AWS TorchNeuron runs PyTorch apps on Trainium and pushes a chip-neutral PyTorch

AWS released TorchNeuron, a PyTorch-based tool that Amazon says lets AI apps written in PyTorch run on its Trainium chips with a one-line configuration change and no code rewrite, TechCrunch reports. Amazon puts PyTorch’s share of AI apps at about 85%, and NVIDIA has spent years making CUDA the path of least resistance inside that framework. Trainium is built by Annapurna Labs, Amazon’s custom chip unit. PyTorch was created at Facebook and has been governed under the Linux Foundation since 2022.

The bigger move is upstream: Amazon says it will contribute changes to PyTorch itself to remove code that ties the framework to specific chips, aiming for a more hardware-agnostic PyTorch for everyone, not only Trainium users. A day later Google open-sourced TorchTPU (Google joins Amazon), and Hugging Face added native TPU support (TorchTPU in Hugging Face).

NVIDIA’s advantage is as much software habit as silicon: developers inherit CUDA because everything works. Amazon is attacking that inertia through the open framework most of the industry already uses, which is more practical than asking developers to learn a new stack. The one-line port is not unique, since PyTorch users can already target AMD GPUs and Apple chips with a configuration switch. Amazon’s pitch is the destination: Trainium handles training and inference and comes wrapped in the rest of AWS, which it argues beats renting from NVIDIA-aligned neoclouds like CoreWeave or Nebius. AWS points to Deepgram, which it says processes 70% more audio per dollar after moving, a vendor-supplied claim rather than an independent benchmark.

A PyTorch that is genuinely neutral about hardware would be good for developers and prices, making switching credible and competition real, in one of tech’s most expensive line items.

Hardware-agnostic can also mean more configuration. PyTorch’s value is hiding the details of getting code onto chips; if neutrality adds configuration burden or subtle performance differences that surface only in production, developers pay. Large upstream changes from a chip vendor will need careful review by maintainers. And escaping one vendor’s moat by moving deeper into one cloud is not full freedom; Amazon sells choice and gravity at once. NVIDIA, meanwhile, keeps deepening its own stack (DOCA GPUNetIO). More in our Nvidia coverage, tools coverage and business coverage.