By the International Energy Agency’s count, data centres used about 415 terawatt-hours of electricity in 2024. That is roughly 1.5% of everything the world used, and the agency expects it to more than double by 2030, mostly because of AI. Numbers that size make it sound like a problem for governments and the biggest cloud companies. Some of it is. But a lot of what makes a data centre efficient is plain, and some of it is decided by people like me, writing the software that runs inside.
A server turns almost all the power it draws into heat. Getting that heat out of the room is the second biggest job in the building, after the computing itself.
The industry measures this with a ratio called PUE. You take all the power the building draws and divide it by the power that actually reaches the computers. A perfect building would score 1.0. Uptime Institute’s 2024 survey put the average at about 1.56, and it has barely moved in five years. Put simply, for every watt that does computing, a typical facility spends about half a watt more on cooling, power conversion and lights. The best new buildings get close to 1.3.
The picture above shows one of the oldest fixes, and one of the cheapest. The racks stand in rows facing each other, so they all blow their hot air into the same aisle. That aisle can be closed off and cooled directly, and the aisles in front stay cool. There is nothing clever about it. It is plumbing for air, and it saves a great deal of energy.
Where the power comes from matters just as much. A very efficient building on a coal-heavy grid can still put out more carbon than an average one on a clean grid. The cost of a computation depends on where it runs and when. Many cloud providers now publish how clean each of their regions is, and for work that is not tied to one place, picking the cleaner region is often the biggest decision on the table.
For smaller server rooms, the kind I work with through Snowleaf, the question comes even earlier. Many run on a mix of the grid, a generator and sometimes solar. Before anyone can improve anything, they need to know how much power came from each source, and when. Most do not, because the answer is sitting in spreadsheets nobody checks.
The building sets the ratio. The software decides how much work gets multiplied by it. Servers that sit idle still draw a large share of their full power, so moving light work onto fewer machines and switching the rest off saves more than it sounds like it should. Reports, backups and model training can often wait a few hours for a cleaner grid, or for the sun on a site with panels. Fewer database calls, smaller responses and sensible caching take load off every machine behind them. And none of these is worth claiming until it is measured, next to latency and cost, where the team already looks.
It is tempting to treat this as a facilities problem, someone else’s department. In practice the two meet in the middle. The best cooling in the world cannot help software that keeps a thousand servers busy doing work nobody needed. The leanest code still runs on whatever power the building buys.
So whichever side of that line you sit on, the question is the same. How much of this energy actually did something for somebody?
Sources. International Energy Agency, Energy and AI (2025). Uptime Institute, Global Data Center Survey 2024.