CScale exits stealth with $145M for optical interconnect at AI scale-up

Palo Alto startup CScale said Wednesday it exited stealth with $145 million in Series C funding to build resilient optical interconnect for AI scale-up — the high-bandwidth, ultra-low-latency links that let accelerators act like one larger computer — bringing total funding to $188 million.

When AI clusters span racks, the wires between chips matter as much as the chips. CScale is raising a large Series C, with NVIDIA and Intel Capital as strategic investors, to sell optical links designed so a failed laser does not stall the whole factory. That is an infrastructure bet on reliability at scale, not another model launch.

On Wednesday, 30 September 2026, Palo Alto-based CScale announced it exited stealth with $145 million in Series C funding. A Series C is a later round of venture funding, after earlier money, for a company raising to build and sell. The $145 million is the new money. The company page is dated that day and does not print an hour. The dateline is Palo Alto, Calif. The lede says the round is to accelerate development and commercialization of optical interconnect for AI scale-up. Optical interconnect, here, means links that carry data as light, through fiber, instead of as electrical signals on a copper wire. Scale-up, as CScale defines it in that sentence, is the high-bandwidth, ultra-low-latency connections that let accelerators work together like a single, much larger computer. An accelerator is the chip that does the AI math. Bandwidth is how much data can move at once. Latency is the wait before a message arrives. Ultra-low latency means that wait is short enough for the chips to stay in step. The round brings the company’s total funding to $188 million. The new round is $145 million of that $188 million total. The difference is $43 million, which is the earlier funding if those two figures are meant to add. The announcement does not list the earlier rounds. Those lines are CScale’s.

Who put the money in, as the company page states it. The round was led by Atreides Management, together with Valor Equity Partners, and Premji Invest joined as a co-lead. Sutter Hill Ventures, which has backed CScale since its inception, continued, alongside existing investor Maverick Silicon. NVIDIA and Intel Capital joined as CScale’s first strategic investors. A strategic investor, in that sentence, is a company in the industry taking a stake, as distinct from a fund whose business is investing. The page calls NVIDIA and Intel Capital strategic investors. It does not say their chips will ship with CScale’s links. The deck under the headline names Atreides Management, Valor Equity Partners, Premji Invest, and Sutter Hill Ventures. Maverick Silicon, NVIDIA, and Intel Capital are in the body of the release, not in that deck. Business Wire, as Financial Content reprints the same announcement, says the round was co-led by Atreides Management, Valor Equity Partners, and Premji Invest. The company page’s wording is “led by,” with Premji Invest named as a co-lead. The wire’s wording is “co-led” by all three. Both are CScale’s release. Financial Content stamps the reprint September 30, 2026, at 9:00 a.m. Eastern.

What CScale says those links are for. In the gigawatt-class AI data centers it expects by the end of the decade, a single scale-up domain will span thousands of tightly coupled accelerators across dozens of racks. A gigawatt is a billion watts. A gigawatt-class site draws power on the scale of a large power plant. The end of the decade, from a page dated 2026, is the close of the 2020s. A scale-up domain is one group of accelerators tied tightly enough to act as one computer. Tightly coupled means they share work as it happens, rather than handing off a job and waiting. A rack is a cabinet of machines. Dozens of racks is a row of those cabinets, not one box. Thousands of accelerators is the count inside one of those domains. Those lines are CScale’s.

The same paragraph is about what happens when a link fails. CScale says these systems demand high bandwidth, predictable low latency, and continuous communication. Across hundreds of thousands of accelerators, failures that are rare on any one link become a continuous condition for the fleet, and portions of the AI factory degrade unless the failures can be contained. Hundreds of thousands is the wider fleet. Thousands across dozens of racks is one scale-up domain. The two counts are not the same number. CScale says it is architecting the interconnect to contain an optical failure without interrupting compute. Contain, here, means a dead link stays a dead link, instead of stalling the computers around it. The about box says the company is building an integrated light engine to deliver reliable optical connectivity for AI infrastructure at gigawatt scale. A light engine, in that sentence, is the piece that makes and handles the light those links use. Those lines are CScale’s. The announcement does not say the light engine is shipping.

Martin Lund, chief executive, is quoted on the company page. He said that as AI scale-up domains extend across dozens of racks, optical interconnect becomes essential. He said that at that scale, reliable, predictable communication is fundamental to system economics. He said easier part replacement improves serviceability, not continuity. Serviceability is how quickly a broken part can be swapped. Continuity is the work never stopping. He said CScale is designing the interconnect for continuity, because lasers will fail but compute should not. A laser, here, is the light source in an optical link. That quotation is his, in the company statement.

Two investors are quoted in the same release. Gavin Baker, managing partner and chief investment officer at Atreides Management, said AI infrastructure is no longer just a compute problem. It is a systems problem. He said that as scale-up systems move toward gigawatt-class deployments, interconnect bandwidth means nothing without system reliability, and that every optical failure is a compute failure. He said that at this scale, failures are not hypothetical. They are inevitable. He said CScale is designing reliable scale-up optics for next-generation AI factories, so the links can run at full optical performance without tradeoffs. Sandesh Patnam, managing partner at Premji Invest, said CScale’s strength is architectural judgment: which technical choices create value for the whole system. He said the team combines photonics, electronics, and software with experience deploying the work, so optical capability becomes dependable infrastructure. Photonics is the engineering of light. He said Premji Invest partners with teams that think big and have the discipline to build enduring companies, and that the firm is proud to co-lead this financing. Those quotations are theirs, in CScale’s release. They are not a measured test that a failed laser stays contained.

The about box describes the company. CScale says it was founded in 2023, is led by chief executive Martin Lund and founder and chief technology officer Sanjai Kohli, is headquartered in Palo Alto, California, and has approximately 85 employees globally. Approximately 85 is the company’s headcount. It is not a count of customers. Lund, the page says, previously built Broadcom’s switching business into what the release calls a billion-dollar franchise, held senior executive roles at Microsoft and Cadence, and most recently led Cisco’s Common Hardware Group, responsible for silicon, hardware systems, and optics, including Silicon One. That billion-dollar line is CScale’s description of Broadcom’s switching business during Lund’s time there. It is not a valuation of CScale. The announcement does not print a valuation for CScale. Kohli, the page says, has built and exited multiple startups in communications, including SiRF, which the page says brought GPS to the mass market. He received the 2010 European Inventor Award, and later founded Inovi, which Facebook acquired in 2014. Those lines are on the company page. The Business Wire about box, as Financial Content reprints it, also says Lund brings three decades of silicon and network leadership from Cisco, Microsoft, Cadence, Broadcom, and Intel, and says the company was founded in 2023 by chief technology officer Sanjai Kohli. The company page calls Kohli founder and chief technology officer and says the company was founded in 2023. It does not use the phrase “founded by.” The media contact on the wire reprint is Nick Gibiser at Wireside Communications, ngibiser@wireside.com.

The release is a funding announcement and a product direction. It says the money is to accelerate development and commercialization. It does not name a customer. It does not print a ship date. It does not print a price. It does not say NVIDIA or Intel will put the links in a product they sell. Development is the building. Commercialization is the work of getting a product to buyers. The page uses both words. It does not say a product is already in a data center.

The picture is CScale’s press announcement. A white field has an orange bar along the left edge. A small label reads NEWS. The headline reads “CScale Exits Stealth With $145 Million to Build the Interconnect for Gigawatt-Scale AI.” The line under it says the company is backed by Atreides Management and Valor Equity Partners, Premji Invest, and Sutter Hill Ventures, and that CScale is building resilient optical interconnect for AI scale-up. It is the company’s headline and deck. It is not a photograph of a data center, and it does not print a calendar date.

In plain terms, CScale said on Wednesday that it left stealth with $145 million in new Series C funding, $188 million in all, to build optical links so the accelerators in a very large AI site can act like one computer, and so a failed laser does not stop the computers around it. Atreides led on the company page. Valor Equity Partners is named with that lead, and Premji Invest joined as a co-lead. The wire’s line is that all three co-led. Sutter Hill, a backer since the start, and Maverick Silicon continued. NVIDIA and Intel Capital joined as the first strategic investors. The company says it was founded in 2023, is based in Palo Alto, and has about 85 people. The announcement does not name a customer, a ship date, or a price for the company.