Rack-Level Power Distribution 2026 — 480V Busway, RPP & PDU Sizing for AI Data Centers

A GB200 NVL72 rack draws roughly 120 kW. A legacy enterprise rack was designed for four to eight. That ten-fold jump breaks the two assumptions most data centers were built on — that power arrives as 208 V to a row, and that a rack is happy with a single 30 A feed. Before you rack the GPUs, you have to fix the power path: the busway, the remote power panel (RPP), the PDU, and the redundancy scheme.

This guide walks through rack-level power distribution for AI and GPU compute, from the 480 V busway down to the PDU outlet, so you can size a deployment correctly the first time.

The 480 V Busway and RPP

At the densities AI racks demand, running individual circuits from a distant panel means voltage drop, copper cost, and a wiring nightmare. A 480 V / 277 V three-phase busway runs a high-ampacity trunk down the row, and each rack taps it through a plug-in tap box feeding a local Remote Power Panel (RPP). The RPP steps the voltage down (e.g. 480 V to 208/120 V or to 415/240 V for high-efficiency PSUs) and distributes branch circuits to the racks. This "busway + RPP" pattern is what lets a single row support 100+ kW of compute and rebalance load as racks change.

Modern AI PSUs are most efficient at 415 V / 240 V (common in EMEA/APAC) or 480 V / 277 V (North America) input, which is why 480 V busway with a 277/480 V output tap is increasingly standard for GPU rows — you skip a transformation stage and gain one to two points of PSU efficiency.

Choosing the Right PDU

For GPU racks, a switched or smart PDU with per-outlet metering and remote reboot is worth the premium: it lets you stage power-up to avoid inrush-current trips and remotely cycle a hung node without a truck roll. Locking outlet types (C19 for higher-current gear) are essential in AI racks.

kW-per-Rack Density and Redundancy

Density tiers matter because they dictate everything upstream:

On redundancy, AI racks almost always run A/B dual feeds into each server (two PDUs, two sources), so one feed or PDU can fail without dropping the rack. Whether the upstream is N+1 (one spare module per group) or 2N (a full parallel path) is a budget-and-availability decision: 2N roughly doubles the upstream gear but removes single points of failure. Training clusters that checkpoint frequently often accept N+1; inference serving that must not drop a single request typically demands 2N.

Sizing the CDU for Liquid-Cooled Racks

Once a rack passes ~60 kW, air stops being practical and direct-to-chip liquid cooling takes over. The Coolant Distribution Unit (CDU) is the liquid-cooling analogue of the RPP: it separates the facility loop from the secondary loop that runs to the cold plates, and it manages flow, pressure, and temperature. Size a CDU by thermal load (kW) and flow rate, roughly 1.5 L/min per kW at a 10 °C delta-T. A 120 kW GB200 rack therefore needs on the order of 180 L/min of coolant flow and a CDU rated for its full thermal load with N+1 pump redundancy — plus a leak-detection and shutoff strategy, since a single fitting failure can take down an entire rack.