Google's Custom Chips Start Running AI Workloads Inside Meta's Data Centers

The dominance of Nvidia in the artificial intelligence hardware market has long been viewed as a near-monopoly, but a recent development suggests that this grip may be loosening as major tech giants seek alternative computing solutions. On October 21, 2026, it was confirmed that Google’s tensor processing units, or TPUs, are now handling production AI workloads for Meta, the parent company of the widely used PyTorch framework.

This integration is made possible through a new, PyTorch-native version of the SGLang inference engine, a tool used to serve large language models efficiently. Previously, running these models on Google’s hardware required developers to use JAX, Google’s proprietary programming framework, which created a barrier for teams invested in the PyTorch ecosystem. The new engine, along with an open-source version of vLLM, allows Meta to run its models on TPUs natively without the significant code rewrites that were previously necessary.

The shift is not merely theoretical but is already active in Meta’s operations. Zongwei Zhou, Meta’s lead for PyTorch-TPU integration, stated that the SGLang-TorchTPU system is running in production today. He described the achievement as the result of a close three-way collaboration between Meta, Google, and the SGLang team. Zhou noted that the new setup provides Meta with a remarkably mature serving stack, allowing them to leverage their existing PyTorch infrastructure on a different hardware architecture.

Meta’s leadership emphasized the strategic importance of hardware diversity. Jana van Greunen, director of PyTorch engineering at Meta, explained that the company cannot be constrained by a single hardware architecture. By enabling the deployment of PyTorch workloads directly on Google TPUs, Meta gains flexibility in where and how it runs its AI models. Google co-developed this TorchTPU solution with Meta’s PyTorch team, signaling a deeper partnership beyond simple hardware sales.

This is not the first time Meta has experimented with Google’s chips, but it is the first time they are doing so at scale through a native PyTorch interface. In the past, Meta used a bridge called PyTorch/XLA to run TPUs from 2020 to 2023, a period during which they eventually canceled those contracts. The previous approach was largely aimed at Meta because the company did not wish to migrate its entire workflow to JAX. The new native support removes that friction, making TPUs a viable, long-term option for Meta’s infrastructure.

While the exact location of these production TPUs has not been disclosed, industry reports from last year suggested that negotiations included rental agreements for Google Cloud TPUs in 2026, with potential purchases planned for 2027. It is reported that Meta later signed a multi-year rental agreement, indicating a commitment to using this hardware for a sustained period rather than a short-term test.

The timing of this announcement is significant, as it came one day after Amazon announced it was open-sourcing TorchNeuron, its own PyTorch backend for its Trainium chips. Both moves appear designed to break Nvidia’s software lock-in by ensuring that the most popular AI framework, PyTorch, runs efficiently on competing hardware. If successful, this could force Nvidia to compete more aggressively on both price and performance to retain its market share.

Performance benchmarks suggest that the new setup is competitive. A tuned version of the vLLM engine reportedly achieves over 700 tokens per second for a single user when serving a massive 2.8-trillion-parameter model on the seventh-generation Ironwood TPU. This figure was tested by Inferact, a startup involved in building the new engine. The test utilized 16 Ironwood chips with speculative decoding, a technique that uses a smaller model to draft responses, outperforming a setup on 16 Nvidia GB200s that reached 452 tokens per second. At smaller batch sizes, the new kernels also deliver 1.4 to 2 times the throughput of the Nvidia baseline.

Looking ahead, the SGLang team plans to add support for new models, starting with DeepSeek V4.1 Flash and Reflection Beam. RadixArk, the startup behind SGLang, aims to provide same-day TPU support for new open models by the first quarter of 2027. Google, Meta, and RadixArk will share further details about their collaboration at the SGLang Summit, scheduled for November 12-13 in San Francisco.