Data center power calculator

A megawatt is not a megawatt-hour

These are different quantities, and almost every article about data center energy consumption uses them as if they were interchangeable. They are not, and the difference is a factor of however long the thing ran.

A megawatt is a rate: what the meter reads at this instant. A megawatt-hour is an amount: a megawatt sustained for an hour, which is the kind of quantity a bill counts. Saying a facility "used 500 MW last year" is the same category of mistake as saying a car "travelled 70 miles per hour last year".

The conversion is one multiplication, and it is the only one on this page. A site drawing 500 MW continuously for a year consumes 500 × 8,760 = 4,380,000 MWh, or 4.38 TWh. Read literally, a headline about a "500 MW data center" that then quotes 500 MWh has described a facility that ran for one hour.

Announced capacity is not consumption

When a new facility is announced as 500 MW, that figure is almost always the grid connection or the design capacity: what the building is provisioned to be able to draw once every hall is fitted out and every rack is full. It is a planning number, and it is the largest honest number available, which is why it is the one in the announcement.

Real facilities do not run at nameplate. Halls are commissioned in phases over years. Racks are half-populated because the cooling in that aisle is the constraint, or because the hardware has not arrived. Servers idle. A site provisioned for 500 MW might be drawing 200, and a multi-year build might be drawing 40 in its first year with the rest as an option on grid capacity nobody has used yet.

That is what the utilisation field above is for. Set it below 100% and the installed capacity stays where it was while the draw, the energy and the bill all come down with it, which is the gap between the number in the press release and the number on the invoice.

PUE is a ratio, not a rating

Power usage effectiveness is total facility energy divided by IT equipment energy. A PUE of 1.5 means that for every kilowatt reaching a server, half a kilowatt is spent on cooling, power conversion, lighting and losses. That is all it means.

Two things follow that get lost when it is treated as a score.

A good PUE on a huge load is still an enormous amount of power. A 300 MW site at a PUE of 1.1 draws 330 MW. A 5 MW site at a dismal 2.0 draws 10 MW. The efficient facility is using thirty-three times the electricity of the wasteful one, and its ratio is better. PUE compares a facility to itself, not to another facility, and never to a smaller one.

The floor is 1.0 and it is unreachable. The ratio includes the IT load in its own numerator, so it cannot go below one. Figures below 1.0 do occasionally appear, and they are either a measurement of something else, or a different metric: energy reuse effectiveness, which credits waste heat sold to a district heating scheme, genuinely can be less than one. PUE cannot.

There is a third thing, which is that a PUE quoted without a period is worth very little. A ratio measured on a mild afternoon in a temperate climate will be a great deal better than the same site's annual average, and the annual average is the only version comparable to anybody else's. The industry-wide figure reported in Uptime Institute's annual survey has hovered around 1.5 to 1.6 for roughly a decade, which is the honest default for a facility you know nothing else about.

The same ratio is also published inverted, as DCiE: a PUE of 1.5 is a DCiE of 67%. The calculator gives both, because reports use both.

How much electricity does a data center use?

There is no single answer, and the range is five orders of magnitude wide, so the useful version is per rack and multiplied up, which is what the calculator above does.

• One rack: 5 to 10 kW is an ordinary enterprise cabinet. Dense compute runs 15 to 30 kW. A rack of current AI accelerators is 40 to 130 kW, which is more than an entire small server room used to draw and is why liquid cooling has stopped being exotic.
• A small server room: ten racks at 5 kW is 50 kW of IT, about 75 kW at the meter, roughly 660 MWh a year. That is around sixty US households.
• An enterprise data center: 1 MW of IT load at a PUE of 1.5 is 1.5 MW drawn, 13,140 MWh a year, about 1,250 households.
• A hyperscale facility: commonly 20 to 100 MW. Campuses of several buildings are announced at 500 MW and above, though see the section on capacity above before repeating that figure as consumption.
• Everything, everywhere: the IEA has put global data centre electricity use at roughly 1 to 1.5% of world demand, on the order of 400 to 500 TWh a year, with projections that rise steeply and disagree with each other. Treat any figure in this row as an estimate with a wide error bar and a publication date attached.

The thing worth taking from that list is that data center power usage scales almost entirely with the IT load. Cooling design moves the multiplier between about 1.1 and 2.0; the number of racks and what is in them moves the answer by factors of a thousand.

The reverse: a PUE calculator from two meter readings

The second mode does the division rather than the multiplication. Give it the facility total from the utility meter and the IT total from the UPS or PDU output, and it returns the ratio, the overhead in the same unit, and the share of the bill that never reached a server.

Two practical notes, since this is where a PUE calculator usually gets a wrong answer. The unit cancels, so MWh against MWh and kWh against kWh both work, but the two readings have to be the same unit and the same period, and mixing them is the reason a facility occasionally reports a PUE of 1,000. And the IT reading has to be the IT equipment only: if the meter you are reading sits upstream of the UPS, it already includes the conversion losses that PUE exists to measure, and the ratio will look better than the facility is.

What the calculator assumes

• 8,760 hours. Annual energy assumes the load runs continuously all year. For a data center that is close to true, and it is what an annual figure means in every published source.
• A flat rate. Real bills have demand charges, time-of-use rates and negotiated industrial tariffs. The cost here is energy at one price, which is the right order of magnitude and not an invoice.
• A constant PUE. Real PUE moves with the weather and with load: it is worse at low utilisation, because the cooling and UPS overheads do not shrink as fast as the IT load does. A part-loaded facility set to its design PUE here will be flattered.
• 10,500 kWh per household per year. The comparison figure, roughly the US residential average. A UK household uses closer to a quarter of it, so the "homes" number is a sense of scale rather than a count of houses.

Electricity is not the only resource this argument is had about, and the two are linked: 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 takes the overhead figure above and turns it into a heat rejection load and a water consumption, and the article on data center water usage explains why two sites of the same size can differ by thirty-fold.

The same conversion applied to a whole state, rather than one site, is on data centers in Minnesota: every announced megawatt figure in the state turned into annual energy and households, which is the step the announcements always skip.

Sources

• How much electricity does an American home use?: US Energy Information Administration, the household figure
• Electricity 2024: IEA, global data centre demand and projections
• Power usage effectiveness: the metric's definition and its DCiE inverse

Figures are ranges from published surveys read in September 2026. Where sources disagree, the wider range is given rather than the more quotable one.