Meta Runs Google TPU Chips in Production for the First Time

700 tokens per second. That is the reported speed achieved by a single user on Google’s seventh-generation Ironwood Tensor Processing Units (TPUs) while running Moonshot AI’s massive 2.8-trillion-parameter Kimi K3 model. This performance figure, highlighted in recent developer announcements, underscores a significant shift in the artificial intelligence hardware landscape. It demonstrates that Google’s custom chips are no longer just an internal tool for the tech giant but are now a viable, high-performance option for external competitors. The speed benchmark was generated by Inferact, the startup founded by the creators of the vLLM inference engine, which assisted Google in building the new integration. The test utilized 16 Ironwood chips with speculative decoding, a technique that uses a smaller draft model to accelerate large language model inference. When compared to a similar setup using 16 Nvidia GB200 chips, which reached 452 tokens per second, the TPU configuration showed a distinct advantage in raw throughput for this specific model.

The catalyst for this shift is a new PyTorch-native version of the SGLang inference engine. Announced on Google’s developer blog, this update, along with a matching open-source version of vLLM, allows these widely used software tools to run natively on TPUs. Previously, using these engines on Google hardware required relying on Google’s proprietary JAX framework. The new versions eliminate that dependency, allowing AI developers to use PyTorch directly on TPUs with minimal code rewriting. This change is significant because PyTorch has become the de facto standard for AI research and development, meaning that removing the JAX requirement lowers the barrier to entry for companies that have already built their infrastructure around PyTorch.

Meta, the creator of PyTorch, is the first major company to adopt this new stack in production. Zongwei Zhou, Meta’s lead for PyTorch-TPU, confirmed that the SGLang-TorchTPU system is currently live. He described the achievement as the result of a close three-way collaboration between Meta, Google, and the SGLang team. Zhou works at RadixArk, a startup founded by the creators of SGLang, which developed the engine in partnership with the two tech giants. Zhou noted that the new setup provides Meta with a remarkably mature serving stack that is fully compatible with their existing PyTorch workflows. This marks a concrete step forward from previous, less integrated attempts to bridge the gap between Meta’s software and Google’s hardware.

Jana van Greunen, director of PyTorch engineering at Meta, emphasized the strategic importance of this move. She stated that Meta cannot afford to be constrained by a single hardware architecture. By enabling the deployment of PyTorch workloads directly onto Google TPUs, Meta gains flexibility in where and how it runs its AI models. Google confirmed that it co-developed the TorchTPU initiative with Meta’s PyTorch team, signaling a deepening partnership that goes beyond simple vendor-client relationships. This collaboration is part of a broader effort by Google to make its TPUs a more attractive option for the wider AI community, rather than just a specialized tool for its own cloud services.

This is not the first time Meta has experimented with Google’s silicon. Google refers to PyTorch/XLA, an older bridge that allowed PyTorch to run on TPUs, as the predecessor to the new TorchTPU system. According to reports from SemiAnalysis, Meta research teams used TPUs through this older bridge between 2020 and 2023 before canceling those contracts. The current push by Google to improve PyTorch support on TPUs was largely aimed at winning back Meta, which had previously shown a preference for staying within the PyTorch ecosystem rather than migrating to JAX. The new native integration addresses the friction points that may have contributed to the earlier cancellation.

While the exact location of Meta’s production TPUs has not been specified, industry talks last year suggested a phased approach. Reports indicated that Google Cloud TPU rentals were discussed for 2026, with potential purchases planned for 2027. Meta is also reported to have signed a multi-year rental agreement, suggesting a long-term commitment to the hardware. This timeline aligns with the current production rollout, indicating that the infrastructure is already in place and being actively utilized for real-world AI workloads.

The timing of this announcement is notable. It came just one day after Amazon revealed that it was open-sourcing TorchNeuron, its own native PyTorch backend for its Trainium2 and Trainium3 chips. Both companies presented their updates at the same PyTorch Conference. These parallel moves appear to be coordinated efforts to loosen Nvidia’s software lock-in. By providing native PyTorch support on their respective custom silicon, both Google and Amazon are giving developers more choices, reducing the dependency on Nvidia’s CUDA ecosystem, which has long dominated the AI hardware market.

Looking ahead, the SGLang-TorchTPU ecosystem plans to expand its model support. The first additions will include DeepSeek V4.1 Flash and Reflection Beam. RadixArk is also working toward Day 0 TPU support for new open models, meaning that support will be available on the same day a new model is released. This goal is targeted for the first quarter of 2027. To share more details about these developments, Google, Meta, and RadixArk are scheduled to present at the SGLang Summit 2026, taking place on November 12 and 13 in San Francisco. This event will likely provide further insights into the growing alliance between these tech giants and the startups driving the next generation of AI infrastructure.