Amazon is in talks to sell its custom Trainium AI chips directly to companies for use in their own data centers, a major strategic shift from years of selling only computing capacity through AWS. Amazon AI Chief Peter DeSantis and CEO Andy Jassy confirmed the move in recent interviews. Jassy stated the company could eventually target a $50 billion market for standalone chip sales—a figure comparable to AMD's annual revenue.
Trainium2 chips have sold out, and Trainium3 capacity (shipping since early 2026) is 'nearly fully subscribed' despite launching only months ago. Jassy says Amazon's chip business has a current $20 billion annual revenue run rate with triple-digit percentage growth year-over-year, but that figure understates the true market: if chips were sold as a standalone product rather than bundled into AWS compute, the run rate would be roughly $50 billion. OpenAI has committed to consume approximately 2 GW of future Trainium capacity, and Anthropic is already training on these chips.
For customers, Trainium offers superior cost economics. World-model startup Odyssey achieved 80% model flop utilization on Trainium—roughly double the industry average of 40–50%. Jassy also cited 'price-performance' advantages over Nvidia and signaled AWS is not concerned that external sales will cannibalize cloud revenue given the underconsumption of AI compute globally.
Architects evaluating AI infrastructure should note: Amazon's move signals confidence in Trainium's competitive position and validates a shift toward custom silicon across cloud providers (Google has TPUs, Microsoft backs Maia). The $50B opportunity sizing suggests Amazon believes it can capture meaningful Nvidia market share. However, Trainium adoption still trails Nvidia GPUs significantly, and the chips' real-world adoption rate outside specialized workloads (world models, video generation) remains unproven at scale.