CScale develops interconnect technology meant to link together the accelerators that make up an AI cluster. SiliconANGLE reports that it will move data between chips over optical links, which are significantly faster than traditional copper wiring.[1]
Technical detail is thin. CEO Martin Lund declined to give specifics, telling Reuters (as carried by the Economic Times) only that the company will integrate lasers onto the chips in a way that does not interrupt the flow of data when a laser fails. That is a real engineering problem in optical links, since a failed laser can otherwise take a link down, but there is no public design to evaluate yet.[2]
The company frames the problem as scale. It is targeting future gigawatt-class AI data centers, where scale-up domains could span thousands of tightly connected accelerators across dozens of racks. CScale says that while failures may be rare on individual optical links, across fleets of hundreds of thousands of accelerators they become a continuous operational issue, so it is designing its architecture to contain optical failures without interrupting compute workloads. It describes its product as an integrated light engine. Its CEO's summary of the pitch: "Lasers will fail. Compute shouldn't."[3]
SiliconANGLE adds context on why this matters. Today a scale-up domain is usually a single server rack, and data moves more slowly between domains than within one. A malfunction in one optical link can disrupt many other components in a cluster. SiliconANGLE also cites a CScale job posting as indicating the company plans to ship its technology as a system-on-chip with an onboard CPU, which it speculates may run network automation software; that is the outlet's inference, not a company statement.[1]
The company has said its chips are not yet on sale. Lund said CScale plans to ship them by 2028, which puts the product two years or more out from the launch announcement.[2]