San Francisco, January 14, 2026, 23:20 (PST)
- OpenAI has committed to purchasing up to 750 megawatts of computing capacity from AI chipmaker Cerebras over the next three years, according to a source.
- The capacity targets speeding up “inference,” the phase when an AI model produces answers for users.
- The deal intensifies the rush among AI companies and chip makers to secure data center capacity, known for its heavy power demands.
OpenAI is set to acquire up to 750 megawatts of computing capacity from chip startup Cerebras over the next three years, in a deal exceeding $10 billion, according to a source close to the matter who spoke to Reuters. The tech giant will deploy Cerebras hardware to power ChatGPT and other AI applications. (Reuters)
The deal comes as AI firms pivot from building ever-bigger models to focusing on continuous operation for users. This next step—known as inference, when models generate responses—is turning into the routine workload and the main expense.
Speed matters in the sales pitch. Chatbots that respond fast usually see more use, with users relying on them for tougher tasks: lengthy answers, code generation, images, and multi-step “reasoning” that requires the system to spend extra time solving problems.
OpenAI plans to phase in Cerebras capacity across various inference workloads. Sachin Katti from OpenAI framed the goal as creating “a resilient portfolio that matches the right systems to the right workloads.” Cerebras CEO Andrew Feldman added, “Just as broadband transformed the internet, real-time inference will transform AI.” (OpenAI)
Cerebras announced a multi-year deal to deploy 750 megawatts of its wafer-scale systems for OpenAI clients, with rollouts beginning in 2026. The company claims its chips can produce responses “up to 15× faster than GPU-based systems,” referring to traditional graphics processors common in AI data centers. (Cerebras)
Neither company revealed pricing details in their announcements. The power figure stands out—megawatts are a blunt measure in an industry that usually breaks things down by racks and chips. Essentially, megawatts show how large a data center expansion can be, and how fast it can fill up.
The Wall Street Journal reported that OpenAI has struck a multibillion-dollar computing deal with Cerebras, aiming to boost its capacity beyond what traditional suppliers can offer. (Wall Street Journal)
The Financial Times put the Cerebras deal at $10 billion, stretching through 2028, as part of OpenAI’s push to broaden its hardware sources. It valued Cerebras at $8.1 billion and highlighted that its chips power France’s Mistral for the Le Chat chatbot. (Financial Times)
Cerebras is among the challengers aiming to chip away at Nvidia’s dominance in the AI chip market, where Nvidia’s GPUs hold sway over much of the training and inference stack. AMD is also making significant moves to expand its presence in this arena.
The downside scenario is all too familiar: massive figures, extended timelines, and an enormous amount of concrete to pour. Phasing in capacity through 2028 allows for potential hold-ups in data center construction and power connections. Plus, a dip in AI demand could saddle suppliers with costly equipment and power agreements they’d still have to honor.
OpenAI is wagering that quicker responses will drive higher usage, which in turn justifies securing industrial-scale compute resources. The competition is shifting focus from chips to power, and this deal highlights where the real bottleneck lies.
