Amazon Gains 5% as TorchNeuron Targets Nvidia's CUDA Advantage

Amazon (AMZN) closed up 4.98% at $262.45 on Monday, the day it unveiled TorchNeuron, a developer tool aimed at one of Nvidia's (NVDA) biggest advantages: the CUDA software that keeps AI developers on Nvidia chips. TorchNeuron is meant to let PyTorch-based AI applications run on Amazon's own Trainium chips with a single configuration change and no code rewrite. We cannot say how much of the stock's move came from this news.

Amazon's move

Amazon closed at $262.45 as of 4:34 PM ET, up $12.45 on the day, and traded at $262.60 after hours. We did not have reliable Monday moves for Nvidia, CoreWeave (CRWV) or Nebius (NBIS), so we are not reporting them.

What Amazon announced

TorchNeuron is Amazon's native PyTorch backend for its Trainium chips. In an Amazon demo, the same training script ran on a conventional chip and on Trainium, with only a setting telling the code which chip to use, and the model learned at the same rate on both, according to Amazon. It supports Trainium2 and Trainium3. Amazon plans to release it as open-source software under the Apache 2.0 license at the PyTorch Conference on Tuesday, October 20, and some reports describe that release as a public beta. Early users include Cartesia, Deepgram and the Lieber Institute for Brain Development, and Hugging Face has added native Trainium support in its tools.

The Nvidia angle

Nvidia's advantage is not only its chips. PyTorch code runs on Nvidia GPUs through CUDA, so moving to another chip normally means reinvesting time and engineering work in a new software stack. Amazon is also contributing changes to PyTorch to remove code that ties the framework to specific hardware, which could over time weaken that lock-in. Switching costs, rather than chip speed alone, are what Amazon is going after.

Analysts are cautious about how fast that happens. The Motley Fool argued the effect on Nvidia is likely to be gradual: TorchNeuron covers only PyTorch work, systems built on custom CUDA code will not port easily, and the tool is still in beta. The same piece noted Alphabet is reportedly building a similar tool, TorchTPU, and that customers do not need to abandon Nvidia entirely for alternatives to give them leverage in price talks.

The cost argument

AWS says Trainium supports both training and inference, which improves overall economics. It cites AI startup Deepgram, which it says processed 70% more audio per dollar after switching. We could not find that figure in Amazon's announcement or in the news coverage we reviewed, so treat it as an Amazon-attributed claim until it is confirmed. TNW noted that Amazon's performance result comes from a single demo, not a broad benchmark, and that some PyTorch features, such as dynamic input shapes, are not yet supported.

Pressure on the neoclouds

Amazon is positioning its integrated AWS infrastructure against neocloud providers that depend on Nvidia, such as CoreWeave and Nebius. If developers can run the same PyTorch code on Trainium inside AWS, Amazon can offer an alternative to renting Nvidia GPUs from those providers. We did not find comments from CoreWeave or Nebius on TorchNeuron. For related coverage, see our reports on Nebius's GPU pricing and Nvidia's buyback.