Google has joined Amazon in attacking Nvidia’s stronghold with AI app developers

Google has released TorchTPU, its long-awaited PyTorch version, as an open-source project. The pitch is simple: let developers run AI apps on Google’s Tensor Processing Units instead of Nvidia GPUs, without tearing their code apart first.

That matters because PyTorch has become the default toolkit for a huge chunk of AI work. By some accounts, it powers 85% of AI apps, and the ecosystem was built around Nvidia hardware from the start. PyTorch itself dates back to 2016 at Meta, when the current AI boom was still years away, and in 2022 Meta handed it to the Linux Foundation.

Google says TorchTPU can move existing PyTorch apps to TPUs and, by extension, Google Cloud with a simple configuration change and no rewriting. Amazon made a similar play on Monday with TorchNeuron, which ports apps to its Trainium chips and AWS without rewriting. Amazon also said it is pushing code into the main PyTorch project to make the framework less tied to Nvidia by default.

Google is trying to go a step further. It says it worked with Meta, Hugging Face, Inferact, RadixArk, and others so TorchTPU fits into the tools developers already use. It also launched the TPU agent suite to help with porting and debugging, and updated LiteRT so it plays better with PyTorch apps on phones, laptops, and embedded devices. None of that guarantees developers will flee Nvidia. But the pressure is real, and now there are two giant clouds trying to loosen the same grip.