Anthropic expands on AWS

Anthropic said on April 20 it will commit more than $100 billion over 10 years to Amazon Web Services and secure up to 5 gigawatts of new capacity to train and run Claude. (anthropic.com) The expansion centers on Amazon’s in-house chips: Anthropic said the deal spans Graviton processors and Trainium2 through Trainium4, with significant Trainium2 capacity coming online in the second quarter and larger Trainium3 capacity expected later in 2026. (anthropic.com) Anthropic said it already uses more than 1 million Trainium2 chips to train and serve Claude across direct AWS usage and Amazon Bedrock, Amazon’s managed service for outside AI models. (anthropic.com) Project Rainier is the centerpiece of that buildout. Amazon said in October 2025 that the cluster had nearly half a million Trainium2 chips in service and was expected to scale past 1 million by the end of 2025, with Anthropic as the anchor customer. (aboutamazon.com) Training an advanced AI model requires huge numbers of chips linked together like a single giant computer. Running the model after it is built, known as inference, needs its own capacity, and Anthropic said the new agreement covers both training Claude and expanding Claude inference in Asia and Europe. (anthropic.com) Anthropic and Amazon have been tightening this relationship since September 2023, when Amazon said Anthropic would use AWS Trainium and Inferentia chips and AWS would become its primary cloud provider for mission-critical workloads. (aboutamazon.com) Amazon deepened that tie in November 2024 with another $4 billion investment, bringing Amazon’s total planned investment in Anthropic to $8 billion and naming AWS as Anthropic’s primary training partner. (aboutamazon.com) The customer side of the partnership has grown alongside the infrastructure deal. Anthropic said more than 100,000 customers now run Claude on Amazon Bedrock, and AWS last week added Claude Opus 4.7 to Bedrock’s lineup. (anthropic.com) (aws.amazon.com) The immediate effect is that Anthropic is tying more of Claude’s future growth to Amazon’s custom silicon instead of relying only on the Nvidia-based systems that dominate much of the AI market. Amazon, in turn, gets a flagship buyer for the chips it has been pitching as a lower-cost alternative for large AI workloads. (anthropic.com) (aboutamazon.com) For Anthropic, the April 20 agreement turns a cloud partnership into a long-term infrastructure bet: more power, more Trainium, and more Claude capacity built inside Amazon’s stack. (anthropic.com)