AWS customers trying to lock capacity
AWS customers are reportedly attempting to buy out large blocks of AWS AI capacity as demand for training and inference outpaces available cloud resources. Amazon is promoting its Trainium chips to improve economics, while outside coverage flags power and data‑centre limits as part of the capacity squeeze. (cio.com/article/4157494/ai-demand-is-so-high-aws-customers-are-trying-to-buy-out-its-entire-capacity-2.html, youtube.com/watch?v=Vh-ZNVLQlwU)
Amazon Web Services customers are asking to lock up huge blocks of artificial intelligence capacity before it is built. (aboutamazon.com) Amazon chief executive Andy Jassy said in his April 9, 2026 shareholder letter that some customers want “significant portions” of future capacity, and that Amazon still has “unserved demand” even after adding 3.9 gigawatts of power capacity in 2025. He said Amazon expects to double total power capacity by the end of 2027. (aboutamazon.com) The immediate bottleneck is not just chips. Jassy said the constraints run through power, servers, networking gear, and data centers, the physical pieces needed to train models and serve responses at scale. (aboutamazon.com) That scramble reflects how artificial intelligence is bought today. Training builds a model by running it across giant clusters for weeks, and inference is the day-to-day work of generating answers, images, or code for users, which can consume vast computing capacity once a model is in production. (cnbc.com) Amazon is using the shortage to push its own chips. Jassy said the AWS custom silicon business, including Graviton central processors, Trainium training chips, and Nitro networking cards, has reached a more than $20 billion annual revenue run rate, up from $10 billion disclosed with fourth-quarter results. (aboutamazon.com) On price, Amazon is arguing that in-house chips can ease the pressure. Jassy said Trainium offers better price-performance than other graphics processing unit options for some workloads, and Amazon has framed lower chip costs as one way artificial intelligence costs can fall over time. (aboutamazon.com; cnbc.com) Amazon has already carved out dedicated capacity for one customer. In October 2025, AWS said Project Rainier, built for Anthropic, went live with nearly half a million Trainium2 chips, one of the largest artificial intelligence compute clusters Amazon has disclosed. (aboutamazon.com) That deal shows what “locking capacity” can look like in practice. AWS said Anthropic was already running workloads on Project Rainier, and outside data-center coverage said Amazon planned to expand Anthropic’s access to more than one million Trainium2 chips by the end of 2025 for training and inference. (aboutamazon.com; datacenterdynamics.com) The spending behind the build-out is also getting larger. Jassy said Amazon plans to invest about $200 billion in capital expenditures in 2026, tying that outlay to customer commitments already signed, deals still in process, and demand that is arriving faster than AWS can install capacity. (aboutamazon.com; ciodive.com) The closing point from Amazon is simple: customers are no longer shopping for cloud servers one rack at a time. They are trying to reserve power, chips, and data-center space years ahead, and AWS is telling investors the line for that capacity is already longer than the supply. (aboutamazon.com; networkworld.com)
