Customers Reserving AWS Capacity
Some AWS customers are trying to book huge chunks of the cloud provider’s AI capacity rather than rely on elastic on‑demand access, turning cloud supply into a rationed resource. Reports say this behaviour is driven by intense model demand and is prompting AWS to lean on its Trainium chips as a cost and capacity lever. (cio.com)
Amazon Web Services customers are no longer treating cloud computing as unlimited utility power; some are trying to lock up future capacity before anyone else can. (aboutamazon.com) Amazon chief executive Andy Jassy said in his April 9, 2026 shareholder letter that Trainium3 started shipping at the start of 2026 and is already nearly fully subscribed. He also said a significant chunk of Trainium4, which is still about 18 months from broad availability, has already been reserved. (aboutamazon.com) That is a change from the old cloud model, where customers typically rented computing on demand instead of booking years of supply in advance. Reuters, via Yahoo Finance, reported that Trainium2 capacity is effectively spoken for and that reservations have filled nearly all available Trainium3 supply. (finance.yahoo.com) Trainium is Amazon’s in-house artificial intelligence chip, built to train and run large models at lower cost than general-purpose cloud servers. Amazon says first-generation Trainium instances can cut training costs by up to 50% versus comparable Amazon Elastic Compute Cloud instances. (aws.amazon.com) Jassy said Trainium3 is 30% to 40% better on price-performance than Trainium2, and he argued that cheaper chips and better model design will lower the cost of artificial intelligence over time. CNBC reported on April 10 that Amazon plans to spend up to $100 billion in 2026 capital expenditures, with most of that tied to artificial intelligence infrastructure. (aboutamazon.com) (cnbc.com) Amazon is also building more dedicated capacity for customers that do not want to wait in the public cloud queue. In December 2025, Amazon introduced AWS AI Factories, which place dedicated AWS infrastructure in a customer’s own data center and operate it exclusively for that customer. (aboutamazon.com) The pressure is not limited to artificial intelligence chips. The Register reported on April 10 that Jassy said two large Amazon Web Services customers asked to buy all of the company’s Graviton instance capacity for 2026, and Amazon declined those requests. (theregister.com) Amazon is using this demand surge to push custom silicon harder against Nvidia, which still dominates artificial intelligence training. Jassy said Amazon’s homegrown chip business is running at more than $20 billion in annual revenue, and he framed Trainium as a lower-cost alternative as customers try to secure scarce compute. (aboutamazon.com) (finance.yahoo.com) The immediate question is whether Amazon can add enough power, data centers, and chips fast enough to keep cloud capacity from turning into a reservation market. For now, Amazon’s own disclosures show customers are already booking supply well before the hardware is broadly available. (aboutamazon.com)
