Google TPUs Begin Serving Meta's Production AI Workloads

Google's custom AI hardware is now running inside Meta's data centers. This development marks a significant step for Google's chip business and could challenge Nvidia's dominance in the AI market.

The core of this shift is new software. Google has released a PyTorch-native version of its TPU software stack. This allows Meta to run its AI models on Google's tensor processing units (TPUs) for production workloads. Previously, running on TPUs required using Google's own JAX framework. The new software removes this barrier, allowing for a native option with minimal code changes. Meta's PyTorch-TPU lead, Zongwei Zhou, stated that SGLang-TorchTPU is running in production at Meta today. He described it as a remarkably mature, PyTorch-native serving stack. Jana van Greunen, director of PyTorch engineering at Meta, noted that Meta cannot be constrained by a single hardware architecture. This new capability, she said, enables them to deploy their PyTorch workloads to Google TPUs. Google confirmed it co-developed this software with Meta's PyTorch team.

This is not Meta's first attempt to use TPUs. Google's older PyTorch-to-TPU bridge was used by Meta research teams from 2020 to 2023 before those contracts were canceled. The current push is aimed at Meta, which did not want to migrate to the JAX framework. The location of these production TPUs is unclear. Reports from last year indicated discussions about renting Google Cloud TPU capacity in 2026 with potential purchases in 2027. Meta later reportedly signed a multi-year rental agreement for this capacity.

The announcement from Google and Meta comes one day after Amazon announced its own similar software. Amazon is open-sourcing TorchNeuron, a native PyTorch backend for its Trainium2 and Trainium3 chips. Both moves signal a coordinated effort by major cloud providers to loosen Nvidia's software lock-in. They aim to make it easier for customers to use their respective AI chips.

On performance, a tuned vLLM setup on Google's seventh-generation Ironwood TPU reportedly exceeds 700 tokens per second for a single user. This test served Moonshot AI's 2.8-trillion-parameter Kimi K3 model. The comparison was against a vLLM setup on 16 Nvidia GB200s, which reached 452 tokens per second. The performance figure comes from Inferact, a startup that helped Google build the vLLM-TorchTPU software. Its kernels for certain models deliver roughly 1.4 to 2 times the GB200 baseline's throughput at small batch sizes.

Google, Meta, and RadixArk plan to share more details about this work at SGLang Summit 2026. The event is scheduled for Nov. 12-13 in San Francisco. The SGLang-TorchTPU engine will continue adding model support, starting with DeepSeek V4.1 Flash and Reflection Beam. RadixArk also stated it is working toward Day 0 TPU support for new open models by the first quarter of 2027.