Google TPUs now run AI workloads in Meta's data centers

Google TPUs are now serving AI in production at Meta via PyTorch-native versions of the SGLang and vLLM inference engines built with RadixArk. On 16 Ironwood chips, vLLM reportedly hit over 700 tokens/s on Kimi K3 versus 452 tokens/s on 16 Nvidia GB200s.

• SGLang-TorchTPU is already running in production at Meta
• vLLM on 16 Ironwood: over 700 tokens/s on Kimi K3
• 16 Nvidia GB200s reached 452 tokens/s in comparison
• Day 0 TPU support for new models targeted for Q1 2027