Cooling: Why "Use Less Water" Can Mean "Burn More Gas"
The single most-misunderstood technical fact in the AI-water discourse: on-site water use and total environmental impact are not the same direction. A data center that consumes zero on-site water is usually consuming more total water (and CO₂) than a comparable one with a cooling tower. To see why, you have to understand the three cooling regimes.
The three regimes
1. Evaporative cooling (cooling towers)
The dominant technology in older / hyperscale facilities. Hot water from the chiller loop is sprayed across packing material; some of it evaporates, carrying away the heat. The cycle then needs makeup water to replace what evaporated.
- On-site water: High. Roughly 0.5–1.5 L per kWh of IT load (varies by climate; arid + hot is worst).
- Electricity overhead: Low. Typical PUE 1.2–1.4.
- Why it's used: Cheapest at scale, especially in temperate climates with cheap water.
2. Closed-loop liquid cooling
Coolant circulates between server cold-plates and an external heat rejection system (dry cooler or chiller). The loop itself is sealed — water in, no water out except small purge volumes for water quality.
- On-site water: Very low. Effectively zero in steady-state, ~50–100 mL per kWh for makeup.
- Electricity overhead: Modest. ~3–5%.
- Why it's used: Increasingly standard for high-density GPU racks (H100, B200, GB200) — the heat density is too high for air, and water is too contentious in many siting contexts.
3. Air cooling
Outside air pushed across the racks (with or without an evaporative pre-cool stage). Heat is rejected directly to the atmosphere.
- On-site water: ~Zero (pure air-side) or low (with adiabatic assist).
- Electricity overhead: Notable. ~10% extra power for the same IT load.
- Why it's used: Cold-climate sites (Iceland, Nordics), or where water is prohibitively expensive or politically untenable.
The trade-off table
| Cooling | On-site water | Power overhead | Total water (US grid avg) | CO₂ |
|---|---|---|---|---|
| Evaporative | High | 0% | High direct + moderate indirect | Low |
| Closed-loop liquid | Very low | ~5% | Low direct + moderate indirect | Slightly higher |
| Air-cooled | ~0 | ~10% | ~0 direct + higher indirect | Notably higher |
Why the indirect column matters. The US grid in 2026 is still ~56% fossil-fired (EIA Electric Power Monthly). Each extra kWh of cooling overhead pulls additional water consumption out of a coal or gas plant somewhere upstream — roughly 0.5–1 L per kWh of fossil generation, per Macknick et al.. So if you switch a hyperscale site from evaporative to air-cooled to "save water" on the local utility bill, you are typically adding water consumption at the power plant — and adding ~10% to its CO₂ footprint. (OECD.AI on this trade-off.)
This is the scope_2_water.md story in microcosm: the on-site water number is a poor proxy for the environmental impact you actually care about.
The real-world migration
Hyperscalers are largely moving toward closed-loop and air-side cooling — not because of activist pressure, but because:
- GPU thermal density (B200 at ~1,200 W per chip; rack densities of 100+ kW) is too high for traditional CRAC + cold-aisle air alone.
- Water permitting in arid regions has become slow and politically expensive.
- As grids decarbonise, the indirect-water penalty of air cooling shrinks toward zero — making it strictly dominant on water terms.
Microsoft has publicly committed to "zero-water-consumption" cooling for new sites starting 2024. Google and Meta have made similar moves for new builds in water-stressed regions.
What this means for the editorial
- "AI uses water for cooling" is true but trivial. The interesting question is which water, and what the alternative would be.
- "Switch to air cooling" is not the obvious win it sounds like. It typically means more CO₂ today, and only neutral water savings once the grid is clean.
- The most meaningful single intervention to reduce AI water consumption is decarbonising the electricity supply — the same intervention that addresses essentially every other AI-environmental concern.
Sources cited on this page
- Macknick et al. (NREL TP-6A20-50900) — fossil-plant water consumption per kWh
- EIA Electric Power Monthly — US grid fuel mix
- OECD.AI — Ways to minimise water use — the cooling-mode tradeoff
- Per-cooling-mode water/energy numbers in the table above are industry rules-of-thumb (Uptime Institute, LBNL); pin to specific studies before publication
- Full bibliography: sources.md
Loose ends to track down before publication
- Pin per-cooling-mode water/energy figures to a specific Uptime Institute or LBNL study.
- Microsoft's "zero-water" commitment — fetch the original announcement, confirm what "zero" means (truly closed-loop vs adiabatic with negligible makeup).
- Confirm the GB200 / B200 thermal envelopes from NVIDIA's published specs.