Power Capping

Power capping is the practice of limiting the electrical power a server, rack or whole cluster may draw to a fixed ceiling. The cap is enforced in firmware, the operating system or the rack PDU. It lets operators fit more equipment under a contracted or provisioned power limit and avoid tripping breakers when load spikes. In AI clusters, GPU power caps, often set at 70-90% of rated board power, trade a few percent of throughput for a larger cut in energy use, which raises usable compute per megawatt. Colocation providers also use rack-level caps to enforce contracted kW per rack. Capping is monitored through DCIM and BMC telemetry, and it is a common tool when grid connection or cooling capacity, not floor space, is the binding constraint.