Google unveils open‑source TorchTPU to shift PyTorch apps from Nvidia GPUs to its own TPUs

Google has released TorchTPU, an open‑source version of the PyTorch framework that runs on the company’s Tensor Processing Units. The launch puts Google alongside Amazon in efforts to reduce Nvidia’s dominance over AI workloads that typically depend on Nvidia GPUs. TorchTPU allows developers to shift existing PyTorch models to TPUs and Google Cloud simply by changing a configuration, without rewriting code. To aid the transition, Google also unveiled the TPU Agent Suite, a tool that automates migration and helps fix compatibility errors. The suite works together with an updated LiteRT runtime, which now supports PyTorch models on phones, laptops and embedded devices. While the new tools make moving off Nvidia hardware easier, they do not guarantee that developers will abandon Nvidia chips overnight. The move signals a broader industry push to give cloud users more hardware options and to challenge Nvidia’s market advantage. ## Context Nvidia’s GPUs have become the default platform for training and running most AI models, giving the company a strong market position. Google’s TPUs are a separate class of AI accelerators that the cloud provider offers to customers as an alternative. By providing open‑source software that eases migration, Google and Amazon aim to broaden the choice of hardware for AI developers.