Amazon Web Services is in early talks to sell its Trainium AI chips to outside data center customers, AWS AI chief Peter DeSantis told Bloomberg, opening a new front against Nvidia's dominance of AI silicon. Amazon CEO Andy Jassy said in his early April shareholder letter that if the chips operation were spun out as a merchant business, it would run at roughly $50 billion annually. Nvidia, for context, is currently on a $326 billion revenue run rate. The talks are preliminary, and AWS has not named the prospective buyers.
The pitch is a direct response to demand AWS cannot internally absorb. Jassy said the current generation of Trainium sold out almost instantly inside AWS, and that capacity for Trainium4 — which is more than a year from availability — is already gone. The OpenAI workloads AWS recently took on were booked after that sellout.
Until now, Amazon has refused outside requests to buy the chips directly, and the economics of that refusal are not subtle.
“If our chips business was a standalone business, and sold chips produced this year to AWS and other third parties (as other leading chips companies do), our annual run rate would be ~$50 billion. There's so much demand for our chips that it's quite possible we'll sell racks of them to third parties in the future.”— Andy Jassy, Amazon CEO
Key facts
- 01AWS AI chip chief Peter DeSantis confirmed early-stage talks to sell Trainium chips to outside data center customers.
- 02Amazon CEO Andy Jassy pegged the chips business at a ~$50B annual run rate if sold standalone.
- 03Nvidia is currently on a $326B revenue run rate, more than 6x the size of Amazon's prospective merchant chip business.
- 04Trainium4 capacity is already sold out and won't be available for more than a year.
- 05Jensen Huang has separately staked out a new $200B CPU-for-AI market, moving Nvidia into Intel and AMD territory.
AWS captures far more than the silicon price when a customer runs AI on its cloud. It bills for the tokens those chips process, then layers on storage, networking, security, and monitoring revenue on top. Selling raw chips or racks to a third party collapses that bundle into a one-time hardware sale. Spokesperson Doron Aronson, who recently hosted a tour of the AWS chip design facility, framed the shift as a change in posture rather than a formal product launch.
The supply problem is the harder constraint. To serve outside buyers without stranding existing AWS customers on waitlists, Amazon would need a surplus from TSMC, which has just supplanted Apple as the foundry's largest customer — largely on the back of Nvidia orders. Elbowing into that allocation is a separate fight from designing the chips themselves.
Scale matters here. A $50 billion merchant chip business would land roughly at Intel's annual revenue and would be the most credible alternative to Nvidia that any hyperscaler has put forward. It would not dent Nvidia's quarter, but it would change the conversation for any customer who has so far had a single realistic vendor for frontier training silicon.
The Trainium pivot also lands in a market where Nvidia itself is widening its addressable surface. Jensen Huang recently declared a new $200 billion opportunity for Nvidia in selling CPUs for AI workloads, pushing into Intel and AMD territory. Jassy is moving in the opposite direction — out of the captive cloud and into Nvidia's core GPU market.
There are reasons to be cautious about the framing. The talks are early. No buyer has been named. The $50 billion figure is a hypothetical run rate Jassy used to argue the chips would be a credible standalone business, not a booked revenue line. And AWS would still rather sell tokens than chips when it can — every rack shipped to a third party is a rack not generating bundled cloud revenue for the next five years.
The interesting wrinkle is what this signals about Amazon's read of the AI compute market over the next 24 months. Selling Trainium externally only makes sense if AWS believes it can manufacture enough silicon to serve both its own cloud and outside buyers — which means either a much larger TSMC allocation or a bet that internal demand growth slows. Neither implication is small. If Amazon is willing to risk seating customers on a waitlist to put Trainium racks in someone else's data center, it is also willing to bet that being a second source to Nvidia is worth more long-term than protecting the AWS bundle. That is the first time a hyperscaler has made that trade out loud.
Working on something we should cover, or seeing a story we missed? Send leads, documents, or feedback to hello@aichatdaily.com. For sensitive tips, see our secure tips page for Signal and PGP options.
Spotted an error? Email hello@aichatdaily.com with the URL and the issue, or read our full corrections policy.
