Analysis
Google released TorchTPU, an open-source version of PyTorch built for its own Tensor Processing Units, the company said Tuesday via a Google Cloud blog post covered by TechCrunch. The move comes one day after Amazon shipped its own PyTorch fork, TorchNeuron, which Pulse covered Monday as a direct challenge to Nvidia's roughly 85% grip on the framework most AI apps are built with.
What's new since that coverage: Google is now making the same kind of move, not just Amazon. TorchTPU lets existing PyTorch applications move to Google's TPUs and Google Cloud through a configuration change rather than a rewrite, and Google says it worked directly with Meta, Hugging Face, and the makers of vLLM and SGLang to keep compatibility with popular AI tooling intact. Google also launched a TPU agent suite to help developers migrate and debug that move, and updated LiteRT, its on-device AI runtime, to work better with PyTorch apps.
“What's new since that coverage: Google is now making the same kind of move, not just Amazon.”
Together, the two moves amount to a two-hyperscaler push against Nvidia's default position inside PyTorch, which Meta built in 2016 and handed to the Linux Foundation in 2022. Amazon's version goes further structurally: it's submitting code upstream to the main PyTorch project that decouples features from Nvidia-specific optimizations, an attempt to change the framework's defaults rather than just offer an alternative alongside it.
However, easier switching doesn't guarantee switching happens. Developers have built years of tooling, debugging habits and performance tuning around Nvidia's CUDA stack, and neither fork removes that inertia overnight -- AMD's ROCm effort has spent years trying to loosen Nvidia's grip with limited success. The real test is whether either hyperscaler's internal AI teams move meaningfully off Nvidia for their own workloads first; outside developers will likely follow proof, not announcements.