Meta runs Google AI chips in production

Google’s custom silicon is now powering artificial intelligence workloads in production at Meta.

The move was confirmed via a new version of the SGLang inference engine, which has been adapted to work natively with the PyTorch framework. This software update, announced on Google’s developer blog, eliminates the need for extensive code rewriting that previously required the use of Google’s proprietary JAX framework. A parallel open source update for the vLLM engine was also released to achieve similar compatibility. Zongwei Zhou, the PyTorch-TPU lead at Meta, stated that the new system is running in production today following a close collaboration between Meta, Google, and the SGLang team. He described the result as a remarkably mature serving stack. Jana van Greunen, director of PyTorch engineering at Meta, emphasized that the company cannot be constrained by a single hardware architecture. She noted that this new tool enables the deployment of PyTorch workloads directly onto Google TPUs.

This is not the first time Meta has experimented with Google hardware. An older bridge called PyTorch/XLA served 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 current push appears aimed specifically at Meta, which had previously resisted migrating to the JAX framework. While the exact location of the production chips remains unclear, industry talks from last year suggested rentals in 2026 and purchases in 2027. Meta reportedly signed a multi-year rental agreement following those discussions.

The announcement arrived just one day after Amazon revealed it was open-sourcing TorchNeuron. This is a native PyTorch backend for its own Trainium chips. Both companies appear to be working toward loosening the software lock-in associated with Nvidia’s dominant market position. Performance data suggests the new setup is highly competitive. A tuned configuration serving a massive open-weight model reportedly exceeded 700 tokens per second for a single user on seventh-generation hardware. This benchmark was achieved using speculative decoding, a technique that accelerates inference with a smaller draft model. The test outperformed a comparable setup on Nvidia GB200 chips, which reached 452 tokens per second. At small batch sizes, the new kernels deliver roughly 1.4 to 2 times the throughput of the Nvidia baseline without speculative decoding.

The development team plans to expand model support in the coming months. Initial additions will include specific versions of DeepSeek and other models. The goal is to achieve Day 0 support for new open models by the first quarter of 2027. This means the hardware will be ready to run new software on the day of release. Google, Meta, and RadixArk, the startup behind the engine, plan to share further details at the SGLang Summit in San Francisco on November 12 and 13.