Amazon has spent years building its own AI chips quietly, keeping them inside its own cloud business. That strategy is now changing. CEO Andy Jassy has publicly floated the idea of selling those chips directly to outside companies, and Amazon's AI chief Peter DeSantis has already held early conversations with potential buyers.
The chip in question is called Trainium. It is a processor Amazon designed specifically for running and training AI models. Customers who use Amazon's cloud today can rent computing time on these chips, but they cannot buy them outright. The new proposal would change that, letting other companies purchase the hardware itself and run it in their own data centers.
Why does this matter to non-tech businesses? Because it signals a coming shift in who controls the supply of AI computing power, and at what price. Today, roughly 80 to 90% of the market for AI processors is owned by Nvidia. That near-monopoly means high prices and long waiting lists. Real-world data suggests Trainium already costs about half what comparable Nvidia chips cost to run, and major companies including Anthropic, Apple, and Uber have already moved workloads onto it.
Jassy's own math is striking. He estimates that if Amazon sold its chip business the same way Nvidia does, it would run at around $50 billion a year. That figure covers three chip lines: Trainium for AI, Graviton for general computing, and Nitro for server management. For context, Nvidia's data center business alone just posted $51 billion in a single quarter.
There is a real tension at the heart of this plan. Amazon's chips are sold out. Every generation has been fully reserved almost as soon as it became available, and the next generation after Trainium3 already has substantial pre-orders despite not shipping for another year or more. Selling to outside buyers means either building far more chips or taking supply away from existing cloud customers.
Manufacturing capacity is the wall Amazon would have to climb. The world's most advanced chip factory, TSMC in Taiwan, has already committed roughly 60% of its most critical production capacity to Nvidia through 2026. Amazon gets some of what remains, but expanding that allocation fast enough to serve a new class of external buyers is not a quick project.
Jassy has pointed to Amazon's earlier success with a different custom chip called Graviton, which handles general computing tasks. Today, 98% of Amazon's large enterprise customers use Graviton. He is betting that Trainium follows the same adoption curve, first proving itself inside Amazon, then spreading outward.
For business operators, the practical takeaway is this: the AI chip market is beginning to look less like a Nvidia monopoly and more like a competitive supply chain with multiple serious players. That means AI computing costs are likely to fall over time, the same way cloud storage costs fell once competition arrived. If you are currently locked into high-cost AI computing arrangements, it is worth watching how this develops over the next two years. Amazon is not the only one moving in this direction: Google has its own chips, Meta has its own, and OpenAI is building chips with Broadcom. Nvidia's position is strong today, but the pressure on pricing is real and building.