infrastructure//data center//power chain//busway

A busway is a rigid enclosed run of copper or aluminium bars that distributes power along a row of racks in a data center, with plug-in points every short distance where a **tap-off unit** takes power down to a rack. It replaces the bundles of individual cables that used to run from a floor PDU under the raised floor to each rack, and it is one stage of the power chain between the switchgear or UPS and the rack PDUs.

A busway is a rigid enclosed run of copper or aluminium bars that distributes power along a row of racks in a data center, with plug-in points every short distance where a tap-off unit takes power down to a rack. It replaces the bundles of individual cables that used to run from a floor PDU under the raised floor to each rack, and it is one stage of the power chain between the switchgear or UPS and the rack PDUs.

The tap-off unit is the part that makes it useful. It clips onto the live busway, contains its own breaker (so a fault in one rack trips only that rack's feed), often meters the current, and ends in a cable or socket for the rack PDU. Adding a rack, or giving one rack more power, means fitting another tap-off at the right spot instead of pulling a new cable through the floor.

Busway is chosen for flexibility and density. A hall that will be reconfigured as hardware generations change (and AI hardware changes the power per rack every couple of years) is cheaper to adapt with an overhead busway than with fixed cabling, and a busway carries far more current in less space than cables, which suits the high power density of GPU rows.

It concentrates risk. Many racks hang from one run, so a fault in the busway itself, or maintenance on it, affects the whole row; that is why rows are fed by two busways from separate paths (A and B), matching the two power supplies in each server (redundancy level).

Breakers in tap-offs, busway and switchgear must be coordinated, so that a short circuit in one rack opens the nearest breaker and not the one feeding the row; this selectivity is calculated when the system is designed and checked when it is changed.