Meta's Production AI Runs on Google's Chips

700 tokens per second. That is the headline speed metric reported for a single user on Google's seventh-generation Ironwood TPU, a figure that has arrived just as Meta begins running production AI workloads on Google's custom hardware. This development marks the first time Google has sold its tensor processing units to an external customer for such heavy-duty, live operations. The move is widely seen as a direct challenge to Nvidia's long-standing dominance in the data center market.

The timeline of this integration reveals a long and complex history. From 2020 to 2023, Meta research teams utilized TPUs through an older bridge called PyTorch/XLA, but those contracts were eventually canceled. The current push represents a fresh start, driven by Meta's reluctance to migrate its entire stack to Google's JAX framework. Instead, both companies co-developed TorchTPU, a new native bridge that allows PyTorch workloads to run on TPUs without requiring extensive code rewrites. This effort involved a close three-way collaboration between Meta, Google, and the SGLang team from RadixArk, a startup founded by the creators of the SGLang inference engine.

Zongwei Zhou, Meta's PyTorch-TPU lead, confirmed that the SGLang-TorchTPU engine is now running in production at Meta. He described the result as a remarkably mature, PyTorch-native serving stack. Jana van Greunen, director of PyTorch engineering at Meta, emphasized that the company cannot be constrained by a single hardware architecture. She noted that TorchTPU enables Meta to deploy its PyTorch workloads to Google TPUs, providing the flexibility the company has long sought. This strategic alignment suggests that Meta is diversifying its hardware dependencies, a move that could reshape the market dynamics.

The announcement arrived one day after Amazon open-sourced TorchNeuron, its own native PyTorch backend for its Trainium2 and Trainium3 chips. Both moves appear to be aimed at loosening Nvidia's software lock-in. The competitive pressure is evident in the performance benchmarks. A tuned vLLM setup serving the 2.8-trillion-parameter Kimi K3 model reportedly tops 700 tokens per second on the Ironwood TPU. This test used 16 Ironwood chips with speculative decoding. In comparison, vLLM's published setup on 16 Nvidia GB200s reached 452 tokens per second. The new kernels for Kimi K3 and Qwen 3.8 27B deliver roughly 1.4 to 2 times the GB200 baseline's throughput at small batch sizes without speculative decoding.

What happens next will be closely watched. SGLang-TorchTPU plans to keep adding models, starting with DeepSeek V4.1 Flash and Reflection Beam. The team is also working toward Day 0 TPU support for new open models by the first quarter of 2027. Google, Meta, and RadixArk intend to share more details at the SGLang Summit 2026, scheduled for November 12 and 13 in San Francisco. The location of Meta's production TPUs remains unclear, though reports suggest talks included Google Cloud TPU rentals in 2026 with potential purchases in 2027. Meta reportedly signed a multi-year rental agreement, solidifying this new hardware relationship.

This shift is not just about hardware; it is about software compatibility and ecosystem control. By making PyTorch native on TPUs, Google is lowering the barrier to entry for its chips. For Meta, it is a hedge against future pricing or supply issues from its primary partner. The performance figures, while impressive, are just one part of a larger narrative about diversification in the AI infrastructure space. As these new engines roll out, the question remains whether other major players will follow suit or stick with the established Nvidia ecosystem. The coming months will provide the first real-world data on the reliability and cost-effectiveness of this new setup.