adtestbench Higgsfield

AI news · Google · chips

Google open-sources TorchTPU to bring PyTorch apps to its chips

Developers can move PyTorch code to Google’s TPU chips with a settings change, a day after Amazon did the same for its own Trainium chips.

By Alexandre S. , 05:45 UTC

Google released TorchTPU as free, open-source code on Tuesday 20 October 2026, its cloud blog says, at PyTorch Conference 2026. PyTorch is a free toolkit for building and serving AI models, used by 85 per cent of AI apps by some accounts, and much of it grew up around Nvidia’s graphics chips. TorchTPU lets that same code use Google’s own AI chips, called TPUs, by changing a setting, with the app left as written, TechCrunch reports.

Retro-futurist illustration: a striped sunset over a grid horizon under a starry sky, with a chip motif standing on the horizon.
Drawn by adtestbench from “TorchTPU is now open source: Run native PyTorch on TPUs”,

Google built it together with Meta’s PyTorch team. Partners include the makers of SGLang and vLLM, two popular engines that serve AI models to users, and Hugging Face, whose open models now load straight onto TPUs. Alongside the release, Google launched a TPU agent suite that does the move for you and repairs what breaks, SiliconANGLE reports.

The launch came a day after Amazon released TorchNeuron, which does the same job for its Trainium chips. Both target the same habit: AI apps are written for Nvidia hardware first, which keeps customers on Nvidia.

For teams paying for AI computing, two of the biggest clouds now take PyTorch code as it stands, which gives you more room to shop on price.