Meta Activates Google AI Chips for Major Workloads

Major AI operators in the Pacific Northwest and beyond now have a new hardware option for running large language models. Meta has officially brought Google’s custom tensor processing units into its production environment. This move allows the company to run its core AI workloads on silicon that is not manufactured by Nvidia, which has dominated the market for years. For developers and enterprises, this signals that the software barrier to using non-Nvidia chips has significantly lowered.

The technical change centers on two popular inference engines, SGLang and vLLM. Previously, these tools required users to adopt Google’s JAX framework to utilize TPUs. The new versions, announced on Google’s developer blog, are native to PyTorch. This means engineers can deploy models without rewriting their existing code. A tuned vLLM setup serving the 2.8-trillion-parameter Kimi K3 model reportedly exceeds 700 tokens per second for a single user on seventh-generation Ironwood TPUs. This outperforms a comparable setup on Nvidia GB200 chips, which reached 452 tokens per second.

Zongwei Zhou, Meta’s PyTorch-TPU lead, confirmed that SGLang-TorchTPU is running in production at Meta today. He described the result as a remarkably mature, PyTorch-native serving stack. This work is the product of a close three-way collaboration between Meta, Google, and the SGLang team. That team operates at RadixArk, a startup founded by the creators of the SGLang engine. Jana van Greunen, director of PyTorch engineering at Meta, emphasized that the company cannot be constrained by a single hardware architecture.

This is not Meta’s first interaction with Google silicon. The company used an older bridge called PyTorch/XLA from 2020 to 2023 before canceling those contracts. Google’s new push was aimed specifically at Meta, which preferred to stay within the PyTorch ecosystem rather than migrate to JAX. While the exact location of Meta’s production TPUs is not public, reports indicate that talks last year included Google Cloud rentals for 2026 and purchases for 2027. Meta later reportedly signed a multi-year rental agreement to secure this capacity.

The announcement arrived just one day after Amazon open-sourced TorchNeuron. That is a native PyTorch backend for its Trainium2 and Trainium3 chips. Both moves appear designed to loosen Nvidia’s software lock-in. By making their hardware more accessible to the dominant PyTorch community, both Google and Amazon are challenging the status quo. This could lead to more competitive pricing and options for AI developers in the coming years.

SGLang-TorchTPU will continue to expand its model support. The first additions include DeepSeek V4.1 Flash and Reflection Beam. RadixArk says it is working toward Day 0 TPU support for new open models by the first quarter of 2027. This means new models will be supported on the day they are released. Google, Meta, and RadixArk plan to share more details at the SGLang Summit 2026, scheduled for November 12 and 13 in San Francisco.