Anthropic plans roughly 5GW of AWS compute capacity, sources say
- Anthropic and Amazon said last week they are expanding Project Rainier to secure up to 5 gigawatts of AWS capacity for Claude training and deployment. - The deal pairs a new $5 billion Amazon investment, with up to $20 billion more possible, on top of $8 billion already invested. - Weeks earlier, Anthropic also locked in multi-gigawatt Google TPU capacity, showing frontier labs now buy resilience across clouds.
AI model companies are starting to look less like software startups and more like power-hungry industrial tenants. That is the real story here. Anthropic did not just sign another cloud contract — it said last week that it will secure up to 5 gigawatts of new AWS compute capacity for training and serving Claude, while Amazon put in another $5 billion and left room for as much as $20 billion more later. That scale matters because the bottleneck in frontier AI is no longer mainly ideas. It is chips, power, and who can reserve both years ahead. (anthropic.com) ### What actually got announced? Anthropic and Amazon expanded Project Rainier, with Anthropic saying it will use up to 5 gigawatts of AWS capacity and spend more than $100 billion on AWS technologies over the next 10 years. Amazon said it is investing $5 billion immediately, building on the $8 billion it had already invested, and it may add up to $20 billio(anthropic.com) soon, which turns the infrastructure deal into a distribution deal too. (anthropic.com) ### Why is 5 gigawatts such a big number? Because this is utility-scale infrastructure, not normal “cloud usage.” A gigawatt is the kind of number you use for large data centers or power plants. Five gigawatts implies a buildout measured in massive clusters, dedicated energy planning, and long lead times. Anthropic also said nearly 1 gigawatt of Trainium2 an(anthropic.com)ust a distant option — some of it is meant to land fast. (anthropic.com) ### Why does Amazon care this much? Amazon is trying to turn Anthropic into the flagship workload for AWS custom AI chips. Back in November 2024, Anthropic named AWS its primary cloud and training partner and said it would use Trainium to train and deploy its largest models. The new deal pushes that much further. If Claude becomes a default model family for (anthropic.com)infrastructure revenue and the strategic proof that its silicon can carry frontier models at scale. (anthropic.com) ### So is Anthropic now all-in on AWS? Not really — and that is the interesting part. Anthropic said in April that it also expanded its partnership with Google Cloud and Broadcom for multiple gigawatts of next-generation TPU capacity expected to come online in 2027. It explicitly said it trains and runs Claude across AWS Trainium, Google TPUs, and Nvidia GP(anthropic.com)e for performance, resilience, and bargaining power. (anthropic.com) ### Why does multi-cloud matter here? Because the failure mode for a frontier lab is no longer just “the model is worse.” It is “we cannot get enough compute where and when we need it.” Different chips are better for different workloads, and different clouds fail in different ways. Anthropic framed that hardware mix as a way to match workloads to (anthropic.com)s really about uptime, costs, and not getting trapped by one supplier. (anthropic.com) ### What does this say about the market? It says hyperscalers are no longer just hosting AI labs — they are financing them, reserving power for them, and fighting to become the home of their inference business. Amazon says more than 100,000 customers already run Claude models on AWS. That helps explain why these deals are getting so large. The pri(anthropic.com)rs. (aboutamazon.com) ### What is the bottom line? Anthropic’s 5-gigawatt AWS plan is not a vanity number. It is a signal that frontier AI has entered the infrastructure era — where the winners are the companies that can lock in chips, power, and cloud leverage before everyone else does. (anthropic.com)
