Amazon Considers Selling Proprietary AI Chips to Third Parties

In an annual shareholder letter released in April 2026, Andy Jassy announced that Amazon is considering selling its proprietary AI processors, specifically the Trainium line, to third-party customers. Currently available only via Amazon Web Services (AWS), the company's chip operation—which includes Trainium, Graviton, and Nitro—already generates a $20 billion annual revenue run-rate. Jassy stated that if the business sold to both AWS and external buyers, the annual run rate would approach $50 billion.

The move comes as AWS faces a capacity crunch, with demand for AI accelerated computing outpacing delivery. Trainium2 is sold out and Trainium3 is nearly fully subscribed, while Trainium4 has already attracted pre-orders. By selling chips "by the rack-load," Amazon aims to provide customers with standardized access to accelerators and reduce its dependence on external GPU suppliers like Nvidia and AMD.

To support this infrastructure expansion, Amazon plans to spend approximately $200 billion on capital expenditures in 2026, primarily for AWS data centers. While the company maintains a partnership with Nvidia, the shift toward in-house silicon is intended to lower capital costs and improve operating margins. In response to this emerging competition, Nvidia continues to leverage its industry-standard CUDA platform and plans to release the Vera Rubin platform in late 2026 to support agentic AI.