There is no reliable universal figure for the power or water used by an AI data center. A defensible estimate starts with the facility’s actual IT equipment and operating schedule, then applies a stated power usage effectiveness (PUE) for total electricity and a site water usage effectiveness (WUE) for on-site water. Water consumed to generate electricity should be calculated separately.
Use scenario ranges rather than a single number. Utilization, accelerator efficiency, cooling design, weather, grid mix, and deployment timing can move the result by orders of magnitude.
As an Amazon Associate I earn from qualifying purchases.
The calculation in four equations
For an early planning estimate, keep the calculation auditable by showing every input and unit:
What’s actually slowing this PC down?
Pick the symptom - the matching free tool is one click away.
| What you are estimating | Calculation | Unit |
|---|---|---|
| Annual IT energy | Sum of (average device kW × operating hours) for servers, accelerators, storage, and networking | kWh or MWh |
| Annual facility energy | Annual IT energy × PUE | kWh or MWh |
| Annual site water | Annual IT energy × site WUE | Liters |
| Annual electricity-related source water | Electricity consumed × an applicable generation-water-consumption factor | Liters |
Report site water and source water as separate lines. If you add them, label the result as a broader source-inclusive or lifecycle estimate and state exactly what is included.
Energy and demand are different. Kilowatt-hours (kWh) and megawatt-hours (MWh) describe energy over time; kilowatts (kW) and megawatts (MW) describe instantaneous or peak demand. A utility interconnection, transformer, UPS, or generator is sized for demand, while annual energy and water budgets depend on operating hours.
Why a single “water per AI query” number fails
A 2025 review by Lei, Lu, Shehabi, and Masanet found more than a 10,000-fold spread in workload-level water-use results across the literature it examined. The spread reflects different hardware efficiency, utilization, cooling systems, climates, electricity supplies, and accounting boundaries. It is evidence that scenarios are necessary, not a prediction for any particular query or facility.
The same caution applies to a watts-per-GPU shortcut. An accelerator’s nameplate rating is not the data center’s average load. Host CPUs, memory, storage, network equipment, voltage-conversion losses, idle periods, redundancy, and cooling all affect the result. Two sites with identical accelerator inventories can have different annual electricity and water use.
Quick wins for a faster PC:
Repair Windows errors before they cause bigger problemsFix Now →Fix the driver behind crashes, sound loss and screen glitchesFind Drivers →Define the metrics before collecting numbers
Power, energy, and load factor
- Power (kW or MW): the rate of electricity use at a moment in time. Use it for peak, average, and coincident-load planning.
- Energy (kWh or MWh): power accumulated over a period. Annual energy equals average power multiplied by the hours operated.
- Utilization: the fraction of available compute capacity actually doing work. A fleet can have high nameplate power but a much lower average draw if utilization is low.
- Coincident peak: the highest demand reached by loads at the same time. Summing each device’s individual maximum can overstate or understate the facility’s actual peak.
PUE: how to calculate total facility electricity
Power usage effectiveness is defined as:
PUE = total facility energy ÷ IT equipment energy
The theoretical minimum is 1.0, where every unit of facility energy reaches IT equipment. Real facilities consume additional electricity in chillers, pumps, fans, cooling towers, air handlers, lighting, UPS systems, switchgear, and other infrastructure. Use a PUE appropriate to the design, climate, and operating load; a national or industry average is not a substitute for a site-specific design value.
Site WUE: on-site water normalized by IT energy
Site water usage effectiveness is:
Site WUE = annual site water use in liters ÷ annual IT energy in kWh
It generally captures water used at the data center, especially for heat rejection and cooling. Confirm the boundary with the engineering team: some studies include only cooling-tower makeup, while others include additional facility water. State whether the number represents consumption, withdrawal, or another measure.
Source water: water used to make the electricity
Power plants and other generators consume water, and the intensity varies by technology, cooling method, grid mix, and geography. Estimate it separately by multiplying electricity use by a source-water factor whose definition and geographic boundary are known. A facility with very low on-site WUE can still have substantial electricity-related source water if its electricity supply is water-intensive.
Do these 3 things before closing this tab:
1Scan for outdated or missing drivers - takes under a minute2Repair Windows errors before they cause bigger problems3Fix the driver behind crashes, sound loss and screen glitchesStep 1: Set the scope, boundary, and time period
Write a one-page scope statement before doing arithmetic. Include:
- The site or sites and the weather data or climate zone used.
- The planning year and whether the result is a full-year operation or a partial-year ramp.
- Whether the boundary covers only the data hall, the entire campus, or colocated offices and support buildings.
- Whether water means consumption, withdrawal, or cooling-system makeup water.
- Whether electricity-related source water is included and which supply region is represented.
- Whether the estimate is for normal average operation, a design peak, or both.
Without these choices, two apparently precise estimates may be measuring different things.
Step 2: Build the IT load inventory
Use an equipment list rather than a single “GPU watts” assumption. Gather the following for every device class:
| Input | Why it matters |
|---|---|
| Accelerator and server model, quantity | Different models have different typical and maximum draw. |
| CPU, memory, storage, and network equipment | These loads remain part of IT energy even when accelerators are busy. |
| Average power and idle power | Separates useful-work demand from standby and low-utilization demand. |
| Expected utilization by period | Converts nameplate capacity into an operating profile. |
| Operating hours and duty cycle | Captures continuous operation, maintenance windows, and seasonal schedules. |
| Deployment and shipment schedule | A partially populated hall should not be modeled as a fully installed fleet. |
| Refresh or replacement schedule | Changes the equipment mix and annual energy over the planning horizon. |
For each class, calculate average operating kW, multiply by the hours in each operating period, and sum the periods. If you have measured telemetry, use it; otherwise label vendor figures as assumptions and distinguish typical from maximum draw.
The Tool Desk
Outbyte PC Repair FREEClear out junk files and repair common Windows errorsFree Scan →Outbyte Driver Updater FREEFix the driver behind crashes, sound loss and screen glitchesFind Drivers →A simple IT-energy formula
For equipment class i:
Annual IT energy = Σ (average poweri in kW × operating hoursi)
Apply utilization explicitly when the power figure represents full activity. For example, if a 1 MW accelerator fleet is expected to draw 70% of that level on average for 8,760 hours, its accelerator energy is 6,132 MWh before adding host, storage, and network loads. That is an illustration of the method, not a benchmark for a particular accelerator.
Step 3: Convert IT energy to total facility electricity with PUE
Once annual IT energy is known, multiply by the design PUE:
Annual facility energy = annual IT energy × PUE
Do not multiply peak demand by 8,760 hours unless peak is genuinely sustained. For demand planning, create a separate load profile with average, maximum expected, and coincident peak values. Include UPS losses, power distribution losses, cooling plant auxiliaries, pumps, fans, lighting, and controls in the PUE assumption.
Free tools Windows power users keep installed
One-click scans. No signup required.
What PUE should you use?
Use a modeled value from the proposed design when available. PUE changes with outside temperature, IT load, redundancy mode, and maintenance status. A lightly loaded facility can have a worse PUE because fixed overhead is spread over fewer IT kilowatts. If the design is not complete, use low, central, and high PUE cases and identify which systems drive the difference.
Step 4: Estimate on-site water with WUE or a cooling model
For a first pass:
Annual site water (liters) = annual IT energy (kWh) × site WUE (L/kWh)
If the cooling design is known, replace a generic WUE with a system model. The model should account for heat load, heat-rejection technology, operating temperatures, weather, economizer hours, water-treatment practice, and blowdown.
How cooling choices change water demand
- Air-cooled chillers: can be attractive when minimizing on-site water is a priority, although their electricity performance and capacity depend on outdoor conditions.
- Water-cooled chillers and cooling towers: can handle substantial cooling requirements efficiently, but evaporation and blowdown create makeup-water demand.
- Direct liquid cooling: moves heat close to the chips and can reduce air-moving requirements; the facility still needs a heat-rejection system, so water use depends on the downstream design.
- Dry or adiabatic systems: may reduce routine water use, while adiabatic operation can consume water during hot periods.
- Economizers: use suitable outdoor conditions to reduce mechanical cooling. Their benefit depends on local temperature, humidity, filtration, and control limits.
DOE Federal Energy Management Program guidance notes that cooling-tower consumption is linked to heat load and the efficiency of heat removal. Temperature and humidity control, airflow management, economizing, and maintenance affect both cooling energy and water. Evaporation is not the only demand: blowdown to control mineral concentration also requires makeup water.
Step 5: Calculate electricity-related source water separately
Start with the electricity figure that matches your boundary—usually total facility electricity—and multiply it by a documented water-consumption factor for the utility, balancing authority, or behind-the-meter supply:
Annual source water = annual electricity use × source-water factor
Check whether the factor is reported as consumption or withdrawal, and whether it covers the delivered grid mix or only generation at selected plants. A power purchase agreement or onsite generation can change the applicable factor. The 2024 Lawrence Berkeley National Laboratory analysis uses balancing-authority and power-plant data and notes that individual power purchase agreements and behind-the-meter generation are not incorporated automatically; handle those supplies explicitly in a project estimate.
Worked example with transparent assumptions
The following example demonstrates the arithmetic. It is not a forecast for a real facility.
Recommended Free Tools
| Case | Average IT load | Annual IT energy | PUE | Annual facility energy | Site WUE | Annual site water |
|---|---|---|---|---|---|---|
| Low | 1 MW | 8,760 MWh | 1.20 | 10,512 MWh | 0.10 L/kWh | 876,000 L |
| Central | 2 MW | 17,520 MWh | 1.30 | 22,776 MWh | 0.30 L/kWh | 5,256,000 L |
| High | 4 MW | 35,040 MWh | 1.60 | 56,064 MWh | 1.00 L/kWh | 35,040,000 L |
The central case assumes a steady 2 MW average IT load for 8,760 hours. Its 17,520,000 kWh of IT energy becomes 22,776,000 kWh of facility energy at PUE 1.30. Multiplying IT energy by 0.30 L/kWh gives 5,256,000 liters of on-site water. A source-water result would require a separate, documented factor; it cannot be inferred from the site WUE.
The spread between these cases comes from assumptions, not arithmetic. Presenting all three lets decision-makers see how utilization, cooling design, and climate affect the result.
Build a range and test the drivers
At minimum, vary these inputs:
- Average IT load and utilization.
- Coincident peak and ramp schedule.
- PUE at representative low, typical, and hot-weather conditions.
- Site WUE or cooling-system water model.
- Weather year and operating temperature limits.
- Grid or supply-specific source-water intensity.
- Equipment refresh timing and expected efficiency improvements.
Use sensitivity analysis to show which assumption changes annual water or electricity the most. LBNL’s June 2026 U.S. report uses equipment shipments, per-device annual electricity use, cooling performance, facility type, location, and compounded-uncertainty scenarios rather than one deterministic forecast. A project estimate should follow the same principle.
Account for peak demand and infrastructure capacity
An annual energy estimate does not tell you whether the site needs a 5 MW, 20 MW, or larger grid connection. Create a separate hourly or sub-hourly profile that includes:
- Accelerator ramp-up and batch-training peaks.
- CPU, storage, and network concurrency.
- Cooling response during hot weather.
- UPS, battery, and generator operating modes.
- Redundancy cases, maintenance transfers, and failure scenarios.
Report at least average IT load, average facility load, maximum expected facility demand, and the design electrical capacity. Label whether each figure is measured, simulated, or a nameplate limit.
Independent reader supportYour contribution helps us test, update, and keep practical guides available for everyone.Use location and electricity supply data, not national averages
Weather affects free-cooling hours, chiller efficiency, tower evaporation, and adiabatic operation. Geography also affects the water intensity of electricity generation. Gather local hourly weather data, utility tariffs and demand rules, the balancing-area region, planned onsite generation, and any contracted supply.
National figures are useful context only. LBNL’s 2026 U.S. projection estimates 649 TWh of data-center electricity in 2030 in its reference case, with a compounded-uncertainty range of 521–843 TWh. The reference estimate equals 11.8% of total U.S. electricity, with a scenario range of 9.5%–15.3%. These are U.S.-wide data-center projections, not AI-only forecasts and not a substitute for a facility model.
Tools for an early estimate
Lawrence Berkeley National Laboratory provides a PUE Estimator and DC Pro. The PUE Estimator focuses on estimating PUE; DC Pro offers a broader profile and recommendations. They are useful scoping aids when a design is still forming. They do not replace detailed electrical studies, equipment telemetry, cooling simulations, water-quality analysis, or a site-specific engineering assessment.
Common mistakes and how to correct them
Using accelerator nameplate power as annual facility energy
Why it fails: nameplate maximum ignores utilization, idle draw, non-accelerator IT, and cooling.
Correction: build a class-by-class inventory, apply an operating profile, then apply PUE.
Conflating MW with MWh
Why it fails: a demand figure is not an annual energy total.
Correction: report peak or coincident MW separately from annual MWh and show the hours behind every annual calculation.
PC Slower Than It Used to Be?
A free scan shows the junk files, broken settings and background clutter dragging Windows down - then fixes them in one click.Free scan · Windows 10 & 11Outdated Drivers Are Slowing You Down
One free scan finds every outdated or missing driver and matches the right update for your exact hardware.Free scan · exact hardware matchApplying PUE to water
Why it fails: PUE is an energy ratio. It does not convert electricity directly into liters.
Correction: use site WUE with IT kWh, or model the cooling system, and calculate source water from electricity with a separate factor.
Adding site and source water without labeling the boundary
Why it fails: readers cannot tell whether the total includes cooling only, power generation, or both.
Correction: publish separate direct/site and indirect/source lines, then label any combined total.
Quoting one WUE for every climate and design
Why it fails: cooling operation changes with weather, heat load, controls, water chemistry, and heat-rejection equipment.
Correction: use a site-specific cooling model or disclose a low-central-high WUE range.
Ignoring deployment timing
Why it fails: a hall that fills over two years will not consume a full fleet’s annual energy in year one.
Correction: model shipments, installation dates, commissioning, and refreshes by month or quarter.
What’s actually slowing this PC down?
Pick the symptom - the matching free tool is one click away.
What a credible estimate should publish
A decision-ready power-and-water estimate should include:
- An equipment inventory with model, count, average and peak power, utilization, and operating hours.
- Annual IT energy and the load profile used to derive it.
- Peak and coincident facility demand in MW.
- The PUE value, operating conditions, and included overhead systems.
- The site-WUE value or cooling model, including evaporation and blowdown treatment.
- The electricity supply region and source-water factor, with consumption-versus-withdrawal terminology.
- Low, central, and high cases with a sensitivity explanation.
- The planning year, deployment schedule, geography, climate data, and accounting boundary.
- Any exclusions, such as offices, construction energy, transmission losses, or unmodeled behind-the-meter supplies.
This format lets a utility, water manager, finance team, and data-center engineer audit the same estimate and update it when the design changes.
Quick Recap
Product prices and availability are accurate as of the date/time indicated and are subject to change. Any price and availability information displayed on Amazon at the time of purchase will apply.
