Google TorchTPU goes open source: native PyTorch on TPUs

Most AI software runs on PyTorch, and most PyTorch runs on Nvidia. That lock-in is why GPU pricing sets the floor on what your AI vendors charge you. On October 20, Google open-sourced TorchTPU, a native PyTorch backend for its TPU chips, built with Meta's PyTorch team. It won't change your stack this week. It does give the market a credible second lane, and a second lane is what puts pressure on price.

What actually happened

Per Google Cloud's announcement at PyTorch Conference 2026, TorchTPU lets existing PyTorch workloads run on TPUs "with minimal code changes." You debug in eager mode, the way PyTorch normally runs, then switch to torch.compile for production speed.

The GitHub repository is licensed Apache 2.0. The current public release is 0.2.0 on PyPI, and it needs Python 3.11 or later.

Google also announced TPU support across the tools teams actually ship with:

• Serving: vLLM, SGLang, and llm-d 1.0
• Training: TorchTitan and Ray, plus a VeRL reinforcement-learning backend that is still in preview
• Hugging Face: in Transformers v5, you pass device_map="tpu" to load a model on a TPU

The performance numbers are Google's own: up to 52% model FLOPs utilization training Qwen 3 8B, with near-linear scaling on up to 8,192 Ironwood TPUs. Meta says it runs LLM, ads, and recommendation workloads on TPUs through TorchTPU. Google's TPU agent suite for AI coding tools is in early preview. The announcement says nothing about TPU pricing.

Why TorchTPU matters for your business

You probably don't rent chips. Your vendors do. If you pay for an AI API, a hosted model, or a chatbot platform, someone's GPU bill is inside your invoice. When the same PyTorch code can run on two chip families, hosting providers get options, and options compress margins. That's the slow-moving win here.

If you do self-host a model, portability is now testable. A team running an open-weights model on vLLM can price the same workload on a TPU instance without a rewrite. Run the benchmark on your real traffic, not Google's chart. Throughput per dollar on your prompts is the only number that counts.

"Minimal code changes" is not "no code changes." Custom CUDA kernels, quantization libraries, and niche dependencies are where ports break. Version 0.2.0 is early software. Keep your production path where it is until the test passes.

Exit options are worth having before you need them. We wrote the same thing about AWS TorchNeuron on Trainium. Now there are two non-Nvidia PyTorch paths backed by hyperscalers. Your architecture should be able to use either.

Sources: Google Cloud, TorchTPU on GitHub.