Google AI Chips Now Power Meta Services

As competition for computing power heats up, Google is delivering custom AI silicon to rival Meta's production systems.

Google's Tensor Processing Units (TPUs) are now handling live AI workloads inside Meta's data centers. This shift is enabled by a new version of the SGLang inference engine designed specifically for PyTorch. Previously, running these chips required the use of Google's JAX framework. The release of a matching open-source version for the vLLM engine signals that these chips can now serve as a native option for PyTorch without needing major code changes.

Google and Meta developed the TorchTPU software together. According to Zongwei Zhou, Meta's PyTorch-TPU lead, the new engine provides Meta with 'a remarkably mature, PyTorch-native serving stack.' Jana van Greunen, Meta's director of PyTorch engineering, noted that the company 'cannot be constrained by a single hardware architecture' and that this tool enables the deployment of workloads to Google TPUs.

This deployment is not the first attempt by Meta to utilize these chips. Meta research teams previously used TPUs through an older bridge from PyTorch between 2020 and 2023 before those contracts were cancelled. Google's push for PyTorch-on-TPU support was driven by the need to accommodate Meta, which did not want to migrate to JAX.

Questions remain regarding the location of Meta's production TPUs. Discussions last year reportedly involved renting Google Cloud TPU capacity in 2026 with potential purchases in 2027. Meta reportedly signed a multi-year rental agreement following those talks.

The announcement arrived one day after Amazon stated it was open-sourcing TorchNeuron, its own native PyTorch backend for its Trainium2 and Trainium3 chips. Both initiatives appear designed to reduce reliance on Nvidia's software lock-in.

Performance reports indicate that a tuned vLLM setup serving Moonshot AI's open-weight, 2.8-trillion-parameter Kimi K3 model exceeds 700 tokens per second for a single user on the seventh-generation Ironwood TPU. That test, conducted by Inferact, ran 16 Ironwood chips with speculative decoding against vLLM’s published setup on 16 Nvidia GB200s, which reached 452 tokens per second.

SGLang-TorchTPU plans to expand support for additional models, beginning with DeepSeek V4.1 Flash and Reflection Beam. RadixArk stated that the engine is working toward Day 0 support for new open models by the first quarter of 2027. Google, Meta, and RadixArk plan to discuss these developments further at SGLang Summit 2026, Nov. 12-13 in San Francisco.