Overview
- AWS said in mid‑June that it is in early talks to sell its Trainium AI chips and possibly full racks to outside companies for use in their own data centers.
- Amazon’s custom‑silicon group has already passed a roughly $20 billion annualized run rate and CEO Andy Jassy said the chips business could scale to about $50 billion if sold broadly.
- Large customers have already committed heavy Trainium use, with Anthropic operating a reported 500,000‑chip Rainier cluster and public reporting of OpenAI plans to add Trainium capacity starting in 2027.
- Significant obstacles remain because Nvidia’s CUDA software creates high switching costs for model developers, TSMC foundry capacity is limited, and many workloads need engineering to run efficiently on Trainium.
- If Amazon can resolve supply and software hurdles the move could cut training and inference costs for some users and force rivals and chip makers to speed capacity and tooling investments.