AWS AI capacity crunch
Customers are trying to buy up AWS's AI capacity as demand outstrips supply, and AWS is pressing its Trainium chips as a way to change training and inference economics. The pressure on cloud AI capacity is being reported as a current constraint behind model costs and operational choices. (networkworld.com)
Amazon Web Services is telling investors that demand for artificial intelligence computing is so tight that customers have asked for all of its available chip capacity. (aboutamazon.com) Amazon Chief Executive Andy Jassy said in his April 9, 2026 shareholder letter that Amazon’s chip business is “on fire” and that demand for Trainium has outpaced supply in every generation so far. He said Trainium2 capacity is effectively spoken for, Trainium3 started reaching customers in early 2026, and reservations have filled nearly all of that supply too. (aboutamazon.com; qz.com) A cloud provider rents out computing the way a utility sells electricity, and artificial intelligence training and inference are the most power-hungry jobs in that system. Amazon says customers are now reserving so much of that capacity that it is considering eventually selling racks of its in-house chips to third parties, not just renting access through Amazon Web Services. (aboutamazon.com; theregister.com) Amazon is pushing Trainium as a cheaper alternative to Nvidia-based systems at a moment when model makers are trying to cut both training bills and the cost of each answer their systems generate. Jassy wrote that “AI does not have to be as expensive as it is today,” and said lower-cost chips, model distillation, and better infrastructure should push costs down over time. (cnbc.com; aboutamazon.com) That pitch lands as the big cloud companies are pouring record sums into data centers and networking gear to keep up with generative artificial intelligence demand. Dell’Oro Group data cited by Network World said hyperscaler capital spending rose 57% last year to a record high, with backlog figures at Amazon and Google pointing to more buildout ahead. (networkworld.com) Amazon says the newest Trainium3 UltraServers can deliver up to 4.4 times more compute performance, 4 times greater energy efficiency, and almost 4 times more memory bandwidth than Trainium2 systems. In its December 2025 launch post, Amazon said some customers were cutting training and inference costs by up to 50% on Trainium. (aboutamazon.com) The customer list is part of Amazon’s argument that its chips are moving beyond internal use. Amazon has said Anthropic uses Trainium, and The Next Web reported this week that Uber is expanding its Amazon Web Services deal to pilot artificial intelligence training on Trainium3 while also using Amazon’s Graviton4 processors for ride-matching workloads. (aboutamazon.com; thenextweb.com) Nvidia still sets the pace in artificial intelligence chips, and some customers continue to prefer its software tools and broader ecosystem. Network World reported that companies including Cohere and Stability AI have favored Nvidia systems in part because of more mature tooling and complaints about Amazon Web Services availability. (networkworld.com) Amazon is spending accordingly. CNBC reported that Amazon has earmarked up to $100 billion in capital expenditures this year, with most of that tied to artificial intelligence infrastructure, while Jassy argues the spending follows clear customer demand rather than speculative bets. (cnbc.com; aboutamazon.com) For now, the shortage is the story: cloud customers want more artificial intelligence capacity than Amazon can currently deliver, and Amazon wants its own chips to be the way it catches up. (networkworld.com; aboutamazon.com)
