Google made TorchTPU, its version of PyTorch for its Tensor Processing Units, available as open source on 20 October 2026, announcing it at PyTorch Conference. The pitch is that an existing PyTorch application can move to TPUs with a configuration change instead of a rewrite. The day before, Amazon had put out TorchNeuron, a comparable layer for its Trainium chips.
The timing matters because PyTorch, which Meta created in 2016 and handed to the Linux Foundation in 2022, is used by roughly 85% of AI applications by some counts. Its features were long built around Nvidia GPUs, which is the root of Nvidia's software advantage and the lock-in rivals want to loosen. SiliconANGLE and TechCrunch both frame the two releases as an attempt to erode that advantage.
“it remains to be seen how much they can chip away at Nvidia’s strongly entrenched position”
Holger Mueller, analyst, Constellation Research, on the TorchTPU release · siliconangle.com ↗
Google built TorchTPU with Meta's PyTorch team and checked it for compatibility with Hugging Face, vLLM through Inferact and SGLang through RadixArk, the serving tools many teams already use. The code is public under google-pytorch on GitHub with an Apache 2.0 licence, and torch-tpu 0.2.0 reached PyPI on 19 October. Google first showed the project in a developer blog post on 7 April, so this week's release follows a preview of about six months.
Around it sit several smaller releases. A TPU agent suite to help migrate applications and fix errors is in early preview, with no dates given. LiteRT was updated to work with PyTorch apps. TorchTitan training was validated on up to 8,000 Ironwood TPUs, reaching up to 52% model FLOPs utilisation on Qwen 3 8B. And llm-d 1.0 shipped with Red Hat and IBM as a CNCF Sandbox project.
Meta's Jana van Greunen, director of PyTorch engineering, welcomed the chance to deploy PyTorch workloads on TPUs and give engineers direct feedback. Amazon went a step further than Google by contributing code to core PyTorch to decouple features from Nvidia GPUs; Google did not.
Constellation Research's Holger Mueller says rebuilding PyTorch is a good way to ease moves onto Google and Amazon compute and gives developers more choice, but that any dent in Nvidia's position is unproven. Both SiliconANGLE and TechCrunch expect developers to keep using Nvidia rather than leave en masse.
Sources · 3