Google and Meta have released TorchTPU, an open-source backend that runs PyTorch natively on Google’s TPU chips, and the authors say a basic port can mean swapping one device name in your code. The PyTorch blog described the project on 20 October. Most AI teams write PyTorch, and the cost of rewriting that code has long been a quiet reason to stay on Nvidia GPUs and its CUDA software. TorchTPU goes after that cost.

Google first shipped a PyTorch route to TPUs in 2020, called PyTorch/XLA. The PyTorch blog admits it fell behind: developers had to edit their source to mark compilable regions, compile times were severe, and it had no support for PyTorch’s normal step-by-step eager mode. TorchTPU began at Google in mid 2025 with Meta’s PyTorch team. It carries no JAX dependency at runtime, so PyTorch users no longer have to think in Google’s other framework.

The compatibility claims are specific. The team says it supports about 450 core operations and has validated out-of-box runs for the 5,000 top-downloaded models on Hugging Face. It also reports working integrations with vLLM, SGLang, TorchTitan, Ray, and TorchRec. According to the post, some of Google Cloud’s largest customers already use it in production for inference and training.

The performance numbers come from Google’s own tests. On a FLUX image model, training step time fell from 11.49 seconds to 2.78 seconds after a series of changes, with full torch.compile alone reaching 3.02 seconds. Llama 3 405B ran on 1,024 TPU v7x chips at 500 tokens per second per chip with better than 50 percent hardware utilization, and a 235-billion-parameter Qwen3 model reached 28 percent on 8,192 chips. On a recommendation model across 32 chips, TorchTPU matched Google’s JAX version at 18.3 milliseconds per step.

None of those figures compares against Nvidia hardware. The post also says plainly that changing the device name only gets a model running. Peak speed needs changes, because a model’s shapes and sharding may be tuned for a GPU’s design. A TPU can also serve only one process at a time, so reinforcement-learning setups that share a chip do not work out of the box yet.

CUDA’s hold has rested on two things: the code teams already have, and the tooling around it. TorchTPU attacks the first by making the initial port cheap and the second by plugging into tools PyTorch teams already run. It does not erase the tuning work, and Nvidia’s lead in mature kernels is untouched by this post. What changes is the price of finding out. A buyer can now test a TPU against real code instead of a rewrite estimate.

Google and Meta say the next six to 12 months go to production hardening, lower-precision FP8 and FP4 support, and reinforcement-learning loops. They plan to open-source four repositories at the October 2026 PyTorch Conference. Google also worked with Meta’s PyTorch team to upstream a TPU backend into TorchRec, the recommendation-model library.

Any team renting GPUs by the year should run its own PyTorch workload on a TPU instance before the next renewal, because the benchmark that decides the price is now cheap to produce.

Reported by the PyTorch blog on 20 October 2026.