Google has released TorchTPU, a version of the PyTorch software toolkit that runs on its own AI chips, as a free open-source project, according to TechCrunch.

PyTorch is a toolkit that developers use to build, train and run AI models. TechCrunch said some estimates put it behind 85% of AI apps. Meta first built PyTorch in 2016, when Nvidia's graphics chips were already the standard for AI research, so many of its features were designed around them, the site reported. Meta handed PyTorch to the Linux Foundation in 2022.

TorchTPU lets PyTorch apps run on Google's Tensor Processing Units, or TPUs, and makes it easier to move apps to Google Cloud. TechCrunch said developers can do this with a simple setting change instead of rewriting their apps. A post on the PyTorch blog said the work began at Google in mid-2025 with Meta's PyTorch team. It said porting should be as simple as switching a model from "cuda" to "tpu," though top speed may take more tuning.

Google said it worked with Meta, Hugging Face and the makers of the vLLM and SGLang tools so that TorchTPU works with them, TechCrunch reported. Google also released a set of AI agents to help developers move their apps and fix errors.

The launch came a day after Amazon released TorchNeuron, which moves PyTorch apps to Amazon's Trainium chips, according to TechCrunch. Amazon said it is also sending code to the main PyTorch project to loosen its ties to Nvidia hardware.

TechCrunch said none of this means developers will drop Nvidia's chips in large numbers.

What we know: TorchTPU is now open source and is meant to let PyTorch apps run on Google's chips with small changes.
What we don't: How many developers will switch. Only TechCrunch has reported this so far; NewsHabit hasn't seen independent confirmation.