Amazon's Annapurna Labs has introduced TorchNeuron, an open-source software backend that allows developers to run PyTorch code on Trainium AI chips. The tool enables migration from Nvidia hardware to Amazon silicon by changing a single line of code from .to(cuda) to .to(neuron).
The public beta was launched on October 20, 2026, at the PyTorch Conference, following an announcement on October 19. TorchNeuron is released under the Apache 2.0 license and supports Trainium2 and Trainium3 chip generations, integrating them as native devices within the PyTorch framework.
More than 30 organizations, including Cartesia, Deepgram, and the Lieber Institute for Brain Development, participated in the testing phase. While early users reported cost savings, Amazon has not disclosed specific financial data regarding these gains.
The division responsible for Amazon's custom chips currently reports an annual revenue run rate of $25 billion, with projections suggesting this figure could grow to $50 billion. The current beta version of TorchNeuron skips approximately 22,000 tests in the PyTorch framework, as it lacks support for certain features like dynamic input shapes.
