Amazon's AI capacity pinch

Amazon’s cloud division is now running into a strange problem: customers want so much artificial intelligence computing power that some have asked to buy entire pools of Amazon capacity, not just rent a few servers at a time. Chief executive Andy Jassy said Amazon Web Services is already at more than a $15 billion annual artificial intelligence revenue run rate, and demand is still ahead of supply. (cnbc.com) (finance.yahoo.com) That is unusual because cloud computing was built on the idea of shared capacity. Amazon Web Services normally slices giant data centers into rentable pieces, but Jassy said two large customers asked whether they could take all of Amazon’s custom Graviton central processing unit capacity for 2026. (finance.yahoo.com) The bottleneck is not just software. Artificial intelligence systems need huge amounts of electricity, networking gear, and specialized chips, and Jassy said Amazon added 3.9 gigawatts of capacity in 2025 and plans to double that by 2027. (finance.yahoo.com) Amazon is spending at a scale that would have looked absurd a few years ago. The company said it expects to spend about $200 billion on capital expenditures in 2026, with most of that going to artificial intelligence infrastructure such as data centers, chips, and networking equipment. (cnbc.com) This is happening inside a cloud business that is already enormous. Amazon’s 2024 annual report said Amazon Web Services revenue rose 19% year over year to $108 billion, which gives Amazon room to fund a much bigger buildout than smaller rivals can manage. (aboutamazon.com) (s2.q4cdn.com) The reason Amazon wants its own chips is simple: renting someone else’s best chip is like running a taxi fleet with leased cars during a shortage. Jassy said Amazon still uses Nvidia chips, but customers want better price-performance, so Amazon has been pushing its own Trainium line for artificial intelligence work and Graviton line for general computing. (finance.yahoo.com) Amazon says those chips are no longer a side project. Jassy said Amazon’s custom chip business, which includes Graviton, Trainium, and Nitro, is at more than a $20 billion annual revenue run rate and growing at triple-digit percentages year over year. (cnbc.com) The sales pitch is lower cost. Amazon says first-generation Trainium can cut training costs by up to 50% versus comparable Amazon Elastic Compute Cloud instances, while Trainium2 offers 30% to 40% better price performance than certain graphics processing unit-based cloud instances. (aws.amazon.com) That cost gap is why Jassy is now talking about selling the hardware itself. He said Trainium2 is sold out, Trainium3 is nearly fully subscribed after starting shipments in early 2026, and Amazon could eventually sell racks of chips to outside buyers instead of only monetizing them through Amazon Web Services. (finance.yahoo.com) (aws.amazon.com) He is making the same argument about robots. Amazon said in October 2025 that it had deployed more than 1 million robots across its operations network, and Jassy signaled that if Amazon keeps improving those systems for its own warehouses, selling robots to third parties could also become a business. (aboutamazon.com) (finance.yahoo.com) So the story is shifting from “Amazon rents cloud servers” to “Amazon is trying to control the full stack.” When customers are fighting over capacity, the company that owns the data centers, the chips, the networking, and eventually maybe the robots gets to decide both the price and who gets served first. (cnbc.com) (finance.yahoo.com)