How an AI Data Centre Is Powered and Cooled: Every Watt Comes Back as Heat

A data centre makes nothing you can hold. Physically it does one thing: it takes electricity off the grid and gives back the same quantity as heat. Everything else in the building exists to move those two flows in and out of rooms full of chips. The hard engineering of an AI data centre is not computing. It is power delivery and plumbing.

The rack is the unit

Data centres are counted in racks: steel cabinets about the height of a door, filled with servers in horizontal slots. What matters about a rack is how many watts it draws, because that is also how many watts of heat must be taken back out of it. Across the industry these are modest numbers. The Uptime Institute’s 2025 survey, answered by 681 operators, found more than 80 percent of them have no rack anywhere in their facility drawing above 30 kW. About one in eight reports racks in the 30 to 59 kW band, and facilities averaging under 4 kW a rack are still common.

An AI rack is a different object. NVIDIA’s GB200 NVL72 is one liquid-cooled rack holding 72 Blackwell GPUs and 36 Grace CPUs. NVIDIA’s power engineering documents describe the range it now designs for as 100 kW to over 1 MW per rack — and a megawatt in one cabinet is more than thirty times what the densest rack in four-fifths of the world’s data centres draws.

Getting the current to the chip

The electricity arrives as high-voltage alternating current, is stepped down by transformers at an on-site substation, passes through uninterruptible power supplies and distribution boards, and reaches power shelves in the rack, which convert it to 54 volts of direct current for the copper busbars. The chips run at around a volt. Every conversion on the way gives up a few percent as heat before any computing happens.

At megawatt scale this chain stops working, for an unglamorous reason: metal. Current rises as voltage falls, and a 1 MW rack fed at 54 volts needs up to 200 kilograms of copper busbar, by NVIDIA’s reckoning — 200,000 kilograms across a one-gigawatt site, for the in-rack bars alone. The power shelves would then occupy up to 64 rack units: most of the cabinet, with no room left for computers.

The answer being standardised is to bring 800 volts of direct current into the row. The same conductor then carries 85 percent more power, copper falls by about 45 percent, and NVIDIA claims up to 5 percent better efficiency end to end, with full production dated to 2027. Note what the fix is: not a faster chip, a higher voltage.

Every watt comes back as heat

A chip does not consume electricity. It converts it, and almost all of it leaves as heat — conservation of energy, with no exemption for cleverness. A 1 MW rack is a one-megawatt electric heater that happens to do arithmetic, so cooling is not a support service: it is the same size as the computer.

Air was the default for thirty years. But a litre of air holds very little heat, so you need enormous volumes moving fast, and the fans doing the moving are themselves a load. Water carries far more heat per litre, which is why dense racks are plumbed rather than ventilated: cold plates sit on the chips, coolant runs through them out to the building’s loop, and the room can be merely warm. The NVL72 is not optionally liquid-cooled — at that density there is no air-cooled version to choose.

The number that stopped improving

The industry’s efficiency measure is PUE, power usage effectiveness: total facility power divided by the power reaching the computers. A PUE of 2.0 means a watt of overhead — cooling, fans, conversion losses — for every watt of computing. The Green Grid introduced it in 2007; it is crude by design, and says nothing about whether the computing was worth doing.

The number has barely moved. Uptime’s 2025 weighted average of 1.54 was the sixth consecutive year virtually unchanged. The averages do separate by class: facilities of 20 MW and above come in at 1.44 globally, those built since early 2020 at 1.48, and 15 percent of respondents report 1.3 or better.

At the top the gap is wide. Google reports a fleet-wide average annual PUE of 1.09 for 2025 — nine watts of overhead per hundred watts of computing, against the survey’s 54 — which its own footnote works out as 83 percent less overhead energy than the industry average. No trick in it, just a decade of building your own buildings.

Water, and the price of not using it

The cheapest way to reject heat is to evaporate water: evaporation carries off a great deal of heat for very little energy spent. It is also why a data centre appears on a municipal water balance sheet, which is where most local opposition begins.

Microsoft’s answer, published in December 2024, is a closed loop: fill the system once during construction, then circulate the same water between the servers and the chillers indefinitely. It puts the water this avoids at more than 125 million litres a year per datacentre, and reports a fleet water use effectiveness of 0.30 litres per kilowatt-hour for its 2024 financial year, against 0.49 in 2021. Every new datacentre has been built this way since August 2024; the first of them, in Phoenix and in Mt Pleasant, Wisconsin, come online late in 2027.

The trade-off is in Microsoft’s own post: replacing evaporative cooling with mechanical cooling will increase its PUE. Heat not carried off by evaporation has to be pumped away, and pumps take power. Water and electricity are the two currencies, and mostly you choose which to spend.

How much electricity, altogether

The best public arithmetic for a whole country is the Berkeley Lab report the US Department of Energy published in December 2024. American data centres used 58 TWh in 2014 and 176 TWh in 2023, by then about 4.4 percent of all US electricity. Its range for 2028 is an increase of 325 to 580 TWh, putting data centres at 6.7 to 12 percent of consumption.

What is being committed to now outruns that record. In the quarter ending 30 September 2026, Alphabet spent $51.4 billion on property and equipment, $132 billion over nine months. Microsoft spent $39.4 billion in the same three months and reported $677 billion of contracted commercial work not yet delivered, up 73 percent. Those are not forecasts. They are buildings, transformers and cooling plant, already ordered.

What that is in Georgian units

Georgia’s whole electricity system, by the market operator ESCO’s 2025 balance, consumed 14,884 GWh. Domestic plants generated 13,838 GWh of it — 10,987 from hydro, 2,767 from thermal, 79 from the Kartli wind farm — with 1,558 GWh imported and 511 exported.

Set a data centre against that. A site drawing 100 MW continuously uses 876 GWh a year: about 5.9 percent of everything Georgia consumed in 2025, from one building. The gigawatt campuses routinely discussed in the United States would need 8,760 GWh — 59 percent of national consumption, and four-fifths of what every hydro plant in the country produced last year.

This cuts both ways. Cheap hydroelectricity and a cool climate are what such projects shop for, and Georgia has both. But the same arithmetic makes any serious proposal here a grid question first and a technology question second.

Where the next watt comes from

When power becomes the binding constraint, the search for it gets strange. Google Research published a design study in November 2025 for putting the computers in orbit, on the grounds that the Sun emits more than 100 trillion times humanity’s total electricity production, and that a solar panel in the right orbit can be up to eight times more productive than one on the ground, almost continuously. Two prototype satellites, built with Planet, launch by early 2027.

The engineering so far is unromantic: Google put a Trillium TPU in a 67 MeV proton beam to see what radiation does to it. Its high-bandwidth memory proved the most sensitive part, showing irregularities only after 2 krad — about three times a shielded five-year dose — and nothing failed outright up to 15 krad. The chips are the easy part.

When the cheapest way to cool a computer starts to look like leaving the planet, the problem was never the computer.