Google is making TorchTPU available as open source. This makes it easier for developers to run AI models built with PyTorch on Google’s own Tensor Processing Units (TPUs).
Google and Meta collaborated to develop the technology. TorchTPU removes a barrier to using alternatives to Nvidia GPUs. Google announced it during the PyTorch Conference 2026. The software ensures that existing PyTorch applications can run on TPUs with minimal modifications. Developers can keep using the programming interfaces and development tools they know.
Until now, using PyTorch on Google’s AI chips usually required an additional software layer, namely PyTorch/XLA. That layer limited how developers can develop, test, and run models. TorchTPU offers more direct support for PyTorch. Among other things, the software lets developers test models interactively before optimizing them for production.
Google and Amazon challenge Nvidia
For Google, this announcement is strategically important. It’s not just GPU development that accounts for Nvidia’s market dominance. Nvidia also owes its strong position in the AI processor market to the extensive CUDA software ecosystem. This makes Nvidia hardware attractive to developers.
Google is not alone in its effort to reduce this dependence. Amazon introduced TorchNeuron earlier this week. This allows PyTorch applications to run on Trainium processors without requiring developers to rewrite the code. According to TechCrunch, Amazon also plans to contribute to the core PyTorch project to reduce the software’s dependence on Nvidia hardware.
As mentioned above, Meta played a key role in developing TorchTPU. The company wants to expand its AI infrastructure without relying on a single-chip architecture. According to Google, Meta is now using the technology for production applications, including running large language models.
Training on thousands of chips
Google reports that TorchTitan with TorchTPU has been tested on up to 8,192 Ironwood TPUs. According to the company, it achieved up to 52 percent model FLOPs utilization while training the Qwen 3 8B model. This metric indicates what portion of the theoretical computing power is actually used to train an AI model.
To simplify the transition, Google is also introducing the TPU Agent Suite. This toolset helps developers migrate applications to TPUs and resolve technical issues. Google has also updated its LiteRT software to better support PyTorch models on devices such as smartphones, laptops, and embedded systems.
The availability of TorchTPU does not automatically mean that Google’s chips are faster or cheaper than Nvidia’s GPUs for every AI application. However, the software does lower the barrier to entry. In doing so, Google and Amazon are trying to reduce one of Nvidia’s key competitive advantages: the tight integration of AI software with its hardware.