Meta puts Google chips to work in its data centers

Meta is now running its production AI workloads on Google's tensor processing units (TPUs), marking the first time these chips have been sold to an external customer for this purpose. This development, detailed in a Google developer blog post, is a significant step for Google's custom AI hardware. The move could alter the competitive landscape for AI processing power, which has been dominated by Nvidia. By providing a way for Meta to run AI models on its TPUs, Google is offering a compelling alternative to Nvidia's GPUs.

The core of this change is a new version of the SGLang inference engine, designed to work natively with PyTorch, the AI framework created by Meta. Previously, to use Google's TPUs, developers had to use Google's own JAX framework. The new PyTorch-native version of SGLang allows for the use of TPUs without needing to rewrite existing code. A similar open-source version of another inference engine, vLLM, was also announced.

This collaboration represents a new chapter for Meta and Google's relationship regarding AI hardware. Zongwei Zhou, Meta’s PyTorch-TPU lead, stated that the new SGLang-TorchTPU is “running in production at Meta today.” This follows a period of close collaboration among Meta, Google, and the SGLang team. Jana van Greunen, director of PyTorch engineering at Meta, emphasized the strategic importance of this development, saying that Meta “cannot be constrained by a single hardware architecture.”

It is important to note that this is not Meta's first attempt to utilize Google's TPUs. Google's older bridge from PyTorch to TPUs, a tool called PyTorch/XLA, was used by Meta research teams from 2020 to 2023 before those contracts were canceled. The current push by Google to support PyTorch on TPUs was aimed specifically at Meta, which was unwilling to switch to Google's JAX framework. This history highlights the ongoing efforts to create a more flexible and compatible system for running AI models on different types of hardware.

The location of Meta’s production TPUs has not been specified. However, discussions last year reportedly included the possibility of renting Google Cloud TPUs in 2026 and purchasing them in 2027. It was later reported that Meta signed a multi-year rental agreement. This arrangement would allow Meta to expand its use of TPUs without the significant upfront cost of purchasing the hardware.

This move by Google and Meta follows a similar announcement from Amazon, which open-sourced its own native PyTorch backend for its Trainium chips. Both initiatives are designed to loosen Nvidia's control over the market by making it easier to run AI models on different types of processors. The development of these tools is a direct response to the growing demand for alternatives to Nvidia's GPUs.

In terms of performance, a tuned vLLM setup on Google's seventh-generation Ironwood TPU reportedly achieved speeds of over 700 tokens per second for a single user when serving Moonshot AI’s Kimi K3 model. This figure comes from Inferact, the startup founded by vLLM’s creators, which helped Google build the vLLM-TorchTPU tool. For comparison, a similar setup using 16 Nvidia GB200s reached 452 tokens per second. This suggests that Google's TPUs can offer competitive or even superior performance for certain AI tasks.

Looking ahead, the SGLang-TorchTPU engine will continue to add support for more models, including DeepSeek V4.1 Flash and Reflection Beam. RadixArk, the startup founded by SGLang’s creators, is working to provide support for new open models on their release day by the first quarter of 2027. Google, Meta, and RadixArk plan to share more details at the SGLang Summit 2026, which is scheduled for November 12-13 in San Francisco.