OpenAI is spending big to move away from Nvidia.
The company has agreed to purchase servers powered by Cerebras Systems’ wafer-scale chips, adding 750 megawatts of high-speed AI compute capacity in a deal that reportedly exceeds $20 billion over three years. According to a report by The Information, total spending could reach $30 billion, though that ceiling figure has appeared primarily in secondary coverage and hasn’t been confirmed directly. The agreement also reportedly includes equity warrants giving OpenAI a stake of up to 10% in Cerebras, according to reports citing The Information.
This isn’t OpenAI’s first move in this direction. A prior agreement with Cerebras was valued at more than $10 billion and carried the same 750MW compute commitment. The new deal potentially doubles that prior commitment, and arrives against a backdrop of documented frustration. OpenAI has previously expressed dissatisfaction with some Nvidia chips and has sought alternative compute sources, according to Reuters reporting.
Cerebras builds a different kind of chip. Its wafer-scale architecture integrates the entire silicon wafer into a single processor rather than dicing it into separate dies. That design delivers high memory bandwidth and low latency for specific AI workloads, advantages that matter for inference at scale. It’s not a drop-in Nvidia replacement. It’s a strategic complement, and OpenAI is betting heavily on that complementarity.
Why does this matter beyond the dollar figure? OpenAI runs some of the most compute-intensive workloads in the industry. When its infrastructure choices shift, suppliers, competitors, and enterprise customers all take notice. A $20 billion-plus commitment to a non-Nvidia vendor is the most concrete signal yet that OpenAI views compute diversification as a structural priority, not a contingency plan.
For developers and enterprise buyers, the Cerebras deal has a less obvious but real implication. Inference costs and latency depend partly on what chips sit behind the API. As OpenAI rebalances its compute stack, the performance and pricing characteristics of its products may shift, and not uniformly across all workload types.
This deal also arrives within weeks of Meta’s reported ~$21 billion compute agreement with CoreWeave, a separate infrastructure play that signals the same underlying dynamic: the largest AI labs are locking in alternative compute at scale before the next generation of model training begins. Two mega-deals in two months isn’t coincidence. It’s a pattern.
What to watch: whether Cerebras can deliver the promised compute capacity on the timelines the deal implies, how Nvidia responds to its largest customer diversifying this aggressively, and whether the equity warrant structure triggers any regulatory scrutiny given OpenAI’s position in the market.
The headline number gets attention. The warrant structure is the more interesting story. An equity stake of up to 10% in a chip company, held by the AI lab purchasing that company’s chips, creates a financial alignment that goes well beyond a typical vendor relationship – and that’s worth tracking as AI infrastructure consolidation accelerates.