Amazon Web Services said Friday it will put processors from Cerebras inside its data centers under a multiyear partnership focused on AI inference.
The deal gives Amazon a new way to speed up how AI models answer prompts, write code, and handle live user requests. AWS said it will use Cerebras technology, including the Wafer-Scale Engine, for inference tasks.
The companies did not share the financial terms. The setup is planned for Amazon Bedrock inside AWS data centers, putting the partnership right inside one of Amazon’s main AI products.
AWS said the system will combine Amazon Trainium-powered servers, Cerebras CS-3 systems, and Amazon’s Elastic Fabric Adapter networking.
Later this year, AWS also plans to offer leading open-source large language models and Amazon Nova on Cerebras hardware. David Brown, vice president of Compute and ML Services at AWS, said speed is still a major problem in AI inference, especially for real-time coding help and interactive apps.
David said, “Inference is where AI delivers real value to customers, but speed remains a critical bottleneck for demanding workloads like real-time coding assistance and interactive applications.”
Amazon splits prefill and decode across separate chips
AWS said the design uses a method called inference disaggregation. That means splitting AI inference into two parts. The first part is prompt processing, also called prefill. The second part is output generation, also called decode.
AWS said the two jobs behave very differently. Prefill is parallel, compute heavy, and needs moderate memory bandwidth. Decode is serial, lighter on compute, and much more dependent on memory bandwidth. Decode also takes most of the time in these cases because every output token has to be produced one by one.
That is why AWS is assigning different hardware to each stage. Trainium will handle prefill. Cerebras CS-3 will handle decode.
AWS said low-latency, high-bandwidth EFA networking will connect both sides so the system can work as one service while each processor focuses on a separate task.
David said, “What we’re building with Cerebras solves that: by splitting the inference workload across Trainium and CS-3, and connecting them with Amazon’s Elastic Fabric Adapter, each system does what it’s best at. The result will be inference that’s an order of magnitude faster and higher performance than what’s available today.”
AWS also said the service will run on the AWS Nitro System, which is the base layer for its cloud infrastructure.
That means Cerebras CS-3 systems and Trainium-powered instances are expected to operate with the same security, isolation, and consistency that AWS customers already use.
Amazon pushes Trainium harder as Nvidia faces another threat
The announcement also gives Amazon another opening to push Trainium against chips from Nvidia, AMD, and other big chip companies. AWS describes Trainium as its in-house AI chip built for scalable performance and cost efficiency across training and inference.
AWS said two major AI labs are already committed to it. Anthropic has named AWS its primary training partner and uses Trainium to train and deploy models. OpenAI will consume 2 gigawatts of Trainium capacity through AWS infrastructure for Stateful Runtime Environment, frontier models, and other advanced workloads.
AWS added that Trainium3 has seen strong adoption since its recent release, with customers across industries committing major capacity.
Cerebras is handling the decode side of the setup. AWS said CS-3 is dedicated to decoding acceleration, which gives it more room for fast output tokens. Cerebras says CS-3 is the world’s fastest AI inference system and delivers thousands of times greater memory bandwidth than the fastest GPU.
The company said reasoning models now make up a larger share of inference work and generate more tokens per request as they work through problems. Cerebras also said OpenAI, Cognition, Mistral, and others use its systems for demanding workloads, especially agentic coding.
Andrew Feldman, founder and chief executive of Cerebras Systems, said, “Partnering with AWS to build a disaggregated inference solution will bring the fastest inference to a global customer base.”
Andrew added, “Every enterprise around the world will be able to benefit from blisteringly fast inference within their existing AWS environment.”
The deal adds more pressure on Nvidia, which in December signed a $20 billion licensing agreement with Groq and plans next week to unveil a new inference system using Groq technology.
