The companies introduced TorchTPU, an open-source PyTorch backend for Google’s TPU chips, with reported compatibility for thousands of popular Hugging Face models.
On October 20, 2026, Google and Meta announced TorchTPU, an open-source PyTorch backend for Google Cloud TPUs, at PyTorch Conference North America. It makes TPU a native PyTorch device, aiming to let developers run existing workloads with minimal code changes rather than porting them to a separate framework. The project was tested across about 5,000 Hugging Face models and integrates with Transformers v5, Diffusers, TorchTitan, vLLM and SGLang; in Transformers, `device_map="tpu"` selects the hardware. Google reports TorchTitan training Qwen 3 8B at up to 52% model FLOPs utilization across as many as 8,192 Ironwood TPU v7x chips. That is a company-reported training result, not a direct price or speed comparison with Nvidia GPUs. The release could lower switching costs for teams already using PyTorch and widen Google’s appeal to AI developers, while strengthening competition among accelerator platforms.