Amazon has taken one of its boldest swings yet at Nvidia’s grip on artificial intelligence hardware, and it did so with software rather than a new chip. The company has introduced TorchNeuron, a tool that lets developers run their existing PyTorch code on Amazon’s own Trainium chips with almost no changes.
PyTorch is the programming framework most AI researchers use to build and train models. Amazon says around 85 percent of AI developers write their code in it. Until now, moving that code from Nvidia graphics processors to other hardware usually meant weeks or months of rewriting and testing, which is one of the main reasons Nvidia and its CUDA software platform have become the default choice.
TorchNeuron is meant to remove that barrier. According to Amazon, developers simply change the device setting in their code from “cuda” to “neuron”, and the same training script runs on Trainium. The tool supports quick experimental runs, compiled production workloads and the standard tools for spreading training across many chips. It works with Trainium2 and the newer Trainium3.
Amazon is releasing TorchNeuron as open source under the Apache 2.0 licence, with a public beta launching at the PyTorch Conference North America on 20 October. Early users already include startups working on voice AI, robotics, video and drug discovery.
For large companies the appeal is flexibility as well as price. Amazon argues that the same Trainium fleet can serve customers during busy hours and switch to training jobs when demand drops, so expensive hardware sits idle less often.
The move matters beyond Amazon’s own cloud. If developers can switch between chips without a costly rewrite, chipmakers have to compete harder on price and performance, and startups gain real bargaining power. Nvidia still dominates the market, but every tool that weakens software lock-in makes that position a little less comfortable.
