Lightmatter has raised $400 million to push photonic computing deeper into AI infrastructure. The funding highlights a practical issue inside modern data centers: adding more GPUs is not enough if those chips cannot exchange data fast enough to stay busy.
The company’s pitch is centered on optical interconnects. By using photonic chips, Lightmatter says it can let hundreds of GPUs work synchronously and reduce one of the costly barriers to training and running AI models at very large scale.
The bottleneck behind bigger AI clusters
AI has increased demand for compute across the data center industry, but the source article makes clear that scale is not simply a matter of installing another thousand GPUs. In high-performance computing, a fast processor can still be underused if it spends too much time waiting for data from other parts of the system.
That is why the interconnect layer matters. It is the networking fabric that helps many CPUs and GPUs behave less like separate boxes and more like one large machine. If that layer is slow, the wider data center cannot fully benefit from the speed of each individual node.
Lightmatter’s answer is to use photonic chips it has been developing since 2018. Instead of relying on conventional switching alone, the company routes large amounts of data through fiber and a purely optical interface.
How Lightmatter frames the optical advantage
Nick Harris, Lightmatter’s CEO and founder, described the challenge in direct terms to TechCrunch: “Hyperscalers know if they want a computer with a million nodes, they can’t do it with Cisco traditional switches. Once you leave the rack, you go from high-density interconnect to basically a cup on a string,”
In the source article, Harris pointed to NVLink and especially the NVL72 platform as the state of the art. That platform wires together 72 Nvidia Blackwell units in a rack and can reach a maximum of 1.4 exaFLOPs at FP4 precision. But the rack still has to send data through 7 terabits of “scale up” networking.
That number is large, but the source frames it as a constraint when clusters need to connect racks to other racks. Harris said that, for a million GPUs, multiple layers of switches are needed, which increases latency. He also described repeated conversions from electrical to optical and back again as costly in both power and waiting time.
Lightmatter’s alternative is fiber at much higher density. The article says the company can support up to 1.6 terabits per fiber by using multiple colors, and up to 256 fibers per chip. Its current photonic interconnect does 30 terabits, while its on-rack optical wiring is designed to let 1,024 GPUs work synchronously in specialized racks.
Why hyperscalers are paying attention
The market Harris is describing is made up of major data center companies with a strong appetite for compute. The source names Microsoft, Amazon, xAI and OpenAI as examples of organizations tied to that broader demand.
Harris said many hyperscalers are already Lightmatter customers, though he did not name them. He compared Lightmatter’s role to a foundry such as TSMC, saying the company does not choose favorites or attach its name to other companies’ brands.
That positioning matters because the company is not presenting the optical interconnect as a single branded supercomputer. It is presenting it as a platform and roadmap that others can build around as AI clusters become larger and more complex.
The funding round also signals how investors are reading the opportunity. Lightmatter’s $400 million D round values the company at $4.4 billion. Harris said that makes it “by far the largest photonics company.”
The round was led by T. Rowe Price Associates, with participation from existing investors Fidelity Management & Research Company and GV.
What comes after interconnect
Lightmatter is not stopping at the interconnect layer described in the article. The company is also developing new substrates for chips, with the aim of handling more networking tasks using light at a closer level inside computing systems.
Harris suggested that power per chip will become a major differentiator. He also predicted that wafer-scale chips will become common across the industry within 10 years because, in his view, there is no other path to improving performance per chip.
The broader point is that chip performance is increasingly tied to movement of data, not only raw calculation. As clusters grow, the cost of waiting, switching and converting signals can become a defining limit.
That is why Harris ended with a larger claim about where the industry is going: “Ten years from now, interconnect is Moore’s Law,”
The takeaway for AI infrastructure
Lightmatter’s raise is important because it puts money behind a specific view of the future data center. In that view, the winning systems are not just those with the most GPUs, but those that can keep those GPUs coordinated at scale.
Photonic data centers remain a specialized subject, but the logic is straightforward. If AI models need larger clusters, and larger clusters create more pressure on networking, then optical interconnects become a strategic part of compute performance.
Lightmatter’s bet is that fiber, photonic chips and high-capacity optical interfaces can remove enough friction to change how hyperscalers build AI systems. The $400 million round shows that investors and customers are taking that bet seriously.