Google announced on Tuesday that its long-awaited PyTorch version TorchTPU has now been released as an open-source project. With this move, Google joins Amazon in attempting to break NVIDIA's dominant position in the AI industry – currently, most AI applications are defaultly written to run on NVIDIA chips, namely GPU.

PyTorch is a widely popular open-source toolkit that AI application developers and startups use to build, train, and run AI models. According to some estimates, it is used by 85% of AI applications. PyTorch was initially developed by Meta (then known as Facebook) in 2016, a time when no one realized that LLM would change the world. At that time, NVIDIA's GPU had already become the standard hardware for AI research, so many features of PyTorch were specifically designed around it. In 2022, Meta handed over PyTorch to the Linux Foundation, a non-profit organization that manages many important open-source projects.

Google's version TorchTPU allows AI applications to run on Google's AI chips TPU ( Tensor Processing Units ). It also enables existing PyTorch applications to be easily migrated to TPU, and thus to Google Cloud. Developers only need to make simple configuration changes, without having to rewrite the applications.

Amazon also issued a similar announcement on Monday, launching its own PyTorch version TorchNeuron. This TorchNeuron allows AI applications to be ported to Amazon's self-developed AI chips Trainium as well as AWS, without the need for rewriting. Amazon also stated that it has made further progress in its efforts to reduce NVIDIA's control over PyTorch: the company is submitting code to the main PyTorch project, decoupling multiple functions from NVIDIA chips, making them more open by default to any hardware.

However, Google may go a step further than Amazon in some aspects. Google states that it collaborates directly with Meta, Hugging Face, Inferact (the developers of vLLM), RadixArk (the developers of SGLang), etc., to ensure that the features and codebase of TorchTPU can work seamlessly with these popular AI development tools, just like the original PyTorch.

In addition, this search giant has also launched specialized tools TPU, agent, and suite to help developers migrate their applications to Google's chips and cloud services, while fixing errors along the way. Google has even updated its popular LiteRT tool – which is used to run AI on mobile phones, laptops, and embedded devices – to better support PyTorch AI applications.

This does not mean that AI application developers will automatically abandon NVIDIA hardware on a large scale just because it is now easy to migrate to alternative chips and cloud services. However, both Amazon and Google are clearly making efforts to weaken NVIDIA's advantages.