Data Center Power and Busway Infrastructure

Overview

Data center power and busway infrastructure defines the upper bound of GPU density achievable in any given facility. According to the Uptime Intelligence Global Survey 2023, approximately 75–80% of data center racks globally operate under 20 kW per rack — a structural gap relative to AI GPU deployments requiring 55–140 kW per rack. Standard data halls are built with two busway systems (one active, one redundant); Supermicro's Johnson Jia highlighted at OCP Global Summit 2025 that NVIDIA's GB200/300 NVL72 configurations require four busways, a design available only in purpose-built AI data halls.

This infrastructure constraint is a major driver of Supermicro's modular product strategy: the 8U Air-Cooled B300 SuperCluster (576 B300 GPUs across 20 racks at ~55 kW per rack) is specifically designed to deploy within standard two-busway facilities, enabling high-density AI without a purpose-built data hall. Enterprise AI factory rack tiers span from ~31 kW (64×NVIDIA RTX PRO 6000, PCIe, air-cooled) through ~55 kW (32×HGX B300, air-cooled) and ~110 kW (64×B300, liquid-cooled) to ~140 kW (72×B300 NVL72, liquid-cooled). Jia advised customers to start from the available power infrastructure and select the appropriate tier, rather than working forward from a desired GPU count.