AI chipmaker Cerebras is betting that the fastest way to serve AI models is to split the work. On November 17 it said it now runs "disaggregated inference" in production: GPUs process the incoming prompt, and Cerebras systems generate the answer.
In its live setup, AMD Instinct GPUs handle the prefill stage and Cerebras CS-3 systems handle decode, linked by a KV cache transfer over RDMA. Measured on a GPT-OSS-120B endpoint on November 14, the company reports five times more tokens per megawatt than a CS-3-only setup, with no loss in generation speed.
Cerebras plans to roll the approach out across hundreds of megawatts in its own cloud and partner sites by the end of 2027. Partners can take the full stack, built on vLLM or SGLang for prefill, or a decode-only SDK. Early access partners include Amazon Web Services, Fireworks AI, Gimlet Labs and General Compute.
Why it matters: power, not chips, is now the main limit on AI data centers. Any design that serves more answers per megawatt changes the economics for cloud providers and the enterprises paying for AI.
