Google open-sources TorchTPU a day after Amazon releases TorchNeuron for PyTorch

Google announced on Tuesday that TorchTPU, a native PyTorch backend for its Tensor Processing Units, is available as an open-source project. According to TechCrunch, it lets existing PyTorch apps run on TPUs, and therefore on Google Cloud, with a simple configuration change and no rewriting. The GitHub project is licensed under Apache 2.0 and lists version 0.2.0 on PyPI.

The announcement followed Amazon's release on Monday of TorchNeuron, which does the same for its Trainium chip and AWS. Amazon says PyTorch apps can run on Trainium by changing one line of configuration. It also says it is contributing code upstream to PyTorch to remove ties to specific chips, so the framework is more hardware-agnostic by default.

Google says it worked with Meta, Hugging Face, Inferact (maker of vLLM), RadixArk (maker of SGLang) and others so TorchTPU works with popular AI tools. It also launched a TPU agent suite to help developers migrate and fix errors, and updated LiteRT, its tool for running AI on phones, laptops and embedded devices, to work better with PyTorch apps.

Hugging Face published posts on both backends. It says Transformers, Diffusers and Kernels run natively on TPU, with device_map set to tpu. For Trainium, it says 89% of the top 100 text-generation architectures are supported, against 46% with the older Optimum Neuron library. Its post lists TorchNeuron as a public beta from Oct. 20, 2026.

Nvidia's advantage has rested partly on software. TechCrunch notes that Nvidia spent years making CUDA work smoothly with PyTorch, which Amazon says underlies about 85% of AI apps, so most apps default to Nvidia GPUs. PyTorch users could already port to AMD GPUs and Apple chips with a configuration change; Google and Amazon now offer the same for their own silicon.

The pitch from both companies is also about cloud lock-in. AWS argues Trainium gives startups a full cloud and one chip for both training and inference, and says it is a better choice than neoclouds tied to Nvidia hardware. It cites Deepgram, which it says processes 70% more audio per dollar on Trainium. These are Amazon's claims and were not independently verified in the coverage.

TechCrunch cautions that developers will not necessarily abandon Nvidia en masse just because moving is easier. It also asks whether decoupling PyTorch from Nvidia will give developers more choice and lower costs, as Amazon promises, or simply more configuration headaches.

The main open question is adoption: whether startups move production workloads to TPUs or Trainium, and whether the economics hold up outside vendor examples.

Also worth following is Amazon's upstream work in PyTorch, and how quickly Google and Hugging Face close the remaining gaps. Hugging Face says TPU test failures are being tracked as issues on the TorchTPU repo, and it is working toward a nightly TPU CI.