Product Launch read

Cerebras launches disaggregated inference pairing GPUs with its systems

Cerebras has put disaggregated inference into production, with GPUs handling prompt processing and Cerebras systems handling token generation. The company says the setup delivered five times the throughput per megawatt in a production test.

AE
Aidose Editorial
Nov 18, 2026

Cerebras announced on November 17 that disaggregated inference is now in production. In this setup, GPUs handle prefill, the processing of the input prompt, while Cerebras systems handle decode, the step that generates output tokens.

The production deployment used AMD Instinct GPUs for prefill and Cerebras CS-3 systems for decode. In a test on a shared GPT-OSS-120B endpoint on November 14, Cerebras says the combined configuration delivered five times the throughput per megawatt of a CS-3-only setup, with no loss in token generation speed. The company notes that performance depends on workload, configuration and model.

Cerebras is offering two software options to infrastructure partners. The first is an end-to-end stack with a disaggregated routing framework, a prefill engine based on vLLM or SGLang, the Cerebras decode engine and KV Portal, which moves model state between systems over RDMA. The second is a decode-only SDK for partners that want to bring their own prefill engine, with a mock decode replica and a reference integration.

Early-access partners are Amazon Web Services, Fireworks AI, Gimlet Labs and General Compute, and Cerebras says it is expanding access. The company plans 700MW of data center capacity and expects to deploy disaggregated inference across hundreds of megawatts by the end of 2027.

Splitting prefill and decode across different hardware lets each part of inference run on the chips best suited to it. For developers using inference providers that adopt this setup, the practical effect Cerebras is pitching is more fast-decode capacity per unit of power, which matters most for agentic workloads that generate long outputs.

AE
Aidose Editorial
Nov 18, 2026 · read
Back to News