What happened
In a blog post dated 17 November 2026, Cerebras described a disaggregated inference design: GPUs (AMD Instinct) handle prefill, and Cerebras CS-3 systems handle decode. The figures below are Cerebras’s own.
Key details
- Result claimed: “5X the throughput per megawatt of a CS-3-only configuration, with no loss in token generation speed”, measured on a shared GPT-OSS-120B endpoint on 14 November.
- Rollout: Cerebras plans to deploy across hundreds of megawatts by the end of 2027 and mentions a planned 700MW of data center capacity and the recently announced CS-4.
- Partners and software: early-access partners are Amazon Web Services, Fireworks AI, Gimlet Labs and General Compute. The stack is offered end to end or as a decode-only SDK; KV Portal moves KV caches over RDMA.
- Not stated: pricing, availability dates for general customers, and independent measurements.
Why it matters
The design targets power efficiency per megawatt, a main constraint for inference capacity. The 5X figure comes from a single Cerebras measurement on one model, so it is a claim to watch for independent confirmation.
Disclosure: Claude, made by Anthropic, was one of the AI tools used to research and draft this article. Cerebras is not affiliated with Deyron Labs; all claims are Cerebras’s own.
Sources
- Cerebras launches disaggregated inference Primary source , Cerebras, 17 November 2026