Trainium prefill, CS-3 decode via Bedrock

This partnership turns Cerebras from a niche hardware vendor into a managed capacity layer inside AWS’s buying path. In practice, Bedrock lets an enterprise developer call a model through AWS while the heavy first pass, prefill, runs on Trainium and the token by token generation, decode, runs on Cerebras CS-3 systems. That matters because prefill and decode stress different parts of the stack, so splitting them can raise throughput and lower latency without asking the customer to assemble a custom serving system.1 2 3 4

• Bedrock is the distribution unlock. Cerebras already sells cloud access through its own inference data centers and API, but Bedrock puts that capacity behind AWS procurement, security, and developer workflows. That makes Cerebras easier to adopt inside large companies that already standardize on AWS.1 3 5
• The stack split is operational, not cosmetic. Prefill is the expensive step where the model reads the prompt and builds the working state. Decode is the repetitive step where it emits each next token. AWS keeps prefill on Trainium, while Cerebras handles decode on CS-3, where wafer scale hardware is tuned for very high token output.2 3 4
• There is a broader pattern here. Writer’s Bedrock availability shows how AWS distributes third party AI products through a common control plane, while Cerebras extends that model one layer lower into inference infrastructure itself. That is rarer, and it gives a specialist chip company hyperscaler reach without owning the entire customer relationship.3 5 6

The next step is for inference to be bought less like hardware and more like a managed utility inside cloud platforms. If Cerebras can prove better speed and cost on decode workloads, Bedrock can turn that technical edge into recurring usage revenue, while Amazon Nova on Cerebras later in 2026 would deepen AWS’s dependence on outside accelerators for differentiated model serving.1 2 3 4