Meta Shifts AI Workloads to Google's Chips

Meta is now running its production AI workloads on Google's custom tensor processing units (TPUs), a significant development announced October 22 that marks the first time these chips are being sold for external customer use. This move allows the social media giant to utilize Google's hardware directly for its AI services, a capability enabled by a new version of the SGLang inference engine specifically designed for PyTorch, the AI framework Meta itself develops. The release of a corresponding open-source version of the vLLM inference engine further expands the tooling available for developers looking to use TPUs with PyTorch.

Previously, running PyTorch models on TPUs required using Google's JAX framework, a workaround that Meta was not willing to adopt. The new TorchTPU software stack eliminates this requirement, allowing Meta to deploy its PyTorch workloads to Google TPUs without needing to rewrite its code. According to a statement from Meta's PyTorch-TPU lead, Zongwei Zhou, 'SGLang-TorchTPU is running in production at Meta today' following a collaboration between Meta, Google, and the SGLang team. Jana van Greunen, director of PyTorch engineering at Meta, stated that the company 'cannot be constrained by a single hardware architecture' and that the new software 'enables us to deploy our PyTorch workloads to Google TPUs.'

This is not the first time Meta has attempted to use Google's TPUs. Previous efforts to bridge PyTorch to TPUs were unsuccessful, and Meta canceled those contracts in 2023. While the location of Meta's production TPUs is not specified, reports indicate that discussions last year included renting Google Cloud TPUs in 2026 with potential purchases in 2027, and Meta has since reportedly signed a multi-year rental agreement. The announcement comes just one day after Amazon released its own PyTorch backend for its chips, a move that, like Google's, is aimed at reducing reliance on Nvidia's software ecosystem.

Performance claims suggest the new software stack is highly competitive. A test by Inferact, a startup founded by the creators of vLLM, reported that a tuned vLLM setup serving Moonshot AI's 2.8-trillion-parameter Kimi K3 model can achieve over 700 tokens per second for a single user on Google's seventh-generation Ironwood TPU. For comparison, the same test setup using 16 Nvidia GB200 chips reached 452 tokens per second. The test also showed that the TPU setup delivers significantly higher throughput than the Nvidia baseline for certain models at small batch sizes.

The SGLang team is continuing to add support for more models, with plans to include DeepSeek V4.1 Flash and Reflection Beam. RadixArk, the startup that built the engine with Google and Meta, 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 provide further details on these developments at the upcoming SGLang Summit 2026, scheduled for November 12-13 in San Francisco.