Amazon secures 5GW for Anthropic with up to $25B AI compute pact

This deal is as close as we’ve seen to a sovereign‑grade compute treaty between a hyperscaler and an AI lab. By committing up to $25 billion in equity and securing $100 billion in future Claude spend plus 5 gigawatts of Trainium capacity, Amazon is effectively underwriting Anthropic’s entire training and inference roadmap for the late 2020s. That scale of dedicated power and silicon dwarfs what most national governments are assembling and puts Anthropic in the same compute league as OpenAI’s Microsoft‑backed clusters and Google DeepMind’s internal TPU fleets.([anthropic.com](https://www.anthropic.com/news/anthropic-amazon-compute?ss_ad_code=s_01_soracom&utm_source=openai))

Strategically, this is Amazon’s answer to Microsoft–OpenAI and Google–Anthropic–Broadcom: if you can’t own the leading lab outright, own the rails it runs on. For Anthropic, the trade is clear—massive, relatively cheap capacity in exchange for long‑term platform dependence and engineering effort to target Trainium. For the broader race to AGI, the implication is that compute concentration is hardening into a three‑horse contest dominated by Microsoft, Amazon and Google. Smaller labs and open‑source projects will find it increasingly difficult to match the sheer volume of training runs these consortia can afford. That dynamic doesn’t guarantee who will reach AGI first, but it does mean that, in practice, only a handful of players will have the hardware to seriously try.([axios.com](https://www.axios.com/2026/04/21/anthropic-amazon-compute-wars?utm_source=openai))