Chapter 1 of Hardware: Chips, Cloud, and the AI Compute
Founded by MIT researchers in 2017, Lightmatter is a photonic computing company that utilizes light to accelerate AI workloads, addressing the speed, power and energy limitations of traditional silicon-based GPUs in AI computing. It is a scale‑up photonics player now explicitly recognized by NVIDIA’s NVLink Fusion ecosystem, sitting in a strategically important layer between GPUs and switches.
Industry context
At COMPUTEX 2025, Jensen Huang’s keynote formally positioned NVLink Fusion as the interconnect backbone for AI factories, with an ecosystem roadmap that includes co‑packaged optics (“CPO”) at rack scale. This marks the moment when “scale‑up” networking inside AI racks begins to open a door to external optics, and explains why Lightmatter joining that ecosystem is strategically important.
There are two different kinds of connections inside an AI data center:
- Scale‑up: tightly coupling many GPUs into one “giant chip” that behaves as if it shares memory; distances are short (inside or across adjacent racks), latency and bandwidth requirements are extreme, and historically this has been almost entirely copper (NVLink backplanes and cables).
- Scale‑out: connecting those “giant chips” into clusters across racks using Ethernet or InfiniBand; distances are longer, node counts large, and pluggable optical modules have long dominated this layer.
Historically “scale‑up” remained a closed copper world owned by NVIDIA’s own NVLink ecosystem, until COMPUTEX 2025. Since then, independent packaged‑level optics vendors like Lightmatter are publicly listed as NVLink Fusion “optical interconnect partners” for scale‑up.
There are five layers of optical value:
- Scale‑out pluggable modules (many Chinese players)
- Scale‑out optical switches (NVIDIA, Broadcom)
- Scale‑up cross‑rack CPO/NPO (NVIDIA, Marvell)
- Packaged‑level optical I/O chiplets (Lightmatter, Ayar Labs)
- Photonic fabric / memory interconnect (Marvell–Celestial)
Layers 3–5 (especially 4) are where the real structural change is; opening NVLink Fusion to optics is “a crack in a once‑closed door,” but NVIDIA still chooses who gets through and at which layer.
In June, Nvidia and Lightmatter jointly announced Lightmatter’s entry into the NVLink Fusion ecosystem while Lightmatter remains as a dedicated, pure‑play supplier.
Lightmatter: Funding history
Lightmatter reached a $4.4bn valuation during its Series D funding round in 2024. The company has raised roughly $850mn of capital from prominent investors including Sequoia, Google Ventures, Matrix Partners, Viking Global, Lockheed Martin, HP, Fidelity, T. Rowe Price, and others.
Business and product model
- A scale‑up photonics player now explicitly recognized by NVIDIA’s NVLink Fusion ecosystem, sitting in a strategically important layer between GPUs and switches.
- A technology leader in single‑fiber bandwidth density (16λ, BiDi, 800G/1.6T per fiber), aiming to radically shrink the fiber and connector footprint in large AI clusters.
- A proponent of 3D photonic interposers and edgeless I/O, addressing the fundamental perimeter vs area scaling mismatch in chip I/O.
- An innovator in VLSP laser integration (Guide), tackling the “power wall” of discrete laser modules and enabling dense CPO/NPO deployments. CEO Nick Harris pointed out that AI “world‑model” and MoE infrastructure needs semiconductor‑grade integration not only in compute and interconnect but also in light sources; scalable lasers are essential for scalable CPO.
- A company that has moved beyond academic demos toward an HVM‑capable stack, but still has to prove full BiDi PAM4 performance, complete engineering margins, and long‑term reliability at hyperscaler volumes.
Lightmatter has established a broad network of strategic partnerships focused on scaling its photonic interconnect technology, particularly for next-generation AI and high-performance computing.
Why It Matters Now
Into 2026, a full ecosystem is emerging: foundries (TSMC, ASE, Tower, etc.), OSATs, ASIC vendors, XPU and switch vendors, and optical component suppliers. 3D silicon photonics is now technically and industrially mature enough to support the “next 1000×” in AI compute.
Silicon photonics and CPOs (“co‑packaged optics”) are entering a commercial adoption window in 2026: Nvidia Spectrum-X / Quantum-X Photonics ship H2 2026, Broadcom positions a 200T optical AI era at OFC 2026, the OCP CPO collaboration is published, and Lightmatter has now packaged Passage into an NPO/OBO + 3D-CPO product family. Market size for silicon photonics is projected to grow from $2.2B to $9.7B annually by 2030 (~30% CAGR).
Key Observations
- Revenue Shifting from Experimental to Infrastructure
Lightmatter’s revenue has moved out of pure R&D and into early infrastructure deployment in 2024, with estimated annual revenue of about $116mn (+130% year‑on‑year). - Structural Business Model Advantage
Rather than building GPUs, Lightmatter sells the photonic interconnect and optical‑engine layer that every large AI cluster needs, positioning it to ride demand from multiple accelerator vendors and cloud providers simultaneously. The company’s Passage platform spans Near‑Packaged Optics (NPO), Co‑Packaged Optics (CPO), and 3D photonic interposers, enabling bandwidth densities and energy per bit that conventional electrical signaling cannot match at multi‑thousand‑XPU scale, and reducing fiber/connector counts by up to about 50% in NVLink Fusion configurations. This “picks‑and‑shovels” position, combined with retained IP on silicon‑photonics devices and packaging (“Passage” and “Guide” as reusable platforms), offers a path to datacenter‑class hardware economics without directly competing with NVIDIA and similar giants on core compute. - Ecosystem Validation + Scaling Capital
In June 2026, Lightmatter joined NVIDIA’s NVLink Fusion ecosystem, adapting its bi‑directional Passage CPO/NPO architecture to NVIDIA’s SerDes and optical specs and becoming, by NVIDIA’s own ecosystem commentary, a reference supplier for photonic interconnect in semi‑custom AI factories. This combination of a $4.4bn valuation, $850mn of cumulative funding by prominent investors, and integration into NVIDIA’s flagship interconnect roadmap is the core “why now” for the story. - AI Infrastructure Spend Moving Toward Photonic Interconnect
The projected cumulative AIDC spend this decade runs into trillions of dollars. The AIDCs are increasingly constrained by power and I/O, not raw FLOPs, and NVLink Fusion’s move to explicit photonic support is a concrete signal that large customers are budgeting around optical interconnect. Lightmatter’s passage‑plus‑guide stack is designed to let customers build scale‑up domains of “thousands of XPUs” per cluster within the NVLink Fusion standard, which is precisely where hyperscalers are concentrating capex for frontier model training.
As frontier AI labs race toward ever larger “world‑model” and MoE infrastructures, the bottlenecks are shifting from raw compute to power, bandwidth, and I/O topology. In that world, companies like Lightmatter don’t compete to build the smartest GPU; they compete to make sure thousands of GPUs can actually talk to each other within realistic power budgets. NVLink Fusion’s embrace of photonic partners is an early, concrete signal that scale‑up interconnect is moving from a closed, copper‑only domain to a more open, optical‑first fabric. For investors and operators, the implication is straightforward: if AI factories are the new data centers, then photonic scale‑up interconnect is part of the “steel and concrete” that everything else rests on.
As a leading player in photonic computing, Lightmatter sits in a structurally interesting spot. It is a picks‑and‑shovels supplier to the AI capex cycle with deep IP in silicon photonics, a growing base of infrastructure revenue rather than pure experimentation, and a seat at NVIDIA’s interconnect table without competing head‑on in GPUs. The story is still early — the company must prove long‑term reliability at hyperscale and navigate an ecosystem where the platform owners ultimately decide who gets through the door. But as AI infrastructure spend tilts toward solving bandwidth and energy walls, photonic interconnect players like Lightmatter are positioned to become quiet, critical beneficiaries of the next leg of AI’s growth.
