Google launches TorchTPU to rival Nvidia's PyTorch grip
Original: Google has joined Amazon in attacking Nvidia’s stronghold with AI app developers
Why This Matters
Both moves directly target Nvidia's default hardware lock-in across the AI developer base.
Google open-sourced TorchTPU on Oct 20, 2026, letting AI developers port PyTorch apps to Google TPUs and Cloud with no code rewrites—one day after Amazon released its own PyTorch fork, TorchNeuron, for AWS Trainium chips.
PyTorch, the AI development toolkit originally built by Meta in 2016, is used in roughly 85% of AI applications. Its features were designed around Nvidia GPUs from the start, giving Nvidia a structural advantage that has persisted even as rival chips have matured. Google's TorchTPU, now available as an open source project, lets developers run existing PyTorch apps on Google's Tensor Processing Units via a simple configuration change—no rewriting required. The release follows Amazon's Monday announcement of TorchNeuron, a PyTorch fork targeting its Trainium chip and AWS infrastructure. Amazon went a step further by contributing code to the main PyTorch project that decouples core features from Nvidia-specific implementations. Google countered with broader ecosystem coordination: it partnered with Meta, Hugging Face, vLLM maker Inferact, and SGLang maker RadixArk to ensure TorchTPU integrates smoothly with popular AI tooling. Google also released a TPU agent suite to help developers migrate and debug apps, and updated LiteRT—its on-device AI runtime—to better support PyTorch workloads on phones, laptops, and embedded hardware.