Google TPUs Enter Meta Data Centers in Landmark Shift for AI Hardware

Google’s custom tensor processing units (TPUs) are now running in production at Meta, marking a significant shift in the AI hardware landscape. This development represents the first instance where Google’s chips have been sold to an external customer for large-scale AI workloads, directly challenging the market dominance currently held by Nvidia. The move signals a break from the industry’s reliance on a single GPU provider, offering major technology firms a credible, high-performance alternative for their inference needs.

The core of this transition is a new, PyTorch-native version of the SGLang inference engine, which allows Google’s hardware to integrate seamlessly with the software ecosystem that Meta and much of the AI community prefer. Previously, running workloads on TPUs required the use of Google’s proprietary JAX framework, a hurdle that often deterred external adoption. By making PyTorch a native option, Google removes the need for extensive code rewriting, allowing existing models to run on TPUs with minimal changes. A corresponding open-source version of vLLM, another major inference engine, was also released to support this broader ecosystem shift.

Meta’s engineering leadership emphasized the strategic importance of this flexibility. Zongwei Zhou, Meta’s PyTorch-TPU lead, confirmed that the SGLang-TorchTPU stack is live in production, describing the result as a “remarkably mature, PyTorch-native serving stack.” He noted that the achievement stems from a close three-way collaboration between Meta, Google, and the SGLang team, which is part of RadixArk, a startup founded by the creators of the SGLang engine. Jana van Greunen, director of PyTorch engineering at Meta, added that the company cannot be constrained by a single hardware architecture and that this new backend enables the direct deployment of PyTorch workloads onto Google’s chips.

This is not Meta’s first interaction with Google’s silicon, but it is the first to reach full production status via a native interface. Google’s previous bridge, PyTorch/XLA, was used by Meta research teams between 2020 and 2023 before contracts were canceled. The new TorchTPU initiative, co-developed with Meta’s PyTorch team, addresses the specific friction points that existed in the older bridge, which was primarily aimed at Meta because the company had resisted migrating its entire stack to JAX.

While the physical location of these production TPUs remains unspecified, recent reports suggest a structured commercial relationship. Negotiations reportedly included Google Cloud TPU rentals for 2026 and potential purchases in 2027, culminating in a multi-year rental agreement. This timeline aligns with the current production rollout, indicating a formalized, long-term partnership rather than a temporary experiment.

The timing of this announcement is notable, as it landed one day after Amazon revealed it was open-sourcing TorchNeuron, a native PyTorch backend for its own Trainium2 and Trainium3 chips. Both moves appear to be part of a broader industry effort to loosen Nvidia’s software lock-in. By making their respective hardware more accessible through the dominant PyTorch framework, Google and Amazon are positioning their chips as viable alternatives for developers who wish to avoid the complexities of proprietary ecosystems.

Performance benchmarks underscore the competitiveness of the new setup. A tuned vLLM configuration serving the 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, provided by Inferact, the startup behind vLLM, outperforms a comparable setup on 16 Nvidia GB200s, which reached 452 tokens per second. Without speculative decoding, the new kernels deliver roughly 1.4 to 2 times the throughput of the GB200 baseline at small batch sizes for models like Kimi K3 and Qwen 3.8 27B.

Looking ahead, the SGLang-TorchTPU ecosystem is set to expand rapidly. The team is prioritizing support for new models, including DeepSeek V4.1 Flash and Reflection Beam. Furthermore, RadixArk has committed to providing “Day 0” TPU support for new open models, meaning compatibility on the day of release, by the first quarter of 2027. Google, Meta, and RadixArk plan to share further details at the SGLang Summit 2026, scheduled for November 12-13 in San Francisco, which will likely serve as a key forum for discussing the future of multi-vendor AI infrastructure.