# Water Use and PUE in Liquid-Cooled Data Centers

Understand Water Usage Effectiveness (WUE), Power Usage Effectiveness (PUE), and the trade-offs between water-free and evaporative cooling for AI data centers.

Updated 2026-11-01. Also on the [web page](https://liquidcooling.fyi/guides/water-use-and-pue).

Liquid cooling changes both power efficiency (PUE) and water efficiency (WUE). This guide explains the metrics, trade-offs, and design choices for AI data centers.

## Key metrics

### Power Usage Effectiveness (PUE)
**PUE = Total Facility Energy / IT Equipment Energy**

- Ideal: 1.0 (all energy goes to IT)
- Typical air-cooled: 1.4-1.8
- Modern air-cooled hyperscale: 1.1-1.3
- Liquid-cooled (direct-to-chip): 1.05-1.15
- Liquid-cooled (immersion): 1.01-1.08

Source: Wikipedia PUE (2026); industry reported values.

### Water Usage Effectiveness (WUE)
**WUE = Annual Water Consumption (L) / Annual IT Energy (kWh)**

- Units: L/kWh (liters per kWh of IT energy)
- Air-cooled with evaporative cooling: ~2.6-4.3 L/kWh (PUE 1.3-1.6, COC 3-10)
- Air-cooled with dry coolers: ~0 L/kWh (but higher PUE)
- Liquid-cooled with dry coolers (closed-loop): <0.1 L/kWh
- Liquid-cooled with cooling towers: ~2.1-3.2 L/kWh (PUE 1.03-1.2, COC 3-10)

Source: The Green Grid; industry reported values; Brazil REDATA regime (MP 1.318/2025).

## Water use by cooling architecture

| Architecture | Heat rejection | Water use | PUE range | Notes |
|--------------|----------------|-----------|-----------|-------|
| Air + cooling tower | Evaporative | High (2.6-4.3 L/kWh) | 1.3-1.6 | Legacy standard |
| Air + dry cooler | Sensible (air) | Zero | 1.4-1.8 | High fan energy, climate-limited |
| Air + adiabatic cooler | Hybrid | Medium (0.2-0.5 L/kWh) | 1.2-1.4 | Water only at peak |
| Direct-to-chip + dry cooler | Sensible (liquid-to-air) | **Zero (closed-loop)** | 1.05-1.15 | Microsoft Fairwater model |
| Direct-to-chip + cooling tower | Evaporative | High (2.2-3.2 L/kWh) | 1.1-1.2 | Open facility loop |
| Immersion + dry cooler | Sensible (liquid-to-air) | **Zero (closed-loop)** | 1.01-1.08 | Highest density, best PUE |
| Immersion + cooling tower | Evaporative | High (2.1-3.0 L/kWh) | 1.03-1.10 | Fluid loop closed; facility loop open |

## The PUE-WUE trade-off

```
                    PUE
    1.8 ┤ Air + dry cooler (zero water, high PUE)
        │
    1.6 ┤ Air + cooling tower
        │
    1.4 ┤ Air + adiabatic
        │
    1.2 ┤ ◄── Direct-to-chip + cooling tower
        │         │
    1.1 ┤         │
        │         │
    1.05┤         ◄── Direct-to-chip + dry cooler (sweet spot)
        │
    1.02┤──────────────────► Immersion + dry cooler (best PUE, zero water)
        │
    1.0 ┤
        └──────────────────────────────────► WUE (L/kWh)
             0        0.1      0.5      1.0      2.0
```

**Sweet spot**: Direct-to-chip or immersion with dry coolers (closed-loop) achieves both low PUE (~1.05-1.15) and near-zero WUE.

## Design choices for water-free liquid cooling

### 1. Closed-loop facility water (dry coolers)
- Dry coolers reject heat to ambient air via sensible cooling
- No evaporation = zero water consumption
- Requires W-class high enough for dry cooler approach temperatures
- **W27-W32**: Dry coolers viable in most climates
- **W32-W45**: Dry coolers viable even in hot climates; adiabatic assist only at peak

### 2. Adiabatic assist (minimal water)
- Dry coolers with evaporative pre-cooling at peak ambient
- Water used only 50-200 hours/year in typical climates
- WUE ~0.01-0.05 L/kWh
- Good compromise for W27-W32 in warm climates

### 3. Cooling towers (high water, lower facility Capex)
- Evaporative heat rejection; lower Capex than dry coolers
- Water use: ~2.1-3.2 L/kWh for liquid-cooled systems, scaling with PUE and cycles of concentration (see worked example below)
- Requires water treatment, blowdown management, legionella control
- **Not compatible** with water-positive/net-zero water commitments

## Real-world examples

| Operator | Site | Cooling | PUE | WUE | Source |
|----------|------|---------|-----|-----|--------|
| Google | Fleet (2025) | Water and air cooling, chosen per site with a water risk framework | 1.09 fleet-wide average (report PDF p. 15) | WUE not stated; 10.9 billion gallons (41 billion liters) of water consumed across data centers and offices in 2025, up 34% from 2024 (report PDF p. 35) | Google 2026 Environmental Report |
| Switch | SUPERNAP 7, NV | Air + advanced HVAC | 1.18 | Not disclosed | Wikipedia PUE, 2026 |
| Allied Control | Hong Kong | Two-phase immersion (3M Novec 7100) | 1.02 | Low (closed-loop) | Wikipedia PUE, 2026 |
| OVH | Europe/NA | Proprietary liquid cooling | 1.09-1.29 | Low | Wikipedia PUE, 2026 |
| Research Institutes of Sweden | Luleå | Direct-to-chip, free cooling | 1.015 | ~0 | Wikipedia PUE, 2026 |

## Regulatory landscape

| Jurisdiction | Requirement | Impact on cooling choice |
|--------------|-------------|--------------------------|
| **Brazil (REDATA, MP 1.318/2025)** | Water-efficiency conditions for tax incentives | Favors dry coolers or high-COPT cooling towers |
| **Germany (EnEfG)** | Annual Annex 3 reporting for ≥300 kW | Drives efficiency transparency; favors low PUE/WUE |
| **EU (proposed)** | Minimum data center performance standards (Q2 2027) | Likely to set PUE/WUE thresholds |
| **China (GB 40879-2021)** | Covers air-cooled; liquid cooling scope expansion proposed | May mandate liquid cooling efficiency metrics |
| **US (FERC/interconnection)** | Large-load tariff reform | Indirect; favors efficient designs that reduce peak demand |

Source: Mordor Intelligence, 2026.

## Calculating your WUE

**For closed-loop dry cooler system:**
```
WUE = (Makeup water for facility loop + Humidification + Irrigation + Domestic) / IT kWh
    ≈ 0.001-0.01 L/kWh (mainly humidification and minor leaks)
```

**For cooling tower system:**
```
WUE = (Evaporation + Blowdown + Drift) / IT kWh
Evaporation = Heat rejection × 1.8 L/kWh_th
Blowdown = Evaporation / (COC - 1)
    where Heat rejection ≈ IT load × PUE (all facility power ends up as heat)
    and COC = Cycles of Concentration (typically 3-10)
```
Example: 1 MW IT load, PUE 1.15, heat rejection ≈ 1.15 MW_th, COC=5
→ Evaporation = 1.15 MW × 1.8 L/kWh_th × 8760 h = 18,133,200 L/yr
→ Blowdown = Evaporation / (COC-1) = 4,533,300 L/yr
→ Total = ~22.67 ML/yr → WUE = 22.67 ML / (1 MW × 8760) = **~2.6 L/kWh**

## Recommendations for AI data centers

1. **Target closed-loop dry coolers** for new builds. Zero water, PUE ~1.1, future-proof for regulation.
2. **Specify W32 or higher** to maximize dry cooler hours. Match to OEM server spec (NVIDIA GB200 NVL72, AMD Helios).
3. **If retrofit with cooling tower**: Maximize cycles of concentration (COC ≥ 10 with advanced treatment), monitor WUE monthly.
4. **Report both PUE and WUE** at facility and rack level. Rack-level PUE enables optimization.
5. **Design for waste heat capture**: Liquid cooling at W32-W45 produces 40-50°C return water usable for district heating (Mordor Intelligence, 2026 notes Brescia, Italy district heating pilot).

## Sources

- Wikipedia. "Power usage effectiveness." Last edited 17 October 2026. https://en.wikipedia.org/wiki/Power_usage_effectiveness
- Wikipedia. "Water usage effectiveness." https://en.wikipedia.org/wiki/Water_usage_effectiveness
- Mordor Intelligence. "Data Center Liquid Cooling Market Report." 2026. https://www.mordorintelligence.com/industry-reports/data-center-liquid-cooling-market
- ASHRAE TC 9.9. "Thermal Guidelines for Data Processing Environments, 5th Edition." 2021. https://www.ashrae.org/technical-resources/bookstore/thermal-guidelines-for-data-processing-environments
- Microsoft. "Made in Wisconsin: The world's most powerful AI datacenter." September 2025. https://blogs.microsoft.com/on-the-issues/2025/09/18/made-in-wisconsin-the-worlds-most-powerful-ai-datacenter/
- Google. "Google 2026 Environmental Report." Fleet-wide PUE on PDF p. 15 (also p. 71); water consumption on PDF p. 35. https://sustainability.google/files/google-2026-environmental-report.pdf
- Brazil. Medida Provisória 1.318/2025 (REDATA tax regime): sustainability and water-efficiency conditions for qualifying data centers. https://www.planalto.gov.br/ccivil_03/_ato2023-2026/2025/mpv/mpv1318.htm
- The Green Grid. "Water Usage Effectiveness: A Green Grid Metric." White paper.

## Questions

### Does liquid cooling always save water?

Not necessarily. Direct-to-chip with dry coolers (closed-loop) eliminates evaporative water use. But liquid cooling with cooling towers can use more water than air-cooled systems if not designed for water efficiency. Microsoft's Fairwater AI datacenter relies on a closed-loop liquid cooling system filled once during construction; only a small portion of the facility uses water for cooling on the hottest days (Microsoft, 2025).

### What is a good WUE target for a liquid-cooled AI data center?

WUE < 0.1 L/kWh is achievable with closed-loop dry coolers. Brazil's REDATA tax regime (MP 1.318/2025) sets water-efficiency conditions for qualifying data centers. Cooling-tower designs land around 2.1-3.2 L/kWh at design load.

### How does PUE change with liquid cooling?

PUE improves because fan energy is eliminated and heat rejection is more efficient. Allied Control claimed PUE 1.02 with two-phase immersion (Wikipedia PUE, 2026). Google reports a 2025 fleet-wide average PUE of 1.09 (Google 2026 Environmental Report, PDF p. 15). Switch SUPERNAP 7: 1.18 third-party audited (Wikipedia PUE, 2026).

### Can I claim PUE 1.0 with liquid cooling?

No. PUE = Total Facility Energy / IT Equipment Energy. Even with perfect heat rejection, you have pump energy, control systems, lighting, and power distribution losses. Theoretical minimum is ~1.01-1.02 for two-phase immersion with passive heat rejection.