Cerebras puts combined GPU and wafer-scale inference into production

Cerebras has put a split inference system into production that uses AMD Instinct GPUs to process prompts and Cerebras CS-3 systems to generate responses, the company announced on November 17.
The arrangement separates two stages of running an AI model: processing the user’s input, known as prefill, and producing output tokens, known as decode. Moving the first stage to GPUs frees the Cerebras hardware to concentrate on generating answers.
Cerebras reports five times the throughput per megawatt of a CS-3-only setup without reducing token-generation speed. That is a company measurement, based on its shared GPT-OSS-120B cloud endpoint on November 14 before and after the change, not a universal performance result.
Infrastructure partners can use a reference architecture and either Cerebras’ complete software stack or a decode-only SDK. Early-access partners include Amazon Web Services, Fireworks AI, Gimlet Labs and General Compute.
Cerebras plans broader deployment across hundreds of megawatts in its cloud and partner systems by the end of 2027. The significance is capacity: the company is seeking to serve more fast AI responses within a constrained electricity budget, rather than relying solely on additional data-center power.
Sources and citations
Revision history
Published .