Amazon's chip unit locks in $225B in committed demand as Trainium sells out
What's the deal? Amazon's custom AI chip business has secured over $225 billion in multi-year revenue commitments, chief executive officer Andy Jassy disclosed on the company's Q1 2026 earnings call on April 29. The figure, tied almost entirely to Trainium chips, is contracted demand rather than a forecast.
The numbers behind the backlog: The unit — spanning Trainium AI accelerators, Graviton CPUs, and Nitro networking chips — is running at an annualised internal rate of roughly $20 billion. Jassy said that figure could reach about $50 billion if the business operated as a standalone merchant chip company.
Who's buying? AnthropicDealroom has a profile for this one. Try Dealroom →, an anchor customer, has a multi-year deal tied to over $100 billion in AWSDealroom has a profile for this one. Try Dealroom → spend. OpenAI has committed to roughly two gigawatts of Trainium capacity, Uber is adopting Trainium and Graviton, and Meta has committed to tens of millions of Graviton cores.
Why now? Trainium2 is largely sold out, and Trainium3, which began shipping in early 2026, is nearly fully subscribed. Pre-reservations for Trainium4 are already substantial, with broad availability expected in about 18 months.
The pricing case is central. Trainium2 offers around 30% better price-performance than competing GPUs; Trainium3 pushes that another 30% to 40%.
What's the endgame? Amazon is reportedly exploring direct external sales of Trainium chips and full server racks to outside data centres. Until now, Trainium was internal infrastructure powering AWS services; selling the chips themselves, as Nvidia does, is a different business entirely.
What could go wrong? Nvidia dominates AI compute with its H100 and B200 GPUs, and its pricing power rests on the absence of a credible alternative at scale. Amazon, with $225 billion in committed demand and three chip generations in rapid succession, is building exactly that.
The signal: Hyperscalers are increasingly unwilling to stay pure customers of third-party chip vendors. Google has Tensor Processing Units, Microsoft is developing its Maia accelerator, and Meta is building its own training silicon — but Amazon is furthest along in commercialising that bet, with a $225 billion backlog to prove it.
Read more: Crypto Briefing
Image credit: jurvetson (via Openverse)
