Google open-sources TorchTPU to move PyTorch apps onto its TPU chips

Google has released TorchTPU, its long-awaited PyTorch variant for its own AI chips, as an open-source project. According to TechCrunch, the move makes Google the second major cloud provider in two days to try to loosen Nvidia's grip on AI application development, after Amazon announced a similar effort on Monday.

What TorchTPU does

PyTorch is the open-source toolkit that many AI developers and startups rely on to build, train and run models — TechCrunch cites estimates that it powers roughly 85% of AI applications. TorchTPU allows those applications to run on Google's Tensor Processing Units, or TPUs, instead of Nvidia GPUs. Just as important for adoption, TechCrunch reports that existing PyTorch apps can be moved onto TPUs, and by extension onto Google Cloud, with a simple configuration change and no rewriting.

How PyTorch became tied to Nvidia

PyTorch was originally built by Meta, then Facebook, in 2016 — well before large language models transformed the industry. At the time, Nvidia's GPUs were already the standard hardware for AI research, so many of PyTorch's features were designed to work specifically with them. Meta handed the project to the Linux Foundation in 2022, but the Nvidia-first assumptions remained baked in. That legacy is a large part of why most AI apps today are written by default to run on Nvidia hardware.

Amazon moved a day earlier

Amazon unveiled its own PyTorch version, TorchNeuron, on Monday. According to TechCrunch, TorchNeuron ports AI apps to Amazon's homegrown Trainium chip and to AWS without requiring rewrites. Amazon is also pushing the change upstream: the company is submitting code to the main PyTorch project that decouples various features from Nvidia's chips, making the framework more open to any hardware by default.

Google's additional bets

Google appears to be going further in some respects. TechCrunch reports that it worked directly with Meta, Hugging Face, Inferact (the maker of vLLM), RadixArk (the maker of SGLang) and others to ensure TorchTPU's features and code libraries work well with these popular AI development tools, just as standard PyTorch does.

Google has also launched a dedicated tool, the TPU agent suite, to help developers move their apps onto its chips and cloud and to fix errors along the way. And it updated LiteRT, its tool for running AI on phones, laptops and embedded devices, to work better with PyTorch applications.

Why it matters

Nvidia's dominance in AI is not only about hardware performance; it rests on years of software defaults that point every PyTorch project at Nvidia GPUs. If moving an application to alternative silicon becomes a configuration change rather than an engineering project, the calculus around chip choice — and cloud spend — changes with it. Amazon and Google both operate their own AI chips and have a clear commercial interest in making that hardware a first-class PyTorch target. TechCrunch notes that developers are unlikely to abandon Nvidia hardware en masse simply because alternatives are now easier to reach. But with two of the largest cloud providers actively working to dismantle the default, Nvidia's lock-in faces one of its most coordinated challenges yet.