Meta has integrated Google’s TPU hardware into its production environment using the new TorchTPU framework. This development follows a collaboration between Meta, Google, and the startup RadixArk to enable native PyTorch support on Google’s AI accelerators without requiring code rewrites.
The new TorchTPU stack succeeds the PyTorch/XLA bridge, which Meta utilized for research workloads between 2020 and 2023. Reports indicate that Meta has entered into a multi-year rental agreement for Google Cloud TPU capacity, with potential hardware purchases planned for 2027.
Performance testing by the startup Inferact showed that a configuration using 16 seventh-generation Ironwood TPU chips achieved over 700 tokens per second when running the 2.8-trillion-parameter Kimi K3 model. This reportedly outperformed a baseline setup of 16 Nvidia GB200 chips, which reached 452 tokens per second.
The initiative aims to reduce reliance on Nvidia’s software ecosystem, mirroring similar moves by Amazon, which recently open-sourced its TorchNeuron backend for Trainium chips. RadixArk plans to expand model support, including DeepSeek V4.1 Flash, with full day-zero support for new open models targeted by the first quarter of 2027.
