CScale exits stealth with $145m Series C backed by Nvidia and Intel Capital

CScale exited stealth with a $145 million Series C, according to Data Center Dynamics, bringing total funding since the Palo Alto company's 2023 founding to $188 million; Atreides Management, Valor Equity Partners and Premji Invest co-led the round, with Nvidia and Intel Capital investing and Sutter Hill Ventures and Maverick Silicon also participating.

Founded in 2023 by chief technology officer Sanjai Kohli, the company is building an optical interconnect for AI scale-up networking — the links inside a training domain rather than the spans between buildings — but has released few details of how the design works beyond a directional claim that optical failures are contained so they do not interrupt compute, and that the new money accelerates development and commercialization. The coverage does not describe the mechanism, the reach the interconnect supports, or the customers it is being built for.

Chief executive Martin Lund sharpened that into a single distinction: “Easier part replacement improves serviceability, not continuity,” he said. “We’re designing the interconnect for continuity. Lasers will fail. Compute shouldn’t.” The first is a faster repair; the second is the outage that never happens — for an operator selling time on training capacity, economically different products.

Gavin Baker, managing partner and chief investment officer at Atreides, lifts the same argument above the component: AI infrastructure is a systems problem rather than only a compute problem, and at gigawatt-class deployments “interconnect bandwidth means nothing without system reliability.” He credits CScale with designing reliable scale-up optics for next-generation AI factories, “unlocking the full performance of optics without trade-offs,” which is a claim about an unshipped product from an investor whose firm co-led the round.

Two of the three leads appear in PID’s records and sit far apart on the adviser scale — Atreides reports $8.9 billion in registered assets with 26 employees, Valor $59.3 billion with 104, as of Sept. 26 — while the third, Premji Invest, does not appear in those records. Nvidia and Intel Capital took equity, with no supply agreement or named customer attached to their participation. The same records log four Nvidia deal events in the seven days to Sept. 29 — two announced and two closed, the largest a $3.4 billion transaction on Sept. 25 — making an equity stake in an optics startup unremarkable in itself and leaving the co-lead structure as the more revealing feature: two firms of very different size writing the same $145 million check.

Grid access and queue position set the pace of what gets built at this point in the AI buildout, as this publication has argued, and the archive’s running account of the sector — interconnection queues, load classes, electrons — reflects it. Schneider’s factory-built 2.5MW modules compress the most controllable part of the data center timeline while leaving the interconnection queue where it was, and CScale belongs to the same category of bet pointed the other way: money going into a component that raises what an already-powered rack can do rather than into the power itself. If a scale-up domain spans dozens of racks, as Lund describes it, and if every optical failure is a compute failure, then reliability is a utilization problem, and utilization is how a queue-constrained industry extracts more from capacity it has already interconnected. On that reading the two constraints are complements, since a queue position whose racks run more training hours is worth more than one whose racks idle behind a failed link.

None of that disputes the house position that power rights are the deal currency; it describes a market that can be underwritten without a megawatt, in components whose value tracks the utilization of someone else’s interconnection agreement, and whose verdict arrives sooner than a campus’s.

The structure is a softer version of a transaction this publication covered in August, when Nvidia weighed a $3 billion investment in SB Energy behind OpenAI’s Ohio campus and the guarantee, more than the equity, was the piece credited with making the project financeable. Nothing in CScale’s round plays that role: there is no guarantee, no disclosed customer and no shipped product, which leaves a $145 million equity check resting on a demand case the announcement does not quantify. The round also carries nothing an underwriter would mark: no valuation, no capacity, no performance figures, with few technical details provided, as DCD reports. Of the $188 million raised in total, $43 million arrived before this round, so most of the capital is new; a named customer is what data center financing would call the anchor tenant, and the anchor-tenant logic that governs those deals has nothing here to attach to. A name would matter next: a hyperscaler or an AI lab putting its name to a deployment. CScale says the funding accelerates development and commercialization of its interconnect, but a production deployment on a named network would be the first public evidence that the continuity argument holds when a laser fails.