Google Open-Sources TorchTPU, Joining Amazon's Assault on Nvidia's AI Developer Lock-In

Google open-sourced TorchTPU on Tuesday, giving the roughly 85% of AI applications built on PyTorch a path onto Google's Tensor Processing Units that requires no rewrite — only a configuration change.

The timing is not accidental. Amazon released its own PyTorch variant, TorchNeuron, one day earlier, on Monday. TorchNeuron ports applications to Amazon's Trainium chip and to AWS, again with no rewriting. Two hyperscalers, two days, one target: the default assumption that AI software runs on Nvidia GPUs.

Why does PyTorch matter so much?

PyTorch is the open-source toolkit that most AI developers use to build, train and run models. By some accounts it powers 85% of AI applications. Meta, then Facebook, created it in 2016 — years before large language models reshaped the industry. Nvidia's GPUs were already the standard hardware for AI research at the time, so many of PyTorch's features were built specifically around them.

That inheritance is the moat Nvidia now defends. In 2022, Meta handed PyTorch to the Linux Foundation, the nonprofit that stewards many major open-source projects. The framework is vendor-neutral in governance, if not in habit.

What does each company actually offer?

Google and Amazon have taken different routes to the same destination.

• TorchTPU (Google): runs AI apps on TPUs and moves existing PyTorch apps over with a configuration change, pulling them onto Google Cloud.
• TorchNeuron (Amazon): ports apps to the Trainium chip and AWS with no rewriting.
• Upstream push (Amazon): Amazon is submitting code to the main PyTorch project that decouples features from Nvidia's chips, making the framework more open to any hardware by default.

The upstream move deserves attention. Changing PyTorch itself reaches every user, not just those who adopt a cloud vendor's fork. Amazon is aiming at the foundation; Google is aiming at the front door.

Where might Google hold the edge?

Google says it worked directly with Meta, Hugging Face, Inferact (maker of vLLM), and RadixArk (maker of SGLang) to ensure TorchTPU's features and code libraries work smoothly with these popular development tools, just as the original PyTorch does. Compatibility with the surrounding ecosystem, not merely the core framework, tends to decide whether developers migrate at all.

Google also launched a dedicated tool, the TPU agent suite, to help developers move applications to its chips and cloud and 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. The roadmap covers the edge as well as the data center.

Will developers actually leave Nvidia?

Not automatically, and probably not en masse. Portability is a necessary condition for switching, not a sufficient one. Developers weigh hardware performance, pricing, ecosystem maturity and sheer inertia before they move a workload anywhere.

But the strategic direction is clear. Amazon and Google are working deliberately to end Nvidia's advantage, and they are doing it in software — the same layer where that advantage was originally built. Every application that moves is recurring GPU revenue that does not materialize. Expect Nvidia's response, and further PyTorch contributions from both challengers, before the developer default shifts by any measurable degree.

By Nathan Brooks