Google has released TorchTPU, an open source software platform that enables developers to run PyTorch AI models on Google’s TPUs, a move designed to expand the market for AI chips beyond NVIDIA’s dominant GPUs.
Developed in partnership with Meta, TorchTPU allows existing PyTorch applications to run on Google’s specialized AI chips with minimal changes to their code. Google announced the release at the PyTorch Conference 2026, positioning its TPUs as an alternative for companies seeking greater flexibility in AI infrastructure.
The launch supports NVIDIA’s competitors by making their AI chips compatible with the software developers already use. NVIDIA has built a powerful competitive position not only through its next-gen GPUs but also through CUDA, its proprietary software platform for AI development.
PyTorch is one of the most widely used frameworks for AI development, making compatibility with the platform a major competitive advantage for chipmakers. Its close relationship with NVIDIA’s hardware has helped make GPUs the standard choice for AI computing.
TorchTPU is designed to weaken that advantage by allowing developers to retain familiar PyTorch programming tools while deploying models on Google’s chips.
TorchTPU potentially expands options for enterprises when purchasing AI computing capacity. Greater hardware compatibility could also help companies negotiate cloud infrastructure costs and avoid relying on a single processor supplier. But NVIDIA retains an entrenched software ecosystem and a large developer base, advantages that competing chipmakers can’t overcome with a single product release.
Perhaps the best news here for NVIDIA’s competitors is that AI infrastructure is evolving beyond chip performance to the software that determines where AI applications can run.
Diversify Computing Infrastructure
TorchTPU supports PyTorch’s eager mode, which enables developers to test and debug AI models interactively. For production deployments, developers can use torch.compile to optimize performance and reduce memory usage.
Google reported that TorchTPU, working with the TorchTitan training framework, was tested across up to 8,192 Ironwood TPUs. Tests involving the Qwen 3 8B model achieved up to 52% model FLOPs utilization, a measure of how efficiently available processing capacity is used.
While these results demonstrate robust training capabilities, they don’t establish that Google’s chips outperform NVIDIA GPUs on cost or speed.
Meta’s involvement is part of its effort to diversify its AI computing infrastructure. The company says TorchTPU supports key production use cases involving LLMs, advertising and recommendation systems.
Google has also worked with Hugging Face to extend TPU compatibility to widely used AI models. Developers using Transformers v5 can run AI models on Google’s TPUs without making complex changes to their code.
Support also extends to the Diffusers library, which enables generative AI applications, and inference platforms including vLLM and SGLang. These integrations matter because companies need more than chip compatibility to deploy AI models at scale.
To further simplify adoption, Google introduced its TPU Agent Suite in preview, which assists developers with application migration and troubleshooting. The company also introduced LiteRT-Torch, which converts PyTorch models for local execution through LiteRT on smartphones, laptops and embedded devices.
