October 20, 2026 — Google has open-sourced TorchTPU, a backend that lets PyTorch — the framework most AI models are built in — run natively on Google’s TPU chips, with existing code moving over by changing a single device setting. Why it matters: a day after Amazon did the same for its Trainium chips, the two biggest Nvidia challengers are attacking the same weak point — the software lock-in that keeps AI work on Nvidia GPUs.
What happened
Google announced TorchTPU at the PyTorch Conference on October 20, releasing it as open source on GitHub under the Apache 2.0 license. The stack was co-developed with Meta’s PyTorch team. With it, TPUs become a native PyTorch device alongside GPUs and CPUs: developers point their tensors at “tpu” instead of “cuda,” debug in PyTorch’s normal eager mode, and use torch.compile when they want production throughput.
TorchTPU replaces PyTorch/XLA, the older bridge between PyTorch and TPUs, which Google now recommends against for new workloads.
The details
The release spans training, serving, and deployment. It supports roughly 450 of PyTorch’s core tensor operations, plugs into the tools teams already use — vLLM and SGLang for serving, TorchTitan for training, Ray for orchestration, and Hugging Face, where models can be sent to a TPU with a device_map setting — and Google reports that 98 percent of the Hugging Face models it tested ran without crashing.
Scale is the headline number: Google says TorchTPU training reached 52 percent model FLOPs utilization on a Qwen 3 8B model across up to 8,192 Ironwood (TPU v7x) chips, with near-linear scaling. Those are Google’s own reported figures from its announcement, not independent benchmarks.
Meta is the marquee partner. Jana van Greunen, Meta’s director of PyTorch engineering, said in the announcement that TorchTPU lets Meta deploy its PyTorch workloads — including large language models, ads, and recommendation systems — onto Google TPUs and feed feedback directly to both engineering teams, rather than being constrained by a single hardware architecture.
Why it matters
Nvidia’s moat has never been only its chips; it is the decade of software built around them, with PyTorch itself shaped early on by Nvidia hardware. TorchTPU and Amazon’s TorchNeuron — released one day earlier for Trainium — are the cloud giants paying to dissolve that moat: if the same PyTorch code runs well on TPUs or Trainium, switching hardware becomes a pricing decision instead of an engineering project. Neither release guarantees developers will move, and Nvidia’s ecosystem remains far deeper. But when both of Nvidia’s biggest cloud rivals ship native PyTorch support in the same week, the direction of travel is hard to miss.
Frequently asked questions
What is TorchTPU?
TorchTPU is Google’s open-source backend that runs PyTorch natively on Google Cloud TPUs. Co-developed with Meta’s PyTorch team and released under Apache 2.0, it lets existing PyTorch code target a TPU by changing the device to “tpu.”
Does TorchTPU replace PyTorch/XLA?
Yes. Google positions TorchTPU as the production-ready replacement for PyTorch/XLA and recommends it for new TPU workloads.
How well does TorchTPU scale?
Google reports 52 percent model FLOPs utilization training Qwen 3 8B on up to 8,192 Ironwood TPU chips. That figure comes from Google’s own announcement and has not been independently verified.
How is TorchTPU different from Amazon’s TorchNeuron?
They solve the same problem for different chips: TorchNeuron, released October 19, runs PyTorch on Amazon’s Trainium processors; TorchTPU does it for Google’s TPUs. Both aim to loosen Nvidia’s software lock-in on AI development.
Sources: Unite.AI, MarkTechPost, Crypto Briefing, Google Cloud Blog