AWS capacity crunch reported
Some large customers are reportedly trying to buy out AWS capacity for AI workloads as demand explodes, creating short‑term limits on cloud availability. The trend was framed as a procurement constraint for heavyweight training and inference projects. (networkworld.com)
Amazon Web Services is telling investors that demand for artificial intelligence computing is outrunning what it can deliver right now. Two large customers asked to lock up all available 2026 capacity for Amazon’s Graviton chip instances, according to Chief Executive Andy Jassy. (aboutamazon.com) Jassy said Amazon Web Services added 3.9 gigawatts of new power capacity in 2025 and still has “capacity constraints” that leave demand unserved. He said the company expects to double its total power capacity by the end of 2027. (aboutamazon.com) Cloud capacity is the rented computing and electricity behind model training and inference, the step where a model answers prompts after it has been built. When a provider runs short, customers cannot simply click a button and get more high-end instances on demand. (networkworld.com) Amazon is trying to ease that squeeze with its own chips as well as Nvidia hardware. Jassy said the Trainium and Graviton chip businesses now have a combined annual revenue run rate in the “multi-billion dollar” range, and Network World reported that Amazon is pitching Trainium on price-performance against Nvidia. (aboutamazon.com) (networkworld.com) The pressure is showing up in Amazon’s spending plans. Amazon said in February that it expects about $200 billion in capital expenditures in 2026, with most of that aimed at artificial intelligence infrastructure for Amazon Web Services. (finance.yahoo.com) (ciodive.com) Amazon Web Services is also expanding the products that need that infrastructure. At re:Invent in Las Vegas from November 30 to December 4, 2025, Amazon announced new Nova models, Bedrock agent tools, and Trainium3 UltraServers, all of which add to demand for data center power, networking, and chips. (aws.amazon.com) (aboutamazon.com) Industry analysts say the bottleneck is not only chips. Dell’Oro Group data cited by Network World shows data center capital spending rose 57% in 2025 to a record high, while backlogs at Amazon and Google point to demand that is still climbing faster than new supply. (networkworld.com) Amazon is framing the shortfall as a buildout problem, not a demand problem. The next test is whether it can turn that 2026 spending surge into enough live capacity before customers shift more training and inference work to Microsoft Azure or Google Cloud. (networkworld.com)