Hugging Face has released TorchTPU, a native PyTorch backend that enables seamless execution on Google's Tensor Processing Units. Previously, PyTorch developers requiring TPU acceleration had to port their code to torch_xla or learn JAX. TorchTPU eliminates this friction by making TPUs appear as a native device to PyTorch, allowing standard PyTorch code to run directly on Google's custom AI accelerators. The framework integrates with Hugging Face's Transformers and Diffusers libraries, enabling production-grade model inference and training on TPU hardware. Developers can load models and move them to TPU using simple commands like device_map="tpu", with built-in Pallas kernel optimizations and support for both eager execution and torch.compile compilation. The system also supports distributed training across multiple TPU chips, with specialized kernels available through the Kernels Hub. The release significantly expands TPU accessibility to PyTorch's larger developer community and lowers the barrier for teams evaluating Google Cloud's TPU infrastructure. By removing the need for JAX expertise or torch_xla workarounds, TorchTPU strengthens Hugging Face's vision of a universal ML platform and brings TPUs into parity with GPU acceleration in the PyTorch ecosystem.