Amazon explores selling Trainium chips externally as third-gen sells out

Amazon's push to sell Trainium chips to third-party data centers threatens Nvidia's near-monopoly in AI accelerators, with the third-generation chip already sold out and the fourth-generation model drawing strong pre-launch demand.

"We want to be a part of" the coming explosion in AI infrastructure deployment, Peter DeSantis, Amazon's senior vice president overseeing AI, chips and quantum computing, told CNBC. He declined to name specific customers in negotiations.

The third-generation Trainium chip is fully allocated, and the fourth-generation model expected next year has already generated strong interest from potential buyers, according to people familiar with the matter. Amazon's custom silicon strategy — spanning the Trainium AI accelerator and Graviton general-purpose processor — has been more than a decade in development through its Annapurna Labs unit, which the company acquired in 2015 to optimize performance and cost for AWS workloads.

The move could reshape the AI chip market, where Nvidia commands roughly 80% of data center AI accelerator revenue. Amazon's ability to offer a lower-cost alternative — Trainium delivers nearly double the compute efficiency for certain workloads, according to startups using the chip — threatens Nvidia's pricing power and could save AWS customers billions in annual GPU procurement costs.

Amazon's chip strategy centers on co-designing hardware and software in lockstep. "If the chips are not telling the model designers what capabilities are coming and where they can optimize, then we're not doing the science necessary to take advantage of those capabilities," DeSantis said at VivaTech 2026 in Paris. AI startups building world models — systems that simulate physics rather than generate text — are choosing Trainium and achieving nearly double the industry-average compute efficiency, he said.

The company's approach mirrors Nvidia's vertically integrated model, where chip architecture, system design and software stack are developed together. But Amazon adds a cloud distribution advantage: its chips are already deployed across AWS's global infrastructure, giving it a built-in customer base of 50,000 Nova2 model users.

Nvidia's H100 and upcoming B200 GPUs remain the industry standard for large-scale AI training, with the company's data center revenue reaching $62 billion in its latest fiscal year. But Amazon's entry into external chip sales introduces a new competitive dynamic. CEO Andy Jassy said in April the company could consider selling racks of Trainium chips to third parties, though DeSantis said there is no timeline for such a move.

The broader implications extend beyond Nvidia. TSMC, which manufactures both Nvidia's and Amazon's chips, could see its wafer pricing power tested if Amazon's volume grows. And hyperscalers including Microsoft and Google, which also design custom AI chips, face pressure to accelerate their own silicon roadmaps.

Amazon shares trade at roughly 22 times forward earnings, with the AI chip opportunity largely unpriced by the market. If Amazon captures even 10% of the data center AI accelerator market, it could add $6 billion in annual revenue — a high-margin stream that would flow directly to the bottom line. Nvidia, trading at 35 times forward earnings, faces the more immediate risk as its dominant market share creates a larger target for competitors.

This article is for informational purposes only and does not constitute investment advice.