The Chips Connecting GPUs: A New Prize
The AI boom put GPUs in the spotlight, but it’s the chips that wire them together that are now becoming highly valued. Israel’s Xsight Labs has secured over $300 million at a $2.8 billion valuation to develop an open and programmable alternative.
The AI boom made GPUs famous. The unglamorous yet crucial chips connecting thousands of GPUs, quietly becoming just as valuable.
Xsight builds the essential components for AI data centers: the E1 processor and X2 switch. These facilitate data flow between servers, memory, and networks as AI models grow larger. This has created a bottleneck in the industry, making the wiring between GPUs a significant focus.
Taking on Nvidia’s Dominance
Nvidia controls both GPU manufacturing and networking through its acquisition of Mellanox. Xsight Labs offers an alternative: open and programmable chips that run standard Linux and SONiC, allowing for software-based customization after deployment.
This approach is gaining traction as global network operators, like SpaceX’s Starlink, adopt Xsight’s silicon. The company’s processors recently passed tests on SpaceX’s new Starlink V3 satellites, demonstrating their efficiency and performance advantages.
Technical Edge and Serial Backing
Xsight’s E1 is the first 800-gigabit data processor to ship while consuming less than 75 watts. The X2 switch moves 12.8 terabits per second with a 40% lower power draw compared to older models. This energy efficiency is crucial as AI’s electricity demands surge.
The company’s strength lies in its founders and investors: veterans from EZchip, who were acquired by Mellanox; and well-known chip entrepreneurs like Avigdor Willenz, backed by Fidelity, Intel Capital, T. Rowe Price, Battery Ventures, Valor Equity Partners, and more.
Xsight is betting that open networking solutions will prevail in the AI infrastructure race, as operators seek alternatives to Nvidia’s closed system.