Google open-sources TorchTPU, its version of PyTorch for TPU chips

Google announced on October 20 that TorchTPU, its long-awaited version of PyTorch, is now available as an open-source project, TechCrunch reports. The release lets AI applications written in PyTorch run on Google's Tensor Processing Units with a configuration change instead of a rewrite.

Why PyTorch matters here

PyTorch is the dominant open-source toolkit for building, training and running AI models, used by most AI applications. Meta originally developed it in 2016, when Nvidia GPUs were already the standard hardware for AI research, so many of its features were built around Nvidia's chips. Meta handed PyTorch to the Linux Foundation in 2022. That history is the root of Nvidia's current advantage: the default software stack assumes the default hardware.

Amazon moved a day earlier

Google's announcement follows Amazon's release on Monday of TorchNeuron, its own PyTorch variant for running apps on its Trainium chips and AWS without rewriting. Amazon also said it is submitting code to the main PyTorch project that decouples features from Nvidia's chips, which would make the framework more hardware-neutral by default rather than only through vendor forks.

What Google says about compatibility

According to TechCrunch, Google says it worked directly with Meta, Hugging Face, Inferact (the maker of vLLM) and RadixArk (the maker of SGLang) so that TorchTPU's features and libraries behave like the original PyTorch across those tools. Google also launched a TPU agent suite to help developers port apps and fix errors, and updated LiteRT, its on-device AI runtime, to work better with PyTorch apps.

What developers should watch

Easier porting does not mean teams will move off Nvidia hardware in large numbers; performance, capacity, pricing and existing tooling all weigh on real migrations. The practical takeaway is that the switching cost is falling. For a startup, hardware portability is now a realistic evaluation axis rather than a theoretical one: a workload benchmarked on two chip families gives a team negotiating room it did not have before. As with any port, measure end-to-end latency and cost on your own workload before committing.