infrastructure//data center//power chain//server power supply

A server power supply is the converter inside a server that turns the alternating current from the rack PDU into the low, steady direct-current voltages the electronics need, and with the voltage regulators on the board it forms the last two stages of the power chain. The **PSU** (power supply unit) rectifies and converts AC to a DC rail, today usually 12 V or, in newer AI racks, 48 V or higher; the **VRM** (voltage regulator module) next to the CPU or GPU steps that rail down to the core voltage, around one volt, and holds it there.

A server power supply is the converter inside a server that turns the alternating current from the rack PDU into the low, steady direct-current voltages the electronics need, and with the voltage regulators on the board it forms the last two stages of the power chain. The PSU (power supply unit) rectifies and converts AC to a DC rail, today usually 12 V or, in newer AI racks, 48 V or higher; the VRM (voltage regulator module) next to the CPU or GPU steps that rail down to the core voltage, around one volt, and holds it there.

The VRM has the hardest job in the chain. A GPU can jump from idle to hundreds of amperes in microseconds when a kernel starts, and its core voltage must stay within a few tens of millivolts while the current changes; a sag resets or corrupts the chip, an overshoot can damage it. So the VRM is a fast multiphase switching converter surrounded by capacitors that supply the first microseconds of each load step before the regulator's control loop responds. At one volt and hundreds of amperes, even the resistance of the board's copper matters, which is why the regulators sit right beside the chip.

Every stage is a converter, and every converter loses a little.

A PSU at 94 % to 96 % efficiency (the 80 PLUS Titanium band at 230 V, best near half load) and a VRM around 90 % mean several percent of the server's power becomes heat before any computation, on top of the losses upstream in UPS and transformers. Fewer conversions and higher distribution voltages (48 V in the rack instead of 12 V) are how designers claw it back, since the loss in a conductor grows with the square of the current and four times the voltage carries the same power with a quarter of it (Joule heating).

Servers in a data center are dual-corded: two PSUs, each fed from a different PDU and path, each able to carry the whole server alone, so a PSU, a PDU or a whole path can fail without stopping the machine (redundancy level).

A PSU also bridges short gaps. Its bulk capacitors hold the output for roughly ten to twenty milliseconds after the input disappears, which is the margin a line-interactive UPS uses when it switches to battery.