second-generation Tensor core
designed to accelerate “deep-learning” neural net workloads. But Jensen
wasn’t satisfied. He wanted additional material in his speech to captivate
the conference audience.

TWO WEEKS BEFORE THE SHOW, he invited Nvidia executives to

pitch ideas for his keynote. Nvidia Research’s Aaron Lefohn suggested that
he demonstrate the new deep-learning anti-aliasing, or DLAA, feature.
Powered by Turing’s Tensor cores, DLAA used artificial intelligence to
enhance image quality, making high-resolution graphics crisp and objects
appear sharply defined. Jensen wasn’t impressed. He wanted something
more exciting. “A better-looking picture is not going to sell many GPUs.”
But he found inspiration in the suggestion. Instead of deep-learning anti-
aliasing, which improved already great images, what if they could use
Tensor cores to make lower-end cards perform as well as the top of the line?
For example, Nvidia could use the image-enhancement function to sample
and interpolate additional pixels, so that a card designed to render graphics
natively at 1,440p resolution, also known as “Quad HD,” could produce
images at the higher-resolution 4K, “Ultra HD,” at a similar frame rate. AI
would be used to fill in details to take the lower-resolution 1440p image to a
higher-resolution 4K image.
“What would really help,” said Jensen, “is if you could do deep-learning
super sampling. That would be a big deal. Can you do that?”
Lefohn huddled with his team and then told Jensen it could be possible.
They needed to research the idea. A week later, just days before the
keynote, Lefohn reported back to Jensen that the early results were
promising and they would be able to make what came to be known as
DLSS. “Put them on the slides,” Jensen said.
“No one in the world had ever thought of making a system and a
machine-learning model that can infer hundreds of millions of pixels per
second on a home computer,” Brian Catanzaro said.2
Jensen had come up with DLSS on the spot. He had seen the promise
inherent in one technology and transformed that promise into a new feature
with a better business case. Now, if DLSS worked, the company’s entire
product lineup, from the low end to the high end, would become more
proficient, and thus valuable, allowing Nvidia to charge higher prices. “The
researchers had invented this amazing thing, but Jensen saw what it was
good for. It wasn’t what they had thought,” Luebke said. “It shows what a
leader Jensen is and how technical and smart he is.”
Jensen’s keynote address was well received, but the GeForce RTX cards
based on the Turing GPU were not. “We launched ray tracing and DLSS to
a thud,” Jeff Fisher said. The problem was the GeForce RTX offered
negligible gains in frame-rate performance over the previous-generation
Pascal cards. And when gamers turned on ray tracing, which was supposed
to be the killer new feature, the RTX cards suffered a 25 percent drop in
frame rate.
DLSS performed marginally better. When enabled, it allowed cards to
run about 40 percent faster than Pascal, but at a noticeable loss of image
quality. Nvidia also had to fine tune and train DLSS’s AI on footage of each
game they wanted the technology to work with, which was a painstaking
and time-consuming process. Still, by now Nvidia had learned the value of
developing and iterating technology over time and of waiting for the market
demand to catch up. “You solve the chicken-and-egg problem by
bootstrapping,” said Bryan Catanzaro. “You can’t have amazing AI in
hundreds of millions of households without building it first. Both ray
tracing and AI were going to change gaming forever. We knew that this was
inevitable.”
Catanzaro joined the DLSS project after the launch of Turing in 2018. He
worked on DLSS 2.0, which was introduced in March 2020 and didn’t need
to be tuned for each game. It received much better reviews. “We
reconceived the problem and got better results without requiring custom
training data from every game,” Catanzaro said.
The next iteration was even better. Catanzaro left Nvidia for a brief stint
at the Chinese search engine and technology company Baidu but returned to
work on what became DLSS 3.0. The goal was to use deep learning to
create interstitial, AI-generated frames between rendered frames for games.
The thinking was that in every successive frame in a video game there were
patterns and correlations, and if the AI chip could predict these patterns and
correlations, it could take some of the rendering computational load off of
the GPU.
It had taken six years of development to build a sufficiently accurate AI
model for the frame-generation feature, according to Catanzaro. “While we
were working on it, we saw continuous improvement in the quality of
results, so we kept working,” he said. “Most academics don’t have the
freedom to work on one project for six years because they need to
graduate.”
The development of DLSS and real-time ray tracing reveal how Nvidia
came to approach innovation. While it would roll out new chips and boards
on a very fast schedule, it would now, with Nvidia Research and other
groups, pursue “moonshots” at the same time. “When we got to the next-
generation Ampere, we had enough momentum for ray tracing and DLSS to
make that product a home run,” Jeff Fisher said.
It was a further, institutionalized form of protection against the kind of
stagnation that Clayton Christensen warned about in The Innovator’s
Dilemma: the inevitable desire to focus on the company’s core, profit-
generating business at the expense of investing in more exploratory
innovations that might not be commercially viable for years.
According to Jon Peddie Research, Nvidia’s share of the discrete or add-
in board GPU market has stayed at roughly 80 percent over the past decade,
as of this writing. Even though AMD offers better price-to-performance
based on traditional metrics, gamers keep choosing Nvidia for its ability to
innovate. Both ray tracing and DLSS have become must-have features that
developers have incorporated into hundreds of games. And the features
perform the best on Nvidia graphics cards, making it difficult for AMD to
compete effectively.
In the case of ray tracing, the journey from the inception to integration
into GPUs had spanned a decade. Similarly, building successive iterations
of DLSS, such as frame generation, took six years. “It requires vision and
long-term persistence. It requires investment even when the results aren’t
totally clear,” Catanzaro said.
Ultimately, Nvidia Research showed how Jensen’s strategic vision has
changed over time. In the beginning, when the company was in survival
mode, he wanted everyone to focus on concrete projects: delivering the next
generation of chips at the “Speed of Light,” selling the “whole cow,” and
beating competitors through sheer execution. As Nvidia got bigger, Jensen
realized that survival now meant future-proofing the company in as many
ways as possible. Continuous innovation would require a more flexible
approach to Nvidia’s operations, even if that meant pursuing some bets that
a younger Jensen might have dismissed.
This new, more mature Jensen was no longer afraid of making a single
wrong move, not least because the company now had some financial
cushion. “You can’t innovate if you’re not willing to take some chances and
embarrass yourself,” he said.3 “We don’t have an ROI timeline. If you don’t
have an ROI timeline and you don’t have a profitability target, those aren’t
things we’re optimizing for. The only thing we’re optimizing for is this: Is it
incredibly cool, and are people going to like it?”
One former senior industry executive argues that Nvidia sets itself apart
from its rivals in its willingness to experiment and invest over long periods,
successfully monetizing its more open-ended efforts. This contrasts with
larger technology giants such as Google, which often spend heavily on
researching new technologies but have little to show for it commercially.
Notably, all eight Google scientists who authored the seminal “Attention Is
All You Need” paper on the Transformer deep-learning architecture—which
proved foundational for advancements in modern AI large language models
(LLMs), including the launch of ChatGPT—soon after left Google to
pursue AI entrepreneurship elsewhere. “It’s just a side effect of being a big
company,” said Llion Jones, one of the coauthors of the Transformer paper.4
“I think the bureaucracy [at Google] had built to the point where I just felt
like I couldn’t get anything done,” he added, expressing frustration with his
inability to access resources and data.
Nvidia’s second decade began with successful R&D investments in
programmable shaders and then progressed to the industry-shaping
innovation that is CUDA. Next came Nvidia Research’s breakthroughs in
ray tracing, DLSS, and AI, all of which proved critical to the company’s
future. The team is now three hundred researchers strong, led by Chief
Scientist Bill Dally. Nvidia appeared not only to have solved the innovator’s
dilemma but also to have completely overcome it.
 CHAPTER 14

The Big Bang

PROFESSIONAL TRADERS ARE A DYING BREED. Computers,

which are faster and generally more effective at placing bets in response to
releases of corporate financial results and economic data, have decimated
the ranks of human traders over the past two decades.
Connors Manguino is one of the few thousand or so humans who still
make a living by trading on news headlines and earnings announcements.
Armed with decades of experience and a Bloomberg terminal, he still goes
against the algorithms every quarter. He’s good enough at it to stay alive
and make a living.
He needs to react quickly. A sub-second delay in pressing the buy or sell
key could mean the difference between a good entry point and a devastating
loss. The running joke among his friends is that Manguino has an inhuman
ability to not blink during important news periods.
On Wednesday, May 24, 2023, he was waiting for Nvidia’s earnings
report, which was scheduled to arrive after the market close. It was one of
the most widely anticipated reports in years, and as the minutes ticked down
toward the end of the trading day, he found himself staring intently at his
terminal.
OpenAI’s release of ChatGPT in late 2022 had generated enormous
media coverage. The chatbot captivated the public with its ability to create
poems, food recipes, and song lyrics on demand. ChatGPT became the
fastest-growing consumer app in history, surpassing 100 million monthly
active users in just two months. All of a sudden, companies were now trying
to use the purported benefits of AI—its speed, its computational power, and
above all its ability to process and generate natural-sounding language—
any way they could.
Manguino knew that Nvidia was in prime position to take advantage of
the AI boom. The question was, how big of a boom would it be, and how
much of an effect would the boom have on Nvidia? The company’s GPUs
were well known in the academic world, thanks largely to David Kirk’s
efforts to build relationships with top universities. Jensen, for his part, had
spent the prior decade working to transform Nvidia’s reputation from a
graphics company to an AI company. He had achieved some success, with
Meta and TikTok using Nvidia’s GPUs to make their algorithms more
effective in recommending videos and ads. But AI was not a massive driver
of Nvidia’s earnings. During the company’s fiscal year 2023, which ended
in January 2023, data-center revenue, which included AI GPUs, accounted
for about 55 percent of overall sales. But that figure was primarily due to a
25 percent decline elsewhere, in the company’s gaming-card revenue,
following a postpandemic slowdown in demand for games overall.
Then everything changed, twenty-one minutes after the stock market’s
close at 4:00 p.m. Manguino saw the headline flash on his terminal screen.

NVIDIA SEES 2Q REV. $11.00B PLUS OR MINUS 2%, EST.

$7.18B

The financial shorthand was, to a seasoned trader, nothing short of

extraordinary. Nvidia had crushed Wall Street estimates for its second-
quarter revenue outlook—by roughly $4 billion. As he read the earnings
print and guidance, Manguino froze. “$4 billion? How could that be real?”
he thought to himself. “Holy shit. What a raise!”
By the time he came to his senses, it was too late to take advantage of the
gap between the earnings release and the market’s reaction. Nvidia shares
had already spiked by a double-digit percentage in after-hours trading. As a
consolation prize, Manguino bought shares in Advanced Micro Devices,
Nvidia’s main GPU competitor, hoping that the run on Nvidia stock might
also boost its competitors. In this instance, the algorithms had won; unlike
him, they did not hesitate in reacting to an earnings report that was better
than any he had ever seen.
Other Wall Street analysts had similar reactions. Bernstein’s Stacy
Rasgon titled his note: “The Big Bang.”
“In the 15+ years we have been doing this job, we have never seen a
guide like the one Nvidia just put up,” he wrote, adding that the company’s
outlook “was, by all accounts, cosmological.” Morgan Stanley analyst
Joseph Moore reported that “Nvidia guides to the largest dollar revenue
upside in industry history.” And former star Fidelity fund manager Gavin
Baker, who now has his own technology hedge fund with several billions of
dollars under management, compared Nvidia’s forward-looking guidance to
other seminal financial reports in the history of the tech industry. He was
there for Google’s first blockbuster report after its 2004 IPO, during which
it reported a doubling of revenue and profit after just a single quarter as a
public company.1 He was there for Facebook’s second-quarter 2013
earnings, when the company demonstrated for the first time that it could
successfully transition its ad business to mobile, beating Wall Street’s
revenue expectations by $200 million.2 Nvidia’s guidance was better than
either. “I have never seen such a large beat at this scale,” he said.
The following day, Nvidia shares soared by 24 percent and added $184
billion in market value—more than the entire value of Intel, and one of the
largest single-day gains ever for a U.S. public company.
Jensen capitalized on the attention, pressing his advantage the following
week when he gave the keynote address at the Computex technology
conference in Taiwan. In his speech, he announced Nvidia’s new DGX
GH200 AI supercomputer, which incorporated 256 GPUs in one system,
thirty-two times the number of GPUs that were in the previous model. It
meant significantly more computing power for generative artificial
intelligence applications, allowing developers to build better language
models for AI chatbots, create more complex recommendation algorithms,
and develop more effective fraud-detection and data-analysis tools.
But his main message was so simple that even a nontechnical listener
could understand it. Nvidia offered far more computing power at a lower
cost per GPU. He hammered this home over the course of his remarks,
punctuating his readouts of technical specs with a refrain: “The more you
buy, the more you save.”
More broadly, Nvidia’s salespeople had been able to stoke unprecedented
demand by arguing to customers that they had to invest aggressively in
generative AI or face the existential threat of falling behind their
competitors. Jensen himself has called AI a “universal function
approximator” that can predict the future with reasonable accuracy. This
applies as much in “high-tech” fields such as computer vision, speech
recognition, and recommendations systems as it does in “low-tech” tasks
such as correcting grammar or analyzing financial data. He believes that
eventually it will apply to “almost anything that has structure.”
The best way to access this universal function approximator, of course,
was through Nvidia technology. And over the next four quarters, the
company achieved one of the most incredible revenue ramps in technology
history. Its first-quarter data-center business for fiscal 2024 rose by 427
percent from the prior year, to $22.6 billion—driven primarily by artificial-
intelligence chip demand. Unlike software, which is easy to scale at
essentially no incremental cost, Nvidia is producing and shipping complex
high-end AI products and systems, some of which contain up to 35,000
parts. There was no precedent for this level of hardware growth at a
technology company the size of Nvidia.
To those outside of the company, Nvidia’s meteoric rise seems like a
miracle. Those inside it, however, consider it a natural evolution, said Jeff
Fisher. Nvidia wasn’t lucky; it was able to perceive the wave of demand on
the horizon years in advance and had prepared for this very moment. It
went to its manufacturing partners—Foxconn, Wistron, TSMC, among
others—to help them build out production capacity. Nvidia sent out so-
called tiger teams to those partners, doing whatever they could to help them
become more efficient: the teams bought equipment, added factory space,
automated testing, and sourced advanced chip packaging.
In keeping with Jensen’s “rough justice” model, Nvidia was not doing all
this just to make its partners more efficient at their current processes. It
wanted to produce new chip designs more quickly, transitioning from its
previous two-year product cycle to a one-year cadence for its AI chips. In
the 1990s, Nvidia shifted to a faster product cadence by releasing a new
graphics card every six months. Now, it wanted to do the same for AI chips.
“The bigger AI gets, the more solutions that will be needed, and the faster
we will meet those goals and expectations,” said Nvidia CFO Colette
Kress.3
Typically, hardware-production plants have average cycle times of
fourteen to eighteen weeks between phases of the manufacturing process.
Manufacturers build in buffer time between those phases, in case an
upstream problem creates issues down the line. This can leave machines,
materials, and components idle for days. Nvidia’s teams figured out how to
add quality controls early in the process to reduce the risk of unforeseen
problems and removing the need for buffer time. According to Jeff Fisher,
Nvidia’s approach involves “no magic.” It’s just hard work and ruthless
efficiency, all in the service of maintaining competitive advantage. And
everyone who works with Nvidia must embrace it, not just its internal
teams.4 Everything the tiger teams did was expensive and resulted in a drag
on the bottom line. Yet Nvidia has always been willing to use its financial
resources to invest in critical parts of the business—even when that has
meant other companies’ business.
Nvidia has key advantages over other AI chip makers. Similar to Apple’s
approach with the iPhone, the company employs a “full-stack” model that
optimizes the customer’s experience across hardware, software, and
networking. Most of its rivals just make chips. And Nvidia moves faster
than its competitors.
For instance, the core architecture used in modern large language models
is the Transformer introduced in the 2017 paper “Attention Is All You
Need” by Google scientists. The primary innovation is self-attention, which
enables the model to measure the importance of different words in a
sentence and measure long-range dependencies on the basis of their context.
The attention mechanism empowers the model to focus on the more
important information, to train the AI model quicker, and thus to generate
higher-quality results compared with prior deep-learning architectures.
Jensen grasped the need to add support for Transformers in Nvidia’s AI
offerings almost right away. Simona Jankowski, a former finance executive
at Nvidia, remembers Jensen getting into a fairly detailed discussion about
Transformers on a quarterly earnings call just months after the Google
scientists’ paper came out.5 He instructed his GPU software teams to write a
special library for Nvidia Tensor Cores that optimizes them for use with
Transformer operations; the library was later called the Transformer
Engine.6 It was included for the first time in the Hopper chip architecture,
which went into development in the late 2010s and was released in 2022,
one month before ChatGPT was launched. According to Nvidia’s own
testing, GPUs with Transformer Engine could train even the largest models,
in a matter of days or even hours, whereas without Transformer Engine
those same training runs might take weeks or months.
“The Transformer was a big deal,” Jensen said in 2023. “The ability for
you to learn patterns and relationships from spatial as well as sequential
data must be an architecture that’s very effective, right? And so I think on
its first principles, you can kind of think Transformer’s going to be a big,
big deal. Not only that, you could train it in parallel and you can really scale
this model up.”7
When demand for generative AI exploded in 2023, Nvidia was the only
hardware manufacturer ready to fully support it. And it was ready only
because it was able to spot the early signals, productize them in the form of
hardware and software acceleration features, and insert those features into a
line of chips that was only months away from arriving on the open market.
The breathtaking speed Nvidia demonstrated was a sign that it will be
difficult to dethrone, even though several other large technology companies,
including Microsoft, Amazon, Google, Intel, and Advanced Micro Devices,
are developing their own AI chips. Nvidia has proved, as it enters its fourth
decade, that it can still outrun the competition.
Its second, but lesser-known, advantage is pricing power. Nvidia doesn’t
believe in building commodity products, because commodities are subject
to downward pricing pressure as competition increases. Instead, from the
very beginning, its pricing has only gone in the opposite direction: up.
“Jensen has always said that we should be doing things that other people
cannot. We need to bring unique value to the marketplace, and he feels that
by doing work that is cutting edge and revolutionary, it allows the company
to attract good people,” Nvidia executive Jay Puri said. “We don’t have the
culture of just going after market share. We would rather create the
market.”8
A former Nvidia executive recalls how Jensen would become upset if any
other company tried to negotiate pricing with him. Potential customers
would always want to meet him when contract discussions were nearing an
end. “We always try to do our best to prepare the customers,” the former
Nvidia executive said. “Do not discuss price. We are here to close the
deal.”9
Jensen has instilled this mentality across the company. Michael Hara, a
former director of marketing, remembers debating with Jensen how to price
Nvidia’s earliest products. When Hara left S3 to join Nvidia, he was used to
a commodity-like pricing strategy; at the time, S3’s market-leading 3-D
graphics chip sold for $5 (about $11 today). When the RIVA 128 came out
in 1997, Hara was worried if they priced it too high, the buyers would balk.
Ten dollars at most, he replied. Jensen said, “No, I think that’s too cheap.
Let’s go with $15.” The card sold out at that price. The derivative RIVA
128ZX chip that came out the following year was priced at $32. And the
next-generation GeForce 256, which came out in 1999, cost $65.
Jensen understood that gamers who buy Nvidia cards are willing to pay
for performance. “As long as they look at the screen and can see something
radically different than they saw before, they’re going to buy it,” he said. It
was a lesson that has stuck with Hara ever since. When he moved from
marketing to investor relations, he made the same case to Nvidia’s investors
—that Nvidia would be a unique semiconductor company where product
ASPs (average selling prices) rise. “We’re going to be the only guy where
ASPs go up over time, when everybody else’s ASPs will go down,” he said.
The reason is that computation for 3-D graphics is an infinitely complex
problem to solve and thus drives a competition to make better and better
hardware. Hardware will never be powerful enough to perfectly reflect
reality. Even so, when you purchase the latest 3-D graphics card, you can
clearly see an improvement in performance over the past generation—
lighting looks better, textures look more realistic, objects move more
fluidly.
A similar dynamic problem is playing out right now with deep learning
and artificial intelligence. Nvidia’s current-generation hardware has enabled
AI models to grow exponentially in size and ability in only a few years. Yet
the demand for AI computational power is growing even faster, because the
problems that AI can solve get ever more complex. There are step-changes
in between generations of AI models, because the underlying hardware and
software have also improved in tandem with the models. Still, the promise
of true general artificial intelligence remains far off: there is plenty more
work to be done. By remaining on the cutting edge of technology—and by
cleverly positioning itself in highly visible fields where performance
increases are immediately obvious—Nvidia is able to increase its pricing
power and its ASPs.
Today, Nvidia graphics cards cost more than $2,000 apiece. And those
are the consumer-grade prices. In the past decade, the company has started
offering AI-server systems equipped with eight GPUs, with each system
costing hundreds of thousands of dollars. Ross Walker, who went head-to-
head against Nvidia over his use of the cheaper GeForce line to accelerate
his molecular dynamics software AMBER (as we saw in chapter 8),
remembered that at the time, a top-of-the-line Nvidia GPU server cost as
much as a small used car such as a Honda Civic. Now, a similar server
might cost as much as a house.
“I was in the audience when Nvidia announced the DGX-1 for
$149,000,” he said, referring to the first GPU server optimized with Tensor
Cores and Transformer Engine for AI research. “There were audible gasps
in the audience. I couldn’t believe it.”10
And that’s not even close to the most expensive Nvidia product. Nvidia’s
latest server rack system as of this writing, the Blackwell GB200 series,
was specifically designed to train “trillion-parameter” AI models. It comes
with seventy-two GPUs and costs $2 million to $3 million—the most
expensive Nvidia machine ever made. The company’s top-end-product
pricing isn’t merely increasing; it is accelerating.

JENSEN DOES NOT POSSESS SPECIAL visionary powers that

allowed him to predict exactly when AI would take off. One could argue, in
fact, that the company was initially measured in its approach; Nvidia did
not allocate many people or resources toward AI development until he saw
significant signals indicating what could be possible. Then, he moved with
a speed and totality of purpose unmatched by the competition.
Early enough, however, Jensen knew the end game. Consider what Reed
Hastings achieved with Netflix, which he cofounded. Hastings knew that
someday the world would move to streaming video over the internet.
Although he did not know exactly when this would happen, he had the
intuitive sense that it would become the ultimate solution. As CEO, he
managed the DVD-by-mail business only until the technology advanced
enough to make streaming possible, and he made the transition forcefully
when the time came.
 Jensen did something similar with AI and, before that, with video games.
In the early 1990s, he was convinced video games were going to be an
enormous market. “We grew up in the video-game generation,” he said.11
“The entertainment value of video games and computer games was very
obvious to me.” He believed the PC game market would explode soon
enough, within five or ten years or fifteen years—which it did, when
GLQuake was released in 1997.
Jensen is always trying to figure out the next thing and what Nvidia can
do to prepare itself to take advantage of it. In early 2023, he was asked by a
student to predict what will follow, and build upon, AI. “There’s no
question,” he said. “Digital biology is going to be it.”12
Though biology is one of the most complex systems, Jensen explained
that for the first time in history, it could be digitally engineered. With AI
models, scientists can now start to model the structure of biological systems
in greater depth than ever before. They can learn how proteins interact with
each other and with their environments, and use the vast computing power
unlocked by advanced computing to perform computer-aided drug research
and discovery. “I’m very proud to say that Nvidia is at the center of all that.
We made it possible for some of the breakthroughs to happen,” he said. “It’s
going to be profound.”
Jensen sees parallels today between digital biology and almost every
major milestone in Nvidia’s history. When he cofounded the company,
computer-aided semiconductor design was just starting to become feasible.
“It was the combination of algorithms, computers that are fast enough, and
know-how,” he said.13 When those three things reached a certain stage of
development, the semiconductor industry could create larger, more complex
chips, because engineers could now design and simulate chips using higher-
level abstractions in software without having to physically lay out every
signal transistor. The same combination of factors saw Nvidia invent the
GPU in the early 2000s and take over the AI space in the late 2010s—the
“fuel-air mixture” that Bill Dally spoke of.
Nvidia’s vice president of health care, Kimberly Powell, has said that
computer-aided drug discovery will do for drug design what computer-
aided design and electronic design automation did for chip design.
Companies will become more consistent and efficient in finding drugs to
treat diseases and even personalize them for individual people. It will “go
beyond discovery and evolve into design, helping create the conditions to
no longer be a hit-or-miss industry,” she said.14
Generate:Biomedicines is one of the start-ups using AI and Nvidia GPUs
to develop new molecular structures and protein-based drugs that do not
form from natural processes. The biotech company has studied millions of
proteins by using machine-learning algorithms to arrive at a more in-depth
picture of how nature functions, a picture it then uses to create new drugs.
Gevorg Grigoryan, the company’s cofounder and chief technology officer,
was previously a professor at Dartmouth College, where he studied
statistical patterns of proteins and tried to better design and model proteins
by using computing power.
“Using very simple statistics, I was seeing that the patterns in the data
were generalizable. We were finding principles that would go beyond the
dataset,” he said. “It was very clear the next step was to use AI, machine
learning, and large-scale data generation.”15 Grigoryan couldn’t do that in
academia because purchasing the required computing power would be
beyond his institution’s reach. He saw the commercial potential of a new
way of designing molecules, and soon enough, Generate:Biomedicines was
born.
Starting in the early 2000s, Grigoryan observed that many scientists who
ran molecular-dynamics simulations were buying Nvidia gaming GPUs and
coercing them into performing nongraphics computations. He appreciated
how the company catered to and collaborated with the research community,
even though the cards were supposed to be used for video games. “That was
really the beginning of this beautiful marriage between Nvidia and
molecular science,” he said.
When he started to use machine learning himself, it was natural for him
to rely on PyTorch, a free, open-source machine-learning library created by
Meta in 2016, now under the purview of the Linux Foundation. “PyTorch
was something that was very well developed, had a huge community, and
had tons of support from Nvidia,” Grigoryan said. “It was not even a choice
of what type of GPU we’d go with. PyTorch works well with CUDA,
CUDA is Nvidia’s creation. By default, we always used Nvidia hardware
without even thinking too much about it.”
Structural prediction and protein design, once considered impossible
problems, are now solvable. Grigoryan explains that the complexity of a
protein and its possible states surpasses the number of atoms in the
universe. “Those numbers are extremely challenging for any computational
tools to deal with,” he said. But he believes a skilled protein biophysicist
can examine a particular molecular structure and deduce its potential
functions, suggesting there may be learnable general principles in nature—
exactly the sort of operation that a “universal prediction engine” such as AI
should be able to figure out.
Generate:Biomedicines has applied AI to examine and map molecules at
the cell level, and Grigoryan sees the potential to extend the same technique
to the entire human body. Simulating how the human body will react is
orders of magnitude more complicated, but Grigoryan thinks it will be
possible. “Once you see it working, it’s hard to imagine it doesn’t just
continue,” he said, referring to the power of AI.
While it may sound like science fiction, Grigoryan and his team are
already building generative models that optimize molecule functions inside
cells. The ultimate dream is to make drug discovery a software question,
where an AI model can take a disease, including a type of cancer, as input
and generate a molecule that cures it. “It’s not totally crazy. I think it may
even be within our lifetime that we’ll be able to see that kind of impact,” he
said. “Science always surprises us, but man, what a time to be alive, right?”

THERE IS AN ENORMOUS REPOSITORY of data inside corporations

that remains untouched and unstructured by AI: e-mails, memos,
proprietary internal documents, and presentations. Because the consumer
internet has already been mined to near-exhaustion by chatbots such as
ChatGPT, the next significant opportunity lies within enterprises, where
customized AI models can enable employees to access knowledge currently
siloed across a company.
Jensen has said AI will completely change how employees interact and
work with information. Traditional IT systems have relied on a static file-
retrieval system, requiring explicitly written technical searches pointed at a
specific storage device. These requests often do not work because of the
fragile and brittle nature of the query format.
Current AI models can now understand requests via context and because
they can grasp natural conversational language. It is a major breakthrough.
“The core of generative AI is the ability for software to understand the
meaning of data,” Jensen said.16 He believes that companies will
“vectorize” their databases, indexing and capturing representations of
information and connecting it to a large language model, enabling users to
“talk to their data.”
This use case makes obvious sense to me. My first job after college was
in management consulting. The worst part of the role was manually sifting
through file directories on servers, searching PowerPoint or Word
documents for specific pieces of information a partner asked for from years
past. Sometimes it would take hours or even days to find the document.
Large language models powered by AI applications such as Nvidia’s
ChatRTX now allow users to receive contextually relevant answers
instantly from private files on computers. It increases productivity
dramatically. What was a tedious, repetitive task that took significant time
now takes seconds and gives employees more space for more critical, high-
level work. Employees will have a virtual assistant, almost like a brilliant
intern with near-perfect memory, capable of instantly recalling any piece of
knowledge stored on computers and the internet. Instead of simple file
retrieval, the models can generate smarter insights drawn from the entire
pool of a company’s internal data.
In a late 2023 report, Goldman Sachs predicted generative AI-driven cost
reductions could total more than $3 trillion over the next decade across
industries. Nvidia management has repeatedly stated that the $1 trillion that
has been invested in global data-center computer infrastructure over the
years, which is currently powered by traditional CPU servers, will
eventually transition to GPUs capable of the parallel computations
necessary for AI. That transition represents a gold mine for Nvidia. In mid-
2024, J.P. Morgan published survey results from 166 chief information
officers, who are responsible for $123 billion in annual enterprise tech
spending. The report revealed that CIOs plan to increase their AI compute
hardware spending by more than 40 percent annually over the next three
years, going from 5 percent of total IT budgets to more than 14 percent in
2027. One-third of CIOs also said they will defund other IT projects in
order to support the new AI investments. The three biggest categories slated
for defunding included legacy system upgrades, infrastructure, and internal
app development.
Jensen believes that increased expenditure on AI will benefit more than
just executives and investors. “I believe that artificial intelligence is the
technology industry’s single greatest contribution to social elevation, to lift
all of the people that have historically been left behind,” Jensen said at an
event at Oregon State University in 2024.17 He does not often venture into
social commentary, but Nvidia’s size and prominence now almost require
him to offer such opinions.
The only thing that might hinder Nvidia are the so-called AI scaling laws.
There are three components of these laws: model size, computing power,
and data. Large technology companies and start-ups are confident that the
capabilities of AI models will continue to improve in the near term and are
aggressively increasing their AI infrastructure spending into 2025. Yet as
companies keep increasing model size, adding more computing power from
Nvidia GPUs, and incorporating larger datasets, they will eventually
encounter diminishing returns. This would lead to an air pocket in demand
for Nvidia, as the majority of its data-center revenue is related to model
training. In early 2024, Nvidia said about 60 percent of its data-center
GPUs were sold to train AI models, while the other 40 percent of its data-
center GPUs were purchased for inferencing, or the process of generating
answers from AI models.
No one knows when this AI slowdown will happen, whether it is in 2026,
2028, or more than five years from now. But history shows Nvidia will be
prepared for the challenge. It will also be ready to adapt to the next big
computing trend, regardless of what it may be.
 CONCLUSION

The Nvidia Way

EVEN AFTER THIRTY-ONE YEARS leading Nvidia, Jensen Huang

refuses to work in a private office. Instead, he stakes out the Metropolis
conference room in Nvidia’s Endeavor headquarters building, where he
hosts group meetings throughout the day. For smaller meetings, he’ll move
to a five-person room called Mind Meld, a reference to the ability of the
Vulcans in Star Trek to telepathically combine their thoughts with those of
other beings. It’s an apt metaphor, if a little on the nose, for how Jensen has
designed Nvidia—as an extension of his own formidable intellect.
Jensen is a technical founder and CEO, which is part of Nvidia’s
advantage over some of its competitors. But to call him a mere technologist
is to undersell his skill at hiring and developing people who are good fits
for Nvidia’s particular culture. He gives his employees a high degree of
independence over their individual projects, but only if they can keep those
projects perfectly aligned with the company’s core objectives. To reduce
ambiguity, Jensen spends a great deal of his time communicating with his
employees and ensures that everyone at the company knows the overall
strategy and vision. He offers a level of visibility most companies don’t
share outside of the C-suite.
A former senior executive at a large software company said that he was
always struck by how you could talk to multiple Nvidia employees and they
would never contradict one another. The message from the top was
consistent, and Nvidia staff learned it and made it their own. He drew a
contrast with almost every other company he’d ever worked with, whose
representatives sometimes argued with each other in front of external
clients.
“Ultimately, my e-staff is something that I have to know how to work
with. The company’s organization is like a race car. It has to be a machine
that the CEO knows how to drive,” Jensen said.
Hiring raw talent is the first essential component of the Nvidia Way. Y
Combinator cofounder Paul Graham, who once worked for Yahoo, noticed
that once Yahoo started losing the war for the best engineers to Google and
Microsoft, the company began slouching toward mediocrity. “Good
programmers want to work with other good programmers. So once the
quality of programmers at your company starts to drop, you enter a death
spiral from which there is no recovery,” he wrote. “In technology, once you
have bad programmers, you’re doomed.”1
Much of the time, that talent finds Nvidia first. Or Nvidia proactively
finds the best people: more than a third of new hires are referred by current
employees.2
When Nvidia sees an opportunity to poach talent from its rivals, it moves
with aggression and speed. Hock Leow, former chief technology officer of
Creative Labs, witnessed an Nvidia approach first-hand. In 2002, Creative
Labs acquired a company called 3Dlabs, which had an office for graphics-
chip engineers in Huntsville, Alabama. Three years later, Creative
announced that it was going to shut down 3Dlabs and the Huntsville
location altogether.
Intel moved quickly, more quickly than Nvidia at first, attempting to lure
away the former 3Dlabs employees in Huntsville. But the offers it produced
were contingent on relocation to another of Intel’s sites, all of them located
far away from Alabama. Many of the workers were reluctant to uproot their
families or to move to a place with a higher cost of living.
Jensen, upon learning of Intel’s interest, promptly sent his executives to
make offers to the 3Dlabs team that did not include a relocation demand. In
fact, he instructed the executives to open a new office in Huntsville to
accommodate the new team members. “Nvidia moves very fast,” Leow
said. “They aggressively accumulate human and technology assets to win.
Speed of execution and decision-making is an Nvidia trademark.” Nvidia
maintains an office in Huntsville to this day.
 Former Nvidia executive Ben de Waal recounted a similar experience. In
2005, he and his boss, head of software engineering Dwight Diercks,
traveled to Pune, India, to evaluate a potential acquisition: a roughly fifty-
person video-encoder-software company. Upon arriving, they discovered
the owners had gathered the employees in a hotel ballroom to announce the
company’s dissolution. The company had tax problems and was in financial
trouble. “It was rough and emotional. People were crying. They put their
hearts and souls into this company,” de Waal said. “I wondered why we
were even there.”3
Diercks knew that to return to California with nothing would be a missed
opportunity. Nvidia needed larger software teams for new projects, and
these workers were excellent. He had been to India nine times that year on
scouting trips and had identified this specific company as the best available
prospect.
He had an idea: why not hire the employees directly instead of acquiring
the company? He pitched it to Jensen, who approved it right away. “We
changed our trip from an acquisition mode to a hiring mode,” Diercks said.
“We spent all night in the run-down hotel business center printing out about
fifty offer packages, which in India is more complex than a U.S. standard
package.”4
By the end of the first day, fifty-one out of fifty-four employees had
accepted Nvidia’s offers. They became the core of a new Nvidia office in
Pune, which would eventually grow into an essential engineering operation
with more than fourteen hundred employees.
“You always need the best people,” Diercks said, adding that Nvidia saw
hiring talent in bunches as a strategy.
Occasionally, Nvidia will go for the most direct approach possible. Its
executives have no qualms telling top technical architects at other
companies that they are going to lose, so they might as well join the winner.
This was how Nvidia poached Walt Donovan, the chief architect of
Rendition, after they demonstrated the RIVA 128 chip to him at a
conference in 1997.
“Walt was the first chief architect from a rival company who wanted to
be part of the Nvidia team rather than trying to compete with us,” Kirk said.
“It gave me the idea that if we hire the best person from every other
company, we can do so much more and do so much better.”5
 David Kirk, Nvidia’s former chief scientist, became particularly skilled at
poaching. He would ask around to find out who the crucial employee at a
company was, and then call that person with his pitch. “Hey, how’s life?
How’s your job? I heard your name. I have a lot of respect for you,” he
recalled asking his targets. “You guys have been making some great
products. How many architects do you have working on this stuff?”
Usually, the number was one or two architects per company. This was
standard and on one level made sense: an architect usually oversaw an
entire family of chips, and most companies had a handful of chip families in
production at a time. Not at Nvidia, however. Kirk would explain that
Nvidia had twenty architects, and that each was working on groundbreaking
projects and had all the resources they could possibly need. Kirk would say
that he needed people like the person on the other end of the line. “Maybe
you want to come in and do this project with us. It’d be really fun, and we’ll
probably make a lot of money together too, instead of just working by
yourself there. That’s probably not as much fun.”
In later years, Nvidia employees were impressed by how the company
was able to poach and retain so many accomplished architects, who are
known for their egos. But because Nvidia’s chips had become so complex,
they needed as many high-level chip designers as possible. There was more
than enough work to go around. And Kirk was deliberate in his choice of
who to pursue, preferring to hire for complementary skill sets rather than
just whoever came along. Some were leaders and managers, while others
were hired to handle specific areas such as math and graphics algorithms.
“It wasn’t like you could just draw a diagram on the back of an envelope
and have a couple of engineers design a chip together anymore,” Kirk said.
One example of the emphasis on complementarity was Nvidia’s most
famous hire, Silicon Graphics’ John Montrym, who had developed SGI’s
high-end RealityEngine 3-D graphics hardware. He was brought on to work
alongside Donovan, who had joined the company just a few months earlier.
Kirk said Montrym was talented as an overall system architect, in that he
saw how all the components fit together, and Donovan was an expert at
graphics textures and texture filtering—“our pixel quality god” according to
one Nvidia employee. Both would remain at Nvidia for decades.
“We built the all-star team of architects,” said Kirk. “Executives were
bitter we were stealing their good people.”
 Diercks’s own arrival at Nvidia in 1994 demonstrated Jensen’s
persistence when it comes to important, but difficult, hires. Before Nvidia,
Diercks had worked at a graphics start-up company called Pellucid, which
was acquired by Media Vision—a company that later faced allegations of
financial fraud. His former Pellucid colleague Scott Sellers, who later
cofounded 3dfx, had originally talked with Jensen about joining Nvidia, but
nothing came of it. During the interview, however, Sellers was asked about
talent at Pellucid, and he said the two members on the software team,
Diercks and his direct boss, were exceptional. Jensen made a mental note.
Later, Jensen called up Diercks’s boss and said, “I hear you’re one of the
smartest guys in the Valley. You should come and talk to us.” Diercks’s boss
agreed and jumped ship for Nvidia.
Not long afterward, Diercks decided he wanted to leave as well, because
the situation with Media Vision was deteriorating. His former boss reached
out and told him to meet Jensen. After talking with Diercks, Jensen, clearly
impressed, told the former boss: “Dwight is a warrior. If I send you and
Dwight to Vietnam, you would come back on his back.”
Diercks was excited. He resigned the next day and told Pellucid’s top
executive he was going to Nvidia. The executive was livid.
“You can’t go there,” he shouted. “I’m going to sue you and Nvidia.
You’ll never work in the Valley again.” He told Diercks that a legal threat
would scare Nvidia away—the company was only a year old at the time and
had limited funds.
But when he told Jensen about the threat, the CEO wasn’t fazed.
“Bring it on,” he replied. This, Diercks realized, was the type of boss he
wanted to work for. He accepted Nvidia’s offer and has remained at the
company for more than thirty years.

THE COMPANY’S HIRING METHODS are just one component of the

Nvidia Way. Its emphasis on retention is another. Jensen rewards
performance by using stock grants, which are distributed on the basis of
how important an employee is considered to the company.
“Jensen looks at stock like his blood,” said former head of human
resources John McSorley. “He pores over the stock-allocation reports.”
Equity compensation occurs through stock grants called restricted stock
units (RSUs). When an individual starts at the company, that employee
receives a brokerage account. At the end of the first year, the employee
vests and receives one-quarter of his or her initial stock grant in a lump
sum; if the total package was one thousand shares, the employee would get
two hundred fifty. Subsequently, the employee receives one-quarter of his
or her annual grant every quarter.
To avoid the “equity cliff” (when engineers depart after their stock
packages have fully vested over the industry-standard four years), Nvidia
offers annual refresher grants. If an employee receives an “outperform”
rating from his or her manager, that employee may be awarded an
additional three hundred shares that vest over the next four years. In theory,
employees can receive these refresher grants every year—more and more
reasons to remain with the company.
Another wrinkle is the TC, or “top contributor,” designation. Managers
can refer an employee for special consideration to senior executives. Jensen
will review the list of TC candidates and give out special one-off grants that
also vest over a four-year period.
Once such a grant is approved, the employee receives an e-mail from a
senior executive, with Jensen copied. The subject line says “Special Grant,”
authorizing the RSU grant “in recognition of your extraordinary
contributions,” with a clear description of the rationale behind the award.
Jensen can also reach down into the organization at any time and award
stock directly, without waiting for an annual compensation review. This
allows him to ensure that people who are doing great work feel appreciated
in the moment. It is yet another sign of his interest in every aspect, and
level, of the company.
Former senior director of sales and marketing Chris Diskin, who played
an important role in closing the Xbox partnership with Microsoft in 2000,
said Jensen doubled his stock grant within months of him joining Nvidia.
Diskin thanked Jensen but pressed for more, saying, “If you’re really
impressed, you’d more than double it.” When he saw his grant, it had
indeed more than doubled.
Nvidia’s merit-based, adaptive, and agile compensation philosophy has
played a role in keeping turnover exceptionally low. In fiscal 2024, Nvidia
reported a turnover rate of under 3 percent in an industry where 13 percent
is the average, according to LinkedIn. It helps that the stock’s price keeps
going up, giving anyone with unvested stock more reason to stay.
 “The company treats people extremely well, not only in terms of salaries
and benefits but also by treating people as human beings rather than
fungible engineers,” a former Nvidia employee said. “There are many
opportunities for advancement.” This person mentioned Nvidia offering
flexibility on remote work when a family member received a cancer
diagnosis or providing ex gratia payments when an employee’s house
burned down.
“People tend to be loyal to a company that supports them,” he said.
Another senior executive spoke about a time when his spouse had a
major health event. He told Jensen he had to move across the country to be
close to family. “Don’t worry about it,” Jensen said. “Go and call me when
you’re ready to work again.” The employee was kept on the payroll even
though he wasn’t able to work full-time.

A COMPANY CAN RETAIN PEOPLE not only with compensation but

also through a culture of excellence—the third component of the Nvidia
Way. No employee wants to spend years working on products or
technologies that get shuttered or set aside or rendered obsolete. At Nvidia,
engineers work alongside industry luminaries who have deep technical
knowledge and experience while making products that may well reshape
the world.
Many senior executives and engineers tend to stay at Nvidia for the long
haul, even more so than at other major technology companies. Head of
software engineering Dwight Diercks, PC business executive Jeff Fisher,
and head of GPU architecture Jonah Alben have all worked at the company
for nearly three decades. Few senior executives have left for competitors or
attempted to strike out on their own in the start-up world. (Of course, they
may also be intimidated by the prospect of competing with Nvidia.)
For employees at every level, the focus on excellence in one’s work,
rather than on internal politics, is reason enough to commit to the company.
The type of person who jockeys for position more than he or she contributes
to the broader good will struggle at Nvidia. “Some companies will prefer
these kinds of people, but not Nvidia,” former GPU architect Li-Yi Wei
said. “You can be 100 percent focused on the technology side without
worrying about everything else.”6
 In fact, Nvidia actively resists the emergence of the kind of cutthroat
culture that most other organizations foster, intentionally or not. Employees
are encouraged to ask for help if they are struggling to meet a target or are
facing a technical challenge.
“If we’re going to lose, it’s not going to be because you didn’t have help.
We’re going to work together. No one loses alone,” Jensen regularly advises
Nvidia employees.7
For example, if you are a sales executive working in a particular region
and you are falling behind in meeting your quota, you are expected to
inform your team early on so they can assist you. Other resources across the
company, from Jensen to senior engineering staff, may be brought in to
solve the issue.
“ ‘No one loses alone’ is particularly pertinent in the sales organization,”
said Jay Puri, head of global field operations. Referring to the head count on
his sales team, he added that “We are so small compared to our competitors
that when something important is happening, we all need to come
together.”8
Sales executive Anthony Medeiros saw a different mentality when he
worked at Sun Microsystems. He and his peers were expected to figure it
out on their own and justify their salaries; asking for help was seen as
weakness.
“It is critical you speak up,” is how he described the culture at Nvidia.
“You would get in more trouble if you didn’t.”9

IN EXCHANGE FOR THE SUPPORT and the high compensation,

Nvidia demands much of its people. Extreme commitment is critical to the
Nvidia Way. Sixty-hour workweeks are expected as the bare minimum,
even at junior positions. The workweek can stretch to eighty hours or more
during critical periods in chip development—especially for hardware
engineers—or as the result of a major and sudden change in corporate
strategy, such as during the pivot to AI.
Transparency is also critical to the Nvidia Way. In addition to the
standard reporting lines, Nvidia employees must have a separate line of
communication with Jensen himself. Sometimes it takes the form of “Top
5” e-mails. In other instances, it can take the form of a drive-by questioning
in a hallway or even in the bathroom.
It is not possible to hide at Nvidia, even at company events. Former
developer-technology engineer Peter Young was first introduced to Jensen
at a party for new hires. Jensen already knew who he was. “You’re Peter
Young,” Jensen said. “You’ve been here for a year from Sony PlayStation
and 3dfx prior to that.” He had a similar recall of biographical details for all
fifty attendees of the party.
Young was surprised by how much the CEO knew about someone both
relatively new and relatively low-level. He mentioned this to his manager,
who replied: “That’s normal. He’s like this with everybody.” Young found it
inspiring that the CEO of the company with thousands of employees cared
enough to put that much time and effort into connecting with each
employee.10 But it was also a signal that Jensen had his eye on everyone at
the company, knew their potential, and expected them to perform
accordingly.
He expects them to continuously expand the company’s—and Jensen’s
own—knowledge base, as well. His executive staff chuckles about a habit
of Jensen that has remained constant over the decades. Whenever one of
them returns from a trade show, gaming event, or trip to Taiwan, he corners
them and asks, “So, what did you learn?”
“That kind of characterizes Jensen because he always wants to know
what’s going on out there,” Jeff Fisher said. “He just wants to know what’s
going on out in the world, so he can make better decisions.”11
When Jensen feels he can’t make the best decisions possible, he will get
frustrated—which, given Nvidia’s culture of transparency, often turns into a
public spectacle. Yet at least some employees don’t think it is fair to
describe Jensen as quick tempered.
“He can show his temper, but you have to screw up badly for him to get
to that stage,” one employee said. “He wants to be involved and understand
what you are doing. Through that process, he’s going to be very direct and
ask lots of tough questions. If you’re not ready for that type of discussion, it
can be a little alarming, but there’s no malice behind it. It’s all about let’s
get an airtight case before we move forward on something.”
Jensen is also ruthless about prioritizing his time. Adobe CEO Shantanu
Narayen recalls a breakfast with Jensen where they had a great conversation
about business issues, from innovation and strategy to culture.12 When
Narayen checked his watch, Jensen remarked, “Why are you looking at
your watch?” Narayen responded, “Jensen, don’t you have a calendar?”
Jensen replied, “What are you doing? I do what I want.” Narayen
appreciated the advice. Jensen was telling him to focus on the most
important activity at all times and not be beholden to a schedule.
Conversely, when an employee starts rambling, Jensen will say “LUA,”
which he pronounces like a single word: Looh-ahh. Bryan Catanzaro, the
Nvidia executive, explained that LUA is a warning sign that Jensen’s
patience is growing thin. When he says it, Jensen wants the employee to
stop and do three things: Listen to the question. Understand the question.
Answer the question.
“LUA means pay attention because you’re talking about something
important and you need to do it properly,” Catanzaro said. “He does not like
it when people put up an abstraction or sort of puffery to deflect an answer
to a question. Everyone who works for Jensen has heard LUA.”13
It’s a mantra that Jensen turns on himself, too. Everyone I talked to for
this book lists Jensen’s extraordinary capacity to listen, understand, and
answer any question about advanced computing. Eunhak Bae, a longtime
Nvidia investor, values Jensen’s ability to “talk through everything, not just
from a technology standpoint, but from a business standpoint. When I think
of truly well-rounded and deep technology CEOs, Jensen stands out.”14

JENSEN HUANG IS SURELY THE only person who could have gotten
Nvidia to where it is today. He grasps technology and business strategy and
also understands the harder work of actually operating a large business, day
to day. He personally enforces his own high standards and snuffs out the
dysfunction before it can take root. He has structured Nvidia to generate
step-change performance, not the slow crawl of incremental improvement.
The entire business operates at the speed of light, and if Jensen catches you
coasting, he will call you out in front of everyone. Perhaps the most
succinct definition of the Nvidia Way is that it is Jensen’s way, or that it is
simply Jensen himself.
 Curtis Priem, Jensen Huang, and Chris Malachowsky in front of Endeavor.
(NVIDIA)

But this also means Nvidia is almost completely dependent on him; in a

sense, he is its single point of failure. As of this writing, he is sixty-one
years old. It is hard to imagine that he will retire at sixty-five like many
American men do, but there will eventually be a time when he will step
back from Nvidia. Who will take his place at the center of the world’s most
important computer hardware company? Who could possibly run Nvidia as
successfully as he has done the past thirty-one years?
In writing the history of Nvidia, I was struck by the times it verged on
failure and outright destruction. If things had gone just a little bit differently
in a few instances, computing would have taken another course—we would
be living in an altered world. Some of Nvidia’s success was pure
serendipity. Chris Malachowsky might have decided to pursue a career in
medicine after taking the MCAT. He might have gone in for that next
interview with Digital Equipment instead of accepting the offer from Sun
Microsystems, which was supposed to be just a practice run. Curtis Priem
might have decided to make the NV1 chip more like everything else on the
market, and it might have succeeded. But that would have deprived Nvidia
of the chance to learn from its failure and respond with the RIVA 128—the
chip that saved the company. “Nvidia would have been a failure if NV1 had
not failed,” Priem said.15
But much of the Nvidia story is a result of Jensen’s own efforts. He
raised funds to launch Nvidia and then raised new funds when it was the
only way to save the company. He licensed the VGA core to get the RIVA
128 out on time. He kept Wall Street at bay during the CUDA years, when
everyone wanted him to sacrifice his long-term vision for short-term profit
taking. He learned how to set a high bar for performance and talent and to
buck conventional wisdom. His bluntness and directness saved time,
avoided miscommunication, and accelerated Nvidia’s pace at critical
moments. And he has distilled his philosophy into a few stock phrases that
keep people focused on what’s really important. “The mission is the boss.”
“Speed of Light.” “How hard can it be?”
Jensen and the culture he created have kept Nvidia internally aligned
despite the near-death experiences and the company’s exponential growth in
head count and revenue. When I interviewed Jensen, he told me repeatedly
that intelligence and genius had little to do with Nvidia’s success. Instead, it
was hard work and resilience. It didn’t have to be this hard, but it was—and
it was always going to be. The work demanded one thing out of everyone,
including himself: “sheer will.”
Nvidia remains the only stand-alone graphics-chip firm to this day, even
though hundreds of others have thrown their hats in the ring. Jensen himself
is now the technology industry’s longest-serving CEO.
We are sometimes told, by various self-help experts and gurus, that we
can make more money while working less. Jensen is the antithesis of that
notion. There are no shortcuts. The best way to be successful is to take the
more difficult route. And the best teacher of all is adversity—something he
has become well acquainted with. It is why he still keeps going at a pace
that would see most other people, at any age, burn out. It is why he still
says, to this day, and without any trace of hesitation or irony or self-doubt:
“I love Nvidia.”
 APPENDIX

Jensen-isms

“As many as needed, as few as possible”: Invite only the essential

employees, those with relevant knowledge, to meetings, and avoid wasting
people’s time if their presence is not necessary.

“AMAN, ALAP”: As much as needed, as little as possible. Be frugal with

employee time and company resources.

“Always hire the best”: If you hire smart and capable people, they will be
able to solve problems and adapt to new challenges.

“Criticism is a gift”: Providing direct feedback leads to continuous

improvement.

“Don’t worry about the score. Worry about how you play the game”:
Don’t be distracted by stock price volatility. Focus on delivering excellent
work and creating value.

“Early indicators of future success”: Evidence a new project is starting to

get traction.
“Floor sweeping” and “Ship the whole cow”: Design chips with
redundancy, so that even if there are small manufacturing defects, the chips
can still be sold as lower-performance parts, reducing waste.

“Honing the sword”: Spirited debate often leads to the best ideas.

“How hard can it be?”: A refrain to use to avoid feeling overwhelmed by

the amount of work ahead.

“Intellectual honesty”: Tell the truth, acknowledge failure, be willing to

move forward and learn from past mistakes.

“If you measure it, you can improve it, but you have to measure the
right thing!”: Don’t fall for tracking the wrong metrics. Be data driven in a
smart way.

“Is it world class?”: Nvidia’s products, talent acquisition, and business

practices must be benchmarked against the best in the industry.

“Let’s go back to first principles”: Tackle problems with a clean sheet of

paper, not on the basis of how they were approached in the past.

“LUA”: Listen to the question, understand the question, and answer the
question. A warning sign Jensen is getting frustrated about a long-winded
response.

“The mission is the boss”: Decisions are made on the basis of the end goal
of serving the customer, not internal politics.

“No one loses alone”: If you are falling behind, inform your team
promptly, so they can help.

“Nvidia can execute”: Nvidia wins with superior technology and

execution.
“Pilot-in-Command”: Jensen’s designated leader for an important project
who should receive priority support from the entire company.

“Second place is the first loser”: The goal and expectation is to win every
time.

“Small steps, big vision”: Prioritize actionable items and complete the
most important first task to the best of your ability.

“Speed of Light”: Strive to improve performance to the absolute limit of

what’s possible according to the laws of physics rather than comparing
results against previous ones.

“Strategy is about the things you give up”: Sort through everything, pick
the most important thing, and then do that and leave the others aside.

“We don’t steal market share, we create markets”: Nvidia wants to be

the market leader in a new area rather than fight over an existing business.

“You got to believe, what you got to believe”: If you believe in

something, go invest in it. Go do it. Put all your energy there.

“Your strength is your weakness”: Being overly kind and tactful can
hinder progress.
 ACKNOWLEDGMENTS

THE ORIGIN OF THIS BOOK STARTED with a cold e-mail. On May

10, 2023, I received a message with the subject header: “Hello from W. W.
Norton—a book on Nvidia?” It came from an editor, Dan Gerstle. He
reached out at the suggestion of one of his authors (thanks, Matthew Ball),
who thought I could write a book on Nvidia.
I thought there must be several books on Nvidia, as every other major
technology company had at least half a dozen. After searching, I found
none. At that moment, I realized I wanted to write this book.
From there, life moved quickly. I met with Dan at Bryant Park Café in
Manhattan. At the end of the meeting, he said I needed a book agent. On the
recommendation of friends, I met with Pilar Queen, who agreed to represent
me. A month after receiving the first e-mail, I had secured a book deal.
The past year has been a whirlwind. I’m indebted to Dan and Pilar for
taking a chance on a new author and providing invaluable advice and
guidance. I also want to thank my freelance editor, Darryl Campbell, who
worked tirelessly, editing and offering great feedback on the manuscript.
I want to thank Jensen. Although Nvidia was initially hesitant about
cooperating with this book, maybe due to some of my prior unfavorable
coverage, he never discouraged sources from speaking with me. I also want
to thank Curtis Priem and Chris Malachowsky for their contributions, as
well as the Nvidia team: Stephanie, Bob, Mylene, Ken, and Hector.
Finally, I extend my heartfelt gratitude to my sources for taking time out
of their busy lives to share their experiences. Gathering their accounts of the
early decades of computer history, many of which were documented for the
first time, has been an honor. Their generosity enriched this book and made
it possible.
 NOTES

Introduction
1. Hendrik Bessembinder, “Which U.S. Stocks Generated the Highest Long-Term Returns?,” S&P
Global Market Intelligence Research Paper Series, July 16, 2024.
http://dx.doi.org/10.2139/ssrn.4897069.

Chapter 1: Pain and Suffering

1. Lizzy Gurdus, “Nvidia CEO: My Mom Taught Me English,” CNBC, May 6, 2018.
2. Matthew Yi, “Nvidia Founder Learned Key Lesson in Pingpong,” San Francisco Chronicle,
February 21, 2005.
3. “A Conversation with Nvidia’s Jensen Huang,” Stripe, May 21, 2024, video, 10:02.
4. Maggie Shiels, “Nvidia’s Jen-Hsun Huang,” BBC News, January 14, 2010.
5. Brian Dumaine, “The Man Who Came Back from the Dead Again,” Fortune, September 1,
2001.
6. Interview with Judy Hoarfrost, 2024.
7. “19th Hole: The Readers Take Over,” Sports Illustrated, January 30, 1978.
8. Yi, “Nvidia Founder Learned Key Lesson.”
9. “2021 SIA Awards Dinner,” SIAAmerica, February 11, 2022, video.
https://www.youtube.com/watch?v=5yvN_T8xaw8.
10. “The Moment with Ryan Patel: Featuring NVIDIA CEO Jensen Huang | HP,” HP, October 26,
2023, video, 1:47.
11. “Jen-Hsun Huang,” Charlie Rose, February 5, 2009.
12. Interview with Jensen Huang, 2024.
13. “2021 SIA Awards Dinner,” SIAAmerica, 1:04:00.
14. “The Moment with Ryan Patel,” HP, 3:07.
15. “Jen-Hsun Huang, NVIDIA Co-Founder, Invests in the Next Generation of Stanford
Engineers,” School News, Stanford Engineering, October 1, 2010.
16. Gurdus, “Nvidia CEO.”
17. “Jensen Huang,” Stanford Institute for Economic Policy Research, March 7, 2024, video,
38:00.

Chapter 2: The Graphics Revolution

1. Frederick Van Veen, The General Radio Story (self-pub., 2011), 153.
2. Van Veen, General Radio Story, 171–75.
3. Interview with Chris Malachowsky, 2023.
4. Interview with Curtis Priem, 2024.
5. Van Hook was a graphics pioneer in his own right, however. He would later design the graphics
architecture of the Nintendo 64.
6. Interview with Chris Malachowsky, 2023.

Chapter 3: The Birth of Nvidia

1. “Jensen Huang,” Sequoia Capital, November 30, 2023, video, 5:13.
2. “Jen-Hsun Huang, NVIDIA Co-Founder, Invests in the Next Generation of Stanford
Engineers,” School News, Stanford Engineering, October 1, 2010.
3. “2021 SIA Awards Dinner,” SIAAmerica, February 11, 2022, video, 1:11:09.
https://www.youtube.com/watch?v=5yvN_T8xaw8.
4. “Jen-Hsun Huang,” Stanford Online, June 23, 2011, video, 9:25.
5. National Science Board, “Science and Engineering Indicators–2002,” NSB-02-01 (Arlington,
VA: National Science Foundation, 2002). https://www.nsf.gov/publications/pub_summ.jsp?
ods_key=nsb0201.
6. Interview with Jensen Huang, 2024.
7. “Jensen Huang,” Sequoia Capital.
8. Interview with Mark Stevens, 2024.

Chapter 4: All In
1. “Jen-Hsun Huang,” Stanford Online, June 23, 2011, video, 45:37.
2. Interview with Pat Gelsinger, 2023.
3. Interview with Dwight Diercks, 2024.
4. Jon Peddie, The History of the GPU: Steps to Invention (Cham, Switzerland: Springer, 2022),
278.
5. Peddie, History of the GPU, 278.
6. Interview with Curtis Priem, 2024.
7. Interview with Michael Hara, 2024.
8. “Jen-Hsun Huang, NVIDIA Co-Founder, Invests in the Next Generation of Stanford
Engineers,” School News, Stanford Engineering, October 1, 2010.
9. “Jensen Huang,” Sequoia Capital, November 30, 2023, video, 13:57.
10. Jon Stokes, “Nvidia Cofounder Chris Malachowsky Speaks,” Ars Technica, September 3, 2008.
11. “Dean’s Speaker Series | Jensen Huang Founder, President & CEO, NVIDIA,” Berkeley Haas,
January 31, 2023, video, 32:09.
12. Interview with former Nvidia employee, 2023.
13. “3dfx Oral History Panel,” Computer History Museum, July 29, 2013, video.
14. Orchid Technology, “Orchid Ships Righteous 3D,” press release, October 7, 1996.
15. “3dfx Oral History Panel,” Computer History Museum.
16. Interview with Scott Sellers, 2023.
17. Interview with Dwight Diercks, 2024.
18. “Jen-Hsun Huang,” Oregon State University, February 22, 2013, video, 37:20.
19. Interview with former Nvidia employee, 2023.
20. “Jen-Hsun Huang,” Oregon State University, 30:28.
21. Interview with Curtis Priem, 2024.
22. Interview with Dwight Diercks, 2024.
23. Interview with Henry Levin, 2023.
24. Interview with Chris Malachowsky, 2023.
25. Interview with Jensen Huang, 2024.
26. Interview with Eric Christenson, 2023.
27. Personal e-mail from Sutter Hill CFO Chris Basso.
28. Nvidia, “Upstart Nvidia Ships Over One Million Performance 3D Processors,” press release,
January 12, 1998.
29. Interview with Jensen Huang, 2024.

Chapter 5: Ultra-Aggressive
1. Interview with Caroline Landry, 2024.
2. Interview with Michael Hara, 2024.
3. Interviews with Tench Coxe and other former Nvidia employees, 2023.
4. Interview with Robert Csongor, 2023.
5. Interview with Jeff Fisher, 2024.
6. Interview with Geoff Ribar, 2023.
7. Interview with John McSorley, 2023.
8. Interview with Andrew Logan, 2024.
9. Interview with Kenneth Hurley, 2024.
10. Interview with Caroline Landry, 2024.
11. Interview with Sanford Russell, 2024.
12. Interview with Andrew Logan, 2024.
13. Interview with Jeff Fisher, 2024.
14. “Morris Chang, in Conversation with Jen-Hsun Huang,” Computer History Museum, October
17, 2007, video, 23:00.
15. Interview with Chris Malachowsky, 2023.
16. Interview with Curtis Priem, 2024.
17. Interview with Geoff Ribar, 2023.
18. Interview with Michael Hara, 2024.
19. Interview with Michael Hara, 2024.
20. Interview with Jeff Fisher, 2024.
21. Interview with Curtis Priem, 2024.
22. Interview with Nick Triantos, 2023.

Chapter 6: Just Go Win

1. Interview with Ross Smith, 2023.
2. Interview with Scott Sellers, 2023.
3. Interview with Dwight Diercks, 2024.
4. Interview with Michael Hara, 2024.
5. Interview with David Kirk, 2024.
6. Interview with Curtis Priem, 2024.
7. Interview with Dwight Diercks, 2024.
8. Interview with Dwight Diercks, 2024.
9. Interview with Rick Tsai, 2024.
10. Dean Takahashi, “Shares of Nvidia Surge 64% after Initial Public Offering,” Wall Street
Journal, January 25, 1999.
11. Interview with Kenneth Hurley, 2024.
12. Takahashi, “Shares of Nvidia Surge.”
13. Dean Takahashi, Opening the Xbox: Inside Microsoft’s Plan to Unleash an Entertainment
Revolution (Roseville, CA: Prima Publishing, 2002), 230.
14. Interview with Oliver Baltuch, 2023.
15. Takahashi, Opening the Xbox, 202.
16. Interview with George Haber, 2023.
17. Interview with Chris Diskin, 2024.
18. Interview with George Haber, 2023.
19. Interview with Curtis Priem, 2024
20. Interview with Michael Hara, 2024.

Chapter 7: GeForce and the Innovator’s Dilemma

1. Clayton Christensen, The Innovator’s Dilemma: When New Technologies Cause Great Firms to
Fail (Boston, MA: Harvard Business School Press, 1997), 47.
2. Interview with Michael Hara, 2024.
3. Interview with Jeff Fisher, 2024.
4. Interview with Tench Coxe, 2023.
5. “Jensen Huang of Nvidia on the Future of A.I. | DealBook Summit 2023,” New York Times,
November 30, 2023, video, 19:54.
6. Interview with Nvidia employee, 2023.
7. Interview with Sanford Russell, 2024.
8. Interview with Dan Vivoli, 2024.
9. John D. Owens et al., “A Survey of General-Purpose Computation on Graphics Hardware,”
State of the Art Reports, Eurographics 2005, August 1, 2005.
https://doi.org/10.2312/egst.20051043.
10. Interview with David Kirk, 2024.
11. Interview with Jensen Huang, 2024.
12. Interview with two former Nvidia employees, 2023.
13. “Best Buy Named in Suit over Sam Goody Performance,” New York Times, November 27,
2003.
14. Interview with Jensen Huang, 2024.

Chapter 8: The Era of the GPU

1. Interview with David Kirk, 2024.
2. Interview with Jensen Huang, 2024.
3. Interview with Jensen Huang, 2024.
4. Ian Buck et al., “Brook for GPUs: Stream Computing on Graphics Hardware,” ACM
Transactions on Graphics 23, no. 3 (August 2004): 777–86.
5. Interview with Andy Keane, 2024.
6. Anand Lal Shimpi, “Nvidia’s GeForce 8800,” Anandtech, November 8, 2006.
7. “A Conversation with Nvidia’s Jensen Huang,” Stripe Sessions 2024, April 24, 2024, video,
01:04:49.
8. “No Priors Ep. 13 | With Jensen Huang, Founder & CEO of NVIDIA,” No Priors: AI, Machine
Learning, Tech, & Startups, April 25, 2023, video. https://www.youtube.com/watch?
v=ZFtW3g1dbUU.
9. Rob Beschizza, “nVidia G80 Poked and Prodded. Verdict: Fast as Hell,” WIRED, November 3,
2006; Jon Stokes, “NVIDIA Rethinks the GPU with the New GeForce 8800,” Ars Technica,
November 8, 2006.
10. Interview with David Kirk, 2024.
11. Interview with Mark Berger, 2024.
12. Interview with Derik Moore, 2024.
13. “NVIDIA CEO Jensen Huang,” Acquired, October 15, 2023, video, 49:42.
14. Interview with Amir Salek, 2023.

Chapter 9: Tortured into Greatness

1. Nvidia Corporation, “Letter to Stockholders: Notice of 2010 Annual Meeting” (Santa Clara,
CA: Nvidia, April 2010).
2. Interview with Dan Vivoli, 2023.
3. Interview with Anthony Medeiros, 2024.
4. Interview with Jensen Huang, 2024.
5. “In Conversation | Jensen Huang and Joel Hellermark,” Sana AI Summit, June 29, 2023, video,
32:10.
6. “A Conversation with Nvidia’s Jensen Huang,” Stripe, May 21, 2024, video, 11:06.
7. Interview with Tench Coxe, 2023.
8. Interview with Oliver Baltuch, 2023.
9. Interview with Andy Keane, 2024.
10. Interview with Jensen Huang, 2024.
11. Interview with Simona Jankowski, 2024.
12. Interview with Jay Puri, 2024.
13. Interview with Jensen Huang, 2024.
14. Interview with Robert Csongor, 2023.
15. Interview with Michael Douglas, 2024.
16. Interview with Michael Douglas, 2023.
17. Interview with John McSorley, 2023.
18. Interview with former Nvidia employee, 2024.
19. Interview with Mark Berger, 2024.
20. Interview with Jay Puri, 2024.
21. Interview with David Ragones, 2024.
22. Interview with Michael Douglas, 2024.
23. Interview with Jensen Huang, 2024.

Chapter 10: The Engineer’s Mind

1. Carl Icahn, “Beyond Passive Investing,” Founder’s Council program, Greenwich Roundtable,
April 12, 2005.
2. Walt Mossberg, “On Steve Jobs the Man, the Myth, the Movie,” Ctrl-Walt-Delete Podcast,
October 22, 2015.
3. Interview with former Nvidia employee, 2024.
4. Interview with Tench Coxe, 2023.
5. Interview with Ali Simnad, 2024.
6. Interview with Leo Tam, 2023.
7. Interview with Kevin Krewell, 2024.
8. “In Conversation | Jensen Huang and Joel Hellermark,” Sana AI Summit, June 29, 2023, video,
29:20.
9. “Jen-Hsun Huang,” Stanford Online, June 23, 2011, video, 32:41.
10. “Jen-Hsun Huang,” Oregon State University, February 22, 2013, video, 1:15:58.
11. Interview with Tench Coxe, 2023.
12. Interview with Jeff Fisher, 2023.
13. Interview with Bryan Catanzaro, 2024.
14. Maggie Shiels, “Nvidia’s Jen-Hsun Huang,” BBC, January 14, 2010.
15. “Saturday’s Panel: A Conversation with Jen-Hsun Huang (5/7),” Committee of 100, May 18,
2007, video, 5:43.
16. “Jensen Huang—CEO of NVIDIA | Podcast | In Good Company | Norges Bank Investment
Management,” Norges Bank, November 19, 2023, video, 44:50.
17. Alexis C. Madrigal, “Paul Otellini’s Intel: Can the Company That Built the Future Survive It?,”
The Atlantic, May 16, 2013.
18. Interview with Pat Gelsinger, 2023.
19. Mark Lipacis, “NVDA Deep-Dive Presentation,” Jefferies Equity Research, August 17, 2023.
20. “Search Engine Market Share Worldwide,” Statcounter. https://gs.statcounter.com/search-
engine-market-share (accessed August 9, 2024).

Chapter 11: The Road to AI

1. William James Dally, “A VLSI Architecture for Concurrent Data Structures,” PhD diss.,
California Institute of Technology, 1986.
2. Interview with David Kirk, 2024.
3. Brian Caulfield, “What’s the Difference Between a CPU and a GPU?,” Nvidia Blog, December
16, 2009.
4. “NVIDIA: Adam and Jamie Explain Parallel Processing on GPU’s,” Artmaze1974, September
15, 2008, video.
5. John Markoff, “How Many Computers to Identify a Cat? 16,000,” New York Times, June 26,
2012.
6. Interview with Bill Dally, 2024.
7. Adam Coates et al., “Deep Learning with COTS HPC Systems,” in Proceedings of the 30th
International Conference on Machine Learning, Proceedings of Machine Learning Research,
vol. 28, cycle 3, ed. Sanjoy Dasgupta and David McAllester (Atlanta, GA: PMLR, 2013),
1337–45.
8. Jensen Huang, “Accelerating AI with GPUs: A New Computing Model,” Nvidia Blog, January
12, 2016.
9. Interview with Bill Dally, 2024.
10. Coates et al., “Deep Learning with COTS HPC Systems,” 1338.
11. Coates et al., “Deep Learning with COTS HPC Systems,” 1345.
12. Interview with Bill Dally, 2024.
13. Interview with Bryan Catanzaro, 2024.
14. Dave Gershgorn, “The Data That Transformed AI Research—and Possibly the World,” Quartz,
July 26, 2017.
15. Jessi Hempel, “Fei-Fei Li’s Quest to Make AI Better for Humanity,” WIRED, November 13,
2018.
16. Gershgorn, “The Data That Transformed AI Research.”
17. Interview with Bill Dally, 2024.
18. Interview with Bryan Catanzaro, 2024.
19. Interview with Bryan Catanzaro, 2024.
20. Interview with Bryan Catanzaro, 2024.
21. Interview with Bryan Catanzaro, 2024.
22. “NVIDIA Tesla V100: The First Tensor Core GPU,” Nvidia. https://www.nvidia.com/en-
gb/data-center/tesla-v100/ (accessed August 13, 2024).
23. Interview with Bill Dally, 2024.
24. “No Priors Ep. 13 | With Jensen Huang, Founder & CEO of NVIDIA,” No Priors: AI, Machine
Learning, Tech, & Startups, April 25, 2023, video, 16:19. https://www.youtube.com/watch?
v=ZFtW3g1dbUU.
25. “Q3 2024 Earnings Call,” Nvidia, November 21, 2023.

Chapter 12: The “Most Feared” Hedge Fund

1. Michael J. de la Merced, “A Primer on Starboard, the Activist That Pushed for a Staples-Office
Depot Merger,” New York Times, February 4, 2015.
2. “Transforming Darden Restaurants,” Starboard Value, PowerPoint presentation, September 11,
2014.
3. William D. Cohan, “Starboard Value’s Jeff Smith: The Investor CEOs Fear Most,” Fortunate,
December 3, 2014.
4. Darden Restaurants, “Darden Addresses Inaccurate and Misleading Statements by Starboard
and Provides the Facts on Value Achieved with Red Lobster Sale,” press release, August 4,
2014.
5. Myles Udland and Elena Holodny, “Hedge Fund Manager Publishes Dizzying 294-Slide
Presentation Exposing How Olive Garden Wastes Money and Fails Customers,” Business
Insider, September 12, 2014.
6. “Transforming Darden Restaurants,” Starboard Value, 6–7.
7. Interview with Jeff Smith, 2024.
8. Starboard Value letter to Mellanox Technologies, Ltd., January 8, 2017.
9. Interview with Jay Puri, 2024.

Chapter 13: Lighting the Future

1. Interview with David Luebke, 2024.
2. Interview with Bryan Catanzaro, 2024.
3. Interview with Jensen Huang, 2024.
4. Jordan Novet, “Google A.I. Researcher Says He Left to Build a Startup after Encountering ‘Big
Company-itis,’ ” CNBC, August 17, 2023.

Chapter 14: The Big Bang

1. John Markoff, “At Google, Earnings Soar, and Share Price Follows,” New York Times, October
22, 2004.
2. Ben Popper, “Facebook’s Q2 2013 Earnings Beat Expectations,” The Verge, July 24, 2013.
3. Interview with Colette Kress, 2023.
4. Interview with Jeff Fisher, 2024.
5. Interview with Simona Jankowski, 2024.
6. Dave Salvator, “H100 Transformer Engine Supercharges AI Training, Delivering Up to 6x
Higher Performance without Losing Accuracy,” Nvidia Blog, March 22, 2022.
7. “No Priors Ep. 13 | With Jensen Huang, Founder & CEO of NVIDIA,” No Priors: AI, Machine
Learning, Tech, & Startups, video, 16:51. https://www.youtube.com/watch?v=ZFtW3g1dbUU.
8. Interview with Jay Puri, 2024.
9. Interview with former Nvidia executive, 2024.
10. Interview with Ross Walker, 2024.
11. “Jen-Hsun Huang,” Stanford Online, June 23, 2011, video, 9:25.
12. “Dean’s Speaker Series | Jensen Huang Founder, President & CEO, NVIDIA,” Berkeley Haas,
January 31, 2023, video, 49:25.
13. “Download Day 2024 — Fireside Chat: NVIDIA Founder & CEO Jensen Huang and
Recursion’s Chris Gibson,” Recursion, June 24, 2024, video, 1:32.
14. Kimberly Powell Q&A interview by analyst Harlan Sur, 42nd Annual J.P. Morgan Healthcare
Conference, San Francisco, CA, January 8, 2024.
15. Interview with Gevorg Grigoryan, 2024.
16. “Nvidia CEO,” HBR IdeaCast, November 14, 2023.
17. Brian Caulfield, “AI Is Tech’s ‘Greatest Contribution to Social Elevation,’ NVIDIA CEO Tells
Oregon State Students,” Nvidia Blog, April 15, 2024.

Conclusion: The Nvidia Way

1. Paul Graham, “What Happened to Yahoo,” PaulGraham.com, August 2010.
2. Nvidia Corporation, “NVIDIA Corporate Responsibility Report Fiscal Year 2023” (Santa Clara,
CA: Nvidia), 16.
3. Interview with Ben de Waal, 2023.
4. Interview with Dwight Diercks, 2024.
5. Interview with David Kirk, 2024.
6. Interview with Li-Yi Wei, 2024.
7. Interview with Anthony Medeiros, 2024.
8. Interview with Jay Puri, 2024.
9. Interview with Anthony Medeiros, 2024.
10. Interview with Peter Young, 2024.
11. Interview with Jeff Fisher, 2024.
12. Interview with Shantanu Narayen, 2024.
13. Interview with Bryan Catanzaro, 2024.
14. Interview with Jeff Fisher, 2024.
15. Interview with Curtis Priem, 2024.
 Copyright © 2025 by Tae Kim

All rights reserved

First Edition

For information about permission to reproduce selections from this book, write to Permissions, W. W.
Norton & Company, Inc., 500 Fifth Avenue, New York, NY 10110

For information about special discounts for bulk purchases, please contact W. W. Norton Special
Sales at [email protected] or 800-233-4830

Book design by Beth Steidle

Production manager: Lauren Abbate

Library of Congress Cataloging-in-Publication Data is available

ISBN 978-1-324-08671-0

ISBN 978-1-324-08672-7 (epub)

W. W. Norton & Company, Inc.

500 Fifth Avenue, New York, NY 10110
www.wwnorton.com

W. W. Norton & Company Ltd.

15 Carlisle Street, London W1D 3BS