AWS in Talks to Sell Trainium AI Chips to External Data Centers

Amazon Web Services is in early-stage talks to sell its custom Trainium AI chips directly to third-party companies — a significant strategic shift for a company that has historically kept its silicon in-house. AWS chief executive Andy Jassy first floated the idea in his April shareholder letter, and AWS AI chief Peter DeSantis confirmed to Bloomberg on June 18 that active discussions are underway. Trainium4, the next-generation version of the chip, is expected within the next year.

The move would represent Amazon's most direct challenge yet to Nvidia's dominance in the AI accelerator market. Nvidia currently generates around $326 billion in annual revenue, with the bulk driven by data center GPU sales. Amazon's internal estimate puts the potential of an independent chips business at a "$50 billion annual run rate" — substantial, though far from displacing Nvidia's current position.

Why Amazon Is Making This Move Now

Trainium chips were originally developed to reduce AWS's own dependence on Nvidia GPUs for training large AI models. Amazon first deployed Trainium in its own data centers to train models for Alexa and internal ML workloads, and made them available to AWS customers via its EC2 Trn1 and Trn2 instances. The logic was straightforward: custom silicon costs less per compute-hour than buying Nvidia H100s and H200s at market prices.

Selling Trainium externally changes that calculus in two ways. First, it creates a new revenue stream — chip sales rather than just cloud compute fees. Second, it positions AWS as a chip supplier rather than only a cloud provider, a role that could become increasingly important as large enterprises and sovereign AI initiatives look to build their own infrastructure without full dependence on Nvidia's supply chain.

The timing also reflects the broader competitive environment. AMD has made significant inroads in the AI chip market with its MI300X GPUs. Intel is repositioning its foundry around AI workloads. Google has been selling its TPU chips externally since 2023. Amazon selling Trainium externally would align with an industry-wide shift toward more open silicon markets.

What Trainium4 Brings

Amazon has not disclosed detailed specifications for Trainium4, but the company has described it as designed specifically for large-scale AI training and inference at hyperscale. Previous Trainium generations delivered competitive performance on transformer model training benchmarks, though Nvidia's H200 and the upcoming Blackwell architecture have maintained performance leads on most standard tasks. Trainium4's external availability would give buyers an alternative for large training runs where price-performance matters more than peak throughput.

The Competitive Implications

For Nvidia, Amazon entering the external chip market is a signal worth taking seriously even if the near-term financial impact is limited. Nvidia's moat has always rested on three things: hardware performance, the CUDA software ecosystem, and supply availability. Trainium runs on AWS's own Neuron SDK rather than CUDA, which means any customer moving workloads to Trainium takes a migration cost — but for new workloads, or for buyers willing to invest in Neuron optimization, the cost differential may make that migration worthwhile.

No specific buyers or pricing terms have been disclosed. AWS said discussions are at an early stage and declined to comment further beyond DeSantis's Bloomberg interview.