Google launches TorchTPU to run PyTorch on TPUs

Google's TorchTPU announcement is aimed at a practical friction point: most AI applications are built on PyTorch, and moving them to new hardware has historically required code changes. By developing TorchTPU with the PyTorch team, Google is offering a path to run those workloads on TPUs with minimal modification, which it says also optimizes throughput per dollar when combined with Kubernetes-native Fleet Routing across TPUs and GPUs. The Hugging Face partnership extends this to popular open-weight models, which can run natively on TorchTPU—lowering the effort for teams already using those models to try TPUs. The body frames this as an effort to compete with NVIDIA's dominant position in AI hardware, since most AI apps are designed by default for NVIDIA GPUs; making it easier to move PyTorch workloads to TPUs, and thereby to Google Cloud, could help Google attract customers. Amazon has made a similar move with TorchNeuron, its own PyTorch version for Trainium chips and AWS, suggesting hardware vendors see software compatibility as a lever for adoption.