AWS capacity crunch

Amazon Web Services says demand for artificial intelligence computing is now tight enough that some customers want to lock up whole pools of capacity. (aboutamazon.com) Chief executive Andy Jassy wrote on April 9, 2026 that Amazon’s chips business, including Graviton, Trainium, and Nitro, is running at more than a $20 billion annual revenue rate and growing at triple-digit year-over-year percentages. He said customers are asking to “buy up all the available capacity” for Trainium. (aboutamazon.com) Trainium is Amazon’s in-house chip for training and running large artificial intelligence models, the systems behind chatbots, coding tools, and image generators. Amazon says its Trainium2 servers became generally available on December 3, 2024, with 30% to 40% better price performance than its current Nvidia graphics-processing-unit instances. (aws.amazon.com) Amazon is pushing those chips while still expanding its Nvidia footprint. At Nvidia’s GTC conference on March 16, 2026, Amazon Web Services said it would add more than 1 million Nvidia graphics processing units across its cloud regions starting in 2026. (aws.amazon.com) That two-track strategy reflects how cloud providers are building for both supply and cost. Jassy wrote that artificial intelligence “does not have to be as expensive as it is today,” and CNBC reported Amazon plans to spend up to $100 billion in capital expenditures this year, with most of it tied to artificial intelligence infrastructure. (cnbc.com) The pressure comes from a market that changed fast after ChatGPT’s release in late 2022. Since then, Amazon has rolled out Bedrock for third-party models, Nova foundation models, a generative-artificial-intelligence Alexa upgrade, and custom silicon meant to reduce dependence on outside chip suppliers. (cnbc.com) Amazon has also tied Trainium closely to Anthropic, the artificial intelligence company in which it has invested billions. When Amazon introduced Trn2 UltraServers in December 2024, it said Project Rainier would link hundreds of thousands of Trainium2 chips for Anthropic and provide more than five times the exaflops used to train Anthropic’s then-current leading models. (press.aboutamazon.com) The hardware details show why customers are chasing reserved capacity. Amazon says each Trn2 instance uses 16 Trainium2 chips, while a Trn2 UltraServer connects 64 chips with its NeuronLink interconnect so larger models can be split across many processors without as much delay. (aws.amazon.com) Amazon is not presenting Trainium as a replacement for Nvidia so much as a second lane inside Amazon Web Services. The company’s latest message is that demand is high enough to support both: more custom chips to lower costs, and more Nvidia systems to add raw capacity. (aws.amazon.com)