# Power & cooling for AI data centres

URL: https://justuju.org/concept/datacenter-power/
Layer 1 · Physical & Energy

AI data centres are limited by how much electricity they can get and how much heat they can remove.

## D1 · Story

Computers get warm when they work hard, like your legs when you run. AI computers work very hard, all day and all night. So they need lots of electricity and lots of cooling.
Building one is a bit like building the power and air-conditioning for a small town. And it all goes to one room full of very busy computers.
1. Electricity in: from the grid
2. Busy computers: doing the maths
3. Heat: nearly all the power ends up as heat
4. Cooling: fans and cold water carry it away
Like: A room full of runners: they need lots of food (electricity) and they get hot, so you need big fans and cold water (cooling).

## D2 · Picture

Almost all the electricity a chip uses turns into heat, so getting power in and getting heat out are two sides of the same problem.
AI racks pack far more power into the same floor space than traditional server racks, which is why new facilities increasingly move from air cooling to liquid cooling.
Where a data centre can be built is often decided by grid connection capacity and permitting time, not by land.
1. Grid: power supply
2. Substation: plus backup
3. Racks: chips compute
4. Heat: nearly all energy
5. Cooling: air or liquid
6. Released: air or water
Like: Like a stadium on match day: the challenge isn't just building it, it's getting enough power in and enough heat out. Where it breaks: A stadium peaks for a few hours. AI clusters run near full load around the clock, so they need steady supply, not just peak capacity.

## D5 · Frontier

D2 explains power and heat for one building. Summed across a country, the totals, the sources of supply and the side effects are all contested.
How much demand. The IEA's base case has global data-centre electricity rising from about 415 TWh in 2024 to around 945 TWh in 2030, and its 2035 cases span roughly 700 to 1,700 TWh (IEA, 2025).
Efficiency versus growth. US data-centre use stayed near 60 TWh from 2014 to 2016 as efficiency offset growth, then reached 176 TWh, 4.4% of US electricity, in 2023. Whether efficiency can catch up again is unresolved (Berkeley Lab, 2024).
Who supplies it. Builders look at on-site generation or sharing a site with an existing power station. US regulators are still deciding how such co-located loads should pay for, and lean on, the wider grid (FERC, docket RM26-4).
Water. Cooling towers evaporate water on site and power stations use more off site, with no agreed way to count both. One projection puts AI's global water withdrawal at 4.2 to 6.6 billion cubic metres in 2027 (Li et al., 2023).

## Examples as of 2026-10-30

- xAI Colossus (Memphis)
- OpenAI, Oracle and SoftBank's Stargate campus in Abilene, Texas
- Meta Hyperion (Louisiana) and Prometheus (Ohio)
- Microsoft Fairwater (Wisconsin)

## Sources

- Energy and AI (IEA report): https://www.iea.org/reports/energy-and-ai
- 2024 United States Data Center Energy Usage Report (Lawrence Berkeley National Laboratory): https://eta.lbl.gov/publications/2024-lbnl-data-center-energy-usage-report
- IEA: Energy and AI, executive summary: https://www.iea.org/reports/energy-and-ai/executive-summary
- Berkeley Lab: 2024 United States Data Center Energy Usage Report (full report): https://escholarship.org/content/qt32d6m0d1/qt32d6m0d1.pdf
- FERC: Interconnection of Large Loads to the Interstate Transmission System, Docket RM26-4: https://www.ferc.gov/rm26-4
- Making AI Less 'Thirsty': Uncovering and Addressing the Secret Water Footprint of AI Models (Li et al.): https://arxiv.org/abs/2304.03271


---
Justuju · https://justuju.org/ · Content CC BY 4.0 (https://creativecommons.org/licenses/by/4.0/). Quote with a link to the page.