Google Open-Sources TorchTPU, Challenging NVIDIA's Dominance in the AI Ecosystem

According to the official Google Cloud blog, Google announced the open-sourcing of TorchTPU at the PyTorch Conference 2026. This project was jointly developed by Google and the Meta PyTorch team, with the code now available on GitHub. TPUs have thus become a native PyTorch hardware option alongside GPUs and CPUs, requiring only minor modifications to migrate existing code. Frameworks such as vLLM, SGLang, TorchTitan, VeRL, and Ray have already integrated support. In Hugging Face Transformers v5, setting `device_map="tpu"` allows models to be deployed on TPUs.