2026-06-08 — views $NVDA · NVIDIA · NVLink Fusion

Co-packaged optics enters NVLink Fusion: Ayar Labs and Lightmatter join NVIDIA's scale-up ecosystem in the same week

In two announcements one day apart, Ayar Labs (June 2) and Lightmatter (June 3, 2026) joined NVIDIA's NVLink Fusion ecosystem, making their co-packaged-optics engines compatible with NVIDIA's SerDes and switch silicon. It signals photonics moving into the scale-up domain that

What shipped

In the span of two days, the two best-known silicon-photonics startups both planted a flag inside NVIDIA’s scale-up interconnect program. On June 2, 2026, Ayar Labs announced it had joined the NVIDIA NVLink Fusion ecosystem, making its co-packaged optics (CPO) “optically and electrically compatible with NVIDIA optical and SerDes technologies.” One day later, on June 3, Lightmatter announced the same step for its Passage CPO and near-packaged-optics (NPO) products.

Neither of these is a product launch with a datasheet attached. They are ecosystem-compatibility announcements — the unglamorous but load-bearing kind of news in interconnect land. What they signal is more interesting than either company’s individual roadmap: light is being invited into the scale-up domain, the rack-internal GPU-to-GPU fabric that copper and NVLink have owned outright.

A quick frame for why that matters. NVLink Fusion is the program NVIDIA unveiled at COMPUTEX on May 18, 2025, to let partners build semi-custom AI infrastructure that still speaks NVLink. The fifth-generation NVLink platform NVIDIA quotes delivers “a total bandwidth of 1.8 TB/s per GPU — 14x faster than PCIe Gen5.” Fusion opens that fabric to outside silicon. The original partner list was a who’s-who of XPU and IP vendors:

Partner	Role in NVLink Fusion
Marvell	Custom XPUs + scale-up networking ($2B NVIDIA investment, Mar 31, 2026)
MediaTek, Alchip	Custom AI compute / ASIC design
Astera Labs	Connectivity silicon
Synopsys, Cadence	NVLink Fusion IP for custom designs
Ayar Labs (Jun 2, 2026)	Co-packaged optics — TeraPHY + SuperNova
Lightmatter (Jun 3, 2026)	Passage CPO / NPO photonic engines

The bottom two rows are the new entries, and they change the character of the list. Everyone above them moves electrons. Ayar Labs and Lightmatter move photons, inside the package.

Why now: copper is running out of room

The scale-up network is where you get the eye-watering bandwidth numbers, and it has historically been copper because copper inside a rack is cheap, reliable, and low-latency. The catch is reach. Copper electrical signaling at the data rates a 1.8 TB/s-per-GPU fabric demands tops out at roughly one to two meters of usable reach, which is exactly why a scale-up domain has been capped at a rack or two and on the order of dozens of GPUs. Push the GPU count up, push the per-lane rate toward 800 Gb/s, and the bandwidth-per-watt math on copper gets ugly fast.

That is the wall both startups are positioning against. Ayar Labs is bringing its TeraPHY Optical I/O Chiplet and SuperNova Light Source — a chiplet-plus-external-laser architecture meant to sit in-package next to the accelerator. CEO Mark Wade framed it as “introducing co-packaged optics as a foundational building block for customers deploying heterogeneous compute in NVIDIA AI factories.” Lightmatter’s pitch leans on density: it claims its Passage approach delivers a “50% reduction in fiber and connector requirements,” with a bi-directional optical link architecture adapted to be compatible with NVIDIA’s optical and SerDes technologies so semi-custom XPUs can connect to NVIDIA switch silicon. NVIDIA’s Ashish Karandikar gave both the same blessing — more “choice and flexibility” for partners building heterogeneous infrastructure.

The common thread: optics is no longer just the scale-out story (rack-to-rack, switch-to-switch). It is being engineered into the scale-up package, where the IP value of the SerDes and the optical interface meets.

The IP read

For a connectivity-IP watcher, two things stand out.

First, the unit of competition is shifting from the SerDes lane to the electro-optical interface. NVLink Fusion is, at bottom, a compatibility envelope around NVIDIA’s SerDes. By certifying their optics against that envelope, Ayar Labs and Lightmatter are betting the defensible IP is the chiplet-to-laser-to-fiber path, not the copper PHY. That is a different competitive surface than the one Astera Labs (retimers, PCIe switching) and Credo (active electrical cables) have been winning on with copper-extension products.

Second, NVIDIA is curating an optics supply chain the way it curated an XPU supply chain. Pairing these June compatibility announcements with the March 31, 2026 Marvell partnership — a $2 billion NVIDIA investment for custom XPUs and NVLink Fusion scale-up networking — you can see the shape: NVIDIA keeps the fabric protocol, and invites a competitive field of compute and now optics vendors to plug into it. Choice for customers, gravity well for NVIDIA.

Practitioner note

If I were sizing this as an infrastructure buyer rather than a stock, I would treat these announcements as a “watch, don’t reprice” signal. They are compatibility statements, not GA hardware with power and bandwidth numbers I can put in a TCO model — and the public copy from both companies is conspicuously light on lane rates and per-port figures. So my move would be to ask vendors three concrete questions before believing the roadmap: (1) what is the actual per-fiber and per-package aggregate bandwidth, (2) what is the picojoule-per-bit at that rate versus the copper baseline I am replacing, and (3) what is the laser-failure story, because an external light source is a new serviceable failure domain a passive copper cable simply does not have. Until those answers exist on a datasheet, I would keep designing the near-term rack on copper and treat in-package optics as the Rubin-generation refresh, not this year’s build.

Under-considered angle

The quiet risk here is reliability economics, not bandwidth. Copper scale-up links are boring on purpose — they fail rarely and predictably. Co-packaged optics introduces lasers, fiber attach, and optical alignment into the single most thermally and mechanically stressed spot in the system: right next to the GPU. The headline “50% fewer fibers and connectors” is partly a reliability argument dressed as a density argument — fewer mated connections is fewer things to go wrong. But the industry has not yet shown, in volume and in the field, that an in-package optical engine survives years of thermal cycling at the same service rate as a copper trace. Whoever proves field reliability first — measured in FIT rates and laser-replacement procedures, not demo bandwidth — likely wins the scale-up optics socket, regardless of who has the prettiest pico-joule-per-bit slide today. That is the metric I would watch, and it is the one nobody is publishing yet.