Google released TorchTPU on Tuesday as an open source project, allowing PyTorch applications to run on Google's Tensor Processing Units and Google Cloud without code changes. The move directly targets Nvidia's dominance in AI infrastructure, where approximately 85 percent of AI applications rely on PyTorch—the Meta-developed toolkit now maintained by the Linux Foundation.
PyTorch's original architecture locked developers into Nvidia's GPUs, the default standard for AI research since PyTorch's 2016 debut. TorchTPU breaks that dependency by letting developers port existing PyTorch workloads to Google's custom silicon through configuration alone.
Amazon executed a parallel strategy on Monday, launching TorchNeuron for its Trainium chip and AWS. The company is also contributing code to the main PyTorch project to reduce Nvidia hardware dependencies across the ecosystem.
Google secured developer adoption through partnerships with Meta, Hugging Face, vLLM, and SGLang—ensuring TorchTPU compatibility with the most widely used AI development tools. The company launched a dedicated TPU agent suite to assist migrations and updated LiteRT to improve performance with PyTorch applications on mobile and embedded devices.
These moves represent a direct assault on Nvidia's ecosystem lock-in. By decoupling PyTorch from Nvidia's hardware at the infrastructure level, Google and Amazon are creating credible paths for developers to shift workloads—and spending—to their respective cloud platforms and custom silicon.
