Amazon will start selling its Trainium artificial intelligence chips to customers outside its own cloud in late 2027, chief executive Andy Jassy said on the company’s third-quarter earnings call on October 29. Until now the custom silicon has been available only as rented computing power inside Amazon Web Services, so the move turns an internal cost advantage into a product sold head-to-head with Nvidia’s accelerators. Amazon has named no outside buyers, prices, or sales terms yet.
What Amazon actually said
The announcement came in answer to a question about the balance between selling chips and using them inside AWS. Jassy said customers keep urging Amazon to make Trainium available outside its cloud because demand is so strong, and he confirmed the plan directly: Amazon intends to sell the chips externally from late 2027.
He called the external business very substantial, and said Amazon will revisit the balance between outside sales and internal use every year. That annual review matters, because every chip sold outside is a chip Amazon cannot rent through its own cloud, and the company has to choose where each unit earns more.
Why Trainium is ready to leave AWS
Trainium is no longer a side project inside Amazon. Jassy said revenue commitments for Trainium and its sibling processor Graviton each grew more than twentyfold over the past year, and customers using Amazon’s silicon saved more than $2 billion in the third quarter alone. Most of the work on Bedrock, Amazon’s service for running AI models, runs on Trainium hardware.
The trade-off Jassy spelled out
Jassy spelled out the cost of the decision. Every chipmaker uses the same foundry, he said, and each gets a yearly quota, so Amazon must divide its allocation between chips it sells and chips it keeps. A sale books revenue when hardware leaves the warehouse, while a chip kept in AWS earns for five or six years as computing power.
That arithmetic explains why Amazon waited. Jassy noted that only two companies combine a large cloud business with chips strong enough to sell externally, without naming the other one. Google, which has begun selling its own tensor processors outside its cloud, is the obvious comparison, and Amazon now joins a very small club of sellers.
Our read
The timing is the tell. Amazon is installing an AI accelerator roughly every fourteen seconds, Jassy said, and customers absorb new capacity as soon as it arrives. Yet the company still chose to promise chips to outsiders in late 2027.
That only makes sense if Amazon believes its foundry quota can grow, or that outside sales will not starve a cloud business it expects to grow more than forty percent a year. The same quarter that produced this announcement also produced a $220 billion capital spending plan for 2026, with more to come in 2027.
Buyers, not announcements, will decide whether this works. Amazon’s hardware ambitions stretch its balance sheet: it arranged an $80 billion loan facility against hardware purchases in its filing. Rival chipmakers are raising hundreds of millions to sell inference hardware of their own, as FuriosaAI’s round shows. The first named outside Trainium buyer, with terms, is the proof point to watch.
Why It Matters
For buyers of AI hardware, a second merchant supplier at Amazon’s scale would loosen Nvidia’s grip on pricing and supply. For Nvidia, the threat is a customer turned competitor: AWS already buys Nvidia chips in bulk while building the alternative. For Amazon, the risk runs the other way, since every wafer sold outside is capacity its own cloud cannot rent.
Background
Amazon’s cloud division reported $46 billion in third-quarter revenue, up 39.3 percent from a year earlier, with a $578 billion backlog of future commitments. The Trainium announcement came on the same October 29 call, in a quarter when Amazon also disclosed the hardware loan facility and reaffirmed its huge capital spending plans.
Source: Amazon’s Q3 earnings call transcript and The Stack’s Q3 takeaways
0 Comments