Meta Integrates Google Chips Into Core AI Operations

For businesses and developers relying on large language models, the landscape of available hardware is becoming less rigid. Until now, running AI workloads on Google's custom chips required significant software adaptation or reliance on proprietary frameworks. That barrier has effectively dissolved with the release of new, open source inference engines. These tools allow major AI workloads to run natively on Google's Tensor Processing Units without requiring teams to rewrite their existing codebases.

The core of this development is the announcement of SGLang-TorchTPU and a matching open source version of vLLM. Previously, these engines depended on Google's JAX framework to operate on TPUs. The new versions make the hardware a native option for PyTorch, the dominant framework in the AI community. This shift means that organizations can deploy their PyTorch workloads to Google TPUs with far less friction, treating the chips as a first-class citizen in their infrastructure rather than a secondary alternative.

Meta, the creator of PyTorch, has already moved its production workloads to this new stack. Zongwei Zhou, Meta’s PyTorch-TPU lead, confirmed that the system is currently running in production. He described the effort as a close collaboration between Meta, Google, and the SGLang team. The team operates out of RadixArk, a startup founded by the creators of SGLang. Zhou stated that the engine provides Meta with a remarkably mature serving stack that is fully PyTorch-native.

Jana van Greunen, director of PyTorch engineering at Meta, emphasized the strategic importance of this flexibility. She noted that Meta cannot be constrained by a single hardware architecture. By enabling the deployment of PyTorch workloads to Google TPUs, the company gains a powerful alternative to its existing infrastructure. Google confirmed it co-developed this TorchTPU technology directly with Meta’s PyTorch team to ensure seamless integration.

This is not the first time Meta has explored Google's hardware. Google refers to PyTorch/XLA, an older bridge for connecting PyTorch to TPUs, as a predecessor to the current system. Reports from last year indicated that Meta research teams used TPUs through that older bridge between 2020 and 2023 before canceling those contracts. The recent push appears designed to address Meta's preference for avoiding a full migration to JAX, offering a more direct path for their existing teams.

While the exact location of Meta’s production TPUs remains unclear, industry reports suggest a phased adoption strategy. Discussions last year reportedly involved renting TPU capacity from Google Cloud in 2026, with hardware purchases potentially following in 2027. Meta has reportedly signed a multi-year rental agreement, indicating a long-term commitment to integrating this hardware into its operations.

The announcement also fits into a broader trend of loosening Nvidia's software lock-in. It came just one day after Amazon open-sourced TorchNeuron, its own native PyTorch backend for its Trainium chips. Both moves by major tech giants signal a collective effort to diversify the AI hardware market. By making alternative chips easier to use with standard frameworks, these companies are challenging the dominance of Nvidia's ecosystem.

Performance metrics from early tests are also notable. A tuned setup using vLLM reportedly achieved over 700 tokens per second for a single user on Google's seventh-generation Ironwood TPU. This test, conducted by Inferact, outperformed a comparable setup on Nvidia GB200 chips, which reached 452 tokens per second. The new kernels also deliver significantly higher throughput at small batch sizes compared to the Nvidia baseline.

Looking ahead, the developers plan to expand model support, starting with DeepSeek V4.1 Flash. They aim to offer Day 0 TPU support for new open models by the first quarter of 2027. More details are expected at the SGLang Summit in San Francisco in November, where Google, Meta, and RadixArk will share further insights into this evolving infrastructure.