Anthropic buys a decade of Trainium capacity with a $100 billion commitment to AWS

Spread evenly, the commitment is about $10 billion a year, roughly a third of the run-rate revenue Anthropic now reports ^[1][5]. It is priced against a business whose run-rate is about 3.3 times what it was at the end of 2025 ^[4]. Amazon's equity, if every tranche lands, tops out at $33 billion, itself about a third of what Anthropic has agreed to spend with it ^[2][3].

Anthropic dated the relief. Significant Trainium2 capacity comes online in Q2, scaled Trainium3 later in the year, and nearly 1GW of the two combined before the end of 2026 ^[3]. The company said the agreement delivers meaningful compute within the next three months ^[16]. It already runs over one million Trainium2 chips to train and serve Claude, through the Project Rainier cluster built with Amazon ^[9].

"Our custom AI silicon offers high performance at significantly lower cost for customers, which is why it's in such hot demand," said Andy Jassy, CEO of Amazon ^[7]. The announcement does not say how that cost advantage shows up in what Claude customers pay ^[17].

The portability story sits at a different layer from the commitment. Anthropic says AWS remains its primary training and cloud provider for mission-critical workloads ^[14], and that Claude is the only frontier model sold on all three of the largest clouds: Bedrock, Google Cloud's Vertex AI and Microsoft's Foundry ^[12]. More than 100,000 customers already run Claude on Bedrock ^[10]. A team buying through Vertex or Foundry keeps its own cloud choice, while the model it calls is trained on capacity contracted from one supplier for ten years ^[1].

The near-term test is narrow: whether the Q2 capacity and the year-end 1GW figure arrive as stated ^[3]. The nearer decision for buyers is Claude Platform on AWS, which Anthropic says will run in the same account with the same controls and the same billing, needing no additional credentials or contracts, and which it describes as coming soon ^[11]. That lowers procurement friction and puts more of a buyer's Claude spend inside one cloud's billing relationship. Anthropic also said the agreement expands inference in Asia and Europe ^[13].

Anthropic describes its own hardware approach as diversified, with workloads spread across a range of chips ^[15]. The decade of new capacity in this agreement is contracted with one supplier, spanning Graviton and Trainium2 through Trainium4, with an option on future generations ^[1][2]. "Our users tell us Claude is increasingly essential to how they work, and we need to build the infrastructure to keep pace with rapidly growing demand," said Dario Amodei, CEO and co-founder of Anthropic ^[8].