How Much Power Does a Data Center Use?

A data center described as “100 MW” does not necessarily draw 100 MW from the grid every minute of the year. In fact, the label is incomplete until we know what the 100 MW refers to.

It could mean the facility is designed to support 100 MW of IT equipment. It could refer to a utility connection. It could describe contracted capacity that has not yet been fully deployed. And if the 100 MW is IT load, the building itself will require more than 100 MW once cooling and electrical overhead are included.

Those numbers are related, but they are not interchangeable. That distinction is the best place to start when trying to understand how much power a data center actually uses.

Power and energy are different measurements

A megawatt measures power at a point in time. A megawatt-hour measures energy consumed across time. Data center discussions often move between the two so quickly that it is easy to lose track of which one is being used.

Hold a 1 MW load continuously for a full non-leap year and it consumes 8,760 MWh, or 8.76 GWh. A 10 MW continuous load consumes 87.6 GWh. A 100 MW continuous load consumes 876 GWh.

Those figures describe the load being measured. If they represent IT equipment rather than the whole facility, infrastructure overhead still has to be added.

PUE is the bridge between IT load and facility demand

Power usage effectiveness (PUE) is total data center energy divided by IT equipment energy. If a facility runs 10 MW of IT load at a PUE of 1.50, its average total facility load is approximately 15 MW.

The latest Uptime Institute Global Data Center Survey says average PUE continues to improve gradually in 2026, while newer facilities tend to perform better than older infrastructure. Uptime's more detailed 2026 analysis puts the industry-wide average at about 1.52 and the capacity-weighted average at about 1.36. Uptime explains why larger and newer facilities pull the capacity-weighted figure lower.

That difference is financially and electrically meaningful. At a steady 10 MW IT load, PUE 1.52 implies about 15.2 MW at the facility level. PUE 1.36 implies about 13.6 MW.

PUE should not be treated as a fixed property of every hour of operation. It changes with load, weather, cooling configuration and other operating conditions. For budgeting, a representative annual PUE can be useful. For engineering or utility planning, the load profile matters more.

What 1 MW, 10 MW and 100 MW look like over a year

The following examples assume a constant IT load and PUE of 1.52. They are not “typical data center sizes”; they are scale markers that make the arithmetic easier to see.

The 100 MW example is the one that tends to surprise people. At continuous full IT load and PUE 1.52, the facility would use roughly 1.33 terawatt-hours per year. But again, a real facility may spend years ramping toward its design capacity and may never operate every server at maximum power continuously.

Nameplate capacity is not the electricity bill

Servers have power-supply ratings. Racks have design densities. Data halls have planned capacity. Buildings have utility connections. None of those automatically tells us the average operating load.

For an operating-cost model, the bottom of that funnel is what matters most. For utility and capacity planning, the upper layers matter because the system still has to be capable of serving future and peak requirements.

A useful sanity check is to ask whether the number being quoted describes what the facility can use or what it is using. A surprising amount of confusion disappears once that one question is answered.

Rack density is rising, but the extreme numbers are not the whole market

AI has made 50 kW, 80 kW and even higher rack densities part of normal industry discussion. That does not mean every new rack is suddenly operating at those levels.

Uptime Institute's 2026 survey says average modal rack densities are still moving upward gradually, while a growing number of operators now report peak rack densities of 30 kW or more. In a separate 2026 analysis, Uptime notes that racks above 50 kW are becoming increasingly common in AI environments, particularly for sustained training workloads. The survey deliberately distinguishes typical density from peak high-density deployments.

This matters when estimating building power from rack count. One hundred racks at 5 kW represent 500 kW of IT load. One hundred racks at 50 kW represent 5 MW. The floor-space difference may be modest; the power and cooling difference is an order of magnitude.

One hundred racks can mean 0.5 MW or 5 MW

Rack count is therefore useful only when density is attached to it. This is the same reason colocation cost per rack becomes less informative for high-density deployments and pricing moves toward committed kW.

The U.S. total is already large — and the range for 2028 is much larger

Berkeley Lab's U.S. Data Center Energy Usage Report estimates that data centers consumed about 176 TWh of electricity in 2023, equivalent to roughly 4.4% of total U.S. electricity consumption. The report models a 2028 range of 325–580 TWh, or about 6.7%–12% of projected U.S. electricity use. Berkeley Lab presents a range rather than one forecast because GPU shipments, utilization and cooling choices remain uncertain.

The width of that range is important. It is not uncertainty to be hidden; it is the result of variables that genuinely matter. AI server shipments, average operational power, utilization, cooling technology and efficiency can all move national demand materially.

AI changes load shape as well as load size

A conventional enterprise environment often has substantial variation between installed capacity, normal utilization and peak demand. Some AI training clusters behave differently because expensive accelerators can operate close to their power envelope for long periods.

Uptime's 2026 work on AI-era capacity allocation specifically notes that training workloads can remain near peak power for extended periods, while inference can behave more like conventional compute. That difference is one reason the industry is looking beyond simple installed-capacity metrics.

Those sketches are conceptual, not measured traces. The point is that two deployments with identical nameplate capacity can create different average and peak demands depending on workload behavior.

Colocation adds another distinction: committed power versus actual consumption

A tenant may contract 500 kW so the capacity is available when needed while drawing substantially less during normal operation. Whether the economics follow committed capacity, metered consumption or some combination depends on the contract.

This distinction matters for both electricity calculations and colocation pricing per kW. A contract may charge primarily for the reserved infrastructure even when the customer's real-time draw is lower.

Which power number should you use?

A headline such as “100 MW data center” is therefore only a starting point. For capacity planning it may be exactly the right number. For annual consumption it can be wildly incomplete.

The most useful mental model is simple: capacity tells you what the infrastructure can support; load tells you what is being used; PUE turns IT load into facility demand; time turns power into energy. Keep those four ideas separate and most data center power calculations become much easier to audit.