Anthropic reportedly reserving ~5GW-per-cloud across multiple providers to lock compute

AI training now looks less like renting cloud servers and more like pre-buying slices of the electric grid. That’s the real story behind the Anthropic compute spree. Over the past six months — and especially in April and May 2026 — Anthropic has lined up huge capacity commitments across Amazon, Google, Microsoft, Fluidstack, and SpaceX, instead of betting on one exclusive provider. The point is simple: if compute is the bottleneck, reserve it everywhere you can. ### What actually changed? The biggest concrete update landed on April 20, 2026. Anthropic and Amazon said they had signed a new agreement securing up to 5 GW of Amazon capacity for training and deploying Claude, with new Trainium2 capacity coming online in the first half of 2026 and nearly 1 GW of Trainium2 and Trainium3 expected online by the end of 2026. Amazon paired that with another $5 billion Anthropic investment and said the deal could scale further. ### Why is 5 GW such a big deal? Because gigawatts are power-plant numbers, not normal cloud-contract numbers. A single large AI cluster might be discussed in megawatts. Anthropic is now talking in multiple gigawatts with at least one provider. That means this is not just about one training run — it’s about securing enough electricity, buildings, networking, and chips to keep training and inference growing for years. (anthropic.com) ### Is Amazon the whole story? No — and that’s the important part. Anthropic still calls Amazon its primary cloud provider and training partner, but it has also expanded with Google Cloud and Broadcom. In April, Anthropic said it trains and runs Claude on AWS Trainium, Google TPUs, and NVIDIA GPUs, and framed that hardware mix as a resilience and performance advantage. Basically, the company is saying exclusivity is less valuable than optionality. (anthropic.com) ### What does the Google side look like? Anthropic had already announced in October 2025 that it would expand its use of Google Cloud technologies, including up to 1 million TPUs, with “well over a gigawatt” of capacity expected online in 2026. The newer Google-Broadcom announcement deepens that path rather than replacing Amazon. So when people talk about “5 GW per cloud,” the Amazon figure is explicit, while the Google side is clearly enormous but described more through TPU scale and multibillion-dollar infrastructure language than one clean public 5 GW line. (anthropic.com) ### Where do Azure and Fluidstack fit? Microsoft said in November 2025 that Anthropic committed to purchase $30 billion of Azure compute capacity and could contract additional capacity up to 1 GW. Separately, Anthropic announced a $50 billion U.S. infrastructure build with Fluidstack in Texas and New York. Those sites are custom-built for Anthropic workloads, which means this is not only cloud leasing — it’s also dedicated physical infrastructure. (anthropic.com) ### Why add SpaceX too? Because demand is hitting Anthropic’s products right now, not just in some future training cycle. On May 6, 2026, Anthropic said it would use all compute capacity at SpaceX’s Colossus 1 data center — more than 300 MW and over 220,000 NVIDIA GPUs within the month — to raise Claude usage limits. That tells you the shortage is immediate enough that even a giant Amazon deal was not sufficient on its own. (blogs.microsoft.com) ### So what’s the real strategy here? Anthropic seems to be treating compute like a portfolio. Train on Trainium where economics work. Use TPUs where they fit. Add NVIDIA-heavy capacity where deployment needs spike. Build custom sites where owning the footprint matters. The catch is that this only works if you have the capital, partner leverage, and demand curve to justify reserving capacity before you need every watt. Anthropic increasingly does. (anthropic.com) ### Bottom line The headline is not just that Anthropic got a huge Amazon deal. It’s that Anthropic is assembling a multi-provider compute stack at utility scale. That changes the competitive picture. Frontier AI labs are no longer merely racing to build better models — they’re racing to lock in power, chips, and data-center space before those inputs disappear. (anthropic.com)