How to Read Data Center Energy and Water Claims

How to Read Data Center Energy and Water Claims

A data-center number becomes meaningful only after you know its unit, period, geography, system boundary, denominator, operating assumption, and whether it was measured or modeled. Two impressive numbers may be impossible to compare even when both are technically correct.

The fastest test is to rewrite the claim in a complete sentence. Name what was measured, where, when, how, against what denominator, and with what uncertainty.

Begin with the claim boundary

"This data center uses 300 megawatts" sounds precise. It may describe requested capacity, contracted capacity, peak demand, an ultimate phase, or on-site generation. It does not state annual electricity consumption.

"This facility uses five million gallons of water" is also incomplete. The period is missing. So are the source, withdrawal or consumption boundary, direct or indirect boundary, measured or modeled status, operating load, and site identity.

flowchart TD
    A["Published energy or water claim"] --> B["Unit and time period"]
    B --> C["Facility, campus, company, product, or supply-chain boundary"]
    C --> D["Denominator and operating assumptions"]
    D --> E["Water source or electricity supply"]
    E --> F["Measured, estimated, modeled, or projected"]
    F --> G["Comparable claim with stated uncertainty"]

Do not move to comparison until each step is visible.

MW is power and MWh is energy

Watts measure power at a moment. Watt-hours measure energy over time. The US Energy Information Administration explains that one megawatt, or MW, equals one million watts. One megawatt sustained for one hour equals one megawatt-hour, or MWh.

A facility drawing 100 MW continuously for one hour consumes 100 MWh. If it maintained that draw for every hour of a 365-day year, it would consume 876,000 MWh, or 876 gigawatt-hours.

The arithmetic is:

100 MW × 24 hours × 365 days = 876,000 MWh

Real facilities do not necessarily operate at nameplate or contracted capacity every hour. A public estimate should state the assumed load factor or utilization.

Never convert MW directly into homes powered without stating a time period, assumed household use, geography, and facility utilization. A capacity analogy is not an energy comparison.

PUE measures facility overhead

Power usage effectiveness, or PUE, is:

annual total data-center energy ÷ annual IT-equipment energy

The Department of Energy's data-center design guide uses this annual ratio. A PUE of 1.20 means the facility used 1.20 units of total energy for each unit delivered to servers, storage, and networking equipment during the stated period.

If a facility used 120 MWh in total and 100 MWh reached IT equipment, its PUE was 1.20.

PUE helps identify energy used by cooling, power conversion, lighting, and other infrastructure. It does not measure the efficiency or usefulness of the computing workload. It does not establish total electricity, carbon emissions, cost, grid impact, or water use.

A lower PUE can coexist with higher total energy. A campus may improve from 1.30 to 1.15 while installing enough computing equipment to increase total demand.

PUE comparisons also need method alignment. Ask whether the value is measured or a design target, which meters and buildings are included, whether it covers a full year, how shared systems are allocated, and whether the operating load and climate are comparable.

WUE measures water against IT energy

Site water usage effectiveness, or WUE, is commonly expressed as:

annual site water use ÷ annual IT-equipment energy

The result is often liters per kilowatt-hour. A site WUE of 0.40 L/kWh means the stated site-water boundary used 0.40 liters for each kilowatt-hour delivered to IT equipment during the period.

WUE needs the same scrutiny as PUE. Ask which water flows are included, whether "use" means delivery, withdrawal, or consumption, whether potable and reclaimed supplies are separated, and whether the denominator is measured IT energy.

The 2024 Lawrence Berkeley National Laboratory report modeled a US average site WUE just above 0.36 L/kWh through 2023. That fleet-level modeled value should not be assigned to a specific facility.

Source WUE expands the boundary to water associated with energy production under a stated method. Site WUE and source WUE answer different questions. One concerns water at the facility boundary. The other adds upstream supply assumptions that may involve other watersheds.

Withdrawal and consumption are different

The US Geological Survey defines withdrawal as water removed from groundwater or surface water for use. Consumptive use is water that does not return to the immediate local water environment, including evaporation and incorporation into products.

A facility can receive or withdraw water, consume part through evaporation, and discharge another part. A cooling system can also recirculate water many times internally. Gross circulation is not the same as new water delivered to the site.

The words "used," "needed," and "saved" do not tell you which quantity is being reported.

Potable and reclaimed water are not synonyms

Potable water is treated for drinking. Reclaimed water is wastewater, stormwater, or another source treated for an intended reuse. The EPA explains that treatment needs depend on the source and end use.

Replacing potable water with reclaimed supply can reduce pressure on drinking-water systems. It does not mean the facility consumes no water. It may require treatment, pipes, pumps, backup, contracts, monitoring, and energy.

Ask for each source as an absolute annual and peak-day volume. Percentages can hide growth. A facility can reduce the potable share while increasing total water.

Direct and indirect use belong on separate lines

Direct electricity is the energy entering the facility boundary. Indirect effects may include generation, transmission losses, equipment manufacturing, or other supply-chain activity.

Direct site water can support cooling and facility operations. Indirect water may be associated with electricity generation or equipment production.

A life-cycle study can combine these effects under a defined method. A project permit or utility contract may need to keep them separate because different institutions control them.

When a claim says "AI uses water," ask whether it describes one product response, a data center, a company portfolio, electricity generation, equipment manufacturing, or a modeled life cycle. Those are not interchangeable.

Measured, estimated, modeled, and projected

A meter reading records operation within the meter's boundary. An estimate derives a value from partial evidence or assumptions. A model represents a system. A projection extends assumptions into the future.

LBNL's United States Data Center Energy Usage Report: 2025 Update is explicit about its model. Its 2030 reference case is 649 terawatt-hours, with a compound uncertainty range of 521 to 843 terawatt-hours. The model uses planned equipment shipments, per-device energy assumptions, cooling simulations, and facility information.

The reference case should not be written as "data centers will use 649 TWh." It is more accurate to say that LBNL's reference case estimates 649 TWh in 2030 under the report's stated assumptions, with a wider uncertainty range.

For a proposed campus, ask whether an energy or water number is a design value at full buildout, an expected first-year value, a permit limit, a contract quantity, or measured operation. Each has a different use.

Percent reductions need a baseline

"Uses 30 percent less energy" is incomplete. Less than what?

The baseline might be an older facility, a code minimum, a design alternative, a previous hardware generation, an unloaded test, a modeled counterfactual, or the same site before expansion. The percentage may cover cooling energy while total facility energy rises.

Rewrite the statement with the baseline year, boundary, absolute starting and ending values, workload, utilization, weather, and whether the comparison was measured.

An absolute value beside the percentage prevents a common distortion. Reducing WUE by 20 percent while doubling IT load can increase total site water.

Company averages do not decide local projects

A portfolio average can describe a company under its reporting method. It may combine climates, grid regions, cooling systems, owned facilities, leased capacity, and different workloads.

A proposed campus needs site evidence. Ask for the design PUE and WUE, expected operating range, peak and annual energy, peak and annual water, cooling modes, climate assumptions, metering plan, and public reporting method.

The Virginia JLARC report found that data-center water use was a small share of statewide withdrawals in 2023 while representing a larger share at some reviewed utilities. That is a useful warning against scale mismatch. Statewide, utility-area, watershed, and facility values answer different questions.

The seven-part rewrite

The method can be remembered as one sentence:

[Named facility or fleet] [measured, estimated, modeled, or projected] [quantity and unit] during [period], within [system boundary and geography], divided by or compared with [denominator or baseline], under [utilization, climate, phase, and method], according to [dated source].

Here is a weak claim:

This data center uses 300 MW.

Here is a useful rewrite:

The developer requested up to 300 MW of utility capacity for the campus's ultimate phase by 2030; the public record reviewed does not yet state measured peak demand or annual MWh.

Here is another weak claim:

The new cooling design saves 50 percent of water.

Here is a useful rewrite:

The design model projects 50 percent less potable cooling-water consumption per annual IT kWh than the stated baseline system at the same modeled load and climate; total site water, reclaimed supply, and measured operation are not yet available.

The rewritten claims sound less dramatic because they stop pretending that unknowns are facts. They are also more useful.

Use the metric to reach the decision

Metrics are not the end of the analysis. Once the claim is clear, ask what decision it informs.

MW and load ramp affect grid planning and contracts. MWh affects energy supply and cost. PUE helps explain facility overhead. Site WUE helps explain direct water intensity. Withdrawal and consumption affect different parts of the water system. Potable and reclaimed volumes affect different infrastructure and scarcity questions.

[[What AI Data Centers Use Electricity and Water For]] supplies the physical model. [[How Data Center Costs Reach Local Communities]] connects the numbers to public decisions. E076's [[Data Center Cooling Options and Their Water Tradeoffs]] shows why measurement boundaries can change when the cooling architecture changes.

This guide was freshly written from the preserved E075 transcript and authoritative definitions reviewed on July 27, 2026. AI assistance was used for research organization, drafting, and validation. Publication remains unauthorized.