At PyTorch Conference 2026, Google released TorchTPU as open source, a native backend that lets existing PyTorch models run on its TPU chips with minimal changes. The tool was co-developed with Meta's PyTorch team and is supported by Hugging Face, vLLM, SGLang, TorchTitan and Ray. Google also announced llm-d 1.0 for distributed model serving and said its LiteRT on-device runtime is joining the PyTorch ecosystem.

What it means: The hardest barrier for anyone trying to compete with Nvidia is not chips but software. Most developers write in PyTorch, and their tools are tuned for GPUs, so moving to other hardware costs time and carries risk. With TorchTPU, Google is trying to remove that cost, as Amazon did a few days earlier for its own Trainium chips. Working with Meta, the main maintainer of PyTorch, gives the tool a firmer footing than earlier attempts. For companies that train or serve models, the point is choice: if the same code runs on GPUs, TPUs or other accelerators, they can negotiate on price and availability instead of being locked into one supplier. That matters at a time of compute shortages. It remains to be seen how close real workloads, outside Google's own examples, will get to GPU performance. TPUs are also only available through Google Cloud, so the freedom applies to software rather than to where the hardware sits.

Read the original on Google Cloud Blog

Related service: Custom software

Related stories