AWS compute scarcity shows
Some AWS customers are attempting to buy out the cloud provider’s available AI capacity as demand for accelerators spikes, according to reporting of customer behaviour and executive comments. The pressure is driving interest in custom silicon like Trainium2 and is reshaping bargaining dynamics between model providers and hardware vendors. (networkworld.com) (tradingkey.com)
Amazon Web Services says some customers want to lock up so much artificial-intelligence computing that the company is refusing requests for all of its available 2026 capacity. (aboutamazon.com) Chief executive Andy Jassy wrote on April 9 that Amazon Web Services added 3.9 gigawatts of new power capacity in 2025 and still has “capacity constraints that yield unserved demand.” He said two large customers asked to buy all available 2026 instance capacity for Graviton, Amazon’s in-house central processor, and Amazon declined because other customers also need access. (aboutamazon.com) Cloud computing is rented computing: instead of buying servers, customers lease processing power from Amazon data centers by the hour or by long-term contract. For artificial intelligence, that rented capacity increasingly depends on accelerators, the specialized chips that train models and generate answers, and on the electricity and networking needed to keep those chips busy. (aws.amazon.com) Amazon is pushing its own accelerator line into that squeeze. Amazon Web Services says Trainium2 offers 30% to 40% better price performance than its graphics-processor-based Elastic Compute Cloud P5e and P5en instances, and a Trn2 UltraServer links up to 64 Trainium2 chips together. (aws.amazon.com) That sales pitch now sits inside a broader buildout. In October 2025, Amazon said Project Rainier came online with nearly 500,000 Trainium2 chips for Anthropic, and that Anthropic was expected to scale Claude workloads to more than 1 million Trainium2 chips by the end of 2025. (aboutamazon.com) Anthropic tied itself more tightly to Amazon before that cluster went live. On November 22, 2024, Anthropic said Amazon would invest another $4 billion, bringing Amazon’s total investment to $8 billion, and named Amazon Web Services its primary cloud and training partner. (anthropic.com) Anthropic also said its engineers were working with Amazon’s Annapurna Labs on future Trainium hardware, including low-level kernels and changes to the Amazon Web Services Neuron software stack. That means a model company is no longer just renting chips; it is helping shape the chips and software it plans to run on. (anthropic.com) Amazon is not walking away from Nvidia while it builds that alternative. Network World, citing Jassy’s shareholder letter and analyst interviews published April 10, reported that Amazon is betting Trainium can change training and inference costs even as Amazon Web Services maintains a major partnership with Nvidia. (networkworld.com) The immediate constraint is not only chips. Jassy wrote that Amazon Web Services expects to double its total power capacity by the end of 2027, which puts electricity, data-center construction, and networking in the same bargaining chain as semiconductors. (networkworld.com) So the shortage showing up at Amazon Web Services is also a map of where the artificial-intelligence market is headed: fewer off-the-shelf purchases, more reserved capacity, and closer ties between cloud providers, chip teams, and model makers. (networkworld.com)
