Amazon launches TorchNeuron to loosen Nvidia's grip on PyTorch

Amazon Web Services on Monday launched TorchNeuron, a tool that moves AI applications built with PyTorch onto Trainium, the company’s own AI chip, by changing a single configuration line rather than rewriting code. The release targets the software edge that has made Nvidia the default hardware for AI workloads: Amazon says about 85% of AI apps run on PyTorch, the open-source framework Facebook originally built and that has been part of the Linux Foundation since 2022, and Nvidia has spent years making sure its CUDA software works seamlessly with it.

TorchNeuron is built on top of PyTorch itself, so anyone already fluent in the framework can pick it up quickly. Amazon isn’t the first to try this trick — PyTorch apps can already be ported to AMD GPUs or Apple’s chips with a similar configuration tweak — but Amazon argues Trainium’s real edge is bundling that portability with its entire cloud stack, which it says beats relying on Nvidia-tied neoclouds such as CoreWeave, Nebius, and Firmus.

Colin Brace, who runs technology software at AWS Annapurna Labs — the unit behind Amazon’s custom chips — framed the pitch as avoiding dependence on a single-purpose compute provider, since AWS already has the infrastructure to take a project from prototype to production. The company also points to the fact that the same Trainium chip handles both training and inference, which it says lowers costs across a full workload even when its per-token price isn’t the cheapest on the market; Amazon cites Deepgram, a voice-AI company that competes with ElevenLabs, processing 70% more audio per dollar after switching to Trainium with native PyTorch support.

Beyond the chip pitch, Amazon says it’s pushing to make PyTorch itself less Nvidia-centric, contributing code upstream to strip out hardware-specific ties so the framework works equally well across chipmakers, not just Trainium. Nvidia’s hardware is currently embedded in PyTorch almost by default, and unwinding that could matter well beyond this one dev tool. “The real story is removing the lock-in that kept developers from having a real choice at all,” Brace said.

Why it matters

If Amazon can genuinely decouple PyTorch from Nvidia’s hardware, the contest moves beyond chip specs to who controls the software layer nearly every AI developer touches — a bigger threat to Nvidia than any single rival chip.