Meta's AI Models Now Run on Google Chips, Bypassing Nvidia

The long-standing software barrier between Meta’s AI operations and Google’s custom silicon has officially collapsed. For years, Meta engineers have relied on Nvidia’s GPU ecosystem for their PyTorch models, but a new development announced this week places Google’s Tensor Processing Units (TPUs) into active production at the social media giant. This move represents the first time Google’s TPUs are serving external customers at scale for production AI workloads, fundamentally altering the power dynamics in the data center market.

The catalyst for this shift is a new, PyTorch-native version of the SGLang inference engine. Previously, running these models on TPUs required using Google’s JAX framework, a tool that Meta did not want to adopt. The new SGLang-TorchTPU integration allows the hardware to function as a native option for PyTorch without requiring extensive code rewrites. Alongside this, an open-source version of vLLM, another major inference engine, was also released with the same native compatibility.

The performance metrics accompanying this release are substantial. Testing by Inferact, a startup founded by the creators of vLLM, showed that a tuned vLLM setup serving Moonshot AI’s 2.8-trillion-parameter Kimi K3 model achieved speeds exceeding 700 tokens per second for a single user. This test utilized 16 seventh-generation Ironwood TPUs with speculative decoding. In comparison, the same model running on 16 Nvidia GB200 chips reached 452 tokens per second. At smaller batch sizes without speculative decoding, the new TPU kernels delivered roughly 1.4 to 2 times the throughput of the Nvidia baseline.

Zongwei Zhou, Meta’s PyTorch-TPU lead, confirmed that the system is currently live. “SGLang-TorchTPU is running in production at Meta today,” Zhou stated, noting the effort was the result of a “close three-way collaboration between Meta, Google, and the SGLang team.” Zhou’s team operates at RadixArk, the startup behind SGLang, which co-developed the engine with Google and Meta. He described the new stack as “a remarkably mature, PyTorch-native serving stack.”

Jana van Greunen, director of PyTorch engineering at Meta, emphasized the strategic importance of this flexibility. In a post on Google’s developer blog, she noted that Meta “cannot be constrained by a single hardware architecture.” By enabling the deployment of PyTorch workloads directly on Google TPUs, Meta gains a critical alternative to its previous reliance on Nvidia’s software and hardware lock-in. Google confirmed it co-developed this TorchTPU solution specifically with Meta’s PyTorch team.

This is not Meta’s first foray into Google’s hardware, though previous attempts were less direct. Google identifies the older PyTorch/XLA bridge as a “predecessor” to the new TorchTPU. Reports from SemiAnalysis indicated that Meta research teams used TPUs via XLA from 2020 to 2023 before those contracts were canceled. The current push appears to be a more robust, long-term integration aimed squarely at Meta’s refusal to migrate its core operations to the JAX framework.

While the exact location of Meta’s production TPUs remains unspecified, industry reports from last year suggested that Google Cloud TPU rentals were expected in 2026, with potential purchases following in 2027. Meta subsequently signed a multi-year rental agreement, laying the groundwork for the production rollout now underway.

The timing of this announcement is notable. It came just one day after Amazon announced it was open-sourcing TorchNeuron, its own native PyTorch backend for its Trainium2 and Trainium3 chips, at the same PyTorch Conference. Both moves appear designed to loosen Nvidia’s grip on the AI software stack, offering developers viable, high-performance alternatives on competing silicon.

Looking ahead, the SGLang-TorchTPU ecosystem will continue to expand. The team plans to add support for models including DeepSeek V4.1 Flash and Reflection Beam. RadixArk aims to provide “Day 0” TPU support for new open models, ensuring compatibility on release day by the first quarter of 2027. Google, Meta, and RadixArk will share further details at the SGLang Summit 2026, scheduled for November 12-13 in San Francisco.

This development signals a maturing market where software compatibility, not just raw compute power, is becoming the primary battleground. By making its chips native to the dominant AI framework, Google is challenging Nvidia’s dominance with a solution that is already proven in one of the industry’s most demanding production environments.