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

By Marcus Bennett

Google on Tuesday released TorchTPU, its long-awaited version of PyTorch, as an open-source project — a direct strike at the default position Nvidia's GPUs hold across the AI application market. The move puts Google alongside Amazon, which announced its own PyTorch variant, TorchNeuron, on Monday. Both companies target the same structural problem: most AI applications are written by default to run on Nvidia's chips.

The stakes are considerable. PyTorch, the open-source toolkit that developers and startups use to build, train and run AI models, is by some accounts used by 85% of AI applications. Whoever controls how PyTorch interacts with hardware effectively shapes where AI workloads run — and where the associated cloud revenue lands.

Why does PyTorch's history favor Nvidia?

Meta, then Facebook, originally developed PyTorch in 2016 — years before large language models transformed the industry. Nvidia's GPUs were already the standard hardware for AI research at that time, and many of PyTorch's core features were built to work specifically with them. That coupling hardened into an ecosystem advantage over nearly a decade.

In 2022, Meta transferred PyTorch to the Linux Foundation, the nonprofit that manages many major open-source projects. The governance change opened the door for hardware vendors other than Nvidia to integrate more deeply with the framework. TorchTPU and TorchNeuron are the two most consequential results of that opening to date.

What does TorchTPU actually do?

Google's version lets AI applications run on its Tensor Processing Units (TPUs), the company's homegrown AI chips. It also simplifies migration: developers can move existing PyTorch applications to TPUs — and therefore to Google Cloud — with a simple configuration change and no rewriting of their code.

That last detail matters. Rewriting production AI applications for new hardware has historically been the largest barrier to switching chips. Removing it lowers the cost of leaving Nvidia's ecosystem to near zero, at least on paper.

Google says it also worked directly with several influential players in the AI tooling ecosystem:

• Meta, PyTorch's original developer
• Hugging Face, whose model hub is a standard starting point for AI developers
• Inferact, maker of vLLM, a widely used inference engine
• RadixArk, maker of SGLang, another inference framework

The collaborations ensure TorchTPU's features and code libraries work with these popular development tools just as the original PyTorch does, according to Google. Ecosystem compatibility, not raw chip performance, is the battleground here.

What is Amazon doing differently?

Amazon's Monday announcement covered similar ground. TorchNeuron allows AI applications to be ported to Trainium, Amazon's homegrown AI chip, and to AWS with no rewriting — mirroring Google's configuration-change approach.

But Amazon said it is going further. The company is submitting code to the main PyTorch project that decouples various framework features from Nvidia's chips, making PyTorch more open to any hardware by default. That is a structural intervention: rather than offering a fork, Amazon wants the upstream framework itself to stop assuming Nvidia silicon.

The two strategies are complementary rather than identical. Google built partnerships and migration tooling around its fork; Amazon is pushing neutrality into the framework's core.

What tools did Google ship alongside TorchTPU?

Beyond the framework itself, Google launched a dedicated tool called the TPU agent suite. It helps developers move their applications to Google's chips and cloud and fix errors encountered along the way. Migration support of this kind addresses the practical friction that documentation alone never solves.

Google also updated LiteRT, its popular tool for running AI on phones, laptops and embedded devices, to work better with PyTorch applications. The update extends PyTorch compatibility beyond the data center, into the edge devices where Google's software footprint is already large.

Will developers actually leave Nvidia?

The honest answer, on the evidence available: not automatically. None of this week's announcements means AI application developers will ditch Nvidia's hardware en masse simply because moving to alternative chips and clouds is now easier.

But the direction of travel is clear. Two of the three largest cloud providers — Google and Amazon — are visibly working to end Nvidia's advantage, using open-source releases, upstream code contributions, and migration tooling as their instruments. Each PyTorch application that becomes hardware-agnostic by default is one that Nvidia no longer captures through framework lock-in alone.

For Nvidia, the risk is not losing customers this quarter. It is watching the software moat that PyTorch's Nvidia-first history built erode, one commit at a time. For developers and startups, the immediate effect is leverage: 85% of AI applications run on a framework that, as of this week, treats Google's TPUs, Amazon's Trainium and Nvidia's GPUs as interchangeable targets rather than a fixed default.