AI water use: does the water cycle make it harmless?
A popular Reddit thread asks a fair question: if water used by AI in data centres returns to the environment via the water cycle, why is it considered harmful?
“Wouldn’t the water that AI uses be returned to the environment via the water cycle?”
Short answer: the water cycle is real, but timing, location, quality, and quantity matter. Data centres often consume water in ways that remove it locally (via evaporation) when it’s needed most, and electricity generation adds a hidden water footprint too. That combination can stress local ecosystems and infrastructure, even if the global cycle balances out eventually.
Here’s what’s going on, and why it matters in the UK.
How data centres and AI actually use water
Cooling 101: withdrawals vs consumption
Most modern data centres cool servers using one of three approaches:
- Closed-loop chillers (mechanical cooling) that mostly recirculate water, often with minimal losses.
- Evaporative or adiabatic cooling that deliberately evaporates water to carry away heat. Evaporated water leaves the local system immediately.
- Once-through cooling (more common at power plants than data centres) that returns warmed water to a source.
The crucial distinction is between water withdrawn (taken from a source) and water consumed (not returned to the same source in usable form). Evaporation is consumptive use: that water is now in the air, potentially falling as rain somewhere else and much later.
The electricity behind AI also uses water
The water footprint of AI isn’t just cooling. Power generation matters. Thermal power stations (gas, coal, nuclear) use water for steam cycles and cooling; some of that is consumed through evaporation or lost as warm discharge. Wind and solar have comparatively low operational water use. So the water impact of a model run depends on both the data centre’s cooling and the grid mix supplying it.
Why the water cycle doesn’t solve it on the ground
Local and seasonal stress
Water availability is local and seasonal. A data centre ramping up cooling on a hot, dry day can consume large volumes right when rivers are low and households or farms are also under pressure. Even if that water eventually returns as rain, it may not return to that catchment or at the time it’s needed.
Quality matters: discharge, heat, and treatment
When water is returned rather than evaporated, it may be warmer or contain treatment chemicals (within permitted limits). Warmer water can affect aquatic life. Treatment and transport add energy and costs for utilities and communities.
Geography and equity
Evaporated water typically precipitates downwind, not necessarily in the same basin. “Returned to the environment” doesn’t guarantee “returned to the community or ecosystem it was taken from.” That’s the equity gap the water cycle doesn’t close.
Metrics that matter: PUE, WUE, and transparency
Two useful metrics show up in data centre sustainability reports:
- PUE (Power Usage Effectiveness) – ratio of total facility power to IT equipment power. Closer to 1.0 is better.
- WUE (Water Usage Effectiveness) – water used for cooling per unit of IT energy (often litres per kWh). Lower is better.
Some providers report WUE by region and season, and whether they use potable, surface, groundwater, or reclaimed wastewater. Specific numbers for individual facilities are often not disclosed, which makes comparisons tricky. When evaluating vendors or regions, look for WUE, source type (recycled vs potable), and independent verification.
Is AI special? Training vs inference and when it runs
Large model training is power- and cooling-intensive over weeks to months. Inference (serving user requests) is lower per operation but continuous at scale. The water impact from AI spikes during:
- Big training runs scheduled during warmer months or in hotter regions.
- High-traffic inference during heatwaves when cooling loads rise sharply.
On the plus side, AI workloads can often be scheduled or shifted. Training can move to cooler seasons or water-abundant regions; inference can route to data centres with lower water stress, subject to latency and data residency constraints.
UK context: what developers and businesses should know
The UK’s climate helps, but water stress is not uniform. The South East is officially water-stressed, and summers are getting hotter. Data centre clusters around Slough and West London, plus new builds around the M25 and in the Thames corridor, concentrate demand for power and cooling where water systems are already tight.
Expect tighter planning conditions on new facilities: requirements for recycled water, caps on potable use, and more mechanical or dry cooling. The Environment Agency and local authorities can restrict abstraction, especially near sensitive chalk streams. This all feeds into cost, siting, and capacity decisions that ultimately affect cloud pricing and regional availability.
Grid decarbonisation in the UK reduces the water footprint of electricity over time (more wind and solar), but flexible gas plants and nuclear still play a role. Where your cloud region sources its electricity will influence its overall water use.
Good practice: how the industry can reduce AI’s water footprint
- Use recycled or non-potable water (reclaimed wastewater, rainwater harvesting, or seawater where feasible) for cooling.
- Adopt dry or hybrid cooling to reduce consumptive evaporation, accepting higher capex or energy trade-offs during heatwaves.
- Site in cooler, water-abundant regions for training; route inference dynamically to lower-stress sites where latency allows.
- Report WUE transparently by region and season, including source type and treatment.
- Pair with low-water electricity (wind and solar) and minimise warm-water discharge impacts.
Practical takeaways for UK teams using AI
- Choose regions thoughtfully. If data residency allows, pick cloud regions with cooler climates and better water availability.
- Ask vendors for WUE and source type. If it’s not disclosed, request it. Preference for recycled water is a strong signal.
- Batch and schedule heavy workloads. Train at night or in cooler months; avoid peak demand windows.
- Optimise models. Smaller models, pruning, quantisation, and retrieval-augmented generation (RAG) can cut compute, energy, and water.
- Consider sustainability in procurement. Include WUE, energy mix, and reporting in your vendor assessments alongside cost and latency.
Bottom line: the water cycle isn’t a free pass
Yes, water returns to the environment, but not necessarily to the same place, at the same time, or at the same quality. AI’s water use matters because it can intensify local stress precisely when communities and ecosystems are most vulnerable. The good news is that technology and policy give us levers: better cooling, smarter siting, transparent reporting, and shifting workloads can make a meaningful difference.
If you’re building with AI in the UK, treat water as a first-class constraint alongside price and performance. Ask for the data, choose regions wisely, and optimise your workloads.
