Meta Runs Google's AI Chips in Production for First Time

Google’s custom Tensor Processing Units (TPUs) are now executing live artificial intelligence workloads at Meta, signaling a significant shift in the hardware landscape. This development was confirmed through the release of a new, PyTorch-native version of the SGLang inference engine, which allows Meta to run its models on Google silicon without the need for extensive code rewrites. The move represents the first time Google has sold its TPUs to an external customer for production use, directly competing with the established market share of Nvidia.

The performance implications are substantial. According to Inferact, a startup founded by the creators of the vLLM engine, a tuned setup running the open-weight Kimi K3 model achieved over 700 tokens per second for a single user on Google’s seventh-generation Ironwood TPU. This test utilized 16 Ironwood chips with speculative decoding to outperform a similar configuration on 16 Nvidia GB200 chips, which recorded a speed of 452 tokens per second. Additionally, specific kernels for the Kimi K3 and Qwen 3.8 27B models are reported to deliver between 1.4 and 2 times the throughput of the Nvidia baseline at small batch sizes when speculative decoding is not employed.

Key figures from both companies highlighted the collaborative nature of this project. Zongwei Zhou, Meta’s lead for PyTorch-TPU, stated that SGLang-TorchTPU is currently running in production at Meta. He described the engine as a "remarkably mature, PyTorch-native serving stack" resulting from a close three-way collaboration between Meta, Google, and the SGLang team. Jana van Greunen, director of PyTorch engineering at Meta, emphasized that her team "cannot be constrained by a single hardware architecture." She noted that the new TorchTPU system enables Meta to deploy its existing PyTorch workloads directly onto Google’s hardware.

This is not the first time the two giants have attempted to bridge the gap between their ecosystems. Google previously offered PyTorch/XLA as a solution, but it is now considered a predecessor to the current TorchTPU. Industry reports from last year indicated that Meta research teams utilized TPUs through this older bridge between 2020 and 2023 before those contracts were canceled. The current push appears to be a more robust effort to accommodate Meta’s preference for staying within the PyTorch framework rather than migrating to Google’s own JAX system.

The strategic timing of this announcement is notable. It was released one day after Amazon announced it would open-source TorchNeuron, its own native PyTorch backend for its Trainium2 and Trainium3 chips, at the same PyTorch Conference. Analysts view both moves as concerted efforts to reduce reliance on Nvidia’s proprietary software stack. While the exact location of Meta’s production TPUs remains unspecified, earlier discussions reportedly involved Google Cloud rentals starting in 2026 with hardware purchases potentially beginning in 2027, following a multi-year rental agreement.

Looking ahead, SGLang-TorchTPU is expected to expand its model support, beginning with DeepSeek V4.1 Flash and Reflection Beam. RadixArk, the startup behind the engine, is aiming to provide Day 0 TPU support for new open models by the first quarter of 2027. Google, Meta, and RadixArk plan to share further details regarding these developments at the SGLang Summit 2026, scheduled for November 12 and 13 in San Francisco.