Anthropic’s Infrastructure Was Failing. So It Signed a $100 Billion Deal.

Amazon is putting up $25 billion. Anthropic is committing $100 billion over a decade to AWS. Anthropic's own statement admits what neither company wanted to lead with.

Amazon is dropping $5 billion into Anthropic today. Up to $20 billion more after that, tied to commercial milestones, on top of the $8 billion it already put in. Anthropic's side of the trade is bigger and stranger: $100 billion committed to AWS over the next ten years. Trainium chips across three generations, tens of millions of Graviton cores, 5 gigawatts of compute capacity. This makes it seem more like a land purchase rather than deal.

Two companies deepening a relationship that started in 2023. Amazon betting on infrastructure revenue, Anthropic betting it can build its way out of a capacity problem it was already living with.

Also Read: Anthropic Gives Claude Code a Big Desktop Redesign

Even Anthropic's own announcement says the deal came out of necessity. The company's run-rate revenue surpassed $30 billion this year, up from roughly $9 billion at the end of 2025. That growth, extraordinary by any measure, created what Anthropic describes as "inevitable strain" on its infrastructure. Free, Pro, Max, and Team users experienced degraded reliability during peak hours. The company needed more compute, and it needed it fast. OpenAI had spent recent months publicly criticizing Anthropic for exactly this — for not securing enough compute, for letting a gap open while demand was already accelerating. 

Monday's deal closes that argument, and the timing says something even if the causation doesn't.

A hundred billion dollars committed to one vendor over a decade is not a partnership in the casual sense. It is a structural bet that ties model development to Amazon's silicon roadmap — if Trainium3 underperforms, if future chip generations fall short of what training frontier models requires, Anthropic's options narrow in ways that are hard to reverse. The company says it maintains a diversified hardware strategy, with a separate deal covering Google and Broadcom, but the weight of $100 billion points in one direction. Switching costs at this scale don't just mean moving workloads. They mean renegotiating the foundation the models are built on.

Also Read: Anthropic Accidentally Exposes Claude Code Source Code in npm Release      

For Amazon, the structure gets better as Anthropic succeeds. The additional $20 billion is tied to commercial milestones, meaning every new tier of Claude's growth adds to Amazon's leverage, not Anthropic's. Andy Jassy's quote in the announcement focuses on Trainium demand and silicon performance. Not model capabilities, not safety research. The incentives only run one way.

Anthropic needed scale it couldn't build alone, but a ten-year infrastructure commitment to a single cloud provider is also a decade of narrowed room to move — to negotiate, to change direction, to push back. Nearly 1 gigawatt of capacity is coming online before end of 2026, with significant Trainium2 already arriving in Q2.

The compute is real. Anthropic just answered the question of who controls the ground it builds on for the next decade.