Lightmatter Enters NVLink Fusion: Co-Packaged Optics Moves from Roadmap to Bill of Materials

The Signal

On June 3, 2026, at Computex Taipei, Lightmatter announced it has joined the NVIDIA NVLink Fusion ecosystem to accelerate the deployment of high-performance optical connectivity for AI infrastructure. That closes a debate running through every major cluster design review for the past two years: optical interconnects at the package level are no longer theoretical. They are a current procurement decision.

The forcing function is physics. The old data-center scaling playbook — faster GPUs, higher SerDes rates, increasingly aggressive board designs — is breaking down as AI clusters push beyond 100 Tb/s per node. The gap between what silicon can generate and what traditional copper interconnects can deliver is widening. At 400 Gbps and beyond, conventional copper methods are hitting physical and energy efficiency limits.

What Lightmatter Actually Delivers

Lightmatter will deliver Co-Packaged Optics (CPO) and Near-Packaged Optics (NPO) products compatible with NVIDIA's optical and SerDes technologies. The vehicle is the Passage platform. The company unveiled its Passage L20 universal optical engine, featuring 12.8 Tbps bandwidth and 30-watt power consumption, with mass production slated for Q1 2027.

The roadmap is explicit: near-packaged optics in 2027, co-packaged optics in 2028, and photonic interposers and optical engines beyond 2029. NPO first lowers integration risk while still cutting electrical path length compared to pluggable modules.

The efficiency argument is concrete. Lightmatter's bi-directional optical link architecture reduces fiber and connector requirements by 50%. On the laser side, a single Guide 1 module enables 51.2 Tbps of I/O on a 16-wavelength DWDM grid centered at 1310 nm, replacing nine conventional ELSFP laser modules with one chip. Real bill-of-materials line reductions.

Three generations of Lightmatter silicon operate in the company's validation data center: bidirectional 800 Gbps per fiber on Passage EVK50, 1.6 Tbps per fiber on Passage EVK100, and 114 Tbps of aggregate bidirectional bandwidth on the M1000 reference platform at 2.3 pJ/bit, including laser power. That 2.3 pJ/bit figure is the metric operators actually use to compare interconnect options at cluster scale.

Why the NVLink Fusion Endorsement Matters

Co-packaged optics has been technically credible for three years. What it lacked was a reference architecture anyone had to design against. NVLink Fusion provides that.

The technology enables semi-custom XPUs to connect with NVIDIA switch silicon through Lightmatter's products, creating seamless high-bandwidth connectivity for chips from multiple suppliers. NVLink Fusion is NVIDIA's opening to third-party silicon — a way for custom ASIC shops to plug into NVIDIA's switching fabric rather than build an entirely independent interconnect stack. By becoming the validated optical layer for that ecosystem, Lightmatter is not selling components. It is becoming infrastructure.

650 Group analyst Alan Weckel called this "a critical milestone in the maturation of co-packaged optics," noting that Lightmatter's compatibility with NVIDIA's high-speed interconnect significantly expands the addressable market for CPO products, and that the partnership "provides a validated blueprint for hyperscalers to overcome traditional I/O bottlenecks and scale AI clusters."

The ecosystem lock-in runs both directions. Optical suppliers must now certify against NVIDIA's reference rather than build independently, raising the barrier for new entrants and compressing the window for incumbents without qualification.

What the Transition Actually Looks Like

Copper does not disappear. The increase in bandwidth requirements is not leading to the replacement of copper with optical, but rather to a more segmented and optimized network architecture, with passive copper, active copper, pluggable optics, and co-packaged optics each playing a role across the stack. Short in-rack connections stay copper. The optical layer grows as cluster size and inter-rack distances increase.

By placing optical components next to the chip or network switch ASIC rather than in separate pluggable modules, CPO minimizes the distance electrical signals must travel before conversion to light, dramatically reducing power consumption and signal losses while increasing data speeds and lowering latency.

The binding constraint, as Lightmatter's own engineering documentation acknowledges, is electrical bandwidth density at the photonic interface, not the optical medium itself. Photonics is not the limiter: a single fiber carries terabits with dense wavelength division multiplexing, and photonic chip surface density runs well beyond that. The bottleneck is how much electrical bandwidth you can deliver to the photonic interface. That is where the engineering work actually sits.

Execution Risk

Lightmatter's risk shifts from technical validation to manufacturing scale. The company produces its VLSP lasers using a 300mm CMOS process, directly transferring indium phosphide structures onto silicon wafers to boost manufacturing efficiency. Whether that process yields at production volumes is uncertain. CPO assembly is complex; meeting market expectations requires more than performance demonstrations, and CPO adoption depends on proving robust, multi-vendor business models along with clear advantages in cost, power, and scalability at the system level, as key ecosystem elements including standardized optical interfaces and interoperable test and qualification flows are still coming together.

If Lightmatter hits the Q1 2027 NPO mass production date and hyperscalers begin qualifying at scale, optical interconnect requirements could migrate into mid-tier cluster designs — those below 10,000 GPUs — within 18 to 24 months as cost per bit normalizes.

What to Watch

1. Q1 2027 NPO production date: Lightmatter's stated milestone for Passage L20 mass production. Any slip reshuffles the procurement timeline for clusters targeting 2027-2028 deployments.
2. Cloud provider qualification timelines: Which hyperscaler files the first public CPO cluster specification referencing NVLink Fusion optical compatibility.
3. Competing optical qualifications: Whether Broadcom or other photonics vendors pursue NVLink Fusion certification, or build parallel optical interconnect ecosystems outside NVIDIA's reference.
4. CPO as hard requirement vs. option: Watch NVIDIA's specification language — whether Fusion optical becomes mandatory for clusters above a certain GPU count, or remains an optional upgrade.
5. Energy per bit benchmarks in production: Lab figures at 2.3 pJ/bit need to hold in operational clusters with real thermal and signal environments. First production data from a named customer will be the credibility event.