Amazon Weighs Selling Its AI Chips to Take On Nvidia

Amazon Web Services is in early-stage talks to sell its Trainium AI chips to other companies and data centers, a sharp break from its long-standing practice of keeping the silicon for its own cloud. The move would put Amazon into more direct competition with Nvidia, whose chips power the vast majority of AI training and which Amazon has, until now, bought rather than challenged head-on. AWS has declined to say which organizations might buy the chips, so the talks remain confidential.

The numbers give a sense of the ambition. In his April shareholder letter, CEO Andy Jassy said that if the chip business operated as a standalone company, it would generate roughly $50 billion in annual run-rate revenue — comparable to Intel today, though still far below Nvidia's current $326 billion run rate. An AWS spokesperson, Doron Aronson, framed the shift plainly: "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." Amazon's AI chief, Peter DeSantis, has separately discussed selling chips externally.

The bottleneck is supply, not demand

The harder problem is making enough of them. Amazon's current Trainium capacity sold out almost instantly, and capacity for the next-generation Trainium4 — which won't be available for over a year — is already fully allocated. The chips are manufactured by TSMC, the same Taiwanese foundry everyone else relies on, and Nvidia is now TSMC's largest customer. So even if Amazon wants to flood the market with an Nvidia alternative, it is competing for the same finite manufacturing slots as the company it is trying to undercut.

Why it matters

Most people never touch an AI chip, but the price of nearly every AI product downstream is shaped by this one market. Nvidia's dominance is a big reason model APIs and AI subscriptions cost what they do — when a single supplier sets the price of the scarcest input, that cost flows through to you. A credible second source of high-end accelerators is exactly what loosens that grip. It is the same dynamic as any market with one dominant vendor: a real competitor pressures prices and eases shortages, and the savings eventually reach whoever is paying for AI at the end of the chain.

I'd hold the skepticism, though. "In early talks," "quite possible," and "in the future" are not shipping products, and a chip business throttled by sold-out capacity can't discipline Nvidia's pricing no matter how good the silicon is. The interesting signal here isn't that Amazon will dethrone Nvidia soon — it's that the most valuable cloud company has decided its own chips are good enough to sell, and that the scarcest thing in AI right now is not ideas but fab capacity.