# Amazon Moves to Sell Trainium Chips Outside Its Cloud, Opening a Direct Front Against Nvidia

Selling custom accelerators as merchant silicon would put Amazon in Nvidia's core business for the first time, testing how durable the graphics processing unit (GPU) maker's pricing power really is.

- Published: 2026-06-22T10:48:41.833Z
- Canonical: https://polylog.news/ai/2026-06-22/amazon-moves-to-sell-trainium-chips-outside-its-cloud-openin
- Publisher: Polylog (AI desk)
- Section: markets
- Sources: [Polylog editors](https://polylog.news), [TechCrunch](https://techcrunch.com/2026/06/18/amazon-hopes-to-challenge-nvidia-more-directly-by-selling-its-ai-chips/), [MLQ News](https://mlq.ai/news/amazon-in-talks-to-sell-trainium-ai-chips-outside-aws-in-direct-challenge-to-nvidia/)

Amazon is in active negotiations to sell its Trainium accelerators directly to outside operators for use in their own data centers. The step would move the company beyond renting cloud capacity and into Nvidia's core hardware business for the first time. Peter DeSantis, the Amazon senior vice president who oversees the company's AI and semiconductor work, [confirmed the talks](https://techcrunch.com/2026/06/18/amazon-hopes-to-challenge-nvidia-more-directly-by-selling-its-ai-chips/) in an interview reported by Bloomberg.

The economics are central to the move. Amazon's chip division has reached an annual revenue rate of roughly $20 billion, and it says [Trainium3 delivers about four times the performance of Trainium2 at close to half the cost of conventional GPUs](https://mlq.ai/news/amazon-in-talks-to-sell-trainium-ai-chips-outside-aws-in-direct-challenge-to-nvidia/), with the chip selling near its production capacity since its late-2025 launch. Reported Trainium commitments from OpenAI and Anthropic alone total around $225 billion. By comparison, Nvidia's data-center segment generated about $115 billion in its most recent fiscal year.

The skeptical view matters here. Selling silicon as merchant hardware is a different business from renting capacity. Amazon would need a standalone sales organization, enterprise-grade firmware and driver support outside the Amazon Web Services (AWS) software stack, and customers willing to move their workloads off Nvidia's CUDA software platform to a system with a shorter track record. A price advantage alone does not solve any of that. The talks are also, for now, only talks, and DeSantis described intent rather than signed agreements.

What is verified is the direction. Every large buyer of AI computing power is now also a seller of it, or wants to be. That is the structural pressure on Nvidia, whether or not this specific deal closes.

## What this means

Nvidia's margins rest on the assumption that no rival offers a credible high-volume substitute. A hyperscaler willing to sell its own accelerators as merchant silicon challenges that assumption directly, even before the first unit ships. The near-term constraint is software portability, not raw computing power, so watch whether Amazon can separate Trainium from the AWS ecosystem.

## What to watch

- Whether any named external customer signs for Trainium hardware, which would convert intent into a real merchant-silicon line.
- How aggressively Amazon invests in a CUDA-alternative software stack, since portability, not peak performance, decides adoption.
- Nvidia's response on pricing or supply allocation, which would signal how seriously it treats custom silicon as a threat.