Amazon's yearslong effort to build a serious alternative to Nvidia's dominant AI chips is gaining traction, with Anthropic and OpenAI committing to rent large amounts of current and future Trainium capacity, and smaller developers increasingly considering moving workloads to Trainium as software support has improved over the past few months. The custom silicon business including Trainium and Graviton has reached a more than $20B annualized run rate as of last month (or roughly $50B if measured as a standalone chip seller), with one Loka client switching its inference workloads to second-generation Trainium chips earlier this year after tests showed up to 35% cost savings compared to Nvidia's H100. Amazon recently announced that Trainium2 is largely sold out, Trainium3 is nearly fully subscribed, and much of Trainium4 (about 18 months from broad availability) is already reserved, with OpenAI committing to 2 gigawatts of Trainium capacity in February and Anthropic recently committing to spend more than $100B on AWS over the next decade including on Trainium.
