800 伏特與 48 伏特之爭：機架供電標準為何正在瓦解

For most of two decades the inside of a datacentre rack was a solved problem. Power arrived as alternating current, was converted to 48 volts direct current near the rack, and was distributed along copper busbars to the servers. The standard came out of telecom, where 48 volts had been the norm for a century because it is low enough to be safe to touch and high enough to run a telephone exchange. It was a good decision for the loads it was designed for.

Those loads were ten to fifteen kilowatts per rack. The racks being installed for accelerator training today are ten times that, and the ones on the roadmap are ten times again. At that scale the arithmetic stops being a matter of taste. A hundred kilowatts at 48 volts is a little over two thousand amps. Two thousand amps through copper is not a design choice, it is a materials problem, and materials problems do not negotiate.

The physics that ends the argument

Power lost in a conductor is the square of the current times the resistance. Double the current and you quadruple the loss for the same conductor. The only ways to hold the loss constant are to halve the resistance, which means roughly quadrupling the copper cross-section, or to raise the voltage.

At 48 volts, a 100 kW rack draws about 2,080 amps. At 800 volts it draws 125. The copper cross-section needed falls by a factor of roughly 64 for the same loss. That is not an incremental improvement in busbar design. It is the difference between a busbar that can be routed inside a rack and one that dominates the rack's volume, weight and cost.

The secondary effects compound. Less copper means less weight, which matters when a rack is being craned into a hall with a floor loading limit. Less current means less heat inside the rack, which changes the cooling requirement. And lower loss means the efficiency curve of the whole facility improves at exactly the point where the load is largest.

At a hundred kilowatts per rack, 48-volt distribution stops being a standard and becomes a copper procurement problem, and that is the point at which standards change.

What has to be rebuilt

Raising the voltage is not a matter of turning a dial. Every element in the chain was qualified at a lower level.

Protection is the hardest part. A direct-current arc does not self-extinguish at a zero crossing the way an alternating-current arc does, so a fault at 800 volts is a sustained plasma until something actively quenches it. That pushes the industry toward solid-state circuit breakers and hybrid devices with fast detection, which are more expensive and less proven than the mechanical breakers they replace.

Conversion is the second problem. Servers want sub-1-volt rails at hundreds of amps, and the conversion chain from 800 volts down has more stages than from 48. Each stage costs efficiency, board area and money. The gain from distribution has to exceed the loss from conversion, and the crossover depends on how far the power travels and how much of it there is. At 20 kW per rack the maths does not close. At 200 kW it closes easily.

The third problem is organisational. A datacentre is a chain of vendors who each guarantee a component and none of whom guarantees the seam between components. Moving to 800 volts means every vendor in the chain requalifies simultaneously, which is why standards bodies matter more than any single product launch.

The counterargument, and why it is weaker than it sounds

The obvious objection is that alternating current already solved this. Medium-voltage AC distribution is mature, well understood, and every electrician in the world is qualified to work on it. Why not simply bring higher-voltage AC closer to the rack and convert once?

The answer is where the conversions happen: a modern facility already converts AC to DC at the UPS, back to AC for distribution, and to DC again inside every server, so each conversion is both a loss and a failure point. A modern facility already converts AC to DC at the UPS, back to AC for distribution, and to DC again inside every server. Each conversion is a loss and a failure point. Consolidating to a single AC-to-DC step at the hall level and distributing DC removes two conversions from the path. That is worth several points of efficiency at facility scale, and at a hundred megawatts several points is a power plant.

The real barrier is not technical superiority, it is the installed base. Nobody replaces a working hall to gain two percent. The transition will happen in new builds first, in the halls built specifically for accelerator density, and it will spread only as those halls become the majority of new capacity.

What this does to the supply chain

If DC distribution at 800 volts becomes the norm for high-density halls, the value shifts toward the components that make it safe and efficient: solid-state protection, wide-bandgap conversion, high-voltage connectors, and the monitoring that lets an operator find a fault before it becomes an arc.

The plausible range, with reasoning: high-density halls are currently a small share of new capacity but a large share of new power draw. If 800-volt distribution reaches a third of new high-density capacity by 2029 and half by 2031, the incremental market for protection and conversion equipment in that segment lands between three and nine billion dollars a year, with the low end assuming continued hybrid AC/DC designs and the high end assuming full DC distribution to the rack.

The transition will not be decided by efficiency alone but by which vendor can guarantee a fault is cleared in microseconds, because that is the only specification an insurer will accept.

What is actually transferable

The pattern here repeats across infrastructure: a standard is chosen for a load profile, the load profile changes by an order of magnitude, and a physical constraint that was invisible becomes binding. The standard then does not gradually improve, it breaks and is replaced.

The same thing happened when racks moved from 2 kW to 10 kW and cooling moved from room air to in-row units, and when single-phase power gave way to three-phase at the rack. Each transition looked sudden and was in fact the end of a long arithmetic argument that nobody had bothered to run.

The transferable lesson for anyone specifying infrastructure is to compute the constraint at the load you expect in five years, not the load you have today. Standards are written at one scale and retired at another, and the retirement is always a materials calculation wearing the costume of a policy debate.

If this saved you time, a tip is appreciated: HCqPmPJM3NVtS6YrmVaTZYnuP8XZ2PFgQZKT94VK1LkS