On June 18, 2026, Bloomberg published an interview with Peter DeSantis, Amazon’s AI chief: AWS is in discussions with potential customers about selling its in-house Trainium AI chips to other companies’ data centers, not just its own cloud. DeSantis declined to name buyers, saying only: “We view AI infrastructure as rapidly evolving. And we’re constantly looking at ways to get to more customers.” The report lifted Amazon’s stock more than 2 percent, and a company spokesperson later confirmed to TechCrunch that the talks, while real, are at an early stage.
The $50B Origin: Jassy’s Shareholder Letter
The starting point is CEO Andy Jassy’s April shareholder letter. If the chips business were a standalone company selling this year’s output to AWS and third parties, he wrote, “our annual run rate would be ~$50 billion.” Demand is strong enough, he added, that “it’s quite possible we’ll sell racks of them to third parties in the future.” That letter made an internal debate public: chips produced this year, priced for both AWS and outside buyers, would already match the annual revenue of a top-tier semiconductor company. AWS spokesperson Doron Aronson later confirmed the direction to TechCrunch: “While we’ve historically declined requests to sell chips directly, Andy noted it’s quite possible we’ll sell racks of them to third parties in the future.” Read together, the message is that the policy of refusing direct chip sales is softening at the very top of the company. Read together, the message is that the policy of refusing direct chip sales is softening at the very top of the company.
Why AWS Has Resisted Selling Chips
The answer is baked into the business model. AWS monetizes its chips through a waterfall: customers directly pay for the AI tokens the silicon processes, but behind that sits a string of bills for storage, security, networking, and monitoring. Once a chip leaves an AWS building, that attached revenue dries up. That is why, even with Trainium3’s cost appeal, AWS has long treated its custom silicon as a hook to pull customers into the cloud rather than a product line to sell. Notably, selling Trainium externally would not end Amazon’s relationship with Nvidia — in March the company said it would deploy one million Nvidia chips across AWS data centers over the following twelve months. Amazon is simultaneously Nvidia’s customer and its would-be competitor.
Capacity, TSMC, and the Waiting List
The other constraint is supply. In the same letter, Jassy admitted that current Trainium capacity “sold out almost instantly,” and that Trainium4 — which won’t be available for more than a year — is already spoken for. Uber already runs AI workloads on Trainium3, and in February OpenAI committed to 2GW of Trainium capacity on AWS, by far the largest external endorsement of the silicon so far. To ship chips externally without squeezing existing commitments, Amazon would need extra capacity through TSMC — just as Nvidia has supplanted Apple as the foundry’s largest customer, making the queue for advanced nodes even longer. Whoever ends up buying racks of Trainium would also have to assemble the surrounding software stack themselves, a hurdle that has historically kept most operators inside Nvidia’s CUDA ecosystem.
Reshuffling the Chip Market
Scale matters here: Nvidia is on a roughly $326 billion revenue run rate, so a $50 billion challenger would not topple it — but it is akin to an Intel-sized chip company appearing out of thin air. Nor is this an isolated move. Google has begun selling TPUs to a “select group of customers” and launched a TPU cloud service with Blackstone, while Nvidia attacks in the other direction, with CEO Jensen Huang claiming a $200 billion opportunity in CPUs for AI. Custom silicon is shifting from a cloud cost tool to a sellable commodity — and for companies buying AI compute, the vendor list just got longer, and so did their leverage. What to watch next is pricing: if AWS sells racks at anything near its internal cloud economics, the margin structure of the whole AI accelerator market comes under question.
Sources
- Amazon could sell Trainium AI chips to data centers — DCD
- Amazon hopes to challenge Nvidia more directly by selling its AI chips — TechCrunch
AI-assisted summary compiled from the sources above, reviewed by a human before publishing.