· Translated from the Spanish original

Cloud computing

AWS launches TorchNeuron and lets PyTorch code run on its Trainium chips

The public beta arrives October 20 under Apache 2.0. Deepgram measured 60% more audio per dollar than on GPU, but several operators are still missing.

What happened

  • AWS announced TorchNeuron, a native PyTorch backend for its Trainium2 and Trainium3 chips. It enters public beta on October 20, 2026 under the Apache 2.0 license, according to the PyTorch blog.
  • It installs with pip install torch-neuronx. To move a model you just name the device, .to('neuron'). In compiled mode the backend must be named once.
  • It supports eager execution, torch.compile and the usual distributed APIs, such as FSDP2 and tensor parallelism. Hugging Face lists it as a device in Transformers v5 and Diffusers.
  • Deepgram, cited by AWS, reports that Trainium processes about 60% more seconds of input audio per dollar than its GPU baseline, in the benchmarks it evaluated.

Why it matters

  • Until now, using Trainium required adapting code with AWS-specific tools. That engineering cost was the practical reason to stay on Nvidia GPUs. TorchNeuron reduces it, without eliminating it.
  • For a data team in Chile that rents compute, the change is one of bargaining power: moving a workload between providers stops being a months-long project.
  • The consequence the announcement omits: the 60% is a figure from a single company in one case, with support from AWS’s innovation center over several months. It is not the result an SME will get with a week of work.

The number

~22,000 skipped tests. These are PyTorch tests that AWS skipped because some operators lack a native Neuron lowering, according to its own report.

Context

Support arrives through PrivateUse1, PyTorch’s official mechanism for third-party hardware. AWS acknowledges limits in this first version: no dynamic shapes or sparse tensors, and FP8 remains in private beta. It tested 100,000 Hugging Face models in a reduced configuration and publishes the result in a chart, with no figure in the text.

What’s next

  • FP8 and TorchAO will reach general availability in the next release, with no date.
  • AWS says it will close most operator gaps “in future releases”, with no calendar.
  • It remains to be seen how many third-party libraries declare real compatibility, beyond AWS’s own tests.

Bottom line

The dispute over compute is now fought in the software around it. Akamai signed a CPU contract with Anthropic and the Department of Justice requested information on the Nvidia–Groq deal. TorchNeuron aims to let the code AI teams already use run without rewrites on a chip that is not Nvidia’s.

Sources

Edited by Rodrigo Cornejo. How we select and verify: who writes these notes.

Related notes

← All notes