AWS and Cerebras have made an obscure inference split into a product claim, with Cerebras saying the collaboration uses AWS Trainium for prefill and Cerebras CS-3 systems for decode, while Amazon Bedrock supplies the channel customers already know. [1]

The language is technical, but the business consequence is plain: inference is not one undifferentiated act, because a model first processes the prompt and context and then generates output tokens, so hardware that handles those stages more cheaply or quickly turns architecture into part of the sale.

Cerebras' OpenAI partnership page gives the same direction from another customer angle, presenting high-speed inference as the thing being commercialized rather than just a chip photograph for investors to admire. [2]

That is why this belongs in technology rather than corporate decoration, because the AI market keeps talking about models as if the system begins and ends with intelligence, while production buyers also care about latency, throughput, integration and whether a cloud platform can make an exotic accelerator feel ordinary. [1] [2]

The social fight will ask whose chip is winning, but the AWS-Cerebras receipt asks a better question: which part of the inference job is being optimized, by whom, and how does the customer buy it?

-- DAVID CHEN, Beijing