Water consumed by data center operations, including evaporative cooling.
10Resources · Evidence explorer
What does a data center actually need?
Electricity to compute. Cooling to remove heat. Water at some sites—and at the power plants supplying them. Explore the quantities, then change the assumptions to see the scale.
All data centers ≠ AI alone. Every figure below names its geography, year and boundary.
01 / Electricity demand
How much—and how fast?
Global · All data centers · Annual operational electricity · IEA, April 2026
Historical estimate · 2025485 TWh
Central projection · 2030≈950 TWh
≈1.96×projected demand relative to 2025
02505007501,000 TWh
About 3% of global electricity in the 2030 projection. AI-focused facilities reach approximately 465 TWh in 2030 in the same Base Case. This includes facility overhead; it is not electricity used solely for AI computations.
Source: IEA 2026 · Executive summary ↗
What this comparison can tell you
A national or global annual total does not show the peak demand or grid constraints in a particular place. The projection is a modeled outlook, not an observed future meter reading. No AI-only absolute quantity is inferred from these totals.
Solid bars = historical estimates. Hatched bars = projections. These are published models, not a census of facility meters. U.S. and global views use different model assumptions and historical baseline years.
02 / Two water boundaries
The building is only part of the picture.
United States · All data centers · 2023 modeled water consumption
Water consumed in generating the electricity serving those facilities.
0200400600800
Consumption is water removed from the immediate water cycle. Withdrawal is water taken from a source, some of which may return. These numbers measure consumption.
Water data table
| Boundary | 2023 consumption |
|---|---|
| Direct site operations | 66 billion L |
| Electricity generation | Nearly 800 billion L |
Source: Berkeley Lab 2024 · Report pp. 55–58 ↗
Why this is not a local water forecast
The upstream estimate uses regional grid mixes and excludes facility power purchase agreements and on-site generation. Site cooling, weather and water stress vary. The two totals cover different locations; national volumes alone cannot establish the impact on a particular watershed. Neither is an AI-only or per-query footprint.
03 / Put a facility in perspective
Turn megawatts into an annual quantity.
MW describes power. TWh describes energy accumulated over time. Adjust an illustrative facility; these inputs are assumptions, not an estimate of an existing site.
Calculated scenario · 8,760 hours
1,051 GWh / yearTotal facility electricity
Average facility draw120 MWIT electricity876 GWh / yearInfrastructure electricity175 GWh / yearDirect water consumption263 million L / year
IT 876 GWh + infrastructure 175 GWh = facility 1,051 GWh
About 0.55% of the U.S. 2024 data center electricity estimate (192 TWh, Berkeley Lab 2026).
A scenario for operations only. Upstream water, chip manufacturing, building construction and equipment disposal are outside this calculation.
Show calculation and assumptions
IT energy (GWh) = average IT load (MW) × 8,760 ÷ 1,000. Facility energy = IT energy × PUE. Infrastructure energy = facility − IT energy. Direct site water (million liters) = IT energy (GWh) × site water intensity (L/IT kWh). The selected accounting definition determines whether the result is consumption or withdrawal.
All assumptions are annual averages. A 365-day year is used. PUE and site water intensity are independently adjustable for sensitivity analysis; real combinations depend on cooling design and climate. Cooling demand is included in PUE and must not be added again. The resulting average power does not establish peak demand or a grid connection requirement.
04 / Infrastructure moves at a different pace
Computing can grow faster than its grid.
IEA indicative development ranges · Published 2026 · Planning, permitting and construction
New data center
1–3 years
New grid infrastructure
5–15 years
051015 years
Infrastructure data table
| Project | Indicative time |
|---|---|
| New data center | 1–3 years |
| New grid infrastructure | 5–15 years |
Source: IEA, Electricity 2026 · Grids chapter ↗
These ranges describe different kinds of projects. Subtracting endpoints would not establish a facility’s waiting time.
Questions to ask about an announced project
- Is the quoted MW IT power or total facility power?
- Is it peak capacity or average draw, requested or connected?
- Is the project operating, under construction or announced?
- Does the water figure describe withdrawal or consumption, on-site or electricity generation?
- Which reporting year and model assumptions support the quantity?
An Intelligence Capitalism reading: electricity, cooling and connections shape who can expand computing. Ask who controls access, pays for upgrades and bears local costs. The resource evidence alone does not establish which firm earns economic rents.
Read the boundary before the headline
What remains outside these numbers?
AI is a subset.
Servers also provide storage, cloud software and other services. Whole-sector totals cannot be assigned to AI.
Operations are not lifecycle.
These figures do not quantify embodied resources in chips, buildings or replacement equipment.
Global scale is not local capacity.
A small national share can still be a large load on one local grid. Connection requests are not electricity consumption.
Evidence ledger
Sources, definitions & methods
Source editions are fixed below so comparisons remain reproducible. Download the displayed evidence (CSV) ↓
Electricity sources and edition dates
- IEA, Key Questions on Energy and AI, April 2026: global 2025 estimate and 2030 central projection. Uses all data centers. The approximately 465 TWh AI-focused facility projection is from report p. 26 and is not an AI-computation-only quantity.
- Berkeley Lab, United States Data Center Energy Usage Report: 2025 Update, June 2026: 192 TWh in 2024; Reference Case 464 TWh in 2028 and 649 TWh in 2030, with 521–843 TWh compounded uncertainty (pp. 8–11). The interval combines scenario uncertainties, not a statistical confidence interval.
The global view uses the updated IEA 2026 outlook: a 2025 baseline and a rounded 2030 headline. Approximately 950 TWh does not establish a meaningful increase from the prior 945 TWh projection. The U.S. view uses Berkeley Lab’s June 2026 update, replacing its December 2024 electricity outlook. Its 2030 share uses the report’s NERC denominator. The models have different shipment assumptions and are not harmonized; no U.S. share of the global forecast is calculated.
Water, units and missing evidence
Berkeley Lab’s December 2024 report pp. 55–58 supports the direct and upstream water quantities. Both are modeled consumption. The approximate upstream figure is kept approximate. Site intensity uses IT electricity as its denominator; upstream intensity would use facility electricity. No site-specific water forecast or lifecycle total is calculated.
1 TWh = 1 billion kWh. 1 billion liters = 1 million cubic meters. Published historical estimates carry model uncertainty; no numerical confidence bounds are supplied here. The facility calculator has no default AI allocation or emissions factor.
This is a work derived by Core Loop Press from IEA material and Core Loop Press is solely liable and responsible for this derived work. The derived work is not endorsed by the IEA or its Member countries in any manner. IEA reuse terms. Evidence checked 8 October 2026. Adapted quantities: IEA (2026), Key Questions on Energy and AI and Electricity 2026, and Smith et al. (2026), LBNL-2001758, CC BY 4.0. Water: Shehabi et al. (2024), LBNL-2001637. Sources and assumptions can be inspected without completing a quiz.