The global weighted average was 1.54 in the Uptime Institute's 2025 survey, essentially unchanged for six straight years, so anything under about 1.3 is genuinely good and anything under 1.2 is excellent. The best hyperscale fleets publish around 1.09 to 1.15; Google reports a fleet-wide 1.09. One caveat: climate does part of the work, so a 1.2 in a hot climate is a bigger achievement than the same 1.2 somewhere cold.

How much electricity does a data center use?

It scales with the IT load, so there is no single answer, but the arithmetic is simple. One megawatt of IT load at the industry-average PUE of 1.54 uses about 13.5 GWh a year, roughly the electricity of 1,285 average US homes. Small server rooms are a fraction of that; the largest AI campuses now run to hundreds of megawatts of IT, each hundred multiplying that homes figure by a hundred.

How much does it cost to run a 1 MW data center?

At the industry-average PUE of 1.54 and 8 cents per kWh, 1 MW of IT load costs about $1,079,232 a year in electricity alone. At a hyperscale-class PUE of 1.09 the same computing costs about $763,872. Staff, hardware, connectivity, and construction financing are all on top; this page prices only the meter, which is typically the largest single operating cost.

Why do data centers use so much water?

Because evaporating water is the cheapest way to shed heat. Evaporative cooling lowers a facility's electricity use (and its PUE) by spending water instead, and a large site can evaporate hundreds of millions of gallons a year. Dry cooling uses far less water but more electricity, so the two footprints trade against each other. The industry measures the water side with a separate metric, WUE, in liters per kWh.

What share of US electricity do data centers use?

About 4.4 percent in 2023, or 176 TWh, per the Lawrence Berkeley National Laboratory's 2024 report to Congress. The same report projects 6.7 to 12 percent by 2028, and the width of that range is honest: it depends on how fast AI demand actually grows. For scale, 176 TWh is roughly the electricity of 17 million average US homes.

Does a low PUE mean a data center is green?

No, and this is the metric's biggest blind spot. PUE only compares the building to the computers inside it. It says nothing about whether the computing is useful (an idle hall at PUE 1.1 wastes more than a busy one at 1.5), nothing about the carbon of the grid feeding it (a 1.1 on coal can out-emit a 1.5 on hydro), and nothing about water. It measures overhead efficiency, which matters, and only that.

Why do AI racks need liquid cooling?

Density. A traditional enterprise rack draws 5 to 10 kW and air cooling handles it easily. A modern AI training rack like NVIDIA's GB200 NVL72 draws around 120 kW, and air simply cannot move heat out of one floor tile that fast; practical air-cooled limits sit around 30 to 40 kW per rack. Above that, liquid cooling stops being exotic and becomes the price of admission, which is why the AI buildout mostly means new construction.

Data Center Power Calculator: MW, Cost, and CO2

Enter the IT load in kilowatts (or count racks) and pick a PUE tier. You get the facility's total power, its annual energy, the yearly electricity bill, and the CO2, plus a table pricing what better efficiency would be worth at your load.

Data reviewed: August 2026. Figures here come from published sources and change over time. How we verify

How the data center power calculator works

You keep hearing that data centers are straining the grid, and you want to know what one actually takes to run. The honest answer has two parts. The first is the IT load: what the servers, storage, and network gear themselves draw, measured in kilowatts. The second is everything the building spends keeping that gear alive: cooling it, converting utility power for it, lighting the halls around it. The ratio of the whole building to the IT alone is PUE (Power Usage Effectiveness), the most quoted number in the industry. This calculator takes your IT load, multiplies by a PUE you pick (every tier sourced: the 2025 industry average is 1.54 per the Uptime Institute, and Google's fleet runs 1.09), runs the result around the clock for a year, and prices it in dollars and CO2. It also shows you the same IT load at every efficiency tier, so you can see exactly what a better building is worth.

If what you are really asking is "what does my chatbot habit cost," that is a different page: our AI energy calculator prices individual prompts. And if you are sizing a GPU training cluster, chip by chip, the GPU cluster power calculator speaks that dialect. This page stays at the meter: kilowatts in, dollars and tons out.

The formula

facility kW = IT kW × PUE
annual kWh = facility kW × 8,760 hours
annual cost = annual kWh × price per kWh
annual CO2 = annual kWh × grid intensity

IT kW is the computing load (entered directly, or as racks times kW per rack). PUE is total facility power divided by IT power, so it can never be below 1.0. The 8,760 is simply 24 hours times 365 days, and it is the reason one multiplication is honest here: unlike your house, a data center's draw is nearly flat around the clock. Grid intensity comes from EPA eGRID 2023 data, the same regional figures our EV pages use, so the two families of tools can never disagree about the grid.

Worked example

2 MW (2,000 kW) of IT load at the industry-average PUE of 1.54, paying 8 cents per kWh on the US average grid:

Facility power = 2,000 × 1.54 = 3.08 MW, of which 1.08 MW is overhead. Annual energy = 3,080 kW × 8,760 hours = 26.98 GWh. Annual cost = 26,980,800 kWh × $0.08 = $2,158,464. Annual CO2 = about 9,389 metric tons, and the energy is enough to power roughly 2,570 average US homes for the year.

Now run the same 2 MW of IT at Google's fleet PUE of 1.09: 2.18 MW at the meter, 19.1 GWh, $1,527,744, and about 6,646 metric tons. Identical computing, identical work done, and the efficient building saves $630,720 and about 2,744 metric tons of CO2 every year. At 1 MW of IT the same gap is worth $315,360 a year. That is what PUE means in dollars, and it is why hyperscalers chase hundredths of a point.

What PUE actually measures (and what it quietly does not)

PUE is a beautifully simple ratio: total facility power divided by IT power. A PUE of 1.54 means that for every watt of computing, the building spends another 0.54 watts on overhead. A perfect 1.0 would mean zero overhead, which is why a PUE below 1.0 on a spec sheet is a typo or a trick. But the simplicity hides three honest problems. First, climate does half the work: a PUE of 1.2 in Phoenix is a real engineering achievement, while the same 1.2 in Stockholm is partly just cold air showing up for free. Comparing two facilities' PUE without knowing where they sit flatters the one in the better climate. Second, the boundary can be gamed: a facility can quote its best hall instead of the whole site, its best season instead of the year, or draw the line so some overhead lands outside the measurement. The trustworthy numbers, like Google's 1.09, are fleet-wide, trailing twelve months, all overhead included, and say so. Third, and deepest: PUE says nothing about whether the computing is useful. A hall of idle servers at PUE 1.1 wastes more electricity than a busy one at 1.5, and PUE also says nothing about the carbon of the grid feeding it. A 1.1 on a coal-heavy grid can out-emit a 1.5 on hydro. PUE measures the building, not the point of the building.

Where the overhead actually goes

Nearly every watt the IT gear draws becomes heat, and that heat has to leave the building or the servers throttle and die. So the overhead budget is dominated by cooling: chillers, fans, pumps, cooling towers. It is the same arithmetic your window air conditioner does (our BTU calculator runs it for a room), scaled up ten-thousandfold. The next slice is power conversion: every transformer and every UPS between the utility line and the server loses a little as heat, and those little losses run 24 hours a day. Lighting and offices are a rounding error. This ordering explains the whole history of PUE improvement: the industry average fell from around 2.5 in 2007 to about 1.54 by 2020 mostly by getting smarter about cooling (hot-aisle containment, higher allowed temperatures, using outside air), and it has been essentially flat since, because the easy cooling wins are taken.

The water question, answered honestly

The cheapest way to shed heat is to evaporate water, which is why many large facilities use evaporative cooling: it lowers PUE by spending water instead of electricity. That trade is real. A large data center can evaporate hundreds of millions of gallons in a year, and the industry has a separate metric for it (WUE, water usage effectiveness, in liters per kWh). Dry cooling exists and uses little water, but pays for it with more fan and chiller energy, which shows up right back in the PUE. This page deliberately measures only the electricity meter, and we would rather tell you that plainly than pretend the water is not there: if you see a remarkably low PUE in a hot, dry place, some of that efficiency may be flowing out of a cooling tower as vapor.

How big is this, really

The context numbers come from the Lawrence Berkeley National Laboratory's 2024 report to Congress, the closest thing the field has to an official ledger. US data centers consumed about 176 TWh in 2023, which was 4.4 percent of all US electricity. By our own household figure, 176 TWh is the electricity of nearly 17 million homes. The same report projects 6.7 to 12 percent by 2028, and the width of that band is itself the honest headline: nobody, including the national labs, knows how fast AI demand will actually grow, so a range stated plainly beats a point estimate defended loudly. Either way the direction is up, which is exactly why the dollars-per-PUE-point table this calculator prints has gone from a facilities-manager curiosity to a board-level number.

Why AI racks broke the old assumptions

For twenty years the planning assumption was that a rack draws 5 to 10 kW, and whole buildings (power distribution, cooling, floor loading) were designed around it. AI hardware tore that up. A modern AI training rack like NVIDIA's GB200 NVL72 draws around 120 kW, with observed peaks above 130: more than ten enterprise racks in one footprint. Air cooling tops out around 30 to 40 kW per rack even with careful design, which is why liquid cooling went from exotic to mandatory almost overnight, and why the AI buildout means new construction rather than filling old rooms. For this calculator the density question only changes how you enter the IT load (40 racks at 50 kW and 400 racks at 5 kW are the same 2 MW to the meter), but it is the reason the "count racks" mode offers hints an order of magnitude apart. If your racks are the 100 kW kind, the GPU cluster power calculator is built for exactly that math.

Sources and method

Every constant in the calculator is pinned to a published source and tested against it. Figures as of August 2026.

Industry average PUE 1.54: Uptime Institute Global Data Center Survey 2025 (weighted average, essentially unchanged for six consecutive years).
Hyperscale best in class 1.09: Google's published fleet-wide PUE, trailing twelve months across all large-scale sites, all overhead included. Uptime puts hyperscale operators generally at 1.10 to 1.15.
The 1.30 and 2.00 tiers are our editorial anchors for a well-run enterprise facility and a legacy one, labeled as such; Uptime's survey places typical colo and enterprise facilities between those marks.
Electricity price default $0.08 per kWh: the EIA industrial average was 8.62 cents in 2025 and about 9 cents in early 2026; large data centers typically negotiate industrial-class or better rates. Fully editable.
Grid CO2 intensities: EPA eGRID 2023 data (Rev 2, June 2025), identical to the constants on our EV pages by design and by test.
Context figures (4.4 percent, 176 TWh, 6.7 to 12 percent by 2028): LBNL, 2024 United States Data Center Energy Usage Report (December 2024), prepared for Congress; summarized by the Department of Energy.
Average US household 10,500 kWh per year: US Energy Information Administration, the same figure our AI energy page uses.
Rack densities: industry-average enterprise racks run 5 to 10 kW; NVIDIA's GB200 NVL72 draws roughly 120 kW per rack (130+ observed), liquid cooling required.

A real facility's PUE moves with the seasons and its load, and utility tariffs are negotiated, not listed. Treat the output as an honest planning estimate, not a metered bill.

Data Center Power Calculator