AWS pairs Cerebras + Trainium on Bedrock
AWS is combining Cerebras WSE‑3 silicon with Trainium on Amazon Bedrock to push faster, metered inference in the cloud — a direct cloud‑native alternative to buy or build with NVIDIA reported. The move signals an option for customers prioritizing metered Bedrock inference over on‑prem DGX rigs or managed GPU fleets.
AWS and Cerebras announced the multiyear collaboration on March 13, 2026 to deploy Trainium‑powered servers and Cerebras CS‑3 systems in AWS data centers and make the service available through Amazon Bedrock "in the coming months." (press.aboutamazon.com) The joint design uses "inference disaggregation"—Trainium for the parallel, compute‑heavy prefill stage and Cerebras CS‑3 (WSE‑3) for the serial, memory‑bandwidth‑heavy decode stage—linked over Amazon's Elastic Fabric Adapter (EFA) to minimize cross‑stage latency. (press.aboutamazon.com) AWS describes the split as delivering "an order of magnitude faster" inference, Cerebras claims the collaboration can provide up to 5x more high‑speed inference capacity, and multiple press reports note vendor assertions ranging from roughly 5x to as high as ~25x improvement in decode versus GPU baselines. (press.aboutamazon.com) Cerebras's WSE‑3 silicon driving the CS‑3 contains about 4 trillion transistors across ~900,000 AI‑optimized cores with 44 GB of on‑chip SRAM and a cited peak of ~125 petaFLOPS per device, while CS‑3 clusters can scale to 2,048 nodes with up to ~1.2 PB of external memory according to Cerebras datasheets. (cerebras.ai) The companies say AWS will be the first cloud provider to offer Cerebras's disaggregated inference solution exclusively through Amazon Bedrock, and AWS plans to support leading open‑source LLMs plus Amazon Nova on Cerebras hardware "later this year." Cerebras entered 2026 with major commercial momentum—an OpenAI compute agreement for up to 750 MW of capacity reported in January and a $1 billion Series H that valued the company at roughly $23 billion in February—which provides the capacity and capital underpinning the AWS rollout. (money.usnews.com)
