Google Takes PyTorch to TPU

Google used 20 October to make a fairly direct attack on one of Nvidia’s quieter advantages in artificial intelligence: not just the chips, but the software habits built around them.

The company said TorchTPU, its version of PyTorch for Google’s Tensor Processing Units, is now available as an open-source project. The useful claim is simple enough to matter: developers can move existing PyTorch applications to Google’s TPUs with what Google describes as a configuration change rather than a rewrite. In cloud terms, that is less a feature than an invitation to leave someone else’s estate.

Why this matters is the denominator. TechCrunch cites estimates that PyTorch is used by 85% of AI applications. Nvidia’s hold on the market has never been only about selling GPUs at handsome prices. It also benefited from years in which popular AI tools were built first, and often best, around Nvidia hardware. Once developers write for the default, the default becomes a moat.

Google is not alone in trying to drain it. TechCrunch reported that Amazon made a similar move with TorchNeuron for its Trainium chips, and said it was also contributing code to the main PyTorch project to make parts of the framework less tied to Nvidia hardware by default. That makes this look less like a one-off product launch and more like a coordinated push by large cloud providers to loosen Nvidia’s grip on the software layer that feeds demand for its chips.

Google’s extra argument is ecosystem polish. It said it worked with Meta, Hugging Face, Inferact, RadixArk and others so TorchTPU would work well with widely used AI development tools. It also launched a TPU agent suite to help developers migrate applications and diagnose errors, and updated LiteRT to better support PyTorch-based AI applications on devices.

None of that proves developers will move in bulk. Porting friction is only one reason Nvidia keeps winning; performance, tooling maturity and customer familiarity still matter, and often matter more than a press release full of “simple” migrations. But Google is addressing the part of Nvidia’s advantage that compounds over time. If PyTorch stops assuming Nvidia first, Nvidia’s hardware lead has to do more of the work on its own.