By Hamza Faisal
Amazon.com is taking its biggest swing yet at the software advantage that has kept Nvidia firmly on top of the AI computing market. On Monday, the company's Annapurna Labs chip unit unveiled TorchNeuron, a new tool that lets developers run code written in PyTorch — the most popular framework for building AI models — directly on Amazon's custom Trainium AI chips with a single configuration change and no rewriting.
The tool will be publicly released as a free, open-source offering. The public beta opened on October 20, timed with the PyTorch Conference in San Jose, California.
A single line instead of a rewrite
Moving AI software between different chips has historically meant months of engineering work. Developers write their models in PyTorch, but the underlying software that connects PyTorch to the hardware differs from chip to chip. On Nvidia's GPUs, that job is handled by CUDA, a software platform Nvidia has refined for more than a decade — one of the main reasons developers stay in its ecosystem.
TorchNeuron changes that equation for Amazon's chips. The tool registers Trainium as a native device inside PyTorch, so developers can take code they already wrote and point it at Trainium instances with a one-line device change. It supports eager execution, PyTorch's distributed training APIs, and Amazon's Neuron Kernel Interface for writing custom kernels, according to Amazon.
Hugging Face, whose Transformers, Diffusers, Accelerate, PEFT and TRL libraries now work with Trainium natively through the new backend, called the change a significant simplification over the previous compiler-style workflow, which relied on the torch-neuronx package and static shapes.
Wall Street likes the swing
Amazon shares climbed as much as 5% intraday after the announcement, and Morgan Stanley named Amazon a "top pick," according to market coverage of the launch.
Colin Brace, vice president of software for Annapurna Labs, told Reuters the goal was to cut the number of steps developers must take to try Amazon's in-house chips. "This has been about meeting customers where they're at," he said, adding that more than 30 early customers — spanning robotics, world models, voice AI, drug discovery and finance — are already using the software and reporting steep cost savings from the switch.
The battle for AI developers heats up
Amazon is not alone in going after Nvidia's developer lock-in. Google, which also builds its own custom AI chips, is similarly working to make PyTorch run better on its hardware. And Amazon says it is contributing changes back to PyTorch itself to make the framework easier to use across different hardware — a longer-term effort to make AI software genuinely portable.
For now, the immediate pitch is narrower but concrete: cheaper AI compute with far less code surgery. If TorchNeuron delivers on that promise, developers may finally have a reason to look beyond Nvidia's chips — and Amazon's multibillion-dollar bet on its own silicon may start to pay off.
