Amazon doubles down on AI infra

Amazon is preparing to spend about $200 billion on capital expenditures in 2026, and Chief Executive Officer Andy Jassy said most of that money is headed to data centers that feed artificial intelligence demand rather than to warehouses or delivery vans. Amazon made the plan public with its February 2026 results, and Jassy defended it again on April 9 by saying the company is “not going to be conservative” in a market moving this fast. (ir.aboutamazon.com, cnbc.com) That is a huge jump from the roughly $100 billion Amazon said it expected to spend in 2025, with the vast majority of that earlier budget also aimed at new artificial intelligence data centers. The company is effectively saying the first wave of spending did not build enough capacity, so it is doubling down one year later. (aws.amazon.com, cnbc.com) The business writing the checks is Amazon Web Services, which is Amazon’s cloud unit. Amazon Web Services rents computing power the way a utility sells electricity, and artificial intelligence customers now want far more of it because training and running models uses thousands of chips at once. (ir.aboutamazon.com, cnbc.com) Amazon Web Services is not just buying more outside hardware. Jassy said Amazon’s custom chip business, which includes Graviton processors, Trainium artificial intelligence chips, and the Nitro system that handles security and networking tasks, is already running at more than $20 billion a year in revenue. (cnbc.com) Those three products do different jobs inside the same machine room. Graviton is Amazon’s general-purpose central processor for ordinary cloud workloads, Trainium is the specialist chip for building and serving artificial intelligence models, and Nitro is the offload system that takes housekeeping work away from the main server so more of the machine can be sold to customers. (ir.aboutamazon.com, press.aboutamazon.com) Amazon’s fourth-quarter release showed how fast that internal hardware push is growing. The company said Trainium and Graviton together had already reached an annual revenue run rate above $10 billion, that Trainium2 was fully subscribed, and that 1.4 million Trainium2 chips had been landed. (ir.aboutamazon.com) The reason Amazon wants its own chips is the same reason a grocery chain builds warehouses instead of renting every shelf from a rival. If Amazon can design more of the stack itself, it can lower costs, avoid some dependence on Nvidia, and lock customers deeper into Amazon Web Services infrastructure. (cnbc.com, aws.amazon.com) Amazon is also trying to prove the demand is real before Wall Street worries about empty buildings. Jassy said Amazon has already received customer commitments for a substantial portion of the 2026 capital spending and expects to monetize most of it in 2027 and 2028. (cnbc.com) One of those commitments is unusually large. Amazon said in March 2026 that OpenAI would consume 2 gigawatts of Trainium capacity on Amazon Web Services infrastructure, and Amazon said it would invest $50 billion in OpenAI as part of that broader partnership. (aboutamazon.com) This spending spree is possible because Amazon Web Services is already enormous. The cloud unit generated $128.7 billion in sales and $45.6 billion in operating income in 2025, which gives Amazon a profit engine big enough to fund a buildout that would be hard for most rivals to copy. (ir.aboutamazon.com) So this is not Amazon making a side bet on artificial intelligence software. It is Amazon treating physical infrastructure itself, from chips to servers to power-hungry data centers, as the product that could decide who wins the next phase of cloud computing. (cnbc.com, ir.aboutamazon.com)