AWS to offer Cerebras CS-3 inference on Amazon Bedrock

Amazon Web Services and Cerebras announced a collaboration to run disaggregated AI inference on Amazon Bedrock. Trainium servers handle prompt processing and Cerebras CS-3 systems handle token generation, linked by Elastic Fabric Adapter networking. AWS said the offering would be exclusive to Bedrock at launch, and no financial terms were disclosed.

• Split design: Trainium handles prefill, Cerebras CS-3 handles decode
• Offered exclusively through Amazon Bedrock; open-source LLMs and Amazon Nova to run on Cerebras later in 2026
• No financial terms disclosed