Google Releases TorchTPU: PyTorch Without Nvidia

Google has open-sourced TorchTPU, a version of the popular AI toolkit PyTorch that moves existing AI applications onto Google's own Tensor Processing Unit chips without rewriting any code. Only a configuration change is needed. Alongside it, Google released the TPU Agent Suite, a tool that helps developers migrate applications and fix errors that come up during the move. Google also updated LiteRT, its tool for running AI on phones, laptops, and small embedded devices, so that it works better with PyTorch applications.

PyTorch, the toolkit all of this is built for, is used in roughly 85 percent of AI applications. Google got backing from Meta, the company that created PyTorch, Hugging Face, the platform for publishing AI models, Inferact, the maker of vLLM, a tool for running AI models, and RadixArk, the maker of SGLang, another tool for running AI models. Because of that support, TorchTPU works smoothly with their software. A day earlier, Amazon made a similar move for its own Trainium chips.

This story is not about AI agents writing code. It is about whose chips those agents end up training and running on. If a team already runs its models on PyTorch, there is now a legitimate way to shift workloads from Nvidia to TPUs without rewriting the project. That genuinely cuts dependence on a single hardware vendor. But Google did not share numbers on how much faster or cheaper TPUs are compared to Nvidia chips, so the comparison will have to be made on your own workload.