On Monday, October 19, Amazon Web Services released TorchNeuron, a tool that lets AI startups move PyTorch applications onto Trainium, Amazon’s AI chip. According to Amazon, a PyTorch app can run on Trainium by changing a single line of configuration, with no code rewrite. The release is aimed squarely at the software advantage that has made Nvidia the default hardware choice for AI developers.

What TorchNeuron does

PyTorch is a free, open source toolkit used to build, train and run AI models. It was created at Facebook and has been part of the Linux Foundation since 2022. Amazon says about 85% of AI apps are written with it. TorchNeuron is adapted from PyTorch, so developers who already know PyTorch should know how to use it. Instead of asking teams to learn a new framework, AWS hides Trainium-specific details behind the interface they already use.

The target is Nvidia’s CUDA ecosystem. For years Nvidia has made sure its CUDA software works seamlessly with PyTorch, so Nvidia hardware is, in TechCrunch’s words, embedded in the toolkit essentially by default. For a deeper look at that advantage, see our explainer on why Nvidia dominates AI through CUDA.

Not the first attempt

Amazon is not the first to try. PyTorch users can also port applications to AMD GPUs or to Apple chips with a simple configuration change. AWS argues its pitch is different because Trainium comes with the rest of its cloud. Colin Brace, VP of Technology Software at AWS Annapurna Labs, the unit that designs Amazon’s chips, told TechCrunch: “You’re building on AWS. You’re not betting on a single-purpose compute provider and scrambling to build everything else around it.” AWS contrasts that with neoclouds such as CoreWeave, Nebius and Firmus, many of which are tied to Nvidia hardware.

The price-performance argument

AWS says one chip handles both training, teaching a model, and inference, running it for users, and that this integration improves total economics even when AWS is not the cheapest per token. “Trainium’s value is price-performance plus the fact that it’s one platform for both training and inference,” Brace said. As evidence, he cited Deepgram, a voice AI provider and competitor to ElevenLabs, which moved to Trainium using native PyTorch support and is processing 70% more audio per dollar at production-grade accuracy. That figure comes from AWS and Deepgram; it is a single customer example and we have not verified it independently. AWS published a related post about Deepgram on its machine learning blog.

Upstream changes to PyTorch

The most consequential part may be what Amazon will do beyond its own hardware. In its announcement, the company says it is contributing changes upstream to PyTorch to remove code that ties the framework to specific chips, with the goal of making PyTorch more hardware-agnostic for all developers, not just Trainium users. If accepted by the PyTorch community, that could lower switching costs between vendors. TechCrunch cautions that PyTorch exists to hide the difficulty of running code on hardware, so it remains to be seen whether decoupling it from Nvidia gives developers more affordable options or merely more configuration headaches.

Context: Nvidia’s customers become competitors

Amazon is one of several large buyers building their own silicon. Our piece on why big tech builds its own AI chips covers the economics: custom chips reduce dependence on a single supplier and let a cloud capture more of the margin. TorchNeuron addresses the main practical objection, which has been software compatibility rather than raw chip performance.

What to watch next

  • Which of Amazon’s upstream changes the PyTorch maintainers accept and when.
  • Independent benchmarks comparing Trainium with Nvidia GPUs on the same PyTorch code.
  • Whether startups beyond Deepgram publish migration results.
  • Nvidia’s response on software, pricing or PyTorch support.
  • How well “one line” migrations hold up for large, custom training code.

Key takeaways

  • TorchNeuron is a PyTorch-based tool for moving AI apps onto Trainium.
  • Amazon says only one line of configuration changes, with no rewrite.
  • AWS is upstreaming code to make PyTorch less dependent on specific chips.
  • The Deepgram result of 70% more audio per dollar is a vendor-cited example.
  • The real test is whether complex production code ports as easily as the pitch suggests.

Frequently asked questions

What is Trainium?

Trainium is Amazon’s custom AI chip. AWS says it handles both training and inference.

Does this mean I can drop Nvidia GPUs?

Not automatically. Amazon says PyTorch apps can run on Trainium with a configuration change, but performance and cost depend on the workload and have not been independently compared here.

Why does PyTorch matter?

About 85% of AI apps are written with it, according to Amazon, so whoever controls how it maps to hardware shapes developer choice.