Nvidia’s grip on AI hardware has never been only about chips. It is about the software that makes those chips the path of least resistance. Amazon just aimed at that, and it’s a smarter target than another benchmark chart.
On Monday, AWS released TorchNeuron, a tool that lets AI apps built on PyTorch run on Trainium, Amazon’s custom AI chip. According to Amazon, the move takes a single line of configuration change and no code rewrite. PyTorch is the free, open source toolkit developers use to build, train, and run models, and Amazon says roughly 85% of AI apps are written with it.
That number explains the strategy. PyTorch began at Facebook and has sat under the Linux Foundation since 2022. Over years, Nvidia worked to make sure its CUDA software supported PyTorch smoothly, so Nvidia hardware became the default underneath it. Developers rarely chose Nvidia in any active sense. They just never had a reason to choose anything else.
TorchNeuron is adapted from PyTorch, so anyone who knows the framework already knows the tool. Amazon is not alone here. PyTorch apps can already be ported to AMD GPUs and even Apple’s chips with a configuration change. What Amazon adds is the pitch around its cloud. Colin Brace, VP of Technology Software at AWS Annapurna Labs, told TechCrunch that building on AWS means you are not betting on a single-purpose compute provider and then scrambling to build everything else around it.
That is a clear shot at the neoclouds, many of which are tied to Nvidia hardware, including CoreWeave, Nebius, and Firmus. Amazon’s argument is that Trainium handles both training and inference, and that tight integration with the rest of AWS lowers total cost even if it isn’t the cheapest per-token price. Brace pointed to Deepgram, the AI voice provider that competes with ElevenLabs, which reportedly processes 70% more audio per dollar after moving to Trainium with native PyTorch support. That is one customer’s claim, and worth treating as a case study rather than a benchmark.
The more ambitious part is upstream. Amazon says it will contribute code to the main PyTorch project to strip out logic that ties the framework to specific chips, aiming for a more hardware-agnostic PyTorch for everyone. If that lands, it matters well beyond Trainium. AMD, Google’s TPUs, and other chip vendors would all benefit from the same cleanup.
Still, some skepticism is fair. Amazon is obviously not a neutral party, and this approach is good news mainly for companies selling hardware and cloud services that compete with Nvidia. PyTorch exists partly to hide the messy details of running code on different hardware. Whether pushing it further from Nvidia gives developers real choice and cheaper options, or just more configuration headaches, is unproven.
So is this noise? Not quite. A one-line switch is a nice demo, but the real test is performance parity, debugging tools, and library coverage on Trainium, which is where CUDA’s years of polish still count. Startups will find out quickly. But if Amazon succeeds in loosening PyTorch’s ties to Nvidia, the consequences reach past chip competition. As Brace put it, the real story is removing the lock-in that kept developers from having a real choice at all. Nvidia’s moat was always code, and that is where this fight is happening.
