Google TPUs Enter Meta Production

Google’s custom tensor processing units are officially handling production AI workloads for Meta through a newly released, PyTorch-native version of the SGLang inference engine. The announcement, made via Google’s developer blog, marks the first time these chips have been sold to an external customer in such a direct production capacity. Alongside this, an open-source version of vLLM, another prominent inference engine, was also revealed, designed to run natively on the hardware without requiring significant code rewrites from developers.

For years, running workloads on Google TPUs required the use of JAX, Google’s proprietary AI framework. The new SGLang and vLLM versions change this dynamic by allowing the hardware to serve as a native option for the PyTorch ecosystem. Zongwei Zhou, the PyTorch-TPU lead at Meta, confirmed that the new SGLang-TorchTPU stack is live in production. He described the development as the result of a tight three-way collaboration between Meta, Google, and the SGLang team at RadixArk, the startup founded by the engine’s creators. Zhou noted that the new stack provides Meta with a remarkably mature serving environment tailored specifically for PyTorch.

Jana van Greunen, director of PyTorch engineering at Meta, emphasized the strategic importance of hardware diversity in her comments on the release. She stated that Meta cannot be constrained by a single hardware architecture and that the new TorchTPU tools allow the company to deploy its PyTorch workloads directly onto Google TPUs. Google confirmed that it co-developed this specific PyTorch backend with Meta’s internal engineering teams to ensure seamless integration.

This is not the first time Meta has experimented with Google silicon, but it is the first major production deployment. Google describes the older PyTorch/XLA bridge as a predecessor to the current system. Reports from SemiAnalysis indicated that Meta research teams utilized TPUs through that earlier bridge between 2020 and 2023 before canceling those contracts. The recent push was heavily targeted at Meta, which previously resisted migrating to JAX, suggesting this new native support addresses a long-standing friction point between the two giants.

The physical location of these production TPUs remains undisclosed, though industry talks last year suggested Google Cloud TPU rentals would begin in 2026, with potential purchases following in 2027. Meta was also reported to have signed a multi-year rental agreement. The timing of this announcement is notable as it arrived just one day after Amazon revealed it was open-sourcing TorchNeuron, a native PyTorch backend for its own Trainium2 and Trainium3 chips. Both moves appear calculated to loosen the software lock-in that has long favored Nvidia.

Performance metrics for the new setup are impressive. A tuned vLLM configuration serving Moonshot AI’s 2.8-trillion-parameter Kimi K3 model reportedly exceeded 700 tokens per second for a single user on the seventh-generation Ironwood TPU. This figure comes from Inferact, the startup behind vLLM, which assisted in building the new backend. In a comparative test using 16 Ironwood chips with speculative decoding, this setup outperformed a published vLLM configuration on 16 Nvidia GB200s, which achieved 452 tokens per second. At small batch sizes without speculative decoding, the new kernels for Kimi K3 and Qwen 3.8 27B deliver roughly 1.4 to 2 times the throughput of the GB200 baseline.

Looking ahead, SGLang-TorchTPU will continue expanding its model support, starting with DeepSeek V4.1 Flash and Reflection Beam. RadixArk stated it is working toward Day 0 TPU support for new open models, aiming to have support ready on release day by the first quarter of 2027. Google, Meta, and RadixArk plan to share further details about this ecosystem at the SGLang Summit 2026, scheduled for November 12-13 in San Francisco.