OpenAI’s $20B Cerebras pact

OpenAI has agreed to spend more than $20 billion over three years on Cerebras-powered server capacity, according to Reuters, in a deal that could also give OpenAI a stake in the chip startup. (reuters.com) Reuters reported on April 16 that the arrangement would let OpenAI use servers built around Cerebras chips, with the terms described by The Information and not independently verified by Reuters. Cerebras said separately on April 17 that it had given OpenAI a warrant to buy stock and could expand its business with the company over time. (reuters.com; cnbc.com) The reported commitment is larger than the partnership Cerebras announced on January 14, when it said OpenAI had signed a multi-year agreement to deploy 750 megawatts of Cerebras systems for inference, the stage when a model answers prompts. Cerebras said that rollout would begin in 2026 and called it the largest high-speed inference deployment it had announced. (cerebras.ai; reuters.com) Inference is the part of artificial intelligence computing that users feel directly: every chatbot reply, code completion, or image request has to run somewhere after the model is trained. OpenAI has been lining up more of that capacity as usage grows and as it tries to avoid relying on a single supplier for the chips inside those servers. (cerebras.ai; reuters.com) That supplier has largely been Nvidia, whose graphics processing units became the standard hardware for training and serving large AI models. A Cerebras deal would give OpenAI another source of compute at a time when the biggest AI companies are signing long-term infrastructure contracts measured in gigawatts and tens of billions of dollars. (reuters.com; openai.com) OpenAI has also been building out capacity through Stargate, its infrastructure effort with Oracle and SoftBank. In September 2025, OpenAI said five new U.S. sites, plus Abilene, Texas, and projects with CoreWeave, brought Stargate to nearly 7 gigawatts of planned capacity and more than $400 billion in investment over three years. (openai.com) Cerebras is taking a different hardware approach from Nvidia’s standard chip clusters. Its flagship systems are built around wafer-scale processors, which place a much larger computing surface on one chip to reduce the time and energy spent moving data between many smaller chips. (cerebras.ai; cnbc.com) That pitch has become more important as AI customers care less about benchmark wins and more about whether a service can answer quickly and cheaply at scale. Cerebras has been leaning on that argument as it returns to public markets; CNBC reported on April 17 that the company had filed to go public again after scrapping an earlier IPO attempt. (cnbc.com) Neither Reuters nor CNBC described final public confirmation from OpenAI of the full reported $20 billion figure. But if the terms hold, the deal would show that the AI race is no longer just about building better models; it is also about locking up enough servers, power, and chips to keep those models running. (reuters.com; cnbc.com)