CScale Raises $145M for Resilient AI Optical Interconnect

The economics of a gigawatt-scale AI system depend on more than buying enough accelerators. Those processors must remain productively connected across a scale-up domain even when ordinary optical components fail inside an increasingly large machine.

CScale has raised a $145M Series C to work on that problem. Atreides Management, Valor Equity Partners, and Premji Invest co-led the September 30, 2026 financing, which brings CScale's total reported funding to $188M.

The Palo Alto company exited stealth with an integrated light engine and an architectural promise: optical failures should be contained without interrupting the compute around them. If CScale can carry that promise into commercial systems, optical interconnect will compete on continuity and system economics alongside bandwidth, latency, reach, and power.

What CScale Announced

The $145M Series C includes continued support from Sutter Hill Ventures and Maverick Silicon. NVIDIA and Intel Capital joined as CScale's first strategic investors, placing 2 major semiconductor organizations alongside the financial firms backing the company's move from stealth toward commercialization.

CScale did not disclose its valuation, detailed deal terms, named customers, or commercial commitments associated with the strategic investments. The company says the new funding will accelerate development and commercialization of its optical-interconnect technology. It reports $188M in total funding, which means approximately $43M predates this round, although a verified round-by-round breakdown was not provided in the announcement.

The company was founded in 2023 by Sanjai Kohli and is headquartered in Palo Alto, California. CScale says it now has approximately 85 employees globally, giving the business a substantial engineering base before public product shipment.

The Reliability Problem Inside AI Scale-Up

Scale-up interconnect is the communication fabric that allows many accelerators to work together like a much larger computer. As systems extend across thousands of tightly coupled processors and dozens of racks, they require high bandwidth, predictable low latency, and continuous communication. A link problem can leave expensive compute unable to contribute to the work it was purchased to perform.

CScale's thesis begins with fleet math. A failure that is rare on one link becomes a recurring operating condition across hundreds of thousands of accelerators. The company's architecture is designed to contain an optical failure so that the surrounding compute can continue working instead of turning one failed laser into a broader interruption.

That distinction moves optical reliability from maintenance into utilization. A replaceable part may improve serviceability after a failure, but AI infrastructure operators also care about how much useful compute disappears before that replacement occurs. CScale is trying to make continuity part of the interconnect design rather than an after-the-fact repair process.

What CScale Is Building

CScale describes its product as an integrated light engine for optical connectivity at gigawatt scale. In an interview reported by Reuters, CEO Martin Lund said the company intends to integrate lasers onto chips while preserving data flow when an individual laser fails. He also said CScale is using established underlying technologies assembled in a different architecture and plans to ship chips by 2028.

The company has released few implementation details. It has not publicly supplied bandwidth, latency, power, manufacturing, packaging, compatibility, or production-reliability benchmarks that would allow an independent comparison with other optical-interconnect approaches. The 2028 shipment date is a company plan, not a completed commercial milestone.

That evidence boundary matters because the optical-interconnect market is crowded with serious engineering efforts. CScale's public differentiation is not simply that light can move data farther than copper. Its pitch is that an integrated optical system can preserve continuity when laser failures become inevitable at the scale of an AI factory.

The Team Behind the Bet

Martin Lund joined CScale as CEO after leading Cisco's Common Hardware Group, where his responsibilities included silicon, hardware systems, and optics, including Silicon One. CScale also credits Lund with senior roles at Microsoft and Cadence and with building Broadcom's switching business into a billion-dollar franchise.

Sanjai Kohli founded CScale and remains CTO and a board member. Kohli previously co-founded SiRF, associated with bringing GPS technology into mass-market devices, and later founded Inovi, which Facebook acquired in 2014. The leadership pairing puts an experienced technical founder beside an executive who has managed networking products and organizations at commercial scale.

That pairing also explains part of the financing logic. CScale is no longer only proving a technical concept inside a stealth company. It must translate an architecture into a manufacturable product, work through qualification cycles, win system-level trust, and support deployments where downtime carries a large capital cost.

Why the Investor Group Matters

Atreides Management, Valor Equity Partners, and Premji Invest are underwriting the commercialization stage, while Sutter Hill Ventures continues backing a company it has supported since inception. NVIDIA and Intel Capital add strategic relevance because both organizations operate close to the compute, packaging, and infrastructure decisions that shape AI systems.

Their participation should not be mistaken for a disclosed customer contract or product endorsement beyond the investment itself. CScale has not named a deployment partner, hyperscaler, AI lab, or system vendor that has qualified the technology for production. The strategic investors make the company easier to place inside the AI hardware ecosystem, but execution still has to produce the evidence.

The capital does give CScale room to tackle a problem with long development and qualification cycles. Semiconductor and optical products do not move from architectural promise to fleet deployment on a software release schedule. Engineering validation, manufacturing readiness, interoperability, customer qualification, and field reliability each become part of the commercial product.

What the Series C Must Prove

CScale's next proof will come from turning failure containment into measurable system performance. Operators will need evidence that the architecture preserves useful compute under realistic fault conditions, fits the economics of large AI systems, and can be manufactured and deployed at the scale the company describes.

The absence of public customers and benchmarks does not invalidate the engineering thesis, but it defines the work purchased by the Series C. CScale has raised enough capital to move the conversation from what optical interconnect should do toward what its own implementation can repeatedly deliver.

AI infrastructure is becoming a systems problem because processors, memory, power, cooling, networking, and software all have to cooperate under sustained load. CScale is entering that system through the light path, where one small failure can determine whether a much larger pool of expensive compute keeps earning its place in the cluster.