Cerebras says it has moved its disaggregated inference design into production, with GPUs reading the prompt and Cerebras wafer-scale systems writing the answer. In a post on its own blog, the company claims a fivefold gain in tokens per megawatt over a Cerebras-only setup, without slowing generation. That figure is Cerebras’ own measurement, and no outside party has verified it.

When we covered the AMD and Cerebras pairing in July, it was a plan on paper. The new piece is a deployment number. Cerebras measured a shared GPT-OSS-120B endpoint on its Inference Cloud on November 14, comparing output per megawatt with the split off and then on. AMD Instinct GPUs process the prompts there, and CS-3 systems produce the responses.

The split follows the two stages of a request. Prefill digests the prompt. Decode then produces the answer one token at a time, with each new token feeding the next, so every step rereads model data from memory and bandwidth sets the pace. Cerebras argues that a chip holding that data in on-chip SRAM wins this stage, and claims 350 times the SRAM and 5,400 times the bandwidth of a standard GB300 GPU. Moving prompt work to GPUs frees the wafers to spend all their time on decode.

Power is the framing Cerebras chose. Agentic products such as Muse, Copilot, and Codex need fast, interactive responses, and the company points to Goldman Sachs Research (May 20, 2026) and data center capacity estimates reported by Investing.com on November 6, 2026 to argue that the industry is short on megawatts. Cerebras says it plans 700MW of data center capacity, alongside its recently announced CS-4 chip and expanded manufacturing. It intends to run the disaggregated design across hundreds of megawatts, in its own cloud and in partner sites, by the end of 2027.

Partners get two ways in. Both start from a Cerebras reference architecture that places the prefill accelerators and Cerebras systems in one network. The end-to-end route bundles a framework that steers traffic across separate prefill and decode groups while watching replica load and uptime, a prefill engine built on open-source engines like vLLM or SGLang, a Cerebras decode engine, and KV Portal, which moves key-value caches between the two stages over RDMA. The decode-only SDK lets a partner keep its own framework and prefill engine, and adds a mock decode replica for development.

Early-access partners already include Amazon Web Services, Fireworks AI, Gimlet Labs, and General Compute. Cerebras says it is now opening access to more.

The evidence has limits. It covers one endpoint, one model, and one GPU vendor. The baseline is Cerebras’ own CS-3-only configuration, not a GPU-only cluster, so the result says nothing about cost per token against Nvidia-based serving. Cerebras published the gain as a normalized multiple and did not release absolute throughput, latency, or power figures.

For anyone buying fast inference, the date that matters is the end of 2027, when the broad rollout is due. If a fivefold power gain holds up at scale, it should show up as lower prices or more available capacity on Cerebras-served endpoints well before then, and a provider that cannot show either has not yet turned the claim into anything you can use.

Cerebras, in a post on the Cerebras blog (cerebras.ai), November 2026.