Google

AI Infrastructure · Cloud Computing · Semiconductors

Launch

Google open-sources TorchTPU, a native PyTorch backend for TPUs

Switching AI chips has mostly meant rewriting CUDA-tied code, a cost Nvidia's rivals are now racing to erase.

• Google open-sourced TorchTPU at PyTorch Conference 2026 on October 20, a native PyTorch backend for its TPUs co-developed with Meta's PyTorch engineering team.
• TPUs are Google's custom AI accelerators; TorchTPU makes them a native PyTorch option alongside GPUs and CPUs, working with tools like vLLM, SGLang, Ray, and Hugging Face.
• Developers can debug in eager mode then switch to torch.compile for production performance, and in Hugging Face Transformers v5, deploying to TPUs is as simple as setting device_map="tpu".
• PyTorch reportedly underlies more than 85% of AI applications, which is why controlling its compatibility with non-Nvidia chips is such a high-stakes software fight.
• The launch came one day after Amazon released TorchNeuron, its own PyTorch backend for Trainium2 and Trainium3 chips, showing cloud rivals converging on the same anti-CUDA playbook.
• Meta's Jana van Greunen, Director of PyTorch Engineering, said the company cannot be constrained by a single hardware architecture and that TorchTPU lets it deploy PyTorch workloads to Google TPUs for major production use cases.
• By making it nearly free to port existing PyTorch code to non-Nvidia chips, Google and Amazon are chipping at the software lock-in that has kept Nvidia's margins and market share dominant.