Google Open-Sources TorchTPU Software to Challenge Nvidia's Ecosystem

Google has open-sourced a software stack called TorchTPU that allows developers to run artificial intelligence workloads built using PyTorch on its Tensor Processing Units with minimal code changes.

Google announced the release of TorchTPU at the PyTorch Conference 2026 on October 20, aiming to target a key advantage enjoyed by Nvidia's CUDA software platform.

The move follows a similar announcement by Amazon, which launched an open-source PyTorch backend called TorchNeuron for its Trainium chips to reduce software barriers to alternatives.