Anthropic and Amazon expand AWS pact for 5 gigawatts
If you build AI automations on self-hosted or cloud-hosted models, treat compute planning as part of product planning, not a later ops task. Design your agent...
Anthropic and Amazon have expanded their partnership with a new agreement that secures up to 5 gigawatts of compute capacity for training and running Claude, Anthropic said on April 20, 2026. The company said the deal covers new Trainium2 capacity coming online in the first half of this year and nearly 1 gigawatt of total Trainium2 and Trainium3 capacity by the end of 2026.
The scale matters because AI models do not train or serve themselves. Compute, meaning the chips, servers, networking, and power needed to run model workloads, has become one of the main constraints on which AI systems can grow, how fast they can respond, and where they can be deployed. Anthropic said it has worked with Amazon since 2023 and already uses more than one million Trainium2 chips to train and serve Claude.
⚡ New to this?
This matters because AI models need huge amounts of computing power, which is the mix of chips, servers, and electricity that makes them run. Anthropic and Amazon are saying they are spending a lot more to make Claude available at larger scale, with better access for businesses that already use AWS. If you use AI tools, this can affect price, speed, and whether those tools stay available when demand spikes.
🦞 OpenClaw angle
If you run AI agents or self-hosted tools, treat model capacity as a dependency, not a background detail. Add fallbacks for slower or unavailable models, and test what happens when latency rises or a provider throttles requests. If you already use AWS, map your agent permissions, billing, and audit logs now so you can switch between hosted model access paths without changing your governance setup later.
Amazon’s latest investment also puts more money behind that relationship. According to Anthropic, Amazon is investing $5 billion today, with up to another $20 billion possible in the future, building on the $8 billion Amazon has already invested. Amazon CEO Andy Jassy said its custom AI silicon is in high demand because it offers high performance at lower cost for customers, while Anthropic CEO Dario Amodei said the partnership will help the company keep pace with demand.
The deal is not only about raw capacity. Anthropic said it will use incremental capacity for Claude in Amazon Bedrock, Amazon’s managed service for running foundation models inside AWS. The company also said the full Claude Platform will be available directly within AWS, with the same account, controls, and billing, and no extra credentials or contracts needed. That should matter to organizations that want AI tools without adding new vendor sprawl to procurement and compliance processes.
Anthropic framed the agreement as part of its effort to keep Claude available at scale as demand rises. The company said over 100,000 customers now run Claude on Amazon Bedrock, and that enterprise and developer usage accelerated in 2026 alongside sharp growth in consumer usage across free, Pro, and Max tiers. Anthropic said that demand has put strain on reliability and performance, especially during peak hours, which helps explain why it is expanding infrastructure rather than treating capacity as an afterthought.
The hardware strategy is also broader than a single chip line. Anthropic said the agreement spans Graviton and Trainium2 through Trainium4, and includes the option to buy future generations of Amazon custom silicon as they become available. It also includes expansion of inference in Asia and Europe, a sign that Anthropic is trying to serve customers closer to where they operate instead of pushing all traffic through a single region.
For now, the immediate next step is access rather than launch-day fanfare. Anthropic said Claude Platform on AWS is coming soon, and interested customers should contact their account team to request access.
