SAN FRANCISCO, Jan 15, 2026, 08:29 (PST)
- OpenAI agreed to buy up to 750 megawatts of computing power from Cerebras over three years, a source said.
- The deal targets faster AI “inference” — the step where models generate answers — as users demand quicker responses.
- Analysts warn the scramble to mix chip types can add cost and complexity inside data centers.
OpenAI agreed to buy up to 750 megawatts of computing power from AI chipmaker Cerebras Systems over three years in a deal worth more than $10 billion, a source familiar with the matter said on Wednesday. The ChatGPT maker is stacking up capacity as demand rises for “reasoning” systems that take longer to answer. (Reuters)
The squeeze now is not just training new models. It is serving them at scale, fast, and without blowing out cost or power use.
OpenAI framed the Cerebras tie-up as a way to help its systems “respond much faster” during inference, the step where a model turns a prompt into text, code or images. In its post, the company called Cerebras “a dedicated low-latency inference solution,” while Cerebras CEO Andrew Feldman said “real-time inference will transform AI.” (OpenAI)
Cerebras, which builds wafer-scale processors designed to cut bottlenecks, billed the rollout as the “largest high-speed AI inference deployment in the world,” starting in stages from 2026. The company also claimed outputs can be “up to 15× faster” than GPU-based systems, a comparison it has used to pitch itself as an alternative for latency-sensitive workloads. (Cerebras)
The deal also underlines the grip Nvidia has on AI computing, and why customers keep shopping. OpenAI is pushing to keep response times tolerable as rivals such as Anthropic and Alphabet’s Google ship competing models and tools.
Cerebras filed for an initial public offering in 2024 and later delayed it while continuing to raise money, TechCrunch reported. It also noted OpenAI CEO Sam Altman is an investor in the chip startup. (TechCrunch)
On Thursday, OpenAI also launched a request for proposals — a formal call for bids — from U.S. manufacturers across data-center hardware, consumer electronics and robotics. OpenAI said it was looking to strengthen domestic supply chains as it expands infrastructure tied to its Stargate effort and a 10-gigawatt capacity commitment. (OpenAI)
Separately, OpenAI said it is participating in Merge Labs’ seed round, calling brain-computer interfaces an “important new frontier.” The company said Merge Labs’ co-founders include researchers Mikhail Shapiro, Tyson Aflalo and Sumner Norman, alongside entrepreneurs including Altman in a personal capacity. (OpenAI)
But more compute vendors can mean more moving parts. “This move is less about replacing Nvidia and more about diversification as inference scales,” said Prabhu Ram, a vice president at Cybermedia Research; “Running heterogeneous accelerators requires managing multiple software stacks,” semiconductor analyst Manish Rawat at TechInsights warned. (Network World)
For OpenAI, the test is whether this capacity turns into noticeably faster replies — without locking the company into hardware choices that age badly. A lot of money is chasing a simple outcome: shorter waits.