Amazon sets late 2027 target for selling its AI chips outside AWS
The Trainium plan creates a choice between upfront chip sales and years of cloud rental income. Amazon expects to revisit that balance annually.
Loading page…
The Trainium plan creates a choice between upfront chip sales and years of cloud rental income. Amazon expects to revisit that balance annually.
Listen to this story
Amazon is preparing a second route to monetize Trainium: selling physical chips to outside customers rather than only renting their capacity through AWS. CEO Andy Jassy expects the business to begin in late 2027, but Amazon has disclosed no buyers, pricing, volumes, chip generation, or revenue target. Sales would compete with AWS deployments for constrained manufacturing allocations, and the company will revisit that balance annually. Jassy said keeping chips in AWS can produce revenue and free cash flow over five to six years; Trainium and Graviton together run at over $25 billion annually—not a forecast for external sales.
Jassy raised potential third-party sales of chip racks in his April 9, 2026 shareholder letter; the earnings call added a late-2027 start window.
An outright sale books revenue when a chip leaves inventory, while AWS deployment can monetize its capacity over a five-to-six-year life.
Amazon plans to reassess the split between external sales and AWS use probably every year; it has not promised a fixed allocation.
Amazon intends to let other companies buy the AI chips it has kept inside its cloud business. On its October 29, 2026 earnings call, CEO Andy Jassy set a late 2027 starting window for external Trainium sales, according to 24/7 Wall St. The plan would add outright hardware sales alongside AWS access.
Trainium is Amazon’s AI chip, used so far within Amazon Web Services, its cloud business. Selling it externally would change how customers obtain the hardware: they could purchase chips rather than access their computing capacity through AWS. Jassy described external sales as a business Amazon expects to become significant, not an offering already available.
The two routes bring in money differently. Jassy said an outright sale records revenue when the chip leaves inventory. Keeping that chip in AWS allows Amazon to rent its capacity over a five-to-six-year life, potentially generating more revenue and free cash flow. The decision therefore involves both how much a chip earns and when Amazon receives the money.
External sales have been under consideration for months. In his April 9, 2026 shareholder letter, Jassy raised the possibility of selling racks of Amazon’s chips to third parties. The earnings-call statement supplies something that earlier possibility lacked: a planned starting window.
Amazon is not committing to a fixed split between those two revenue paths. Jassy said the company would review the balance between external Trainium sales and AWS use probably every year. That leaves the amount of hardware available to outside buyers subject to recurring allocation decisions.
Manufacturing supply helps explain that choice. Jassy said advanced chipmakers share one manufacturer and each receives an allocation. Within that constraint, a Trainium chip sold to an outside customer is one Amazon cannot use to supply its own cloud service. External sales and AWS expansion would draw on the same limited supply.
The commercial terms remain open. Amazon did not identify an outside buyer, price, sales volume, chip generation or revenue target. The late 2027 date is therefore a planned beginning, not a disclosed order book or a commitment to sell a particular quantity of chips.
Amazon also said Trainium and Graviton together were running at more than $25 billion a year. That combined annualized figure provides context for the scale of its chip operations. It is not a Trainium-only figure, nor a forecast for the external chip-sales business Jassy intends to start.
Loading discussion...
Join the conversation
Explain what would make that relationship dependable or too risky.
Be the first to share a perspective or an experience.
Reader comments
Newest comments first. Replies stay oldest first.