48v

Datacenter power delivery is the unglamorous logistics problem underneath every AI cluster: taking tens of megawatts from the utility grid and delivering it, at high efficiency, all the way down to the sub-volt rails that feed each GPU die. It is a cascade of voltage conversions, from medium-voltage AC at the property line down to roughly seven-tenths of a volt at the transistor, and every stage loses a little energy as heat. Because an AI rack can now draw over a hundred kilowatts, the design of this chain has become as important to total cost and density as the silicon it feeds.\n\n**Power arrives as medium-voltage AC and is stepped down in stages, each of which costs efficiency.** The utility feed enters at medium voltage, passes through transformers and switchgear, rides through uninterruptible power supplies for backup, and reaches the rack through power distribution units. Every conversion in this chain, AC to DC and one voltage to another, is imperfect, and because the efficiencies multiply rather than add, a few points lost at each of several stages compounds into a large fraction of the incoming power turned to waste heat before it ever reaches a chip.\n\n**The industry moved rack distribution from 12 volts to 48 volts to attack resistive loss directly.** The power lost in a conductor is the square of the current times its resistance, so for a fixed amount of power, raising the distribution voltage by four times cuts the current to a quarter and the loss to a sixteenth. This is why hyperscale racks, following the Open Compute model, distribute 48 volts down a vertical copper busbar fed by a shared power shelf of rectifiers that convert incoming AC to 48-volt DC. Servers simply clip onto the busbar, drawing the current they need without long, lossy cables.\n\n**On the board, that 48 volts is converted down to the core voltage right next to the processor.** The final step-down happens as close to the die as physically possible, because the currents involved are enormous: a GPU pulling a kilowatt at under a volt draws well over a thousand amps. Multiphase buck regulators, the voltage regulator modules, take 48 volts, sometimes through an intermediate bus and sometimes directly, down to the core rail as a dense array of point-of-load converters that surround the package and hand off clean, tightly regulated current with minimal remaining copper to lose it in.\n\n**Reliability and end-to-end efficiency are the two numbers the whole chain is judged on.** Redundancy such as N+1 rectifiers and battery backup units gives the cluster ride-through when the grid stumbles, while bulk capacitance absorbs the violent current transients an AI workload creates as thousands of GPUs step their load in unison. Above all, every percent of power lost in conversion becomes megawatts of heat to remove and a permanent line on the electricity bill, so operators push toward higher distribution voltages and better converters, and pair the effort with the same liquid cooling that the resulting power density demands.\n\n| Stage | Typical voltage | Role | Why it matters |\n|---|---|---|---|\n| Utility feed | Medium-voltage AC | Bulk power into the site | Transformed down on site |\n| Facility distribution | ~480 V AC | Feeds UPS, PDUs | Backup + distribution |\n| Rack busbar | 48 V DC | Vertical bar down the rack | 16× less loss than 12 V |\n| Board / VRM | ~48 V → ~1 V | Multiphase step-down | Placed next to the die |\n| Point-of-load | <1 V, >1000 A | Feeds the GPU core rail | Tight regulation, low copper |\n\n```svg\n\n```\n\nRead datacenter power delivery through a cascade-of-conversions-and-losses lens rather than a wires-in-the-wall lens. Once you see the whole path as a sequence of voltage step-downs, each one imperfect and each one shedding heat, the design choices become obvious: distribute at the highest practical voltage to kill resistive loss, do the final conversion to sub-volt rails as close to the die as you can, and treat every point of efficiency as megawatts of cooling and cost that you either pay or avoid.