← Back to feed

Amazon and Anthropic announced April 20 an expanded partnership that dwarfs every previous bilateral AI infrastructure deal by dollar volume: Amazon commits up to $25 billion in equity and Anthropic commits more than $100 billion in AWS cloud spend, locking the two companies together through at least 2036.

The Structure

Amazon is putting in $5 billion today, adding to the $8 billion it had already invested in two prior rounds ($4 billion in September 2023, $4 billion in November 2024). A further $20 billion is available at unspecified commercial milestones. Total potential exposure: $33 billion.

Anthropic’s side of the ledger is the more striking number. Spending $100 billion on AWS over a decade means roughly $10 billion per year in cloud infrastructure costs — a commitment that makes AWS effectively Anthropic’s only viable training and inference platform for the foreseeable future.

Compute Access

The deal grants Anthropic access to up to 5 gigawatts of computing capacity — a combination of AWS Trainium2, the forthcoming Trainium3 (launching H2 2026), and Trainium4. Anthropic already uses more than one million Trainium2 chips, most deployed in Project Rainier, one of the largest AI clusters in the world.

Trainium3 servers will add nearly 1 gigawatt of capacity for Anthropic by end of year. Trainium4, which AWS claims delivers more than 2 exaflops of FP4 performance per server rack, is included in the agreement’s future options. Graviton CPUs in the tens of millions of cores round out the compute commitment.

Product Integration

The partnership’s go-to-market component is notable: AWS customers will be able to access the full Anthropic-native Claude console directly from within the AWS portal, with no additional credentials, contracts, or billing relationships. For the more than 100,000 developers already building on AWS, this removes the last friction point between choosing AWS and choosing Claude.

What It Means

Anthropic’s revenue run rate hit $30 billion earlier this year. A $10 billion per year AWS commitment, against that revenue base, signals that Anthropic is planning for substantially larger compute expenditure as model training scales and inference demand grows. The structure — milestone-based payments from Amazon — also aligns Amazon’s additional investment with Anthropic’s commercial success, making this as much a revenue partnership as a funding deal.

For AWS, the deal cements Trainium as the credible alternative to NVIDIA for frontier model training at scale. Anthropic’s workloads are among the most demanding in the industry; a successful multi-year deployment of Trainium2, 3, and 4 makes the customer acquisition argument for every other lab considering custom silicon.

Dario Amodei: “Our users tell us Claude is increasingly essential… we need to build the infrastructure to keep pace with rapidly growing demand.”