CScale, a Silicon Valley startup backed by Nvidia and Intel, has raised $145 million to develop optical connections that sit closer to AI chips — and keep data flowing even when individual lasers fail.

The round was led by Atreides Management, Valor Equity Partners, and Premji Invest. It brings the company's total funding to $188 million. CEO Martin Lund, previously a hardware executive at Cisco Systems, says CScale aims to ship chips by 2028.

The copper problem

Today's AI servers mostly use copper cables to connect chips inside a single box. Copper is cheap and dependable, but it cannot carry ultra-fast signals very far without degrading. That limitation pushes designers like Nvidia and Advanced Micro Devices to pack processors close together and lean harder on complex cooling systems.

Optical links, which transmit data as pulses of light through fiber, can move far more data over longer distances. The industry therefore wants to bring optics inside the server and, eventually, right up to the chip package itself.

The reliability snag

One obstacle has slowed that migration: lasers wear out. If a laser failure forces a full component swap, downtime and maintenance costs spike. That risk explains why many current designs keep lasers in replaceable modules rather than integrating them directly onto the package.

CScale claims it can place lasers onto chips while maintaining traffic when a single laser fails. A failure would then become a manageable performance hit rather than a full outage — the difference between degraded service and a stopped data center.

What it means for the market

CScale's 2028 timeline will test whether laser failure is a downtime problem or simply a design constraint.

If the approach works, near-chip optics becomes easier for data center operators to live with: fewer truck rolls, less forced replacement of tightly integrated parts, and higher real-world uptime. That would remove one of the biggest practical barriers to swapping short-reach copper links inside racks for optical connections closer to Nvidia and AMD processors.

For chipmakers, server builders, and suppliers of cables, transceivers, and cooling gear, such a shift would reshape where money gets spent across AI hardware. Operators could stop treating interconnects as a hard physical limit that dictates dense layouts and heavy cooling. Instead, they would weigh optics as an architectural choice — one judged on reliability, serviceability, and total system cost.