Meta Deploys Google AI Chips in Production, Potentially Delivering 700 Tokens Per Second

Google's custom AI chips are now running production workloads at Meta, a significant step for Google's hardware division as it seeks a foothold in the competitive market for AI accelerators. The development, announced in a Google developer blog, means Google's Tensor Processing Units (TPUs) are being used to serve AI models within Meta's data centers. This marks the first time these specific chips have been sold to an external customer for this purpose, a move that directly challenges Nvidia's long-standing dominance in the AI hardware market.

The integration was made possible through the creation of a PyTorch-native version of the SGLang inference engine, allowing Meta's AI software to run on Google's TPUs without significant modification. Previously, running such software on TPUs required the use of Google's own JAX AI framework. The new version, SGLang-TorchTPU, was developed through a three-way collaboration between Google, Meta, and the SGLang team. "SGLang-TorchTPU is running in production at Meta today," said Meta's PyTorch-TPU lead Zongwei Zhou, adding that the engine gives Meta "a remarkably mature, PyTorch-native serving stack." A matching open-source version of another inference engine, vLLM, was also announced.

This is not the first time Meta has explored using Google's TPUs. In the past, research teams at Meta used TPUs through an older software bridge called PyTorch/XLA from 2020 until 2023, when those contracts were canceled. Google's current push is aimed squarely at Meta, which has stated its need to not be constrained to a single hardware architecture. "TorchTPU enables us to deploy our PyTorch workloads to Google TPUs," said Jana van Greunen, director of PyTorch engineering at Meta. While the exact location of these new TPUs is unclear, reports from last year indicated that discussions included renting Google Cloud TPUs in 2026, with potential purchases in 2027. Meta later reportedly signed a multi-year rental agreement for these resources.

On performance, Google's hardware appears to be highly competitive. A test conducted by Inferact, the startup that helped Google build a vLLM-TorchTPU version, showed a tuned setup serving Moonshot AI's 2.8-trillion-parameter Kimi K3 model achieved speeds over 700 tokens per second for a single user. This test ran on 16 of Google's seventh-generation Ironwood chips. For comparison, a similar setup using 16 Nvidia GB200 chips reached 452 tokens per second. The announcement from Google and Meta comes just one day after Amazon revealed its own tool for running PyTorch on its custom AI chips.

Looking ahead, the SGLang-TorchTPU engine will continue to add support for more models, starting with DeepSeek V4.1 Flash and Reflection Beam. The team is also working toward providing support for new open models on their release day, a goal it aims to meet by the first quarter of 2027. Further details on these developments are planned to be shared at the SGLang Summit 2026, scheduled for November 12–13 in San Francisco. This collaboration between Google, Meta, and RadixArk signals a growing effort across the industry to provide alternatives to Nvidia's hardware and software ecosystem.