On Oct. 20, Google released TorchTPU as open source. It is software that lets PyTorch, a widely used framework for building AI apps, run on Google's TPU chips. Google built it with Meta's PyTorch team. The news is the release, not the idea: Google described the project in an April engineering post. Amazon put out a similar tool for its own chips this week. Google's release post says existing PyTorch workloads run on TPUs "with minimal code changes." TechCrunch wrote it up as "a simple configuration change and no rewriting." Google's own wording is the smaller claim.
Google's April post is more careful. As of then, it said models may need refactoring to run well on a TPU, such as using attention heads of 128 or 256 dimensions instead of 64. It said inputs that change size trigger a recompile, with full support listed as a 2026 goal. Hugging Face, a launch partner, says 96.5% of its tests pass on TPU and that 34 issues are filed, 19 of them fixed. Google's headline number doesn't settle it. In its own testing, training reached 52% model flops utilization (the share of the chip's math capacity put to work) on Qwen 3 8B across up to 8,192 Ironwood TPUs. There is no Nvidia baseline to compare against.
My read: the code moves easily. What it costs once it runs is not shown. Google's documentation describes running TorchTPU on Cloud TPU hardware, and TechCrunch spelled out where that leads: it "makes it easy to move existing PyTorch apps to TPUs and, therefore, to Google Cloud." The lock-in has changed address. Before you price a move, ask Google to run your own workload next to what you pay today, quoted as cost per unit of work. Then ask what it costs to move back.