Amazon's Trainium AI accelerators are beginning to win over developers who have built their careers on Nvidia hardware, according to The Information. The report lands on the same day Nvidia starts hand-delivering its new Vera CPUs to AI labs and two days before Jensen Huang reports earnings. The timing is pointed.
The shift has been building for a while. AWS CEO Andy Jassy told shareholders in April that Amazon's custom chip business now generates more than $20 billion in annualised revenue and is growing at triple-digit rates. He said the business would be worth $50 billion a year if sold on the open market. Trainium2 is largely sold out. Trainium3, which shipped earlier this year, is nearly fully subscribed. Reservations for Trainium4 are already being taken.
The customer list tells the story
Anthropic runs Claude on more than one million Trainium2 chips. OpenAI committed over $100 billion to AWS infrastructure, including roughly two gigawatts of Trainium capacity. Meta signed a deal for millions of Amazon's custom CPUs. These are not marginal workloads from marginal customers.
AWS claims Trainium offers 30% to 50% better price-performance than comparable GPU instances. For companies burning through billions on inference, that gap compounds fast.
The CUDA problem is fading
Related reading
- Amazon's AI coding agent deleted its own production environment. Then things got considerably worse
- Amazon's AI agent sends software stocks to their worst quarter since 2008
- Nvidia earnings to test Wall Street's AI spending thesis
The historic barrier to Nvidia alternatives was always software. Developers knew CUDA. Everything else required rewriting code. Amazon has been chipping away at that lock-in through its Neuron SDK and by leaning on Anthropic's engineers to build out the software library. TorchNeuron, released last year, lets developers run PyTorch natively on Trainium with minimal code changes.
Nvidia remains dominant and will likely report another blowout quarter on Wednesday. But the direction of travel is clear. The hyperscalers are building their own silicon, their biggest AI customers are adopting it, and the price-performance argument is getting harder to ignore.