Summary: A June 18, 2026 Bloomberg interview reported that AWS was negotiating with potential customers to sell Trainium chips into other companies’ data centers. Amazon said talks were early and declined to identify buyers. The article places the move against Jassy’s earlier $50 billion chip-business estimate, supply constraints, and competition with Nvidia.
An interview with Amazon’s AI chief Peter DeSantis, published by Bloomberg on June 18, 2026, confirms that AWS is negotiating with prospective customers over selling its homegrown Trainium AI chips to data centers run by other companies — a step beyond keeping the silicon exclusive to its own cloud. He would not identify any of the buyers, offering only: “We view AI infrastructure as rapidly evolving. And we’re constantly looking at ways to get to more customers.” Amazon’s shares rose more than 2 percent on the news, and a company spokesperson afterwards told TechCrunch that the talks are real but still at an early stage.
Where the $50B Figure Began: Jassy’s Letter to Shareholders
The idea traces back to the shareholder letter CEO Andy Jassy released in April. Picture the chips operation as an independent company selling this year’s chips to both AWS and outside buyers, he wrote, and “our annual run rate would be ~$50 billion.” On demand, he went further: “it’s quite possible we’ll sell racks of them to third parties in the future.” Putting those numbers in writing turned an internal debate into a public one — this year’s chip output, sold to both AWS and external customers, would by itself rival the yearly revenue of a leading semiconductor maker. AWS spokesperson Doron Aronson later reinforced that 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 as a whole, the signal is that the long-standing refusal to sell chips directly is weakening right at the top of the company.
What Has Kept AWS From Selling Chips
The reluctance is rooted in how the business makes money. AWS earns from its chips in a cascading, waterfall fashion: what customers directly pay for is the AI tokens the silicon processes, yet that payment drags behind it a whole series of follow-on charges for storage, networking, security, and monitoring. The moment a chip leaves an AWS facility, that attached revenue stream is cut off. This is why, Trainium3’s cost appeal notwithstanding, AWS has for years positioned its custom silicon as bait that draws customers onto the cloud rather than as merchandise for sale. Even so, an external Trainium business would not sever Amazon’s ties with Nvidia — in March, Amazon said it would spread one million Nvidia chips through AWS data centers over the next twelve months. Amazon is, at one and the same time, Nvidia’s customer and its would-be rival.
Supply Limits, TSMC, and the Queue of Buyers
The second constraint is supply. In that same letter, Jassy conceded that today’s Trainium capacity “sold out almost instantly,” and Trainium4 — more than a year away from launch — is already fully reserved. Uber is already running AI workloads on Trainium3, and OpenAI pledged in February to take 2GW of Trainium capacity on AWS, easily the largest outside endorsement the chip has received so far. To ship externally without squeezing those existing commitments, Amazon would have to pull extra capacity through TSMC — precisely as Nvidia has replaced Apple as the foundry’s biggest customer and the line for advanced nodes grows longer still. Anyone who does buy racks of Trainium would also need to piece together the surrounding software stack alone, a barrier that has long kept most operators from leaving Nvidia’s CUDA ecosystem.
A New Order Taking Shape in the Chip Market
Scale puts things in perspective: Nvidia is running at a revenue pace of roughly $326 billion a year, which means a $50 billion challenger would not dislodge it — yet that challenger amounts to an Intel-scale chipmaker materializing out of nowhere. And this is no isolated move: Google has started offering TPUs to a “select group of customers” and has teamed with Blackstone on a TPU cloud service, while Nvidia pushes back the other way, with CEO Jensen Huang touting a $200 billion opportunity in AI-purpose CPUs. Custom silicon is turning from a cloud cost tool into a commodity that can be sold standalone — and companies buying AI compute now face a longer list of vendors, and with it more leverage. Pricing is what to watch next: should AWS put its racks on the market anywhere near the economics of its own cloud, margin structures across the entire AI accelerator market would come into question.