How it works · Part 6

The data centre

A gigawatt campus from the substation to the cold plate. Eleven exhibits on what it takes to feed and cool half a million GPUs: the power coming in at 345,000 volts, the batteries and generators, why a rack went from 10 kW to 120 kW, where the gigawatt actually goes, why air stopped working, the cold plates and the CDU, where the water goes, and where the power comes from. Play the tour, or pick an exhibit and drag it.

A gallery, not a single site: each exhibit is one idea, drawn to its own scale. Figures are for a campus of GB200 NVL72 racks unless a step says otherwise.

Playing the tour

Step 1 of 12

What a gigawatt campus is

The unit of AI infrastructure is no longer a building but a campus measured in power. A gigawatt is what a large nuclear reactor produces, and at today's densities it feeds roughly 400,000 to 500,000 Blackwell GPUs in about 7,000 liquid-cooled racks, with its own substation, a yard of generators and a field of coolers.

1 GW: ~750,000 homes
GPUs per gigawatt: 400,000–500,000
GB200 NVL72 rack: 120 kW
Rubin Ultra rack, 2027: 600 kW
Industry PUE: ~1.55
Best sites: 1.1
Water, evaporative, 100 MW: 2–5 million L/day
Grid connection: 3–5 years

The words that come up

Twelve terms from the site plans and the utility filings.

Gigawatt (GW): A billion watts. The output of a large nuclear reactor, the draw of three-quarters of a million homes, and the new unit for an AI campus.
IT load: The power the computing equipment itself draws: GPUs, CPUs, switches, storage. Everything else on the site exists to deliver it and remove it as heat.
PUE: Power usage effectiveness: total site power divided by IT power. 1.0 would be perfect; the industry averages about 1.55; the best sites run 1.1.
WUE: Water usage effectiveness: litres of water used per kilowatt-hour of IT energy. The industry average is around 1.8; dry-cooled sites approach zero.
Substation: Where the transmission voltage (138–345 kV) is stepped down for the site, usually to 34.5 kV. Its transformers are the longest-lead item on the campus.
UPS: Uninterruptible power supply: batteries (and their converters) that carry the load for seconds to minutes while generators start.
Busway / busbar: Rigid copper or aluminium conductors that carry power along a row (busway, 480 V) and down a rack (busbar, 54 V DC, soon 800 V DC).
Direct-to-chip (DTC): Liquid cooling with a cold plate bolted to each hot chip. Captures 70–80 % of a rack's heat as warm water; the rest still goes to air.
Cold plate: A copper block with microchannels inside, through which coolant flows across the chip's surface. One per GPU, CPU and switch in an NVL72.
CDU: Coolant distribution unit: pumps, filters and a plate heat exchanger that join the closed rack loop to the facility water without mixing them.
Cooling tower / dry cooler: The two ways to shed heat outdoors: evaporate water (cheap, uses water) or blow air over finned coils (uses no water, more fan power).
Behind the meter: Generating power on site rather than taking it from the grid: gas turbines, fuel cells, batteries. The fast route when the grid connection is years away.

Figures from Nvidia's GB200/GB300 NVL72 and 800 V DC materials, the Uptime Institute's annual survey (PUE), ASHRAE's liquid-cooling classes, Vertiv's and Schneider Electric's reference designs, the US Department of Energy's 2024 report on data-centre energy use, and company statements on campus sizes and power deals. Everything is rounded; real sites differ, and operators rarely publish their own numbers.

The whole series · Part 5: How 100,000 GPUs talk · Part 4: How a language model works · Part 3: A wafer's three months in the fab · Part 2: How a GPU works · Part 1: How EUV lithography prints a chip