Lumentum is pitching gallium arsenide VCSEL light sources for co-packaged optics (CPO) alongside its higher-power and multi-wavelength indium phosphide lasers.
Adopting VCSELs also addresses an ongoing laser supply constraint. Every laser Lumentum sells into AI data centres is indium phosphide-based, whether externally modulated lasers (EMLs), continuous-wave lasers for silicon photonics-based transceivers, or the ultra-high-power (UHP) lasers used in co-packaged optics.
“Despite best efforts from companies like Lumentum to ramp indium phosphide as fast as possible, it has been constrained,” says Rafik Ward, Senior Vice President, Chief Strategy Officer and Chief Marketing Officer at Lumentum. “And it’s difficult to point to a time frame when those constraints come off.”
The pressure explains Nvidia’s multi-year agreement with Lumentum, combining a $2 billion investment, a multi-billion-dollar purchase commitment and support for a new US manufacturing facility. Nvidia and Lumentum are also collaborating on next-generation optics applications.
“We supply what we call the UHP, the ultra-high-power laser, to Nvidia, which is the prime laser source for CPO,” says Ward. “With all of the demand that we see, there was a need to continue to scale [the output power of] that laser.”
VCSEL-based co-packaged optics
Unlike indium phosphide lasers, VCSELs are fabricated on gallium arsenide, using a manufacturing base that Lumentum has for consumer 3D sensing.
“We have shipped over two billion VCSEL arrays to date,” says Ward. “Each array typically has hundreds of VCSELs inside, so we’re talking about hundreds of billions of VCSELs shipped.”
Moreover, the VCSEL production line has excess capacity available today. “What the VCSEL does, first and foremost, is open up this alternate supply chain,” says Ward.
At the OFC conference and exhibition event earlier this year, Lumentum demonstrated a VCSEL-based co-packaged optics platform to show that a gallium arsenide light source can satisfy the requirements of future co-packaged optical systems.
The VCSELs operate at 1060nm making them tolerant to elevated operating temperatures. The VCSELs can be integrated directly beside the host ASIC rather than relying on remotely located external laser sources.
The VCSELs are used in a “slow-and-wide” architecture: each fibre carries 32 gigabit-per-second (Gbps) per lane rather than a “faster-and-narrow’ 200Gbps or 400Gbps line rates.
Operating at 1060nm rather than the traditional 850nm substantially improves VCSEL reliability, Ward says, given the device’s tolerance of the thermal environment while total energy consumption is around 2–2.5pJ/bit. The reach is some 100m.
Lumentum demonstrated a bridge-type host ASIC surrounded by eight VCSEL arrays. Each array provides 10Tbps of bidirectional bandwidth (5Tbps in each direction), delivering 80Tbps around the package. That is around 160 VCSELs per array, not counting redundancy, and 320 fibres.
The demonstration measured four VCSELs; showing more was largely a question of finding enough oscilloscopes, quips Ward.
The approach inevitably requires significantly more fibre. The trade-off, however, is lower electrical signalling complexity. Relaxing the serialiser-deserialiser (SerDes) requirements reduces overall power consumption compared with high-speed “fast-and-narrow” architectures operating at 200 or 400Gbps.
“There’s more than one way in which customers want to architect these systems,” says Ward. “We see customers looking for both: wide-and-slow, and fast-and-narrow.”
Although the demonstration was not tied to a specific protocol, Lumentum sees UCIe and PCIe interfaces for chip-to-chip links as natural applications.
Ward also positions VCSELs ahead of a third approach based on microLEDs under development for similar applications. MicroLED interconnects, he argues, typically operate an order of magnitude slower, requiring substantially more fibres.
VCSEL arrays also offer system fault tolerance through sparing. By implementing sparing, reliability improves significantly. “When we’ve run the numbers, it’s frankly amazing: the reliability is measured in multiples of a human lifespan.”
For now, however, the work remains a technology demonstration rather than a product.
“For this type of environment, we need to work very closely with customers to co-design a solution like this into their host products,” says Ward.
Lumentum is working with several lead customers.
Doubling laser output power
Lumentum continues to extend its indium phosphide laser roadmap.
The company introduced a 400mW single-wavelength 1310nm ultra-high-power laser a year ago that is now entering high-volume production. Eight of these devices populate an external laser source (ELSFP) module used to power such designs as a co-packaged optics switch chip. Lumentum has now shown an 800mW version.
Such high-power lasers allow you to do two things, says Sri Venkat, senior vice president and general manager at Lumentum.
“It gives you more scalability with the ASIC going to higher speeds, as well as reducing the number of lasers, which reduces the number of fibres, the number of optics, isolators and so on,” he says. In other words, reducing cost per transmitted bit.
However, achieving the power doubling is far from straightforward. Lumentum scaled its continuous-wave silicon photonics laser platform from 100mW to 400mW before enlarging the die, redesigning the quantum-well structure and carefully managing thermal behaviour and wafer-level reliability.
“We found a recipe that allows us to scale power linearly,” says Venkat.
Lumentum expects to scale the power further, but beyond a certain die size, reliability, thermal behaviour, and manufacturability start to compete with one another.
“I don’t think you should expect this to double again in a year [beyond the 800mw]. But there is a possibility that we can extend it further,” says Venkat.
Wavelength count
Lumentum is also extending its external laser source technology to support wavelength-division multiplexing (WDM).
Using the same UHP laser architecture, the company has demonstrated two external laser source modules producing a combined 16 wavelengths around 1310nm. Each wavelength is spaced at 200GHz with a wavelength positional accuracy of ±25GHz.
“What that does is give the co-packaged optics application multiple wavelengths through the same fibre to the ASIC, and it gives you much higher bandwidth density,” says Venkat. “It’s a natural progression in the market.”
There are several ways data rates can be increased down a fibre, but WDM is seen as the main scaling approach for co-packaged optics.
Optical circuit switching ramps
Alongside lasers, Lumentum says optical circuit switching (OCS) is another major emerging AI business.
The company disclosed a $400 million optical circuit switching backlog scheduled for delivery during the second half of calendar 2026, and expects the business to exceed a $1 billion annual run rate during 2027. “Just that math gives you some insight into the kind of continued ramp we’re seeing,” says Ward.
At OFC, Lumentum demonstrated its R300 300×300-port optical circuit switch jointly with Marvell that provided a variety of optical modules and telemetry equipment.
The R300 switch is wavelength agnostic, supporting O-band, C-band and L-band operation, and WDM systems.
“When we put our OCS in front of customers, the key architects start thinking about it as a new tool,” says Ward. “It’s so ubiquitous that it can be used in different networks—everything from scale-across to scale-out to scale-up.”
For scale-out networks, optical circuit switching can potentially eliminate an entire switching layer. In scale-up clusters, it enables dynamic bandwidth allocation, directing additional optical bandwidth toward groups of AI accelerator chips (xPUs) running demanding AI workloads while routing around failed xPUs or links.
Standardisation
Lumentum views recent multi-source agreements and renewed industry interest in near-package optics (NPO) as signs that optical architectures are finally converging.
“There’s developing consensus among the key hyperscalers and AI accelerator vendors on the mechanical and technical characteristics they want to see, and it helps bring the entire technology supply chain along,” says Venkat.
Venkat believes co-packaging is finally entering commercial deployment after years of anticipation.
“Last year was a coming-out year, where we started to see the technology ramp,” says Venkat. “What we’re now seeing is real legs, both for scale-out and scale-up.”
Lumentum’s strategy is to assemble multiple options: more powerful indium phosphide lasers, multi-wavelength external laser sources, and now gallium arsenide VCSELs.
And with VCSELs, the industry now has a second laser manufacturing ecosystem at a time when indium phosphide capacity remains under sustained pressure, says Lumentum.