Cerebras Systems

The conventional wisdom in chip design is to cut silicon wafers into individual chips and discard the ones with defects. Cerebras Systems, founded in Sunnyvale in 2016, rejected this approach entirely. The company's co-founders — led by Andrew Feldman, a serial entrepreneur who had previously sold SeaMicro to AMD for $334 million — bet that the future of AI computing required a fundamentally different architecture: a chip that used the entire wafer as a single processor. Their Wafer Scale Engine (WSE) was the first commercial chip to do so. At 46,225 square millimetres with 4 trillion transistors and 900,000 AI-optimised cores, the WSE-3 is 57 times larger than Nvidia's H100 GPU.

AI training requires moving data between thousands of small GPU chips connected across a network — a process that introduces latency and communication overhead. Cerebras's wafer-scale approach put everything on one chip: no inter-chip communication, no network bottleneck. The architecture was particularly suited to AI inference — generating responses to user queries — where speed directly determined product quality. Cerebras claimed inference speeds more than ten times faster than GPU-based systems for certain workloads. Benchmark Capital, which had backed the company since a $27 million Series A in 2016, raised a dedicated $225 million SPV in 2026 to increase its position.

Cerebras first filed to go public in September 2024. The filing revealed a structural vulnerability: G42, an Abu Dhabi-based AI company backed by sovereign wealth, contributed 85-87% of Cerebras's revenue. The U.S. Committee on Foreign Investment in the United States launched a review of G42's minority stake, concerned about potential technology transfer to China. The IPO was withdrawn. The filing had been premature, the customer concentration was disqualifying, and the geopolitical dependencies were unresolved. G42's stake was restructured to non-voting shares, clearing the path for a second attempt.

In December 2025, OpenAI signed a $10-20 billion compute agreement with Cerebras to supply 750 megawatts of AI compute through 2028, with options for nearly 3 gigawatts more by 2030. OpenAI advanced Cerebras a $1 billion working capital loan and received warrants for 33 million shares. Andrew Feldman told the Wall Street Journal: "Obviously, Nvidia didn't want to lose the fast inference business at OpenAI, and we took that from them." Cerebras also secured an agreement with Amazon Web Services to deploy Cerebras chips in Amazon's data centres. Revenue for 2025 reached $510 million — up 76% — with a $24.6 billion backlog in remaining performance obligations.

Cerebras re-filed its S-1 in April 2026, targeting a Nasdaq listing under the ticker CBRS at a $22-25 billion valuation. The February 2026 Series H, led by Tiger Global, had valued the company at $23 billion — nearly triple its valuation from five months earlier. Morgan Stanley, Citigroup, Barclays, and UBS were leading the underwriting. The company that had built a chip the industry said couldn't work — one chip using an entire silicon wafer — had taken inference revenue from Nvidia, signed a multi-billion-dollar contract with OpenAI, and was preparing to go public as one of the most consequential semiconductor companies to list since Nvidia itself.