Google has released TorchTPU as an open-source project, giving AI developers a way to run PyTorch applications on Google’s Tensor Processing Units (TPUs). The launch joins Amazon’s effort to make it easier to move AI workloads away from Nvidia hardware.
Why PyTorch matters to chip makers
PyTorch is an open-source toolkit used to build, train and run AI models. The source article says some accounts estimate that 85% of AI apps use it, making compatibility with the framework a significant consideration for developers choosing hardware and cloud services.
Meta, then known as Facebook, developed PyTorch in 2016. Nvidia GPUs were already the standard hardware for AI research at the time, and many PyTorch features were designed to work specifically with them. Meta transferred the project to the Linux Foundation in 2022.
That history helped make PyTorch a natural fit for Nvidia’s chips. For developers, moving an application to different hardware can involve more than selecting a new cloud provider: software must also work with the new chip and the tools used to build and run the model.
TorchTPU aims to simplify the move
Google says TorchTPU lets AI applications run on its TPUs and makes it easier to move existing PyTorch applications to those chips and Google Cloud. Developers can make the switch with a configuration change, without rewriting their applications.
Google also says it worked with Meta, Hugging Face, Inferact, maker of vLLM, and RadixArk, maker of SGLang, among others. The goal was to make TorchTPU’s features and code libraries work with popular AI development tools in a way that supports the PyTorch ecosystem.
The company has also introduced a TPU agent suite to help developers migrate applications to Google’s chips and cloud and address errors during that process. Its LiteRT tool, which runs AI on phones, laptops and embedded devices, has also been updated to work better with PyTorch AI apps.
Amazon is pursuing a similar path
Amazon announced TorchNeuron on Monday. Like TorchTPU, it is a version of PyTorch designed to let developers port AI apps to the company’s own Trainium chip and AWS without rewriting them.
Amazon has also said it will submit code to the main PyTorch project to decouple various features from Nvidia chips. That work is intended to make the framework more open to different hardware by default, addressing compatibility at the project level as well as through Amazon’s own tools.
Together, the two announcements show how cloud providers are trying to lower the practical effort of using their AI chips. Compatibility tools, migration support and changes to shared software can all reduce friction for developers considering alternatives.
Availability alone may not shift the market
Making applications easier to move does not mean developers will automatically leave Nvidia’s hardware in large numbers. The launch removes one obstacle, but it does not by itself establish that teams will change their existing infrastructure or where they run their workloads.
Google’s announcement is therefore part of a broader effort to challenge Nvidia’s advantage among AI app developers. TorchTPU gives developers a route to Google TPUs, while Amazon’s TorchNeuron and proposed PyTorch changes offer another route toward alternative chips and clouds.