Power per Rack, and What It Does to a Building

A single rack of the new kind draws what roughly fifteen typical racks do. The Uptime Institute's 2025 global data center survey put the average typical rack density reported by operators at just under nine kilowatts. Hewlett Packard Enterprise, or HPE, in its published specifications for its build of the NVIDIA GB200 NVL72, lists a nominal thermal design power of 132 kilowatts per rack, with about 115 kilowatts removed by liquid and about 17 by air, and a peak electrical design figure of one and a half times the nominal. Divide one figure by the other and the ratio comes out near fifteen.

That ratio is the whole lesson. Every building system was sized for the old number.

Begin with cooling. Air cools by moving volume. The heat a stream of air can carry depends on how much air moves and how much it warms. At under ten kilowatts a rack, fans, raised floors and containment aisles work. At more than a hundred kilowatts in the same footprint, the air would have to move at impractical speed. Water carries far more heat per unit of volume than air. So the high-density rack moves heat into liquid.

The common arrangement is direct-to-chip cooling. Metal cold plates sit on the hottest components, the graphics processing units, or GPUs, and the central processing units, or CPUs. A coolant loop runs through them. That loop feeds a coolant distribution unit, or CDU, a heat exchanger that passes the heat to the building's own water system. HPE's documentation for its GB200 NVL72 build describes exactly this two-loop arrangement. The building now needs piping to the rack, leak detection, water treatment and a plan for what happens when a pump fails.

The standards bodies have followed. The American Society of Heating, Refrigerating and Air-Conditioning Engineers, or ASHRAE, through a technical committee devoted to data centers, publishes thermal guidelines that data center designers use, and those guidelines now cover liquid-cooled equipment and facility water temperatures. When a heating and ventilation society writes rules for chip coolant, the chip has become a building-services problem.

Next, power distribution. Electricity has to reach the rack and be converted, step by step, to the low voltages the chips use. Inside today's racks, power is commonly distributed at 54 volts direct current. NVIDIA published a technical blog in May 2025 arguing that this cannot scale. At 54 volts, the current needed for a very high-power rack requires heavy copper busbars and several conversion stages, each losing energy. NVIDIA proposed distributing 800 volts direct current instead, and said it would lead a transition towards racks of one megawatt and beyond, starting in 2027 with its Kyber rack architecture. Schneider Electric and other equipment makers have published supporting material. Higher voltage means less current for the same power, which means thinner conductors and fewer losses. It also means new switchgear, new safety practice and new skills for electricians.

Then, structure and space. A liquid-cooled rack full of GPUs, copper cabling and coolant is heavier than a rack of ordinary servers. Floors, lifts and loading docks were sized for the old weight. Faraday's read is that this is the least discussed and most stubborn retrofit constraint, because you can add a pipe to an existing building far more easily than you can strengthen its slab.

Finally, the grid. A building full of such racks is a large, concentrated load that can change quickly when training jobs start and stop together. That is why grid operators and the North American Electric Reliability Corporation, or NERC, now study data center demand as its own category of load.

Put the chain together. The chip got denser because the math rewards parallel arithmetic packed close together. The rack got denser because scale-up links reward short distances. The building has to absorb the result: liquid instead of air, higher distribution voltage, stronger floors and a bigger, less predictable grid connection.

For an investor, the useful question about any facility is simple. What rack density was it designed for, and what would it cost to raise that figure? For an operator, it is whether the building's cooling and power can be changed without taking the building out of service. For an engineer, it is where the next conversion stage or heat exchanger will sit. The chip roadmap answers none of these. It sets the deadline by which they have to be answered.