On Monday, Amazon Web Services (AWS) released a tool called TorchNeuron to help AI startups more easily migrate their AI applications to Amazon’s AI chip, Trainium. This move directly challenges NVIDIA’s software advantage—the very advantage that has helped NVIDIA become the default hardware choice for AI applications and solidified its position as the leading AI hardware provider.
TorchNeuron works with PyTorch, a free, open-source toolkit used by developers to build, train, and run AI models. Amazon states that with TorchNeuron, AI applications written in PyTorch can run on Trainium with just a single configuration change, without requiring any code rewrites.
Amazon states that approximately 85% of AI applications are written in PyTorch. Originally developed by Facebook, PyTorch has been part of the Linux Foundation since 2022, which now maintains it as an open-source project. For years, NVIDIA has ensured that its CUDA software—the infrastructure enabling applications to run on NVIDIA GPUs and other hardware—seamlessly supports AI applications written in PyTorch. As a result, NVIDIA’s hardware is now deeply embedded in this toolchain and has become the de facto standard.
TorchNeuron is adapted from PyTorch, so developers familiar with PyTorch will know how to use it. Amazon is not the first company to try making its own chips easier to integrate with this platform. PyTorch users can also migrate their applications to AMD GPUs or Apple chips with simple configuration changes.
However, Amazon believes that migrating applications to Trainium is not only straightforward but also enables AI startups to access Amazon’s full suite of cloud services. AWS clearly considers this a better option for startups than using neoclouds, many of which are tightly coupled with NVIDIA and its hardware, such as Coreweave, Nebius, and Firmus.
“You’re building on AWS. You’re not betting on a provider that only offers single-purpose compute and then scrambling to build everything else around it,” said Colin Brace, Vice President of Technical Software at AWS Annapurna Labs, in an email to TechCrunch. Annapurna Labs is Amazon’s in-house chip development division. He added, “On AWS, the infrastructure needed to go from prototype to production is already in place, allowing founders to move quickly and scale up when needed.”
AWS states that its cloud services can help customers save costs because the same chip can handle both training (teaching AI models) and inference (deploying trained models, such as answering user questions), and is tightly integrated with the cloud infrastructure needed by startups—even though AWS’s per-token price is not the lowest.
Brace said: "The value of Trainium lies in its cost-performance ratio and its ability to handle both training and inference on a single platform, which improves the economics of the entire workload. Take Deepgram—they migrated to Trainium with native PyTorch support and achieved a 70% increase in audio processed per dollar while maintaining production-grade accuracy." Deepgram is an AI speech applications provider and a competitor to ElevenLabs.
But Amazon is not content with merely offering PyTorch application developers an easy path to migrate to its cloud.
Amazon also stated that it is pushing PyTorch to move away from being automatically NVIDIA-centric and will contribute the code to achieve this goal to the main PyTorch project. In the blog post announcing this new tool, the company wrote:
Amazon's approach extends beyond its own hardware. The company is contributing upstream changes to PyTorch itself and collaborating with the PyTorch community to remove code that ties the framework to specific chips. The goal is to make PyTorch more hardware-agnostic for the entire developer community, not just Trainium users.
Although this strategy undoubtedly benefits sellers of hardware and cloud services that compete with NVIDIA, one of PyTorch’s goals is to hide the troublesome details of running AI code on hardware. We still need to wait and see whether pushing NVIDIA further, as Amazon claims, will truly give developers more choices and lower costs—or simply create more configuration headaches.
Regardless, if Amazon truly succeeds in deeply decoupling PyTorch from NVIDIA, the impact could extend far beyond countering chip competitors. As Brace said, “The real story is eliminating the lock-in effect that leaves developers with no real choice.”