How much power does an AI data center need?

“A 100 MW data center” can mean installed IT nameplate, expected IT draw, maximum facility demand, requested utility service, or generation capacity. Those figures answer different questions. Before comparing two projects, ask which boundary the number crosses and whether it is a maximum, an operating estimate, or a future buildout target.

IT power is what servers, network equipment, and storage consume at the equipment input. Facility power also includes cooling and electrical losses. Utility service may be requested for the future full campus even when the first operating hall is much smaller. For example, a campus designed for four halls can energize one hall first while reserving electrical space for the remaining three. Its first-day demand and ultimate service request should be shown as separate lines.

The Lawrence Berkeley National Laboratory 2024 U.S. Data Center Energy Usage Report estimated U.S. data centers used about 176 TWh of electricity in 2023 and modeled 325–580 TWh in 2028. Those are annual energy totals across data centers, not the peak demand of an individual AI campus. The IEA Energy and AI report projected global data-center electricity consumption around 945 TWh in 2030 in its base case. It likewise does not tell a developer how large a particular substation must be.

For an initial electrical sketch, multiply the number of concurrently operating racks by the specified kilowatts per rack. Divide by 1,000 to express the answer in megawatts. Use the vendor’s rack-level input specification for the actual configuration, since a GPU thermal figure omits host CPUs, networking, conversion, and cooling equipment inside the rack.

NVIDIA’s GB200 NVL72 systems FAQ gives a roughly 120 kW rack figure. Schneider Electric’s GB300 NVL72 reference designs describe support for rack densities up to 142 kW. These are different platforms and design scopes; 142 kW should not be silently assigned to every GB200 rack.

A 120 kW rack consumes 0.12 MW at the assumed rack input. One hundred such racks would therefore be 12 MW of IT load if all reach that input together. The multiplication is simple; the engineering judgment lies in which racks can run together, their actual configured limit, and the headroom needed for future equipment.

PUE is total data-center energy divided by IT equipment energy over the same measurement period. The LBNL report uses that definition when modeling facility electricity. A PUE of 1.2 means 60 MW of IT power corresponds to 72 MW at the facility boundary under that operating assumption: 12 MW is non-IT load and losses.

PUE is often reported as an annual measured ratio, whereas a utility asks about maximum demand at a particular point. An annual PUE should not be multiplied by a peak IT figure without checking whether cooling and electrical losses at peak are represented. A facility in a hot afternoon can have a different overhead from the same facility in mild weather. The electrical design needs a coincident peak calculation, not just an annual average.

The Uptime Institute 2025 Global Data Center Survey reported a weighted-average PUE of 1.54 among respondents. That is a survey benchmark, not a design guarantee for a new liquid-cooled AI hall. If a developer assumes 1.2, ask which cooling design, climate, operating hours, and redundancy plan support it. If those inputs change, update the facility estimate.

Rack-scale AI systems concentrate substantial electrical and thermal load in a small footprint. NVIDIA’s GB200 documentation describes the roughly 120 kW class. At that density, an electrical designer must check busway and rack distribution, while a mechanical designer must move the resulting heat out of the rack. The room’s floor area alone does not indicate the required megawatts.

ASHRAE’s AI data-center design guidance treats power, liquid cooling, controls, and water as linked decisions. A water-cooled rack does not make the processor heat disappear. It transfers heat to another part of the facility, where it must be rejected under the site’s weather and water conditions. A 60 MW IT hall creates roughly 60 MW of heat from the IT equipment while it is drawing 60 MW; the additional facility power also ends as heat somewhere in the system.

The rack plan must therefore be paired with a cooling plan. If the project changes from one rack platform to another, update rack input, liquid flow, heat rejection, floor loading, and distribution together. Keeping the same “MW campus” label while changing the rack platform can hide a serious local design change.

A load can fit within an annual energy budget yet cause trouble when it changes quickly. Large AI jobs can synchronize work across many processors. NVIDIA describes power-smoothing features for GB300 NVL72 intended to limit sharp rack-level changes during training and shutdown. The existence of those controls is evidence that the dynamic load matters; it is not proof that every installed cluster has identical ramp behavior.

A utility or on-site generator needs the shape of the load, not only its maximum. For example, a training cluster that climbs from idle to full operation across many racks may create a steeper step than a campus of independent services with the same maximum MW. The design team should supply expected ramps, job-scheduling controls, UPS behavior, and any rack-level power caps. An average meter reading cannot substitute for that profile.

The response plan may involve workload controls, electrical storage, generator controls, or a utility operating agreement. Each has a different duration and response time. Ask what event the proposed equipment actually handles: a sub-second change, a multi-minute ramp, or a long grid outage.

Assume a first phase with 500 racks, each specified at 120 kW input. The installed IT calculation is 500 × 120 kW = 60,000 kW, or 60 MW. With a provisional PUE of 1.2, the facility arithmetic gives 72 MW. The calculation does not say that all racks will run at maximum at all times, and it does not include an unbuilt second phase.

Now create three rows for the project team. First, show installed IT capacity: 60 MW. Second, show modeled peak facility demand: 72 MW under the stated 1.2 assumption, subject to a coincident-peak study. Third, show requested service or on-site plant capacity after engineering evaluates growth, startup loads, resilience architecture, and any operational limits. The third row cannot be inferred from PUE alone.

If only half the racks are installed at opening, the first-phase IT figure is 30 MW at the same assumed per-rack input. The facility estimate at the same PUE would be 36 MW. That does not mean the utility can provide exactly 36 MW on the date the first racks arrive. The utility must evaluate the actual connection and planned stages under its service process. The generator interconnection queue concerns generator projects; it does not assign a large-load facility a queue position by this multiplication.

Carry the assumptions beside the numbers: rack specification, rack count by phase, power cap if any, weather design condition, cooling scheme, PUE basis, expected concurrency, and measurement boundary. A later change to any one of these can move the facility figure.

A utility-facing load request needs location, service voltage, desired energization dates, phase-by-phase demand, load factor, ramping behavior, and the customer’s proposed electrical one-line. The utility can then study its network and identify required facilities and agreements. The detailed route varies by utility and region; a generator application under FERC Order No. 2023 is not the universal process for a data-center load.

The regulatory distinction is live as of November 2026. DOE’s October 2025 Section 403 directive asked FERC to address large-load interconnections, and FERC’s June 2026 action issued show-cause orders to six regional operators. Those steps should not be described as a final nationwide large-load rule.

An on-site generation proposal creates a separate plant design and may also create a generator interconnection matter if the project will export or connect under relevant rules. DOE’s large-load microgrid discussion emphasizes coordination among customers, utilities, and microgrid resources. For the equipment and operating boundary, see behind-the-meter power.

When the rack and facility assumptions are documented, an equipment inquiry can name the required power block, fuel, operating duty, site location, and target schedule. Send those inputs with a quote request for generation or electrical equipment. We can then confirm what our sourcing network can identify and make an introduction where appropriate. The engineering and utility study determine the site’s actual service design.

Do not multiply GPU thermal design power by GPU count and call the result facility demand; it leaves out the rest of the rack and the building. Do not treat a measured annual PUE as a peak design factor without checking weather and concurrency. Do not call a future full-campus service request the first-day load. Do not treat a generator queue statistic as the large-load connection process. Each mistake changes the question that the MW number is meant to answer. Label every figure with its measurement boundary and operating phase before sharing it with a utility or equipment supplier.