The idea, explained
Data centres turn electricity into computation and heat. Cooling systems remove that heat. Some facilities use evaporative cooling, which can reduce electricity used by mechanical chillers but consumes water through evaporation. Other systems use more air cooling, closed loops or reclaimed water. There is no universally best design: climate, water stress, electricity mix, reliability needs and local infrastructure change the trade-off.
The measurement boundary matters. Direct water is used at the data-centre site. Indirect water may be used to generate the electricity feeding the facility. Manufacturing servers and semiconductors also has a water footprint, but that belongs to a broader life-cycle boundary. A facility-level WUE number usually does not capture every indirect or supply-chain use, so two published figures may not be comparable unless their boundaries and time periods match.
A common annual WUE calculation is total on-site water consumption divided by IT-equipment energy consumption. If litres and kilowatt-hours are used, the result is litres per kilowatt-hour. A lower value means less consumed water for each unit of IT energy within that boundary. It does not reveal total water use by itself. A very large efficient facility can consume more water overall than a small inefficient one.
Location turns an engineering metric into a community question. One litre consumed in a water-abundant watershed does not create the same local pressure as one litre consumed during drought in a stressed basin. Seasonal timing, use of potable versus reclaimed water, return-water quality and competition with homes, farms or hospitals matter. The OECD notes that reporting remains incomplete, while the World Bank emphasizes resource-efficient and climate-resilient data-centre design. Responsible evaluation therefore pairs efficiency with absolute volume, source, watershed conditions and transparent boundaries.
Follow a worked example.
Consider a hypothetical data centre. During one year it withdraws 60 million litres for cooling and returns 15 million litres to the same local system. Its on-site consumption is 60 − 15 = 45 million litres. The servers and networking equipment—the IT load—use 50 million kilowatt-hours of electricity.
Step 1: choose the stated boundary: annual on-site consumption divided by IT energy. Step 2: calculate WUE = 45,000,000 litres ÷ 50,000,000 kilowatt-hours = 0.9 litres per kilowatt-hour. Step 3: interpret carefully: within this boundary, each kilowatt-hour used by IT equipment is associated with 0.9 litres of on-site consumed water.
Now suppose an upgrade reduces WUE to 0.7, but demand grows and IT energy rises to 80 million kilowatt-hours. Annual consumption becomes 0.7 × 80,000,000 = 56 million litres. Efficiency improved by about 22%, yet total water consumption rose from 45 million to 56 million litres because workload grew faster. Decision-makers should report both numbers and examine whether the extra 11 million litres occurs in a stressed watershed or can be supplied with reclaimed water.
Where people get stuck
Common misconception: a lower WUE automatically means a smaller local water impact. WUE is an intensity ratio, not total consumption, and it can omit electricity-generation and manufacturing water. Another mistake is comparing facilities without checking whether they report withdrawal or consumption, use the same units, cover the same seasons and include the same operational boundary.
Now try it yourself.
1. Check the core idea
Facility A reports WUE of 0.5 litres per kilowatt-hour and uses 100 million kilowatt-hours of IT energy. Facility B reports WUE of 0.8 and uses 40 million kilowatt-hours. Which consumes less on-site water within the stated boundary, and what must be checked before judging environmental impact?
Reveal the answer & reasoning
A consumes 0.5 × 100 million = 50 million litres. B consumes 0.8 × 40 million = 32 million litres, so B consumes less in absolute terms even though its WUE is worse. Before judging impact, verify that both figures use the same definition and period, then examine water source, watershed stress, seasonal timing, potable versus reclaimed water, returned-water quality, indirect electricity-related water and whether workload or service output differs.
Why this was added
The library explains AI inference and infrastructure reliability but lacks a lesson on water accounting, WUE, measurement boundaries and the difference between efficiency and total local consumption.