How Much Power Does a Data Center Use? Rack, Site and Nation

How much power does a data center use? Search for it and you get a different answer on every page: 1 MW, 50 MW, 650 MW, 5 kW a rack, 120 kW a rack. Most of those numbers have no source. Some mix up power and energy. Others compare the servers alone with the whole building.

This guide gives the answer at three levels, each tied to a public source: a single rack, a single facility, and data centers across the US and the world. Sources: Lawrence Berkeley National Laboratory (LBNL), the Department of Energy (DOE) and the International Energy Agency (IEA). We left out any number we could not verify there.

Power or energy: what does "how much" mean?

Data center power consumption is quoted in two different units, and mixing them is the most common reason figures don't line up. Power is the rate of draw at a moment, measured in kilowatts (kW) or megawatts (MW). Energy is power used over time, measured in kilowatt-hours (kWh), megawatt-hours (MWh) or terawatt-hours (TWh).

A site that draws 1 MW for one hour uses 1 MWh. If it drew a steady 100 MW for a full year (8,760 hours), it would use 876,000 MWh, or about 0.88 TWh. Real sites rarely run flat out all year, so their annual energy is lower than that ceiling.

• MW answers "how big is the site?" It sizes the grid connection or on-site generation, the transformers and the cooling.
• MWh and TWh answer "how much electricity did it use?" National statistics, bills and emissions estimates use energy.
• When someone says a data center "uses 100 MW", they almost always mean its capacity, not what it drew on an average day.

LBNL makes the same point in its 2025 update. It notes that converting between annual energy and power capacity is hard without data on average demand, peak demand and the size of the grid connection. It estimates that data centers use only around half of their interconnection capacity on average, because of redundancy and maintenance needs.

How much power does a data center rack use?

A rack of conventional servers used about 3 to 5 kW, and a rack of AI servers can use as much as 100 kW. That is how Arman Shehabi, who leads LBNL's data center analysis team, described the shift in a 2024 Berkeley Lab interview.

A December 2025 report from Pacific Northwest National Laboratory (PNNL), published by DOE, gives a similar picture for today's halls:

• Modern racks are typically in the tens of kilowatts for traditional applications.
• Racks for AI training run at over 100 kW.
• Present-day GPU racks attain 120 kW.
• Rack prototypes of up to 1 MW have been developed and are planned for 2027.

The exact figure matters less than the jump. A rack that draws twenty times more power also produces twenty times more heat in the same floor space. Air cooling struggles at those densities, which is why power-dense AI halls move to liquid cooling.

How many megawatts does a data center use?

The IEA puts a conventional data center at around 10 to 25 MW, and a hyperscale, AI-focused data center at 100 MW or more. The IEA says a 100 MW AI site uses about as much electricity in a year as 100,000 households.

To size a facility, start from the rack figures above. Multiply the rack count by the power per rack, then add the cooling and power losses covered in the next section. A rough illustration, not a real site:

Large sites are not one giant electrical system. PNNL describes large data centers as built from repeated power blocks, each rated for several megawatts, so a campus grows block by block.

So when you read that a data center "uses" a certain number of megawatts, ask three questions. Is it the IT load or the whole facility? Is it the design capacity or the average draw? Is it one building or a whole campus?

IT load vs facility load: where cooling fits

The servers are not the only load. Cooling, power conversion, lighting and the rest of the building add overhead on top of the IT equipment, and cooling is the largest part of it.

The standard measure is power usage effectiveness (PUE): total facility energy divided by IT equipment energy. A PUE of 1.0 would mean every unit of power reaches the servers. LBNL's figures:

• The US national average PUE improved from 1.55 in 2018 to 1.45 in 2024.
• Over the same years, infrastructure fell from 36% to 31% of total data center electricity.
• For facilities serving AI equipment, LBNL estimates an average PUE of 1.145 in 2024, helped by liquid cooling.
• In its reference case, the aggregate PUE reaches 1.36 in 2030.

In practice: at a PUE of 1.45, a hall with 10 MW of IT load needs about 14.5 MW at the facility level. At 1.145, the same IT load needs about 11.5 MW. When two sources quote different megawatts for the same building, one is often counting IT load and the other the whole facility.

How much energy do data centers use in the US and worldwide?

US data centers used 192 TWh in 2024, or 4.7% of total US electricity, according to LBNL's 2025 update. That was a 14% increase on 2023.

Most pages online still quote the earlier 2024 report. Its figures, as DOE summarized them:

• 58 TWh in 2014, rising to 176 TWh in 2023.
• About 4.4% of total US electricity in 2023.
• Load growth tripled over the past decade.
• A projection of 325 to 580 TWh by 2028, or about 6.7 to 12% of US electricity.

The 2025 update looks further out. Its reference case has US data centers using 649 TWh in 2030, or 11.8% of forecast US electricity, within a range of 521 to 843 TWh (9.5 to 15.3%). It also estimates that 148 GW of interconnection capacity would be needed to serve them in 2030. AI servers make up 55% of all data center energy use in that 2030 reference case.

Worldwide, the IEA estimates data centers used around 415 TWh in 2024, about 1.5% of global electricity. In its base case, that roughly doubles to around 945 TWh by 2030.

Why do the numbers online disagree?

Five things cause most of the conflict. Check them before you trust any figure, ours included:

1. Power vs energy. A MW figure and a TWh figure answer different questions.
2. IT vs facility. Some figures count only servers; others include cooling and power losses.
3. Capacity vs actual draw. Design capacity and grid requests are well above average use. LBNL estimates average use at around half of interconnection capacity.
4. Year of the data. The 2023 figure (4.4%) and the 2024 figure (4.7%) come from two different LBNL reports. Pages quoting one or the other can both be right for their year.
5. Scope. LBNL leaves cryptocurrency mining out of its data center totals. Other estimates may include it, or count a whole campus as one site.

What this means if you are planning AI capacity

For AI capacity, the constraint is now power, not floor space. At 100 kW a rack, a modest AI deployment needs megawatts of firm power, plus the cooling to match. Data center power requirements for AI are set by the hardware's rated draw, so the questions to ask a site are about power first:

• How many MW are available for IT load, and how many in total including cooling?
• Can that power be delivered today, or is it waiting on a grid study or upgrade?
• What backup or spare generation is in place if a unit fails?
• Can the cooling handle your rack density, air or liquid?
• Is there room to add more power blocks on the same site later?

The grid is where many projects wait. Large new loads need studies and upgrades before they connect, which is why the data center interconnection queue is now a planning problem of its own. One response is to put generation on the site itself. Read how behind-the-meter data centers work, and how gas-powered data centers supply firm power on site.

If you bring your own hardware, the model you are probably looking for is GPU colocation: you own the servers and someone else provides the power, cooling and building.