Amazon opens its Trainium AI chips to PyTorch code as shares post best day since July

Amazon said on Monday, Oct. 19, that AI developers can now run code written in PyTorch, the most widely used framework for building AI models, on its Trainium chips without rewriting it. The software, called TorchNeuron, targets one of the main reasons companies stay with Nvidia, and Amazon's stock had its best day in more than two months.

What Amazon released

TorchNeuron is a piece of software that makes Amazon's Trainium2 and Trainium3 chips show up to PyTorch as a standard device. Amazon says code already written in PyTorch can run on the chips with a single configuration change. Until now, moving AI work to a new chip usually meant rewriting and revalidating code that "took months to perfect," the company said.

The software will be released as open source under the Apache 2.0 license on Tuesday, Oct. 20, at the PyTorch Conference North America, where it goes into public beta. It supports PyTorch's "eager" mode, which runs each step immediately and is used for research and debugging, as well as its compiled mode for production work. Developers who want more speed can write their own low-level code for the chips.

Amazon says it tests the software against hundreds of thousands of AI models on every build. The technical post on the PyTorch Foundation blog is more candid about the limits: a continuous sweep of 100,000 models from Hugging Face, which the authors say cover 99.9% of real-world use by download count, still turns up gaps that Amazon aims to close in future releases. Support for lower-precision FP8 math remains in private beta.

PyTorch matters because of how many people use it. Amazon says 85% of AI researchers and developers write their code in it. The framework is open source and stewarded by Meta, Reuters noted.

Why it is aimed at Nvidia

Nvidia's lead in AI computing rests partly on software. The company has spent more than a decade on a tool called CUDA that makes PyTorch code run well on its own chips, Reuters reported, describing TorchNeuron as an attempt to erode the advantage that underpins Nvidia's dominance. Google, which also designs its own AI chips, is working on better PyTorch tools to lure developers away from Nvidia, according to Reuters.

Amazon's pitch is cost. "This has been about meeting customers where they're at," Colin Brace, vice president of software at Amazon's Annapurna Labs chip unit, told Reuters. He said more than 30 early customers, in robotics, world models and voice AI, are using the software, and that switching to Amazon's chips can bring steep savings.

One of those customers put a number on it. Voice AI company Deepgram said in a statement published with the launch that, in the benchmarks it evaluated, Trainium processed about 60% more input audio seconds per dollar than its existing GPU setup. Hugging Face, which runs a popular open platform for sharing AI models, said TorchNeuron makes Trainium a native option in its widely used Transformers and Diffusers libraries. Meta's director of PyTorch engineering, Jana van Greunen, said native support "expands that choice" for developers.

Amazon has been building chips since it bought the designer Annapurna Labs in 2015 and now has four chip families. The company says its custom chip business has passed a $25 billion annual revenue run rate.

What the stock did

Amazon shares closed at $262.45 on Monday, up 5% from Friday's close of $250.00, according to Yahoo Finance data. It was the stock's biggest one-day percentage gain since July 31.

The software news shared the day with a Wall Street call, and their separate effects on the price cannot be untangled. Morgan Stanley named Amazon a top pick on Monday and raised its price target, CNBC reported, crediting the gain to that call, with analyst Brian Nowak pointing to accelerating growth at Amazon Web Services, the potential of its Bedrock AI platform and efficiencies from robotics. Reuters also placed the morning rise alongside the Morgan Stanley call.

Nvidia shares rose 1.4% to $217.09 on Monday, as chip stocks bounced after a long losing streak. The VanEck Semiconductor ETF climbed nearly 3% by midday after nine straight losing days, CNBC reported.