Anthropic mixes Amazon Trainium chips and Nvidia GPUs in multi‑cloud deployment
- Anthropic said this week it added SpaceX’s Colossus 1 to Claude’s infrastructure, while continuing to train and serve models on AWS Trainium, Google TPUs, and Nvidia GPUs. - The new SpaceX deal gives Anthropic all capacity at Colossus 1 — over 300 megawatts and 220,000 Nvidia GPUs — within the month. - The bigger shift is strategic: Anthropic is building Claude across rival chips and clouds instead of betting the company on one stack.
AI labs are turning into infrastructure buyers as much as model builders. That is the real story here. Anthropic did not just add more chips this week — it made unusually explicit that Claude now runs across Amazon Trainium, Google TPUs, and Nvidia GPUs, with capacity coming from multiple clouds and now SpaceX’s Colossus 1 as well. That matters because frontier AI is no longer bottlenecked mainly by ideas. It is bottlenecked by power, data centers, and whoever can actually deliver working compute on time. (anthropic.com) ### What changed this week? Anthropic announced on May 6 that it signed a compute partnership with SpaceX and will use all of the capacity at the Colossus 1 data center. Anthropic said that adds more than 300 megawatts of new capacity and over 220,000 Nvidia GPUs within the month, and that the extra headroom is already letting it raise Claude Code and Claude API usage limits. (anthrop([anthropic.com) is the hardware mix the interesting part? Because Anthropic is not presenting this as a one-vendor build. In its recent infrastructure announcements, the company has said plainly that it trains and runs Claude on AWS Trainium, Google TPUs, and Nvidia GPUs. That is a very different posture from the simple version of the AI race where one lab pairs with one cloud and one chip fam(anthropic.com)t workloads can land on different silicon. (anthropic.com) ### What is Amazon actually providing? Amazon is still the anchor partner. On April 20, Anthropic and Amazon said they were expanding their collaboration for up to 5 gigawatts of new compute, tied to more than $100 billion of Anthropic spending on AWS technologies over 10 years. The commitment spans Trainium2 through Trainium4, with Anthropic saying it already uses more (anthropic.com)primary cloud provider and training partner through Project Rainier. (anthropic.com) ### And what is Google providing? Google is the second major rail. Anthropic said last month that it was expanding its partnership with Google Cloud and Broadcom, building on an October 23, 2025 Google Cloud deal worth tens of billions of dollars and expected to bring well over 1 gigawatt of TPU capacity online in 2026. So this is not a side experiment. Google TPUs are part of the production plan. (anthropic.com) ### Why add SpaceX if Amazon and Google are already huge? Speed. The SpaceX deal looks like near-term Nvidia capacity while the longer Trainium and TPU buildouts come online. Anthropic itself framed the SpaceX agreement as a way to “substantially increase” compute capacity in the near term, and outside coverage tied the deal directly to filling the gap before those other(anthropic.com)s is the fast lane. (anthropic.com) ### Why not just standardize on Nvidia? Because standardizing is simpler, but dependence is expensive. Nvidia still matters a lot — Colossus is all about Nvidia GPUs — but Anthropic’s recent announcements read like a deliberate hedge against single-supplier risk. Mixing Amazon’s chips, Google’s TPUs, and Nvidia clusters gives Anthropic more resilience if one roadmap slips, one cloud ge(anthropic.com)in negotiations, even if Anthropic never says it that bluntly. The leverage point here is an inference from the multi-provider design and the scale of the separate deals. (anthropic.com) ### What does this mean for the AI race? It means the frontier is starting to look like a portfolio, not a platform. The winning labs may be the ones that can juggle custom silicon, merchant GPUs, and multiple cloud relationships without breaking their training pipeline. Anthropic’s move makes that shift visible. Claude is not being built on one machine. It is being assembled across an infrastructure market. (anthropic.com) ### Bottom line? Anthropic is turning compute sourcing into a competitive skill. The model company is now also a power buyer, chip allocator, and cloud negotiator — and that may matter almost as much as the model itself.
