From Server PSU to Silicon

Inside the server the power stack completes its descent, and the numbers get extreme. The two cords at the back of the chassis carry ordinary low-voltage alternating current. The transistors they ultimately feed run at roughly a volt. Between those two points lies a chain of conversions that must shed almost all of the voltage while multiplying the current enormously, and must do it efficiently, because everything lost here becomes heat inside a box that is already fighting to stay cool.

The first stage is the server power supply unit. The PSU rectifies incoming alternating current to direct current and regulates it to an intermediate voltage that the board distributes — historically 12 volts, and increasingly 48 volts in high-density designs. A resilient server has two PSUs, one on each of the independent power feeds, so the loss of an entire upstream path leaves the server running on its surviving supply. The PSU is also a major efficiency checkpoint: modern units convert at well above 90 percent, and industry efficiency programs push the best designs past 96 percent at typical load. Every percentage point here is multiplied across tens of thousands of servers.

From the PSU, direct current travels across the board to the point-of-load converters — the voltage regulator modules, or VRMs, that sit immediately beside the processor. The VRM performs the final, brutal step-down: from the board's distribution voltage to the sub-two-volt level the silicon demands. And here the current explodes. Power is voltage times current, so a processor consuming several hundred watts at roughly one volt draws several hundred amperes; the largest AI accelerators cross a thousand amperes into a single package. Those currents are why the VRM cannot sit anywhere but right next to the chip.

Distance is the enemy at these currents. Voltage drop across a conductor is current times resistance, and with hundreds of amperes flowing, even the milliohms of resistance in a few centimeters of copper trace waste meaningful power and starve the chip of voltage. That physics dictates the entire board layout: the VRM's power stages ring the processor socket as closely as manufacturing allows, and the high-current path from regulator to silicon is made as short and as wide as possible. The final conversion happens at the point of load precisely because the power cannot afford to travel any distance at that voltage and current.

This is why the industry is re-architecting the last meter around 48 volts. Moving the board distribution voltage up from 12 volts to 48 volts cuts the current for the same power by a factor of four, and because resistive loss scales with the square of current, it slashes distribution losses on the board by roughly sixteen times. The same principle that justified high-voltage transmission and medium-voltage campus distribution now reappears at the scale of a circuit board: keep the voltage as high as you can for as long as you can, and step down to the lethal-current final voltage only in the last few millimeters, right at the chip. The power stack is self-similar from the transmission tower to the transistor.

The VRM is also a control system, not just a converter. It must hold the chip's voltage stable while the processor's current demand swings by hundreds of amperes in microseconds as workloads start and stop. It does this with multiple interleaved phases, fast feedback, and banks of capacitors that supply instantaneous current during a transient before the regulator catches up. A VRM that cannot respond fast enough lets the supply voltage sag under a sudden load, which forces designers to run the chip at a higher nominal voltage for safety margin — and higher voltage means more power and more heat. The quality of the final conversion stage directly sets how efficiently the chip itself can be run.

Trace the watt to its destination and the whole stack resolves into one continuous idea. High voltage and low current for distance; a controlled step-down at each boundary; stored energy to bridge disturbances; redundant paths for continuity; and, at the very end, a final conversion pressed right against the load because the currents have become too large to move. The transmission line and the transistor are running the same playbook at wildly different scales. And the total power that transistor is allowed to draw was never set inside the server at all — it was set years earlier, at the gate, by the interconnection the site was granted. The last meter spends the budget the first meter established.