Google Open-Sources TorchTPU for 85% of AI Apps as It Targets Nvidia's Grip
Updated
Updated · TechCrunch · Oct 20
Google Open-Sources TorchTPU for 85% of AI Apps as It Targets Nvidia's Grip
3 articles · Updated · TechCrunch · Oct 20
Summary
TorchTPU became open source on Tuesday, giving PyTorch developers a way to run existing AI applications on Google TPUs with a simple configuration change instead of rewriting code.
PyTorch underpins about 85% of AI apps and was built around Nvidia GPUs, so Google is using compatibility with that dominant framework to steer workloads toward TPUs and Google Cloud.
Google said it worked with Meta, Hugging Face, Inferact and RadixArk so TorchTPU integrates with widely used tools, and added a TPU agent suite to help developers migrate and fix errors.
LiteRT was also updated to better support PyTorch-based apps on phones, laptops and embedded devices, extending Google's push beyond cloud training and inference.
Amazon made a similar move a day earlier with TorchNeuron for Trainium, underscoring a broader effort by major cloud providers to weaken Nvidia's default position in AI development.
Could Google's TorchTPU finally break Nvidia's iron grip on AI, or is the legacy of CUDA too powerful to overcome?
As native PyTorch hits TPUs at massive scale, are we witnessing the beginning of the end for hardware-locked AI development?
Breaking CUDA Lock-In: TorchTPU’s Native PyTorch Backend Delivers 50%+ Lower AI Training Costs on Google TPUs
Overview
On October 20, 2026, Google and Meta launched TorchTPU, an open-source PyTorch backend that lets developers run standard PyTorch code on TPUs without rewriting for JAX. Built on the PrivateUse1 framework, TorchTPU introduces features like DeferAndFuse Eager mode, which boosts performance by grouping operations efficiently. This innovation, combined with cost advantages of TPU hardware over Nvidia GPUs, has shifted industry dynamics—OpenAI even used the threat of switching to TPUs to negotiate better Nvidia pricing. TorchTPU’s native integration and improved developer experience mark a major step toward hardware-neutral AI infrastructure.