Nvidia-backed CScale emerges from stealth with $145M to build optical interconnects for gigawatt-scale AI

CScale emerged from stealth Wednesday with $145 million in Series C funding and backing from Nvidia and Intel Capital, betting that the next bottleneck in giant AI data centers will not be finding more chips. It will be keeping thousands of those chips talking to each other when parts of the network inevitably fail.

The Palo Alto startup is building optical interconnect technology for what the industry calls AI scale-up, linking large numbers of accelerators across racks so they can operate as one tightly connected computing system.

The new round brings CScale’s total funding to $188 million. Atreides Management, Valor Equity Partners, and Premji Invest co-led the financing. Longtime backer Sutter Hill Ventures participated alongside Maverick Silicon. Nvidia and Intel Capital joined as CScale’s first strategic investors.

The timing puts CScale in the middle of one of the biggest infrastructure shifts created by AI. The largest planned AI data centers are moving toward gigawatt-scale deployments, where thousands of accelerators may need to exchange enormous amounts of data with extremely low latency.

That creates a problem that becomes harder as clusters grow. A laser or optical link may fail rarely at the component level. Across hundreds of thousands of components, failures become a normal operating condition.

CScale says its architecture is being built so an optical failure can be contained without stopping the connected compute.

“As AI scale-up domains extend across dozens of racks, optical interconnect becomes essential,” CScale CEO Martin Lund said. “At that scale, reliable, predictable communication is fundamental to system economics. Easier part replacement improves serviceability, not continuity. We’re designing the interconnect for continuity. Lasers will fail. Compute shouldn’t.”

AI’s bottleneck is moving beyond the GPU

That last sentence gets to the heart of CScale’s pitch.

AI infrastructure spending has centered heavily on accelerators from Nvidia, AMD, and other chipmakers. Packing more accelerators into a data center solves only part of the problem. Those processors have to move data between one another quickly enough to stay productive.

As AI clusters spread across more racks, conventional electrical connections face greater pressure from bandwidth, distance, latency, and energy demands. Optical networking is becoming a larger part of the answer.

CScale is developing an integrated light engine to make those optical connections more resilient. The company argues that replaceable lasers address maintenance after something breaks, rather than keeping the computing system operating through the failure itself.

“AI infrastructure is no longer just a compute problem. It is a systems problem,” Atreides managing partner Gavin Baker said. “As scale-up systems move toward gigawatt-class deployments, interconnect bandwidth means nothing without system reliability. Every optical failure is a compute failure. At this scale, failures are not hypothetical—they are inevitable.”

CScale has some experienced hands behind that bet.

Lund previously built Broadcom’s switching business into a billion-dollar operation and later held senior positions at Microsoft, Cadence, and Cisco. At Cisco, he led the Common Hardware Group responsible for silicon, hardware systems, and optics, including Silicon One.

Founder and CTO Sanjai Kohli has spent decades building communications technology. He co-founded SiRF, the semiconductor company credited with helping push GPS into mass-market devices, and later founded Inovi, which Facebook acquired in 2014.

Kohli founded CScale in 2023. The company now employs about 85 people globally.

For CScale, the larger opportunity rests on a simple premise: the economics of enormous AI clusters depend on more than how many accelerators operators can install. They depend on how much of that expensive compute stays working.

At gigawatt scale, a tiny failure rate can become an everyday problem. CScale is betting that keeping those failures from spreading could become one of the defining infrastructure challenges of the AI buildout.