What happened
OpenAI CEO Sam Altman weighed in publicly to say Cerebras is "a close partner," and that the two companies have "a deep engagement pushing on the frontiers of speed." The statement closed out open speculation about the status of the alliance.
Behind the confirmation sits a Master Relationship Agreement, effective December 24, 2025, that lays out the build in 250MW tranches: 250MW by the end of 2026, 500MW by the end of 2027, and 750MW by the end of 2028. It is all inference capacity, not training. A separate SEC filing indicates an OpenAI Codex Spark model was already running on Cerebras infrastructure back in February 2026, so the relationship has been live for some time.
Why speed, and why now
Inference is where AI turns into money. Training builds the model; inference answers every ChatGPT query, every API call, every agent step. Cerebras built its company on this one idea: wafer-scale processors that keep computation on a single giant die instead of shuttling it between dozens of GPUs. The pitch is latency. For OpenAI, faster answers mean products that feel instant and a lower cost per query, which is the difference between an agent people use and one they abandon.
It is also about supply. OpenAI, like everyone else, still leans heavily on Nvidia. Adding 750MW from a different vendor is a second source of firepower — and in a market this tight, leverage.
The fine print
750MW is a capacity contract, not a capability claim. Nothing disclosed so far says which products get the fast lane, how much latency actually drops, or whether any of it changes pricing for API customers. The delivery milestones are end-of-year tranches, so the near-term effect is 250MW by the end of 2026: meaningful, but a fraction of the headline. And the speculation Altman was responding to was never identified, so the confirmation arrives without its context.
The pattern is familiar by now. In recent weeks, Anthropic lined up a $42 billion Broadcom chip loan and Amazon began shopping a GPU sale-leaseback to Wall Street. Nobody is short on ambition; everyone is short on silicon. OpenAI's move is the clearest signal yet that the next thing the labs will fight over is not model size. It is milliseconds.
Sources
- [1] Nation Press: Sam Altman confirms OpenAI's deep partnership with CerebrasRead source
- [2] DEV Community: OpenAI and Cerebras confirm 750MW AI inference deployment through 2028Read source