How much power does a data center use? Ask what the megawatts mean
A developer announces plans for a 100 MW data center.
Soon the project is described as consuming as much electricity as a city, a steelworks, or tens of thousands of homes. The comparison looks precise. It may be wrong before the calculator even starts.
What does the 100 MW describe?
It could be capacity requested from the grid, power intended for IT equipment, the completed campus after several phases, only the first building, or the maximum load under design conditions. Each interpretation leads to a different answer.
This confusion matters because data-center electricity demand is rising quickly. The International Energy Agency’s 2026 outlook estimates that data centers consumed around 485 TWh worldwide in 2025 and could approach 950 TWh in 2030. It also stresses how uncertain that path remains as hardware, software, usage, investment, and infrastructure constraints change.
Those global numbers tell us something about direction. They tell us little about the facility proposed beside a particular town or substation.
To understand that facility, start with the unit.
Power is a rate. Energy is what accumulates.
Megawatts measure power at a particular moment. Megawatt-hours measure energy used over time.
The distinction is the same as speed and distance. A vehicle traveling at 100 kilometers per hour has not necessarily traveled 100 kilometers. We still need to know how long it maintained that speed.
The U.S. Energy Information Administration explains the electrical version directly: watts measure power at a moment, while watt-hours measure energy produced or consumed over a period.
If a data center draws 100 MW continuously for a full year, it consumes:
100 MW × 8,760 hours = 876,000 MWh, or 876 GWh
But a site with a 100 MW grid connection may average 70 MW across the year, consuming about 613 GWh. During an early phase, it might average 35 MW and consume roughly 307 GWh.
All three cases can belong to something described as a “100 MW data center.”
A household comparison therefore needs more than headline capacity. It must convert expected annual consumption into the same unit used for household consumption, using a clearly stated average load and a relevant household figure for the same region.
Even then, the comparison remains limited. Ten thousand homes spread across a region do not necessarily affect one substation in the same way as one facility drawing continuously from a single connection.
Annual energy matters for procurement and cost. Peak power matters for cables, transformers, substations, generation capacity, and grid planning. A responsible account needs both.
One project can contain several different power numbers
A data-center proposal may use the word “capacity” several times without describing the same thing.
| Number | What it describes | What it does not establish |
|---|---|---|
| Grid connection capacity | The electrical envelope approved, contracted, or provisioned at the site boundary | Current demand or annual electricity use |
| Ultimate campus capacity | The intended size after all planned phases are built | What will operate first or whether every phase will proceed |
| Total facility load | IT equipment plus cooling, conversion, controls, lighting, and other supporting systems | How much useful computing work is completed |
| Design IT load | Power intended for servers, storage, and networking under stated design conditions | Installed hardware, utilization, or facility overhead |
| Installed IT load | Equipment physically installed at a given point | Whether it is busy or drawing maximum power |
| Peak demand | The highest measured draw during a defined period | Typical or annual demand |
| Average demand | Average power drawn over a period | Short peaks or maximum required capacity |
| Annual consumption | Energy accumulated across the reporting year | Connection headroom or momentary demand |
| Backup power rating | Capacity intended for outages or ride-through | Normal grid consumption |
These figures are related, but none can replace the others.
Power Usage Effectiveness, or PUE, compares total facility energy with the energy reaching IT equipment. In a simplified design calculation, a facility supporting 80 MW of IT load at a PUE of 1.25 would require around 100 MW in total.
If a project instead describes itself as having 100 MW of IT capacity at the same PUE, its total requirement would be closer to 125 MW.
The phrase “100 MW data center” could describe either one. PUE also needs a reporting period, operating load, physical boundary, and climate conditions.
The arithmetic is easy. Finding out what the original number means is usually harder.
Where the electricity goes
Servers are the largest electrical load in most modern data centers, but they are not the whole facility.
The IEA’s equipment breakdown estimates that servers account for around 60% of electricity demand on average, with storage and networking taking smaller shares. Cooling and environmental control vary much more, from a modest share in efficient hyperscale facilities to more than 30% in some enterprise data centers. Power conversion, controls, lighting, security, and other systems account for the rest.
Two sites with the same IT load can therefore draw different amounts from the grid because their cooling, power distribution, climate, redundancy, and operating conditions differ.
Cooling is particularly sensitive to the site. A system operating through a hot summer may need more fan, pump, compressor, or chiller power than the same equipment in cooler conditions. Higher-density hardware can shift electricity between server fans, liquid pumps, and facility heat-rejection systems. The complete heat path shows where those loads enter.
Electrical conversion creates smaller but persistent losses. Power may pass through transformers, switchgear, uninterruptible power supplies, distribution units, and voltage-conversion equipment before reaching processors. Each stage is necessary for safety, control, resilience, or compatibility. None is perfectly lossless.
Redundancy changes the power envelope again. A facility designed with additional power and cooling equipment available during failures may need to reserve more of its provisioned capacity for supporting systems.
Uptime Institute’s 2026 analysis argues that PUE does not show how much of the original grid connection remains structurally available for computing after auxiliary loads, design margins, and redundancy are accounted for.
A site can have a large electrical connection and still fit less computing hardware than the connection number suggests.
The racks can be installed without being busy
Even IT load has several layers.
A site may be designed to support 80 MW of computing equipment, have equipment installed for 40 MW, and average well below 40 MW because it is still filling, workloads vary, or servers are underused.
Installed capacity is not actual demand. Actual demand is not useful work.
A server draws power while idle. A GPU may consume far less than its maximum rating during one workload and operate close to it during another. Hardware can be installed for resilience or growth and spend long periods underused.
Three forms of utilization should stay separate:
- Site occupancy describes how much planned IT capacity has been filled with equipment.
- Hardware utilization describes how actively processors, accelerators, storage, and networking are being used.
- Electrical utilization describes how much of the available connection or power envelope the facility is drawing.
A full data hall can contain underused hardware. A small amount of densely packed AI equipment can draw a large share of available power. A data center can also reserve more grid capacity than it currently consumes because it expects to fill over several years.
That gap may be commercially rational, but it complicates utility planning and connection queues. An ultimate design number therefore deserves a schedule: expected year-one demand, triggers for later phases, and what happens if customer demand arrives more slowly than forecast.
AI changes the shape of the load
AI increases data-center electricity use, but total energy is only part of the change. Power density is rising inside the facility.
The IEA reports that AI-server power density increased sharply between 2020 and 2025 and may rise further as new systems arrive. The significance is concentration. Cooling, busways, cables, switchgear, and backup systems must support large loads inside a small physical area.
The workload also changes how demand behaves. AI training can keep large groups of accelerators near their power envelope for sustained periods. Inference may be more variable, responding to the number, size, and timing of requests. Some workloads create rapid changes as clusters move between computational stages.
A utility therefore needs more than expected annual GWh. It needs the load profile: how rapidly demand moves, how long the facility remains near its peak, what the minimum load is, and whether any work can be delayed, reduced, or moved. Flexible demand begins with the workload’s real margin.
A 70 MW average made up of steady demand presents one grid problem. The same average produced by repeated movement between 40 MW and 100 MW presents another. Energy is the same over time. The infrastructure required to deliver it may not be.
What a useful power disclosure looks like
A credible power claim should allow the reader to reconstruct the relationship between the site, IT equipment, and time.
It should identify what the headline capacity describes; whether the figure is permitted, contracted, designed, installed, or measured; expected first-phase demand; peak and average draw; annual consumption; the share reaching IT equipment; the workload profile; loads included at the site boundary; capacity reserved for redundancy and growth; and the role of any on-site generation or storage.
A journalist may not publish every answer. A utility, regulator, buyer, or local authority will need many of them.
“100 MW data center” is the beginning of the disclosure, not the end.
Power begins with the site
Policloud develops physical, modular data-center infrastructure for defined sites. A power figure therefore has to remain attached to a configuration and a place.
Available electricity, grid connection, workload, cooling, networking, site readiness, ownership, regulation, and operating model shape what capacity can be installed and how it may be used.
Modularity can allow capacity to be considered in physical increments rather than only as the eventual total of a large campus. It does not make power requirements generic. A maximum specification does not establish average consumption. A grid connection does not prove utilization. One measured deployment does not become a fleet-wide figure.
For a prospective site, the useful process works from available power inward: what connection can be delivered, how much supports cooling and resilience, what IT load remains, which workloads fit, and what annual energy follows from the expected operating profile.
Those questions connect directly to where data-center demand should be located and what makes a viable site.
Ask what the megawatts mean
A megawatt is a real unit. The ambiguity comes from the noun attached to it: connection capacity, facility demand, IT load, installed equipment, peak draw, average draw, backup power, or ultimate buildout.
So when a data center is announced at 100 MW, do not begin by converting it into homes, cities, or power stations.
Ask where the number was measured. Ask whether it describes today or a future phase. Ask how much reaches the computing equipment and how often the facility expects to draw it.
Then multiply by time.
Only after that do 100 megawatts become an electricity bill.
