How much electricity does a data center use?
Published 22 September 2026
Between about 5 kilowatts and a gigawatt, which is a range of two hundred thousand to one and tells you nothing. The more useful answer is that most published data center power consumption figures are not consumption at all: they are design capacity, and the gap between the two is routinely a factor of two.
What the scales actually look like
"Data center" covers a cupboard and a campus. These are the bands that published figures fall into, as continuous IT load:
| Facility | IT load | Annual energy |
|---|---|---|
| One ordinary rack | 5–10 kW | 44–88 MWh |
| A server room or wiring closet | 20–100 kW | 0.2–0.9 GWh |
| Enterprise data center | 1–5 MW | 9–44 GWh |
| Hyperscale facility | 20–100 MW | 175–876 GWh |
| Announced AI campus | 300 MW–1 GW | 2.6–8.8 TWh, if it ever runs full |
Those annual figures assume the load runs continuously, which is the assumption behind nearly every data center energy consumption number you will read. Real facilities do not. More on that below, because it is where most of the error lives.
Work it out for a specific site. The data center power calculator turns racks, rack density and a PUE into IT load, facility load, annual MWh and annual cost, and will also derive PUE backwards from two meter readings.
Announced capacity is not consumption
This is the single largest source of wrong numbers, and it is not anybody lying. When a facility is announced at 500 MW, that figure is almost always the grid connection: what the utility has agreed to deliver if every hall is built and every rack is filled.
It is a planning number, and it is the largest honest number available, which is why it is the one in the press release. Meanwhile:
- Halls are commissioned in phases over years. A site announced at 500 MW might be drawing 40 MW in its first year.
- Racks are not filled to their design density. Power, cooling and floor space rarely run out at the same moment, so something is always stranded.
- Servers idle. A well-utilised fleet averages far below its own nameplate, and most fleets are not well utilised.
So a headline saying a data center "uses as much electricity as a city of half a million people" is usually describing a facility that does not exist yet at a draw it has never reached. The claim is not fabricated. It is a capacity figure doing duty as a consumption figure.
Megawatts and megawatt-hours are different things
A megawatt is a rate: what the meter reads at this instant. A megawatt-hour is a quantity: a megawatt sustained for an hour. "Used 500 MW last year" is the same category of mistake as "travelled 70 miles per hour last year".
The conversion is one multiplication, and it is the only one on this page. A 40 MW facility running continuously for a year uses 40 × 8,760 = 350,400 MWh, or 350 GWh. Confusing the two is what lets a 500 MW grid connection get reported as though the site burned 500 MWh, which would be one hour of its own operation.
Then multiply by PUE
Everything above is IT load: the power reaching servers, storage and switches. The building also runs cooling, power conversion and lighting. Power usage effectiveness (PUE) is total facility energy divided by IT energy, so it is the multiplier between the two.
| PUE | Overhead | Typical of |
|---|---|---|
| 1.1 | 9% of the bill | Best-in-class, purpose-built, good climate |
| 1.5–1.6 | 33–38% | Around the industry average, roughly flat for a decade |
| 2.0 | 50% | Older enterprise room, mechanical cooling |
Note the overhead column: at PUE 1.5 the overhead is 33%, not 50%. It is
(PUE − 1) ÷ PUE, because the PUE is the total rather than the
extra. This trips people up constantly and it is worth a moment.
PUE is also a ratio and not a rating. A 300 MW site at PUE 1.1 draws 330 MW. A 2 MW site at PUE 2.0 draws 4 MW. The first is using its electricity far better and is using eighty times more of it.
What AI actually changed
Not the number of data centers: the rack density inside them. This is the part of the story that is genuinely new rather than merely loud.
| Rack type | Power |
|---|---|
| Ordinary enterprise rack | 5–10 kW |
| Dense virtualisation or HPC | 15–30 kW |
| Current GPU training rack | 40–130 kW |
A single modern GPU rack can draw more than an entire small server room did. That is why liquid cooling stopped being exotic: above roughly 30 kW a rack, moving that much heat with air becomes impractical in a normal hall.
It also breaks the old planning arithmetic. A hall designed for 8 kW a rack cannot be filled with 120 kW racks no matter how much floor space is left: it runs out of power and cooling first, which is why so much retrofit work is really electrical work.
Density drives the network inside the building as directly as it drives the switchgear: cabling, uplink counts and the shape of the fabric all follow from it. Inside a data center network covers how that hall is actually wired.
The global picture, with the uncertainty attached
Aggregate data center electricity usage is genuinely hard to measure, and the published figures disagree. Two estimates are cited most often:
- The IEA put global data centre consumption at roughly 415 TWh in 2024: about 1.5% of global electricity, and projects something near 945 TWh by 2030.
- Lawrence Berkeley National Laboratory put US data centres at about 176 TWh in 2023, or 4.4% of US electricity, with a 2028 range of 325–580 TWh: between 6.7% and 12.0% of forecast US demand.
That LBNL range is the honest part. A projection spanning a factor of nearly two, four years out, from the most careful study available, is a fair statement of how much nobody knows. Treat any single-number forecast beyond about two years as an opinion with a citation attached.
The other thing worth holding onto: the percentages are of a growing denominator, and they are national or global averages. Local concentration is what actually strains a grid. A county where three hyperscale campuses land has a problem that the global 1.5% figure does not describe at all.
The questions that make a figure mean something
- Capacity or consumption? Grid connection, design capacity and measured draw are three different numbers and only the last is consumption.
- Power or energy? MW or MWh. If the unit is missing, the figure is unusable.
- IT load or facility load? The difference is the PUE, and it is typically a third of the total.
- Measured over what period? Peak draw and annual average can differ by more than half.
Electricity is also not the whole environmental question: what that power is worth in carbon depends almost entirely on the grid supplying it, which whether data centers are bad for the environment takes apart. And it is tied to the other resource constraint: the cooling design that lowers a facility's PUE is usually the one that evaporates the most water. The cooling load and water use estimator shows both sides at the same load, and how much water a data center uses covers why two sites of identical size can differ thirty-fold.
Average consumption also hides when the load occurs. A facility that can move batch compute, draw briefly from storage or cap accelerators may offer data center demand response even when its annual MWh barely changes. The useful claim is not "flexible" but a measured number of MW, a notice time and a duration.
Sources
- Energy and AI: International Energy Agency, global consumption and projections
- 2024 United States Data Center Energy Usage Report: Lawrence Berkeley National Laboratory, released December 2024
- Power usage effectiveness: the metric's definition and its DCiE inverse
Figures are ranges from published studies read in September 2026. Where sources disagree (and on projections they disagree enormously) the wider range is given rather than the more quotable number.