Culture & Critics
BO OKS Another day, another new AI large language model
that’s supposedly better than all previous ones.
When I began writing this story, Elon Musk’s xAI
had just released Grok 3, which the company says
performs better than its competitors against a wide
range of benchmarks. As I was revising the article,
Anthropic released Claude 3.7 Sonnet, which it
says outperforms Grok 3. And by the time you read
this, who knows? Maybe an entirely new LLM will
have appeared. In January, after all, the AI world
was temporarily rocked by the release of a low-
cost, high-performance LLM from China called
DeepSeek-R1. A month later, people were already
wondering when DeepSeek-R2 would come out.
The New King of Tech The competition among LLMs may be hard
to keep track of, but for Nvidia, the company
How Jensen Huang built Nvidia that designs the computer chips—or graphics-
processing units (GPUs)—that many of these
into a nearly $3 trillion business large language models have been trained on, it’s
also enormously lucrative. Nvidia, which, as of
By James Surowiecki this writing, is the third-most-valuable company
in the world (after Apple and Microsoft), was
PATRICK T. FALLON / AFP / GETTY; USGS PUBLICATIONS WAREHOUSE; RAWPIXEL
80 ILLUSTRATI ON BY RICARDO SANTOS
started three decades ago by engineers who wanted “Lori,” he says of his wife, “did ninety percent of the
to make graphics cards for gamers. How it evolved parenting” of their two children. For the past 30 years,
into the company that is providing almost all the picks his life has clearly revolved around Nvidia.
and shovels for the AI gold rush is the story at the core
of Stephen Witt’s The Thinking Machine. Framed as Huang’s reluctance to talk about himself makes
a biography of Jensen Huang, the only CEO Nvidia him a challenging subject for Witt to bring to life. But
has ever had, the book is also something more inter- Nvidia’s employees, who almost all refer to Huang by
esting and revealing: a window onto the intellectual, his first name, are effusive. They “worship him—I
cultural, and economic ecosystem that has led to the believe they would follow him out of the window of a
emergence of superpowerful AI. skyscraper if he saw a market opportunity there,” Witt
That ecosystem’s center, of course, is Silicon Valley, writes. He later adds that they see Huang “not just as
where Huang has spent most of his adult life. He was a leader but as a prophet. Jensen was a prophet who
born in Taiwan, the son of a chemical engineer and a made predictions about things. And then those things
teacher. The family moved to Thailand when he was 5, came true.” He has a ferocious temper—referred to in
and a few years later, his parents sent him and his older the company as “the Wrath of Huang”—and is notori-
brother to the United States to escape political unrest. ous for publicly reprimanding, at length, workers who
Eventually, his parents relocated to the U.S. as well, and have made mistakes or failed to deliver. But he rarely
Huang grew up in the suburbs of Portland, Oregon. In fires people and, in fact, inspires intense devotion. One
the early 1980s, after majoring in electrical engineer- of his key subordinates says, “I’ve been afraid of Jensen
ing at Oregon State (which at the time didn’t offer a sometimes. But I also know that he loves me.”
computer-science major), he got a job at Advanced Huang’s greatest strength as a CEO has been his
Micro Devices. The company—then the poor cousin Nvidia’s willingness to make big, risky bets when opportuni-
of the chip giant Intel—was headquartered in Sunny- employees see ties present themselves. The first of those came when
vale, California, near US 101, the highway that runs he changed the architecture of Nvidia’s chips from
from San Jose to Stanford. Since then, Huang’s career
Huang as serial processing to parallel processing. Witt calls this
has unfolded within a five-mile radius of that office. “a prophet move “a radical gamble,” because up to that point,
Huang soon left AMD for a firm called LSI Logic who made no company had been able to make selling parallel-
Corporation, which built software-design tools for predictions processing chips a viable business.
chip architects, and then left LSI in 1993 to start Serial computing is the way your computer’s cen-
Nvidia with the chip designers Curtis Priem and Chris
about things. tral processing unit works: It executes one instruction
Malachowsky: He was right on target “to run some- And then at a time, very, very fast. Witt likens it to telling one
thing by the age of thirty,” as he’d told them he aimed those things delivery van to drop off packages in sequence. By con-
to do. The company was entering a crowded mar- came true.” trast, “Nvidia’s parallel GPU acts more like a fleet of
ketplace for developing graphics cards, the computer motorcycles spreading out across a city,” with the driv-
hardware that’s used to render images and videos. ers delivering each package at roughly the same time.
Nvidia didn’t have a real business plan, but Huang’s The coding required to make parallel processing work
boss at LSI recommended him to Sequoia Capital. was much more complex, but if you could do it, you
One of the Valley’s most important venture-capital had access to enormous amounts of computing power.
firms, Sequoia helped the company get off the ground. Initially, all that power was used mainly to make
The graphics-card business was built on a perpet- computer games look and perform better. But then
ual upgrade cycle that forced developers into a never- Huang took another big risk, remaking Nvidia’s GPUs
ending game of performance improvement: A company so that they could also process massive data sets, of the
was only as good as its last card. At various points in kind scientists might use. As one Nvidia executive puts
those early years, Nvidia was one misstep away from it, “You have a video game card on one side, but it has a
bankruptcy, and its unofficial motto became “Our com- switch on it. So you flick that switch, and turn the card
pany is thirty days from going out of business.” over, and suddenly the card becomes a supercomputer.”
One gets the impression that Huang liked it that The fascinating thing about this decision was that
way. He says his heart rate goes down under pressure, Huang didn’t know who might want to buy a super-
and to call him a relentless worker is to understate mat- computer in the guise of a graphics card, or how many
ters. “I should make sure that I’m sufficiently exhausted such people were out there. He was just betting that
from working that no one can keep me up at night,” if you make powerful tools available to people, they
he once said. His reading diet features business books will find a use for them, and at a scale to justify the
(which he devours). He has no obvious politics (or at billions in investment.
least never discusses them). He’s not a gaudy philan- That use—and it was big—turned out to be artificial
thropist. Though devoted to his family, he’s also honest: intelligence, in particular neural-network technology.
MAY 2025 81
Culture & Critics
As Witt notes, just as parallel processing was revolution- emergence of LLMs, and fending off competitors
izing computing, a similar revolution was happening (many of whom are Nvidia’s customers, now build-
in AI research—though no one at Nvidia was paying ing their own chips). But the foundations for that
attention to it. AI had gone through a series of boom- pivot, and all that ensued, were already in place when
and-bust cycles as researchers tried different techniques, Huang decided to act on his AI insight.
all of which ultimately failed. One of those methods
was neural networks, which tried to mimic the human T h o s e f o u n d at i o n s , The Thinking Machine
brain and allow the AI to evolve new rules of learning makes clear, were not laid by Nvidia alone. Indeed,
on its own. When you train these networks on mas- among Witt’s key contributions is to show that Nvid-
sive databases of images and text, they can, over time, ia’s success can’t be understood apart from the culture
identify patterns and become smarter. Neural networks and economy of Silicon Valley (and of tech more gen-
had long been peripheral, partly because they’re black erally). Take the simple fact of free labor markets. One
boxes (you can’t explain how the AI is learning, or why catalyst of the Valley’s success, as the scholar AnnaLee
it’s doing what it’s doing), and partly because the com- Saxenian has famously argued, was a freewheeling,
puting power required to make a high-performance risk-taking culture that encouraged workers to leave
neural network operate was out of reach. companies for competitors or to start their own firms.
Parallel-processing GPUs changed all that. Sud- And that depended, in part, on the fact that noncom-
denly, AI researchers, if they could write software well pete clauses were unenforceable in California. Nvidia’s
enough to get the most out of the chips, had access to history exemplifies this: not just Huang’s mobility, but
sufficient computing power to allow neural networks that of his early hires as well. Later, one of the com-
to evolve at an extraordinary pace. In 2009, Geoff pany’s favorite tactics was to poach its competitors’
Hinton, one of the godfathers of AI research, told best engineers and coders—bad form, perhaps, but a
a conference of machine-learning experts to go buy good business tactic.
Nvidia cards. And in 2012, one of Hinton’s students, Nvidia also benefited from the research invest-
Alex Krizhevsky, strung together two Nvidia GPUs THE THINKING
ments made by the government and universities. One
and built and trained SuperVision (which he later MACHINE: of the crucial breakthroughs in unlocking the power of
renamed AlexNet). It was an AI model that could, JENSEN HUANG, parallel computing, for instance, was an open-source
NVIDIA, AND
for the first time, identify images with startling accu- THE WORLD’S programming language called Brook, which a gamer
MOST COVETED
racy, largely because, in Witt’s words, “the GPU pro- M I C RO C H I P
and Stanford graduate student named Ian Buck devel-
duced in half a minute what would have taken an Intel oped with a group of researchers in 2003, relying on a
machine an hour and what would have taken biology Stephen Witt Defense Department grant. Alex Krizhevsky and his
a hundred thousand years.” partner Ilya Sutskever (who later helped start OpenAI)
Huang did not immediately recognize the impor- VIKING were grad students at the University of Toronto when
tance of what had happened. When he spoke at Nvid- Krizhevsky devised AlexNet. The contest in which the
ia’s annual GPU Technology Conference in 2013, he model demonstrated its accuracy, the ImageNet chal-
never mentioned neural networks, talking instead about lenge, was designed by a Stanford computer scientist
weather modeling and computer graphics. But a few named Fei-Fei Li. And as that lineup demonstrates
months later, after an Nvidia researcher named Bryan (Krizhevsky and Sutskever were born in the Soviet
Catanzaro made a direct pitch to him about the impor- Union, Li in China), immigration has been central
tance of AI, Huang had what Witt calls a “Damascene to the history of not just Nvidia but AI generally.
epiphany”: He placed another big bet, essentially trans- Practical economic features of the ecosystem mat-
forming Nvidia from a graphics company into an AI tered as well. The most important was the rise of inde-
company over the course of a weekend. This bet was less pendent chip foundries: factories that serve many dif-
risky than his earlier ones, because even though Nvidia ferent companies and make chips on order. Nvidia’s
had competitors who also built GPUs, none of them partnership with Taiwan Semiconductor Manufac-
had really designed theirs to be used as supercomputers. turing Company, the best-known of these factories,
Still, going all in was prescient—developments such as allowed it to become a dominant player by focusing
large language models had yet to take off—and is what on designing and writing software for its chips; Nvidia
has turned Nvidia into a nearly $3 trillion company. didn’t have to invest in actual production, which would
That weekend feels as if it were the compressed have required prohibitive amounts of capital.
culmination of Nvidia’s story, which isn’t empirically Finally, Nvidia benefited from patience, and its
true. The 12 years that followed have been incredibly board’s willingness to put long-term thinking ahead
eventful, and incredibly profitable, as the company of short-term profits. Because of the gaming mar-
has kept improving its chips, servicing the insatia- ket, Nvidia was almost always a profitable company,
ble appetite for computing power created by the but its stock price dropped nearly 90 percent two
82 MAY 2025
BO OKS
different times; it didn’t appreciate for a full 10 years
after the dot-com bubble burst, while the company
was spending billions turning its graphics cards into
supercomputers. One familiar indictment of Ameri-
can capitalism is that it’s too short-term-focused. In
the tech industry, at least, the trajectory of Nvidia
(and many other companies) suggests that’s a bum rap.
To be sure, Huang himself was central to Nvidia’s
success: He has run the company essentially on his
own (as Witt puts it, he has had “no right-hand man
or woman, no majordomo, no second-in-command”),
and he’s made the bold moves. What’s more, he seems Chickadee
to have done so without a trace of doubt. Lots of By Stanley Plumly
people in the AI industry—including the people train-
ing LLMs—have raised concerns about AI’s dangers,
but Huang is not one of them. For him, Witt writes, Margaret remembering in summer how they’d fly
“AI is a pure force for progress.” Huang does not fret into her hand, black-capped, black-masked,
that it may eat all of our jobs, or replace artists, or go bobbing one birdseed at a time—I remember
rogue and decide to wipe out humanity.
In fact, when Witt, stricken with existential anxi- in cold Amherst how they’d fill the lonely feeder
ety about how AI will change the world, asks Huang just outside the kitchen window, especially
whether some of these concerns might be worth pon- when the ice mixed in with snow would slap
dering, he is subjected to one of his legendary tirades: the double glass, shake it a little, and start to sing.
“Is it going to destroy jobs?” Huang asked, his voice
One wearies of the sublime, the great deep thing,
crescendoing with anger. “Are calculators going to the red-tailed kiting hawk sliding down the sky
destroy math? That conversation is so old, and I’m to make the kill, the sky itself changing on its own,
so, so tired of it,” he said. “I don’t want to talk about
depth of feeling depth of field. Margaret sitting still,
it anymore … We make the marginal cost of things
zero, generation after generation after generation, and
pieces of the sun falling in the shadows all around her,
this exact conversation happens every single time!” while my bright chickadees are braced against the wind,
feathers fluffed, each of them so small I could wrap one
You could write this off as an example of Upton in my fist to keep it warm, alive, then suddenly gone.
Sinclair’s adage “It is difficult to get a man to under-
stand something, when his salary depends upon his All winter in the snow depths just outside you live
not understanding it!” But the fact that Huang talks in separations made of glass—I’d never have
about AI in terms of its impact on “marginal costs” the patience to hold out my hand and wait out
shouldn’t be reduced to mere opportunism: It fits right a bird, regardless of how beautiful the weather.
in with the single-minded focus on performance that
has driven him from Nvidia’s beginning. Witt at one
point calls Huang a “visionary inventor.” The vision
Huang has been in thrall to, though, seems to be less
about grand future goals, and more about tools—about Stanley Plumly (1939–2019) was the author of numerous books
making the fastest, most powerful chips as efficiently as of poetry and prose. His posthumous collection Collected Poems
possible. “Existential risk” has no place in that vision. will be published in August.
Huang’s unapologetic stance on AI is bracing in its way,
especially in contrast with the public hand-wringing of
many AI chieftains, fretting about the dangers of their
LLMs while continuing to develop them. But he is in
effect making the biggest, riskiest bet ever—not just
for Nvidia, but for all of us. Let’s hope he’s right.
James Surowiecki is a contributing writer at The Atlantic
and the author of The Wisdom of Crowds.
83
