Amazon plans to sell Trainium AI chips outside AWS in late 2027
Amazon logo (public domain) via Wikimedia Commons
Demand for Amazon's in-house AI accelerators has reportedly outrun what AWS can absorb, so the company is preparing to sell them elsewhere
Amazon plans to start offering its Trainium AI chips outside of Amazon Web Services in late 2027. The company says the move responds to customer demand.
Trainium is Amazon’s custom silicon for training and running AI models. Think of it as Amazon’s house-brand alternative to the GPUs most AI labs buy from Nvidia.
Until now, the only way to use those chips was to rent them as part of AWS’s cloud. Selling them outside AWS would let other data centers run Trainium hardware directly.
Peter DeSantis, who leads AI at AWS, has confirmed that Amazon is in talks with potential buyers. Those discussions were taking place as of mid-2026.
The timeline stays loose. Late 2027 is the window for meaningful external availability. Amazon has not disclosed specific timelines or pricing for these sales.
The reason is simple: there isn’t enough Trainium to go around. Capacity through AWS is reported to be mostly sold out for 2026. It is reportedly largely committed for 2027 and parts of 2028 as well.
The numbers behind the demand
The financial commitments are large. More than $225 billion in revenue commitments tied to Trainium have been made, involving major AI firms such as OpenAI and Anthropic.
Both labs are committed to using Trainium. Each plans to ramp up its capacity significantly in the coming years.
AI, tech, and the markets they move—in one daily briefing.
Daily. Free. Join 34,000+ readers across crypto, finance, and policy.
Amazon CEO Andy Jassy has suggested the broader chip segment could reach more than $20 billion in annual revenue by 2026. That segment covers both Trainium and Graviton, Amazon’s general-purpose processor line.
Trainium3 delivers a 30-40% improvement in price-performance over Trainium2. Trainium2 was itself approximately 30% better than existing GPUs on that measure.
Background: the software problem
Fast chips are only half the battle in AI hardware. The other half is software, and that is where Nvidia has built its moat.
Nvidia’s CUDA platform is the toolkit most AI developers grew up on. Years of code, libraries and engineering habits are tied to it. Building a comparable ecosystem around a rival chip is a formidable challenge, regardless of how good the silicon is.
Amazon is trying to lower that barrier. In October 2026, it launched TorchNeuron and open-sourced it. The tool lets developers move PyTorch code onto Trainium chips with minor adjustments.
What this means
For Amazon, external sales would add a new revenue stream on top of its cloud rental business. Today, Trainium income runs through AWS. Selling hardware to outside data centers would let Amazon earn from demand it currently cannot serve.
For investors, the picture comes with real upside and clear unknowns. The $225 billion-plus in commitments and Jassy’s $20 billion-plus revenue outlook point to a chip business with serious scale.
No pricing has been disclosed, and pricing will decide whether Trainium can win buyers away from Nvidia. The software ecosystem remains a hurdle, since TorchNeuron is still new and CUDA’s head start is measured in years.
There is also a supply question. If Trainium is already reportedly committed into 2028 through AWS alone, external buyers will be competing for the same limited output.
Disclosure: This article was edited by Diego Almada Lopez. For more information on how we create and review content, see our Editorial Policy.

