Summary: OpenAI and Cerebras announced a multiyear agreement on January 14, 2026 for up to 750 megawatts of inference capacity, with deployments running through 2028. CNBC, citing people close to Cerebras, valued the deal at more than $10 billion, while official posts did not state an amount. The word up to matters because rollout depends on Cerebras production expansion, and the deal arrived before the startup’s planned IPO.
On January 14, 2026, OpenAI and wafer-scale chip startup Cerebras announced a multiyear agreement: OpenAI will buy as much as 750 megawatts of compute capacity delivered on Cerebras silicon, with deployments running through 2028. CNBC, citing people close to the company, values the agreement at more than $10 billion — the startup’s flagship customer win ahead of its long-planned IPO.
Infrastructure headlines usually mean stockpiled training compute. This one doesn’t: every one of the 750 MW is earmarked for inference. OpenAI is buying latency rather than training throughput — responses quick enough for real-time interaction — from the one company whose architecture was built to make individual requests fast.
What Got Signed: Up to 750 MW, Running Through 2028
The deal is sized in electricity, not chip counts. At 750 MW this is campus-scale — the draw of a large AI data center, not a few racks — hosted in data centers Cerebras operates at sites across the United States. CNBC also supplied the timeline: a term sheet signed just before Thanksgiving 2025, with the news held until January 14. The $10 billion-plus figure traces back to sources close to Cerebras; the official posts never stated a number.
Three details repay a closer look. The contract reads “up to” 750 MW, so the rollout rides on how fast Cerebras can expand production. The relationship has history: OpenAI evaluated the company’s technology as far back as 2017, and only now has that long interest turned into contracted capacity. And the announcement lands right before the IPO — signing a multiyear OpenAI agreement ahead of listing is an obvious lift for the valuation story.
Inference, Not Training: Where the Capacity Goes
Sachin Katti, who runs compute infrastructure at OpenAI, laid out the reasoning in the announcement: “Cerebras adds a dedicated low-latency inference solution to our platform. That means faster responses, more natural interactions, and a stronger foundation to scale real-time AI to many more people.”
The company’s edge is the wafer-scale engine: a whole wafer manufactured into a single dinner-plate-sized processor, which bypasses the interconnect bottlenecks of GPU clusters and drives per-request throughput to extreme levels. For training frontier models it remains a minority architecture; for the “one user wants an answer right now” shape of inference, it is on home ground.
Compatibility was proven ahead of time. OpenAI’s gpt-oss open-weight models already run on Cerebras silicon, working alongside Nvidia and AMD chips. This is not a port built from zero but proven technology promoted into contracted production capacity.
Cerebras’ End of the Deal: Easing Its Reliance on G42
For Cerebras, the strategic weight of this deal rivals its dollar value. CNBC reported that in the first half of 2024, roughly 87 percent of the company’s revenue came from one customer: G42, the Abu Dhabi AI group. A firm flying the Nvidia-challenger flag while its revenue rests on a single large Middle Eastern client is exactly the story public-market investors punish.
CEO Andrew Feldman kept the framing practical: “The way you have three very large customers is start with one very large customer, and you keep them happy, and then you win the second one.” OpenAI is that second customer. With the IPO still queued, one contract now works on the revenue-concentration question and the growth narrative at the same time.
Where This Sits in OpenAI’s Compute Sourcing
From OpenAI’s side, this is another step away from single-supplier dependence. Nvidia crossed the $5 trillion market cap line in October 2025 before anyone else — a measure of how concentrated the AI silicon market remains — and that December it agreed to a nonexclusive licensing deal with Groq worth $20 billion in cash, which CNBC called Nvidia’s largest transaction to date. The supply side keeps consolidating, and buyers have never been in a bigger hurry to diversify.
Cerebras hands OpenAI an inference path that shares nothing with the Nvidia architecture. gpt-oss already runs on Nvidia and AMD chips; wafer-scale systems now join that formal procurement mix. For the first time at OpenAI, inference-dedicated compute stands as a procurement category of its own rather than an offshoot of training clusters.
Takeaways for Developers and Product Teams
Three effects matter in practice. First, latency-sensitive products — voice, real-time agent interaction, streaming UIs — should put wafer-scale inference on the evaluation shortlist, now that the route carries OpenAI’s own endorsement. Second, compute procurement is splitting along the training-versus-inference line, and cost models need to split with it: training is bought on cluster scale, inference on latency and throughput, and the two can come from different suppliers. Third, supplier diversification has stopped being a slogan and started showing up in contracts. From our 2026 opening outlook to this deal, the opening month of the year keeps sending one consistent signal: choosing your infrastructure is choosing your product.