Google TPUs Go Live Inside Meta's Data Centers

Google's custom artificial intelligence processors are now operating in production environments at Meta, marking the first time the tech giant has sold these tensor processing units to outside customers. The deployment allows Meta to run AI workloads directly on Google silicon, potentially disrupting Nvidia's current grip on the AI hardware market.

The integration was made possible by a new PyTorch-native version of the SGLang inference engine, announced in a Google developer blog post. Previously, both SGLang and vLLM required Google's own JAX framework to operate on tensor processing units (TPUs). The updated software stack enables Meta to utilize this hardware natively without needing to rewrite its existing codebases.

The timeline for this deployment has been marked by years of shifting strategies and changing contracts. According to reports from SemiAnalysis, Meta utilized an older bridge known as PyTorch/XLA to access TPUs between 2020 and 2023. However, Meta ultimately canceled those contracts. Reports from last year indicated that negotiations were underway regarding rental agreements for 2026 with potential purchases in 2027, though the specific location of these production TPUs remains unclear.

Discussions regarding hardware procurement were reportedly finalized through a multi-year rental agreement. Following this, Google, Meta, and RadixArk, the startup formed by SGLang's creators, collaborated on the development of the new engine. Zongwei Zhou, who leads PyTorch-TPU efforts at Meta, stated that SGLang-TorchTPU is running in production at Meta today after a close three-way collaboration between Meta, Google, and the SGLang team.

Jana van Greunen, director of PyTorch engineering at Meta, emphasized the strategic necessity of hardware diversity, stating that Meta cannot be constrained by a single hardware architecture. Google confirmed that it co-developed the TorchTPU backend with Meta's PyTorch team to facilitate this deployment.

This development follows a similar announcement by Amazon, which released TorchNeuron, its own native PyTorch backend for Trainium2 and Trainium3 chips, just one day prior. Both moves signal a concerted effort by major cloud providers to loosen Nvidia's software lock-in by making their proprietary hardware accessible through widely used open-source frameworks.

Performance metrics suggest these chips are highly competitive. A tuned vLLM setup on Google's seventh-generation Ironwood TPU reportedly achieved speeds exceeding 700 tokens per second for a single user while serving Moonshot AI’s 2.8-trillion-parameter Kimi K3 model. This figure, provided by Inferact, outpaced the 452 tokens per second achieved by Nvidia's GB200 hardware under a comparable configuration. Inferact also noted that its kernels for specific models deliver roughly 1.4 to 2 times the GB200 baseline’s throughput at small batch sizes without speculative decoding.

Looking ahead, the software ecosystem for these processors continues to expand. The SGLang-TorchTPU engine is scheduled to add support for DeepSeek V4.1 Flash and Reflection Beam, with plans to achieve Day 0 support for new open models by the first quarter of 2027. Further details regarding these developments are expected to be shared at the SGLang Summit in San Francisco on November 12.