Google unveils TorchTPU in a push to break Nvidia’s stranglehold on AI compute
SiliconANGLE Mike Wheatley ● Covered by 4 sources
Google made a PyTorch version for its TPUs. It’s a direct shot at Nvidia’s AI chip moat.
Based on reporting by SiliconANGLE, Mike Wheatley — read the original for the full story.
Summary, retelling and take written by AI under human oversight; images are AI-generated illustrations. How we work · Report an error
Google is trying a familiar trick with a new target: make the software easier, then pull the hardware along with it. Its new TorchTPU is a specialized version of PyTorch built for Google’s Tensor Processing Units, with the goal of making it easier to create and run AI models on Google’s chips instead of Nvidia’s GPUs.
That matters because PyTorch is everywhere. The source says more than 85% of AI applications are built with it, and that long-running tie to Nvidia hardware has helped keep Nvidia at the center of AI compute. PyTorch itself dates back to 2016, when Meta built it with Nvidia chips in mind. Meta open-sourced it in 2022 and handed stewardship to the Linux Foundation, but the hardware bias never really went away.
TorchTPU is meant to soften that dependence. Google says developers can adapt existing apps and models for TPUs with simple configuration tweaks, rather than rewriting the underlying code. And it isn’t just about getting code to run. Google says it worked with Meta, Hugging Face, and startups Inferact Inc. and RadixArk Inc., the makers of vLLM and SGLang, to keep TorchTPU compatible with their tools and libraries.
Google also used the launch to show a couple of extra pieces that AWS didn’t mention in its own PyTorch move earlier this week. The TPU agent suite is supposed to help move AI apps to Google Cloud and automatically repair errors during migration. LiteRT, Google’s existing tool for building AI apps that run on smartphones, laptops and embedded devices, has also been updated so it works with PyTorch-based apps. None of this guarantees developers will abandon Nvidia en masse. But it does show Google is planning a longer fight, not a one-day PR burst.
My take — AI-written commentary, not fact-checked reporting
This is the right kind of attack on Nvidia: not louder model hype, but boring compatibility work. That’s how platform lock-in gets broken, one annoying migration step at a time. AWS and Google have both figured out the same thing: if developers can move code without pain, the chip war suddenly gets much less glamorous for Nvidia.
Read more about this at: SiliconANGLE
