How Much Power Does a Data Center Need?

Data center power is quoted in megawatts (MW). The range is wide because the facilities are very different. A single-building edge site near a population center has little in common with a multi-building AI training campus, and their power needs differ by two to three orders of magnitude.

These are planning ranges, not rules. A colocation operator can build a 100 MW campus, and a hyperscaler can build a 30 MW facility. What matters for a site is the specific load the buyer intends to put there and the order in which it arrives. Our guide to site needs by data center type covers how these profiles differ beyond power.

Developers often describe a project by its critical IT load: the power consumed by servers, storage and network equipment. The utility has to serve more than that. Cooling systems, fans, pumps, lighting, offices and losses in transformers and UPS systems all add to the total.

The standard measure of that overhead is power usage effectiveness (PUE), a metric developed by The Green Grid. PUE is total facility energy divided by IT equipment energy. A PUE of 1.0 would mean zero overhead. Modern facilities run meaningfully above that, and the figure depends on climate, cooling design and how fully the building is loaded.

• Total facility demand ≈ IT load × PUE.
• Example: 100 MW of IT load at a design PUE of 1.3 is about 130 MW of facility demand.
• Design PUE (at peak, on the hottest day) is what sizes the utility request. Annual average PUE is lower and matters for energy cost, not interconnection.

Utilities also think in MVA rather than MW. MVA is apparent power, which includes the reactive component. Data centers generally operate at a high power factor, but a utility will still ask for both figures, and transformers and lines are rated in MVA.

Within a given building footprint, four things set how many megawatts it will draw.

• Rack density. Traditional enterprise racks have commonly run in the single-digit to low-teens kW range. Racks for AI training with liquid cooling can run many times higher, sometimes above 100 kW per rack. The same floor area can therefore need several times more power.
• Cooling approach. Air cooling, evaporative systems and direct liquid cooling have different energy overheads and different sensitivity to climate. The choice moves PUE and, with it, the facility load.
• Building count and phasing. Campuses are built one building or one data hall at a time. The full-build number may be years away.
• Non-IT loads. Offices, security, mechanical yards and on-site water treatment are small relative to IT load but still part of the request.

Redundancy is often misunderstood. A 2N electrical design installs two complete power paths, but the IT equipment only draws its load once. Redundancy increases the transformers, switchgear, UPS and generators on site. It does not double the demand placed on the grid. What it can change is the request for two independent utility sources, which affects the substation and transmission design.

AI workloads add another consideration. Large training clusters can change their power draw quickly and in step. Utilities and grid operators increasingly ask about load behavior, ride-through during faults and how fast demand can swing, not only how large it is. Our guide to AI data center site requirements goes further into those loads.

A utility does not plan around one final number. It plans around peak demand in each year and how that demand grows. A large-load request is usually expressed as a ramp: for example, 50 MW in the first year of service, 150 MW two years later and 300 MW at full build.

Data centers are also high-load-factor customers. They run close to their peak around the clock, with little daily or seasonal variation compared with most commercial or industrial loads. That makes them attractive to some utilities, because the energy sold per MW of capacity is high. It also means the grid has to carry that load during its own peak hours, with no natural off-peak relief.

A credible ramp matters. Utilities increasingly ask for load ramps tied to financial commitments, and a ramp that overstates early demand can trigger cost obligations the project cannot meet. The large-load interconnection process is where the ramp becomes binding.

Power and land move together, but power is usually the constraint. A 300-acre tract with access to 40 MW is a colocation site, not a hyperscale campus, however large it is. The reverse also happens: a site with a strong transmission path may be held back by too little buildable acreage once setbacks, wetlands and a substation footprint come out.

Higher rack densities are pushing more megawatts onto each acre. That raises the importance of electrical yard space, substation pads, cooling equipment and, for some designs, room for on-site generation. Our guide to how much land a data center needs works through the land side of the math.

1. 01Start with the target product: edge, colocation, hyperscale or AI campus, and the likely IT load at full build.
2. 02Apply a design PUE suited to the climate and cooling approach to get facility demand in MW, then convert to MVA.
3. 03Break the full build into phases with target energization years.
4. 04Decide whether the design needs one utility source or two independent sources.
5. 05Compare that ramp with what the nearby transmission system and substations can plausibly deliver, and when.

The output is a load profile you can take to a utility. It also becomes the test for every site on a shortlist: can this location deliver this ramp on this schedule? When we screen a site, that question comes before almost any other.