Anthropic Locks Up to 5 GW on AWS — $100B+ Decade Bet as Revenue Hits $30B Run-Rate
Anthropic–Amazon expand: up to 5 GW new compute (Trainium2/3 heavy), >$100B over ten years to AWS tech, Amazon +$5B now (up to +$20B later). Anthropic cites >$30B run-rate revenue (from ~$9B end-2025) and peak-hour reliability strain. Multi-cloud Claude stays; AWS remains primary for mission-critical.
Correction (October 14, 2026): Clarified that Claude Platform on AWS was coming soon, matching Anthropic’s April 21 clarification; the agreement did not make it generally available on April 20.
Frontier labs are making decade-long bets on power and custom silicon, not just buying GPU spots. Anthropic’s Amazon expansion prices that reality: up to 5 GW, a >$100 billion ten-year AWS tech commitment, and fresh Amazon equity — while Claude stays multi-cloud for customers who refuse lock-in.
Anthropic (April 20) expanded its Amazon collaboration for up to 5 gigawatts of new training/inference capacity, including significant Trainium2 and Trainium3. Commitment: more than $100 billion over ten years to AWS technologies (Graviton, Trainium2 through Trainium4). Amazon invests an additional $5 billion now, with up to $20 billion more possible (on prior ~$8 billion). New Trainium2 in H1 2026; nearly 1 GW Trainium2/3 by year-end; inference expansion in Asia and Europe. Full Claude Platform on AWS was announced as coming soon, with unified accounts, controls, and billing. Claude remains on Bedrock, Vertex AI, and Azure Foundry; AWS stays primary for mission-critical workloads including Project Rainier and >1 million Trainium2 chips in use (company figures).
Demand context (Anthropic-stated)#
| Signal | Figure (Anthropic primary) |
|---|---|
| Run-rate revenue | >$30B (from ~$9B end-2025) |
| Bedrock Claude customers | >100,000 (partnership framing since 2023) |
| Reliability | Peak-hour strain across free/Pro/Max cited as capacity driver |
Dario Amodei emphasized advancing research while delivering Claude; Andy Jassy described the decade-long Trainium commitment as reflecting progress in custom silicon.
Claims vs checks#
Capacity, investment, and platform terms are Anthropic / Amazon primary. Revenue run-rate and Trainium chip counts are company disclosures, not audited GAAP segments. “One of the world’s largest clusters” for Rainier is marketing language. Complementary Google/Broadcom expansions are referenced as parallel, not measured here.
Limits#
- 5 GW is a ceiling (“up to”), not a same-day energized fleet.
- $100B+ is a multi-year technology commitment description — cash timing opaque.
- Multi-cloud availability does not equal equal capacity or feature parity across clouds.