Amazon puts another $5B into Anthropic. Anthropic promises $100B back to AWS.

Amazon is putting another $5 billion into Anthropic, with up to $20 billion more tied to future milestones. In exchange, Anthropic committed to spend over $100 billion on AWS over the next decade and take delivery of up to 5 gigawatts of new compute. The announcement landed late Monday on both companies’ newsrooms. It’s the second giant circular deal Amazon has signed this quarter.

The shape of the commitment

Anthropic’s own announcement spells out the compute schedule. Significant Trainium2 capacity hits in Q2 2026. Scaled Trainium3 arrives later in 2026. By year-end, the combined Trainium2 and Trainium3 footprint for Anthropic alone will be nearly 1 GW. The contract extends through Trainium4 with options on future generations, plus “tens of millions” of Graviton cores for CPU-bound workloads.

Dario Amodei’s quote in the release is the straightforward version: “Our users tell us Claude is increasingly essential to how they work, and we need to build infrastructure to keep pace with demand.” Translation: Claude run-rate revenue is up big, and the only way to serve it is to lock in wafer-scale capacity years ahead of demand.

Amazon’s side of the deal now totals around $13 billion in direct investment, per TechCrunch. The additional $20 billion is tranched against commercial milestones that neither company has disclosed, though the structure is similar to the tranched Google follow-ons and to Amazon’s February OpenAI deal.

Why the round-trip accounting matters

When a hyperscaler writes an equity check to an AI lab and the lab spends it back on that same hyperscaler’s cloud, both sides post bigger numbers. Amazon books $100B+ of future AWS revenue. Anthropic books $5B to $25B of fresh capital. The cash mostly doesn’t move; it circulates.

That’s not fraud, and it’s not new. It’s how the 2022-2026 AI compute cycle is financed. The thing to watch is the ratio of equity to committed spend. For Anthropic’s original 2023 Amazon deal, it was roughly $4B equity to ~$8B compute. Today’s ratio is closer to 1:8 ($13B in equity against $100B+ spend), and closer to 1:4 if Amazon fully funds the tranches. Either way, Anthropic is now financially locked into AWS the way a homeowner is locked into a 30-year mortgage.

Not the only house Anthropic is renting

Earlier this year Anthropic also signed a multi-year Google Cloud commitment, and a significant Nvidia GPU buy for internal R&D. It’s the same hedging strategy Meta is running with Broadcom: spread demand across silicon suppliers so no single one can squeeze you in 2028. Anthropic’s version is unusual because the equity cap table pulls in the opposite direction. Amazon and Google are both major shareholders and both major compute vendors. At some point “neutral arbiter of model quality” gets harder to say with a straight face.

None of this changes who the headline beneficiaries are. Trainium was a skunkworks project five years ago. It’s now booked as the substrate for Claude. That’s a credibility trade Amazon couldn’t have bought any other way.

What this means for you

If you’re building on Claude through the API or Amazon Bedrock, expect capacity to loosen in the second half of 2026 as the first Trainium2 blocks land. Anthropic’s rate-limit story has been tight all year. The Monday deal is the first clear signal that it stops being tight, and it’s the answer to the “why are Claude Code users hitting caps?” complaints from March. If you evaluated Claude six months ago and got priced out on throughput, it’s worth re-running the numbers in Q3.

If you’re a procurement or architecture lead at a larger shop, the longer play is the custom-silicon story. Trainium3/4 is going to get pushed hard on price-per-token versus H-class Nvidia. The honest comparison needs your own workload, not a vendor benchmark. Run it. If Amazon can’t beat a 40% price delta on comparable tokens/sec by mid-2027, the hardware investment thesis here wobbles.

One thing not to read into this: the deal doesn’t make Anthropic “Amazon’s lab.” Google still sits on a huge stake. So does the founder team, employee pool, and a cluster of sovereign funds. What it does do is make Anthropic’s infrastructure story indistinguishable from AWS’s.