Google's TorchTPU takes aim at Nvidia's AI chip lead
Google released TorchTPU, an open-source PyTorch variant that lets AI applications run on its Tensor Processing Units. The tool is intended to simplify moving existing PyTorch workloads to Google's chips and cloud without rewriting code. It follows Amazon's similar TorchNeuron effort, as both companies try to reduce Nvidia's dominance in AI hardware.
Google’s Tuesday release makes TorchTPU available as open source. PyTorch, the widely used framework for building and running AI models, originated at Meta in 2016 and moved to the Linux Foundation in 2022; many of its capabilities were shaped around Nvidia GPUs, then the standard AI hardware.
TorchTPU lets existing PyTorch workloads target Google’s TPUs and cloud through configuration rather than code changes. Amazon issued TorchNeuron a day earlier for Trainium and AWS, and is also proposing changes to PyTorch that reduce hardware-specific coupling. Google says it collaborated with Meta, Hugging Face, vLLM and SGLang teams and added a TPU agent suite plus LiteRT updates.
The shift could give AI developers and startups more choice over where their models run, potentially lowering migration costs and reducing reliance on one chip supplier. Cloud customers may see more competitive pricing or tooling as Google and Amazon promote alternatives. Nvidia’s ecosystem advantage may persist, though, because existing code, skills, and performance tuning remain deeply tied to its hardware. The broader effect may be a more varied AI infrastructure market rather than an immediate switch.