Google and Meta open-sourced TorchTPU at the PyTorch Conference in San Jose on October 21, a backend that lets PyTorch models run on Google's Tensor Processing Units with a one-line change, swapping model.to("cuda") for model.to("tpu"). Meta, which created PyTorch in 2016, said it plans to move larger-scale language-model workloads onto TPUs using it. The move makes Google's chips a native target in the dominant AI framework rather than a second-class one.
TorchTPU replaces PyTorch/XLA, a 2020 bridge that the announcement says "required intrusive source-code modifications." Because the new backend plugs into PyTorch's low-level dispatcher, Google says Hugging Face models work with it out of the box, and the vLLM and SGLang inference engines run natively. Google reported near-linear scaling past 8,000 TPU chips and a Qwen3 235B run on 8,192 chips at 28 percent model FLOPs utilization. Four repositories were released: TorchTPU, vLLM-TorchTPU, SGLang-TorchTPU and TorchTitan.
The companies were candid that the one-line change only gets a model running, not running well. Google's own note says a model's shapes, sharding and operations may still be tuned for a GPU architecture, so peak TPU performance takes further work. Its proposed fix is more automation: a TPU agent suite, now in preview on GitHub, that plugs into coding harnesses to handle onboarding, diagnostics and tuning.
Google was not alone in loosening the dependence on a single vendor's chips. AWS put TorchNeuron, a native PyTorch backend for its Trainium chips, into public beta on October 20, with .to('neuron') as the only code change. Both backends register through the same PyTorch extension point. TorchNeuron's first release has gaps: The Next Web reported it does not yet support dynamic input shapes and skips about 22,000 PyTorch tests.
The context is a framework layer trying to break free of hardware lock-in at the exact moment demand is exploding. PyTorch Foundation executive director Mark Collier said inference has overtaken training as the largest data-center workload, creating "an incredible hunger for tokens," and described the goal as software for models that "run on any chip in any cloud." Figures cited at the event capture the pace: a DeepSeek Flash model became 60 percent cheaper in 37 days, and agent token consumption passed human consumption on February 6 and has since grown 14-fold, though that last number comes from a single platform, OpenRouter, and should be treated with care.
The strategic target is unmistakable even though no vendor is named as the adversary. Nvidia's CUDA software stack has long been the reason buyers stay with its GPUs, and this publication reported separately today that Nvidia is now trying to charge about $2,000 per GPU a year for agent software. A credible path to run the same PyTorch code on Google or Amazon silicon chips away at that moat. Independent analysts caution it is not gone: writing for Forkast, Blair Hayes argued the lock-in has "moved rather than disappeared," since the optimization and supply layers stay under each vendor's control. One conference attendee was blunter, saying his company had evaluated Google's TPUs and "decided to stick with Nvidia, at least for one more generation."