New to this? Read it in simple words
- Amazon released TorchNeuron, a free and open-source tool, as a public beta on October 20.
- It lets AI code written for PyTorch run on Amazon’s own Trainium chips.
- Amazon says one setting is enough. Its own list shows some features don’t work yet.
- It is a step against Nvidia, whose software makes its chips the usual choice.
- PyTorch
- A free toolkit that most developers use to build and train AI models.
- Backend
- The part of a toolkit that turns the code into work a specific chip can do.
- CUDA
- Nvidia’s software that lets programs use its chips. Much AI code is built around it.
PyTorch is the most widely used toolkit for building AI models, and Meta looks after it. Trainium is the AI chip that Amazon designs for its own cloud. Until now, PyTorch code needed extra work before it could run on Trainium.
What Amazon released
TorchNeuron is a “native PyTorch backend” for Trainium, the PyTorch blog says. It works with Trainium2 and Trainium3, Amazon’s two newest chips. The code is on GitHub under the Apache 2.0 licence, and the project is marked as a beta.
The public beta starts on October 20, the first day of the PyTorch Conference in North America. The tool itself is free and open source. Renting Trainium chips on Amazon’s cloud still costs money.
It is not a brand-new idea. Amazon’s own documents say TorchNeuron was first announced at AWS re:Invent in December 2025. What is new is the open code and the public beta.
From the PyTorch blog, Amazon’s post and its GitHub page. “One setting” is Amazon’s claim.
Amazon’s claims and the limits
Amazon says developers can move their code “with a single configuration change and no rewrite”. In the PyTorch blog’s example, the only change is sending the work to a device called “neuron”.
The same blog lists what does not work yet. The first release does not support sparse tensors, dynamic shapes and some other PyTorch features, and some operations have no Trainium version so far.
Customers in the blog describe different amounts of work. The voice company Cartesia speaks of “minimal changes”. Deepgram, a speech company, says moving its model took several months with help from AWS.
The race with Nvidia
Nvidia has built its CUDA software for more than a decade so that AI code runs well on its chips, Reuters reports. TechCrunch writes that Nvidia’s hardware is now built into PyTorch “essentially by default”.
“The real story is removing the lock-in that kept developers from having a real choice at all,” Colin Brace of AWS’s Annapurna Labs told TechCrunch.
Brace told Reuters that more than 30 early customers already use the software, in areas such as robotics and voice AI. Google is also working on better PyTorch support for its own chips, Reuters reports.
Sources
Every fact in this story comes from the sources below. Open them to check our work.
- 1Primary source · October 19, 2026Introducing TorchNeuron: Native PyTorch backend for AWS Trainium PyTorch (AWS engineers)
- 2Primary source · October 19, 2026Amazon opens its AWS Trainium AI chips to every PyTorch developer, no rewrite needed Amazon
- 3
- 4Primary source · October 19, 2026About PyTorch support on Neuron (TorchNeuron) AWS Neuron documentation
- 5Research · October 19, 2026Amazon is taking its biggest shot at Nvidia yet with new TorchNeuron dev tool TechCrunch
- 6Research · October 20, 2026Amazon unveils tool to challenge Nvidia’s AI chip dominance Reuters (via Daijiworld)
- 7Research · October 19, 2026Amazon introduces toolkit for running PyTorch workloads on AWS Trainium chips SiliconANGLE
That one setting is enough is Amazon’s claim; the customers quoted describe very different amounts of work, and nobody outside has tested it yet. Deepgram’s cost saving is left out, because Deepgram and AWS give different numbers.
Amazon’s free, open-source TorchNeuron opens its Trainium chips to PyTorch code, though some features don’t work yet. How we rate the mood