AWS capacity squeeze

Some Amazon Web Services customers are asking for so much artificial intelligence computing power that Amazon says it has had to refuse requests to lock up all available 2026 capacity. (networkworld.com) Network World reported on April 10 that two large customers asked to buy all of Amazon Web Services’ 2026 instance capacity for Graviton, Amazon’s in-house central processing unit line. Amazon declined because it still has to serve other customers across its cloud business. (networkworld.com) The scramble is spilling into Amazon’s custom artificial intelligence chips. Amazon Web Services says Trainium2 offers 30% to 40% better price performance than its graphics processing unit-based P5e and P5en instances, and Trainium3 doubles compute performance over Trainium2 to 2.52 petaflops of FP8 compute per chip. (aws.amazon.com) In cloud computing, “capacity” is the inventory of servers, chips, networking gear, power and cooling that customers can rent by the hour instead of buying their own machines. When model training and inference demand rises faster than data centers can be built, procurement turns into a queue for physical hardware. (aboutamazon.com) Amazon has spent the past year turning Trainium from a side bet into a supply valve for big model builders. On October 29, 2025, Amazon said Project Rainier had gone live with nearly 500,000 Trainium2 chips, and that Anthropic was expected to scale Claude across more than 1 million Trainium2 chips by the end of 2025. (aboutamazon.com) The customer list widened again on February 27, 2026, when OpenAI and Amazon said OpenAI would consume 2 gigawatts of Trainium capacity through Amazon Web Services infrastructure. The same announcement said Amazon would invest $50 billion in OpenAI and become the exclusive third-party cloud distribution provider for OpenAI Frontier. (openai.com) That leaves Amazon balancing two tracks at once: keep buying Nvidia-based systems that many customers already use, while pushing its own chips as a cheaper way to train models and run them in production. Amazon’s Trainium page says first-generation Trn1 instances can cut training costs by up to 50%, and its December 2025 Trainium3 launch said some customers were reducing training and inference costs by up to 50% with Trainium. (aws.amazon.com, aboutamazon.com) Amazon is also spending against the bottleneck. In his shareholder letter published April 9, Chief Executive Officer Andy Jassy said Amazon Web Services’ artificial intelligence revenue run rate had climbed above $15 billion in the first quarter of 2026 and was still rising. (aboutamazon.sg) For buyers, the immediate problem is less about whether chips exist in theory than whether enough reserved capacity exists on the dates they need. Amazon’s answer so far is to add more data center capacity, keep rationing supply across customers, and use Trainium to stretch each dollar of training and inference spend further. (networkworld.com, aboutamazon.com, aws.amazon.com)