Amazon wants to make swapping out Nvidia chips about as complicated as changing a setting in a config file. That is not hyperbole: the company’s new TorchNeuron software backend lets developers move existing PyTorch code onto Amazon’s Trainium AI chips by changing one line, switching from .to(‘cuda’) to .to(‘neuron’), with no rewriting or revalidation required.
Amazon’s Annapurna Labs unit announced TorchNeuron on October 19, 2026, and the public beta launched the following day at the PyTorch Conference. The timing was deliberate: PyTorch is the dominant framework for AI model development, and landing at its flagship conference plants a flag directly in Nvidia’s territory.
What TorchNeuron actually does
TorchNeuron is designed to remove that friction. Released under the Apache 2.0 open-source license, it functions as a backend that integrates directly with PyTorch, allowing Trainium chips, specifically the Trainium2 and Trainium3 generations, to appear to the framework as native devices. The practical effect is that a team running workloads on Nvidia hardware could, in theory, migrate to AWS Trainium infrastructure without touching its codebase.
More than 30 early adopters participated in alpha and beta testing before the public launch. Those firms include voice AI companies Cartesia and Deepgram, along with the Lieber Institute for Brain Development, which has been using the tool for drug research applications. Early users have reported cost savings associated with the switch, though Amazon has not published specific figures.
The business case behind the software
Amazon’s custom chip division has reached a $25 billion annual revenue run rate, according to Amazon, with the company projecting that figure could reach $50 billion.
What to watch as TorchNeuron matures
The initial release is not without rough edges. TorchNeuron currently skips approximately 22,000 tests within the PyTorch test framework, primarily because it does not yet support certain advanced features including dynamic input shapes. That is a real limitation for teams running models that require flexible input handling, and it signals that TorchNeuron is a beta product in the genuine sense of the term rather than the marketing sense.
Disclosure: This article was edited by Diego Almada Lopez. For more information on how we create and review content, see our Editorial Policy.