AI is doubling the energy data centres use. How they run decides how much.
The largest tech companies spent more than $400bn on capex in 2025 to build AI capacity. Here is what that means for electricity, from the global picture down to a single site.
01 · Demand
Data-centre electricity roughly doubles by 2030.
The IEA puts global data-centre use at 485 TWh in 2025 and about 950 TWh in 2030, around 3% of the world's electricity. In the EU, the European Commission expects about 68 TWh in 2024 to grow to about 114 TWh in 2030.
The two charts have different scales, because the EU is a much smaller part of the global total.
World
TWh a year. Dashed = projection.
IEA 2025, 2026
Global data-centre electricity, TWh
Year
TWh
Status
2024
415
IEA estimate
2025
485
IEA estimate
2030
≈950
IEA projection
European Union
TWh a year. Dashed = projection.
European Commission, COM(2026) 500 (IEA-based)
EU data-centre electricity, TWh
Year
TWh
Status
2024
68–70
EC, IEA-derived
2030
114–115
EC projection, 3.2% of EU demand
02 · The driver
AI servers drive the growth. Cooling grows with them.
Gartner expects power for AI-optimised servers to almost triple between 2025 and 2027, while conventional servers stay flat. Cooling and other infrastructure grow along with the load: from 159 to 243 TWh.
Gartner's totals are lower than the IEA's (447 vs 485 TWh for 2025) because the two use different methods. That is why they are shown in separate charts.
Where data-centre electricity goes
TWh a year, by use. 2026 and 2027 are forecasts (dashed frame).
Gartner, press release, 10 June 2026
Data-centre electricity by use, TWh (Gartner, June 2026)
Year
AI-optimised servers
Conventional servers
Cooling & other infrastructure
Total
2025 (estimate)
95
193
159
447
2026 (forecast)
175
195
195
565
2027 (forecast)
258
200
243
702
03 · Efficiency
Efficiency stopped improving a decade ago.
PUE is total facility energy divided by IT energy. Operators cut it fast until about 2014. In nine surveys since 2018 the average has stayed between 1.52 and 1.67.
7% to over 30%
of a data centre's electricity goes to cooling: about 7% at efficient hyperscale sites, over 30% at less efficient enterprise sites. (IEA, 2025)
Average PUE, 2007–2026
Lower is better. 1.0 would mean every kWh reaches IT.
Uptime Institute, Global Data Center Survey (annual)
Average annual PUE, Uptime Institute survey
Survey year
PUE
2007
≈2.5
2011
1.98
2014
1.65
2018
1.58
2019
1.67
2020
1.59
2021
1.57
2022
1.55
2023
1.58
2024
1.56
2025
1.54
2026
1.52
04 · What better operation recovers
Every step down in PUE is energy you no longer buy.
At the same IT load, a lower PUE means less energy for the whole facility. Uptime's capacity-weighted average is already 1.36, because larger sites run leaner.
Sensverk estimate · not third-party data
€0.8M a year
Cooling energy saved at a single 10 MW IT site, base case.
Facility energy saved by lowering PUE from 1.52
Calculation, not measured data: 1 − target PUE ÷ 1.52, same IT load.
Calculation by Sensverk on Uptime 2026 survey values
Calculation: facility energy saved at the same IT load = 1 − target PUE ÷ 1.52
From → to
Saved
1.52 → 1.36
10.5%
1.52 → 1.30
14.5%
1.52 → 1.20
21.1%
See where your site stands.
Tell Brian how your site runs today. In 30 minutes he'll show you where Sensverk would find energy to save.
