Amazon to Sell Trainium Chips Outside AWS, Ending AWS-Only Model

Amazon announced on its October 29 earnings call that it will begin selling its Trainium AI chips outside AWS starting in late 2027, abandoning a strategy that has kept its custom silicon exclusive to Amazon Web Services. CEO Andy Jassy framed the move as a lever the company can pull each year, but named no outside buyers, prices, volumes or chip generation. The announcement came alongside a quarterly report that missed revenue and earnings estimates and showed trailing free cash flow turning negative $42.4 billion.

A Stated Plan With a Start Window, Not a Deal

Jassy said “the number of companies who have a significant cloud business and a chip strong enough to sell externally is effectively 2,” not naming the other; CNBC identified Google, which already offers its TPU chips outside its own facilities. He added that advanced chip designers share the same manufacturer and receive allocations, so every Trainium chip is either sold or deployed in AWS. Selling a chip books revenue when it leaves inventory; keeping it inside AWS “you actually will end up driving more absolute revenue and free cash flow from those chips because you get to use them over 5 to 6 years.” Jassy said the split would be revisited “probably every year.” [1][3]

No outside customer was named on the call, and no terms, volume targets or chip generation were disclosed. The commitment remains a stated plan with a late-2027 start window. [3]

Trainium Already Powers Most of Bedrock Inside AWS

Inside AWS, the chip is already doing the work. AWS revenue grew 39.3% to $46.0 billion in the third quarter at a 38.7% operating margin, with a $578 billion backlog. Jassy said “the majority and most of Bedrock,” Amazon’s AI inference service, runs on Trainium. Inference already accounts for well over half of AWS’s AI revenue. Revenue commitments for Trainium and Graviton each rose more than 20x year over year. Short-term capacity on prior-generation Trainium now sells at a premium to the first half of the year. [1][2]

CFO Brian Olsavsky flagged the cost structure: “the outsized growth of AI puts margin pressure on the business as a whole because the margins are lower there.” [1][2]

Why the Timing Matters: Cash Flow and Capex

The announcement arrives as Amazon’s capital intensity accelerates. Trailing 12-month free cash flow was a $42.4 billion outflow. The company reiterated about $220 billion of 2026 capital spending and said it will “spend meaningfully more in 2027.” Trainium and Graviton run at more than $25 billion a year at the latest quarterly pace. Selling chips externally pulls revenue forward; keeping them in AWS spreads higher-margin rental income over years. With capex outrunning operating cash flow, the option to monetize some allocation sooner is a financial lever, not just a product one. [3]

The Competitive Picture: Nvidia’s Scale, Google’s Precedent

The merchant-silicon market is dominated by Nvidia, whose data center revenue reached $89.02 billion in one quarter, up 117%. AWS itself is deploying an additional 2 million Nvidia GPUs. Jassy’s “effectively 2” companies comment underscores that Google’s TPU program is the only direct precedent for a hyperscaler selling its own AI chips externally. [3]

Manufacturing supply is the binding constraint. Jassy said every Trainium sold outside is one AWS cannot rent against its $578 billion backlog. Demand for Trainium is also concentrated: Anthropic reportedly pledged more than $100 billion of AWS spending over 10 years for up to 5 gigawatts of Trainium capacity, yet it also buys Nvidia systems. [3]

Software as the Real Lever

Amazon open-sourced TorchNeuron so developers can run “existing PyTorch code” on Trainium without rewriting it, targeting the CUDA lock-in that keeps buyers on Nvidia. [3] A separate announcement detailed a new software layer letting Trainium chips run native AI code without modification, timed with AWS capacity sold out through 2027 and a $25 billion AI run-rate. [4]

The software work is real and shipped. The external chip business is not: it is a plan for late 2027 with no named customers.

What to Watch

The first proof point is a named outside Trainium buyer with real terms before the late-2027 launch. Until then, AWS growth is the scoreboard: Jassy projects above 40% AWS growth in both 2027 and 2028. Growth slipping back below the third quarter’s 39.3% would weaken both halves of the plan—the cloud rental model and the prospective chip sales. [1][3]