Alphabet’s Google has released TorchTPU as an open-source project, enabling native support for PyTorch applications on its Tensor Processing Units (TPUs). Announced at the PyTorch Conference on October 20, 2026, the project was co-developed with Meta’s PyTorch team. It is designed to let developers move existing workloads from GPUs to TPUs with minimal code changes.
Nvidia GPUs currently run an estimated 85% of PyTorch workloads. Google aims to reduce Nvidia’s software advantage and capture a larger share of the AI computing market by giving PyTorch developers a direct path to its custom chips.
The initiative follows a similar move by Amazon, which recently announced improved PyTorch support for its own custom chips.