Amazon Web Services released a developer tool called TorchNeuron on Monday that lets AI startups move applications written with PyTorch onto Trainium, Amazon's custom AI chip, by changing a single line of configuration. The release is AWS's most direct attempt yet to erode the software advantage that has made Nvidia the default hardware choice for AI workloads — an advantage built on years of CUDA integration rather than raw silicon alone.

One line of configuration

TorchNeuron works with PyTorch, the free, open-source toolkit developers use to build, train and run AI models. Amazon says applications written with PyTorch can run on Trainium without a code rewrite. The tool is adapted from PyTorch itself, so developers who already know the framework do not need to learn a new programming model to use it.

That matters because of how much of the AI software stack is written in PyTorch. Amazon estimates roughly 85% of AI applications use the framework. Meta originally developed PyTorch, and since 2022 it has been maintained under the Linux Foundation.

The CUDA moat

Nvidia's grip on AI hardware rests substantially on CUDA, the software infrastructure that lets applications run on its GPUs and other hardware. Over years, Nvidia has made CUDA support PyTorch applications seamlessly, to the point where Nvidia hardware is effectively embedded in the toolkit by default. TorchNeuron is designed to break that default.

AWS is not the first to try. PyTorch users can already port applications to AMD GPUs or even Apple chips through configuration changes. Amazon's pitch is that moving to Trainium combines that ease with access to its broader cloud. The company argues this makes it a better option for startups than neoclouds tied to Nvidia hardware, such as CoreWeave, Nebius and Firmus.

"You're building on AWS. You're not betting on a single-purpose compute provider and scrambling to build everything else around it," Colin Brace, VP of Technology Software at AWS Annapurna Labs, the unit that makes Amazon's custom chips, said in an email. "On AWS, the infrastructure to go from prototype to production is already there, allowing a founder to move fast and scale when they need to."

Price-performance and the upstream play

AWS contends that because the same chip handles both training and inference, and is tightly integrated with the cloud infrastructure startups need, its platform saves money even where it does not offer the lowest per-token price. "Trainium's value is price-performance plus the fact that it's one platform for both training and inference, which improves the economics across a full workload," Brace said. He cited Deepgram, an AI voice provider competing with ElevenLabs, which moved to Trainium using native PyTorch and now processes 70% more audio per dollar at production-grade accuracy.

Amazon is also going further than its own hardware. In the blog post announcing TorchNeuron, the company said it is contributing changes upstream to PyTorch, working with the community to remove code that ties the framework to specific chips. The stated goal is to make PyTorch more hardware-agnostic for all developers, not only Trainium users.

That is a notable strategic move. PyTorch's value partly lies in hiding the difficulty of getting AI code to run on different hardware. If Amazon succeeds in decoupling the framework from Nvidia, the consequences for the GPU giant could extend well beyond competition from rival chips. As Brace put it: "The real story is removing the lock-in that kept developers from having a real choice at all."

What to watch

Whether developers gain real choice and cheaper options, as Amazon promises, or simply more configuration headaches will depend on how the upstream changes land and how well TorchNeuron performs in production outside Amazon's own benchmarks. For now, the tool gives PyTorch teams a low-friction path onto Trainium — and gives Nvidia's software moat its most serious test to date.