AI Data Center Networking

Overview

AI data center networking refers to the interconnect fabric architectures that connect GPUs and accelerators within training and inference clusters. The dominant paradigm divides into three domains: scale-out (rack-to-rack fabric connecting thousands of GPUs over optical links, already dominated by optics), scale-up (GPU-to-GPU connections within a server or rack, historically copper but increasingly limited by signal integrity at high lane speeds), and a hybrid "scale-across" topology emerging as cluster sizes expand.

Optical transceivers are critical to scale-out, and scale-up is approaching copper limits as lane rates reach 112G and move toward 224G per lane. Global data center electricity consumption by accelerated servers is projected to grow exponentially from 2025 toward 2030 according to IEA data, making energy efficiency — the key tradeoff parameter across DSP, LPO, LRO, and CPO transceiver architectures — central to AI infrastructure planning.