OpenAI’s big Cerebras bet

OpenAI has agreed to spend more than $20 billion over three years on servers built with Cerebras chips, extending a January computing pact into one of the biggest disclosed AI infrastructure commitments yet. (reuters.com) Reuters reported on April 17 that the deal also gives OpenAI an equity stake in Cerebras. Cerebras said in January that the companies had signed a multi-year agreement to deploy 750 megawatts of Cerebras systems for OpenAI customers, with rollout starting in 2026. (reuters.com) (cerebras.ai) Cerebras is selling a different kind of artificial intelligence chip system than Nvidia’s graphics processors: it builds wafer-scale chips, which keep more computing on one giant piece of silicon instead of spreading it across many smaller chips. The company pitches that design as a way to speed up inference, the step where a trained model generates answers for users. (cerebras.ai) (finance.yahoo.com) That distinction matters in 2026 because AI companies are now buying not just training capacity to build models, but inference capacity to serve millions of prompts after those models go live. Reuters said the OpenAI-Cerebras agreement comes as demand rises for alternative suppliers to Nvidia in AI computing. (reuters.com) (finance.yahoo.com) Cerebras is trying to use that demand to reopen its path to the public market. The company announced on April 17 that it had filed a new Form S-1 for a United States initial public offering after withdrawing its earlier filing in October 2025. (cerebras.ai) (finance.yahoo.com) Its first IPO attempt stalled after U.S. officials reviewed Cerebras’s ties to Abu Dhabi-based G42, a major customer and investor. Reuters reported in December 2025 that Cerebras was preparing to refile after that delay, and a later report said G42 was no longer listed as an investor. (finance.yahoo.com 1) (finance.yahoo.com 2) (finance.yahoo.com 3) Cerebras’s earlier S-1 showed how concentrated its business was before this OpenAI push. The filing said the company generated a significant majority of its revenue from one customer, G42, and Reuters reported in 2024 that Cerebras’s revenue had more than tripled in 2023. (sec.gov) (finance.yahoo.com) OpenAI, meanwhile, has been lining up compute from multiple providers rather than relying on one chipmaker or one cloud partner. The Cerebras expansion fits that pattern: Reuters described it as part of OpenAI’s effort to secure more server capacity as competition for AI infrastructure intensifies. (reuters.com) If the spending plan holds, Cerebras heads toward its IPO with a marquee customer, a much larger backlog, and a public case that its chips can win business alongside Nvidia rather than only around it. (reuters.com) (cerebras.ai)