Correction — October 10, 2026. Our earlier version printed 12 trillion litres for the IEA’s 2030 projection. The correct figure is about 1.2 trillion litres in its Base Case. We also removed unsupported land-area forecasts and company comparisons.
What the forecast actually says
The IEA’s Energy and AI (2025), printed page 242, estimates annual data-centre water consumption at about 560 billion litres in 2023 and about 1,200 billion litres in 2030 in its Base Case. This is a scenario for data centres, not a measurement of an individual chatbot or of AI alone.
The report includes water associated with primary energy supply and electricity generation, direct cooling and chip manufacturing. Electricity-related use accounts for roughly two-thirds and direct cooling roughly one-quarter in its estimate. Comparing that total with a cooling-only number changes the boundary and produces a misleading comparison.
Reproduce the unit conversion
| Quantity | Calculation | Meaning |
|---|---|---|
| 2023 estimate | 560 billion L ÷ 1,000 billion L/km³ = 0.56 km³ | Annual consumption in the report’s boundary |
| 2030 Base Case | 1,200 billion L ÷ 1,000 = 1.2 km³ | Projected annual consumption, not observed use |
| Ratio | 1,200 ÷ 560 ≈ 2.14 | About 2.14 times the earlier estimate |
| Relative change | (1,200 − 560) ÷ 560 × 100 ≈ 114% | An increase of about 114%, using rounded inputs |
A billion is 10⁹ and a trillion is 10¹². The previous error arose from adding a zero to the billion-litre figure. Keep the original unit next to any conversion so that a reader can reproduce it. Ratios from rounded inputs are approximate; they do not create extra precision.
A worksheet for comparing two facilities
| Record for each facility | Why it changes the interpretation |
|---|---|
| Period and status | Use the same year; mark observed, modelled or proposed. |
| Water measure | Distinguish withdrawal from consumption. A withdrawal figure cannot simply replace a consumption figure. |
| Included activities | List direct cooling, electricity supply and manufacturing separately. |
| Location and water source | Identify the local basin, season, potable or reclaimed source where documented. A global total does not establish local stress. |
| Output denominator | Name the actual denominator: energy, floor area, work completed or another measured output. |
| Uncertainty | Keep missing supplier data and estimates visible instead of treating blanks as zero. |
Suppose a fictional cooling system consumes 1,000 m³ every day for 365 days. Its arithmetic total is 365,000 m³, or 365 million litres. This is only a worked example. It excludes electricity-supply and manufacturing water, assumes constant operation and says nothing about local scarcity. If a second system reports withdrawals, first obtain its consumption figure before ranking the two.
How to read a land claim
“The site uses 20 hectares” is incomplete unless the writer specifies whether this means the fenced parcel, buildings, supporting electrical equipment or a wider supply chain. For a local planning decision, record the planning document, parcel boundary and land-use baseline. A proposed footprint is not evidence that construction occurred.
Water and land also need different denominators. Do not turn a national land estimate into a water-efficiency score or infer a per-query footprint from annual global totals. A defensible request to an operator asks for the facility, period, measured boundary and method together.
Sources and scope
- IEA, Energy and AI (2025), page 242: Supports the 2023 estimate, 2030 Base Case and water-use components. It does not provide a verified per-query figure for this page.
AI-assisted drafting. The linked material supports the stated facts; the examples and worksheets are Luminesca’s own editorial work. We did not conduct facility measurements, commercial project due diligence or model benchmarking for this article. Send a correction with the specific passage and supporting source.