AI Data Centers and Water Use

⬅️ Hub: Artificial Intelligence · Documentary claim: Kenyan data centers · Policy: water protections · Environment: Environment

Bottom line

AI does not inherently require dramatically more water per megawatt than cloud computing. Cooling architecture, climate, workload, and power source determine water intensity. The U.S. fleet-level model used here rises from just over 0.36 L/kWh through 2023 to about 0.45–0.48 L/kWh by 2028, an increase of roughly 25–33% per unit of IT electricity. The much larger local issue is scale: 2026–2027 AI projects are commonly discussed in hundreds of megawatts, and some campuses reach gigawatt scale. A 100 MW water-assisted facility can therefore consume roughly 394–420 million liters per year, while a genuinely non-evaporative closed-loop design can use almost no continuing water for cooling.

Quick answers

Measurement discipline

Water claims become misleading when unlike quantities are treated as interchangeable.

Boundary rule

Every credible claim should state withdrawal or consumption, direct or indirect, potable or non-potable source, annual volume or peak rate, and actual operation or modeled design.

How data-center cooling actually works

flowchart TD
    A[Servers and GPUs create heat] --> B[Internal air or liquid circuit]
    B --> C[Heat exchanger]
    C --> D{Final heat rejection}
    D --> E[Dry cooler or air-cooled chiller]
    D --> F[Hybrid or adiabatic cooler]
    D --> G[Water-cooled chiller and cooling tower]
    E --> H[Very low continuing site water]
    F --> I[Water evaporates during wet mode]
    G --> J[Evaporation, blowdown, and drift]
    I --> K[Makeup water required]
    J --> K

The internal loop

High-density AI racks commonly circulate water or a water–glycol mixture through cold plates attached to CPUs and GPUs. The warmed fluid passes through a heat exchanger, cools, and returns to the racks. That circuit can remain sealed and reuse the same coolant for years.

NVIDIA specifies approximately 120 kW per rack for a GB200 NVL72 system; its rack contains liquid-cooling manifolds because ordinary room-air cooling becomes increasingly difficult at that density. This describes how heat leaves the chips, not how the building finally releases that heat outdoors. NVIDIA hardware documentation

Final heat rejection

The U.S. Department of Energy explains cooling-tower makeup demand as:

Makeup water = evaporation + blowdown + drift

Evaporation removes heat. Blowdown discharges some mineral-concentrated water to control scaling and corrosion. Drift is the smaller loss of airborne droplets. DOE cooling-tower guidance

What “closed loop” proves—and what it does not

“Closed loop” is meaningful only when the speaker identifies which loop:

1. A closed server-coolant loop means the fluid touching the chips is recirculated.
2. A closed data-hall air loop means outside air does not enter the server room.
3. Neither fact establishes how the heat exchanger releases heat outdoors.
4. If heat rejection uses evaporation, the facility still needs continuing makeup water.
5. Only a closed internal loop paired with non-evaporative dry heat rejection supports a near-zero-water cooling claim.

Microsoft describes a newer design in which coolant recirculates between servers and chillers while air-cooled equipment rejects the heat without normal water evaporation. It projects more than 125 million liters avoided per data center per year, based on its FY2024 fleet WUE; pilot sites are scheduled to begin operating in late 2027. Water is still used for sanitation and other non-cooling purposes. Microsoft, “Next-generation datacenters consume zero water for cooling”

Precise formulation

A closed internal loop can coexist with either near-zero continuing water use or substantial evaporative makeup demand. Ask how the facility rejects heat, not merely whether its coolant recirculates.

Why AI changes the physical plant

AI workloads affect infrastructure in two distinct ways:

• Density: rack-scale GPU systems concentrate far more electrical load and heat in a small area, making direct-to-chip liquid cooling attractive or necessary.
• Scale: planned AI campuses increasingly involve hundreds of megawatts or gigawatts rather than a single modest enterprise room.
• Utilization: training clusters can sustain heavy loads for long periods, although nameplate capacity should never be treated as measured annual use.
• Networking: dense AI clusters add high-speed networking and supporting power equipment.
• Efficiency: newer accelerators may perform much more computation per watt, so water per useful computation can decline even while total site demand rises.

OpenAI and SB Energy announced a 1.2 GW data-center lease with initial facilities entering service from 2026, while OpenAI’s broader Stargate announcements described nearly 7 GW of planned capacity across multiple sites in 2025. These are portfolios and multi-phase campuses rather than evidence of a statistically “average” AI facility. OpenAI–SB Energy · Five additional Stargate sites

Comparing 2020 cloud computing with 2026–2027 AI

There is no single valid “2020 cloud WUE” or “2027 AI WUE.” Representative operator disclosures and the best public U.S. modeling establish a range:

• Meta reported a fleet WUE of 0.30 L/kWh in 2020. Meta ESG data index
• Microsoft later reported 0.49 L/kWh in 2021, improving to 0.30 in FY2024 and 0.27 in FY2025. Microsoft efficiency data
• Lawrence Berkeley National Laboratory modeled average U.S. site WUE at just over 0.36 L/kWh through 2023, rising to approximately 0.45–0.48 by 2028 as hyperscale, colocation, and liquid-cooled AI systems expand. This is a fleet projection, not an AI-only meter reading. LBNL, 2024 United States Data Center Energy Usage Report

Using the LBNL fleet series, the normalized increase is approximately:

0.45–0.48 ÷ 0.36 = 1.25–1.33, or about 25–33% more direct site water per kWh.

The range across actual designs is wider: near zero for non-evaporative cooling, roughly 0.1–0.3 L/kWh reported for some hyperscale adiabatic systems, and substantially higher for water-cooled chillers and cooling towers in difficult climates.

Water per continuously operating IT megawatt

The scenario formula is:

Annual liters = IT MW × 1,000 kW/MW × 8,760 hours/year × WUE

Scenario boundary

These calculations assume operation at 100% of stated IT capacity for an entire year. At an 80% average load, multiply every volume by 0.8. They estimate on-site cooling/humidification water from WUE; they do not include chip fabrication, construction, sanitation, or the water used to generate electricity.

Whole-facility scenarios

For an illustrative 2020-era cloud facility, this note uses 35 MW as a transparent midpoint scenario rather than claiming that it was the statistical industry average.

The central lesson is that the largest multiplier is usually facility size, not an inherent AI-water coefficient. A 100 MW center has almost three times the IT capacity of the 35 MW comparison; a 1.2 GW campus has more than 34 times the capacity.

How many people’s household water is that?

Population-equivalent comparisons depend on the human-use benchmark:

Sources: U.S. EPA household-use estimate · England water-resource performance · Nairobi City Water, August 2024 update · WASREB Impact Report 16

A representative 100 MW water-assisted AI facility

At 0.45–0.48 L/kWh, the annual 394–420 million liters equals the household or basic-service volume of approximately:

At a more efficient 0.20 L/kWh, the same 100 MW facility would equal approximately 1,400 Americans, 3,500 English residents, 9,600 people at 50 L/day, or 17,100 people at 28 L/day.

A 1.2 GW water-assisted AI campus

At 0.45–0.48 L/kWh, the annual 4.73–5.05 billion liters equals approximately:

Do not turn an equivalence into a diversion claim

These comparisons communicate volume, not proof that a facility deprived that number of people of water. Household withdrawals may return through sewers; evaporated cooling water leaves the immediate watershed. A center may use reclaimed water rather than drinking water. Seasonal peak demand can also matter more locally than the annual total.

Kenya case study

The documentary claim audited in Ghost in the Machine combines a real scarcity problem with an unproved site-specific causal claim.

Facilities and project status

• iXAfrica NBOX1, Nairobi: operating since 2024. iXAfrica lists 4.5 MW live IT capacity in the first building and 22.5 MW planned across the campus. It specifies autonomous indirect adiabatic cooling. Campus specifications
• Nxtra by Airtel, Tatu City: construction began in September 2025. The official plan is a 44 MW, high-density, GPU-ready facility targeted for 2027. Airtel Africa · Kenyan ICT Ministry
• Microsoft–G42, Olkaria: announced in 2024 as a $1 billion geothermal-powered project with water-conservation technology. Reporting in May 2026 said it had stalled over power capacity and financial guarantees, so it should not be counted as an operating campus. Microsoft announcement · 2026 status report
• Africa Data Centres, Nairobi: the operator says its expansion does not use water in the cooling system, demonstrating that Kenyan data centers do not share one cooling architecture. Operator statement

What NBOX1’s “indirect adiabatic” design implies

The data-hall air or internal coolant can remain closed while a separate outside stream evaporates to pre-cool air at the heat exchanger. The system may operate dry during favorable conditions and activate wet mode during hotter weather or higher loads.

iXAfrica identifies the utility company as its primary water source and a treated borehole as backup, but it does not publish:

• liters withdrawn or consumed per day or year
• WUE
• wet-mode operating hours
• the utility/borehole share
• seasonal peaks
• blowdown or discharge quantities

iXAfrica sustainability disclosures

Hypothetical calculation only—not a disclosure by the Kenyan facilities

If the completed 22.5 MW NBOX1 campus operated continuously at the modeled 0.45–0.48 L/kWh range, the result would be 88.7–94.6 million liters per year, equivalent to 4,860–5,184 people at 50 L/day. If the 44 MW Nxtra facility did the same, the result would be 173–185 million liters, equivalent to 9,500–10,100 people at 50 L/day. Actual values could be much lower or higher; neither operator has disclosed the inputs needed to make this estimate site-specific.

Kenya’s water context

• Kenya’s Ministry of Water classifies the country as water-scarce, with approximately 452 cubic meters of renewable freshwater per person per year, below the common 1,000-cubic-meter threshold. Kenya Ministry of Water
• The 2022 Kenya Demographic and Health Survey reported 68% national access to at least basic drinking-water service, 65% with sufficient quantities when needed, and 99% at least-basic access in Nairobi City. “Basic” does not mean a continuously pressurized household tap. KNBS survey
• Nairobi County’s 2023–2027 development plan estimated demand at approximately 850,000 m³/day and supply at approximately 525,000 m³/day, leaving a deficit of about 325,000 m³/day, or 38% of demand. Nairobi County Integrated Development Plan
• Field research in informal settlements documents intermittent service, reliance on kiosks and vendors, unequal costs, and water-quality concerns. PLOS Water study

Kenya verdict

What regulators and communities should demand

The strongest policy response is disclosure tied to local watershed conditions, not a universal claim that every AI facility has the same footprint. A permit or community-benefit agreement should identify:

1. Nameplate IT MW and expected annual load factor.
2. Rack density and cooling architecture.
3. Monthly withdrawal and consumption, not merely annual totals.
4. Cooling WUE and the exact reporting boundary.
5. Potable, reclaimed, rainwater, surface-water, and groundwater shares.
6. Dry-season and drought-stage operating plans.
7. Wet-mode hours for adiabatic or hybrid systems.
8. Blowdown volume, chemistry, treatment, and discharge destination.
9. Electricity source and source-water footprint.
10. Water permits, borehole yield, aquifer monitoring, and utility-capacity studies.
11. Who pays for mains, wells, treatment, storage, and emergency supply.
12. Actual operating data after commissioning, audited against the design promise.

This evidence sharpens the closed-loop proposal discussed in AI policy positions: require non-evaporative heat rejection or quantified low-water operation, not merely a closed server-coolant loop.

Claim audit

Research conclusions

1. At equal IT power, AI is not inherently a large water multiplier. Cooling system and climate dominate direct WUE.
2. Scale is the central risk. Hundreds of megawatts or gigawatts multiplied by continuous operation can produce city-scale annual volumes even at moderate WUE.
3. Closed-loop language is insufficient. The complete design must identify the final heat sink and whether evaporation occurs.
4. Dry cooling moves part of the burden from water to electricity. Source water and grid emissions must therefore accompany site WUE.
5. Local context matters more than global averages. A liter in a water-stressed basin during the dry season is not equivalent to a liter of reclaimed water in a water-abundant region.
6. Kenya’s risk warrants scrutiny, but present diversion claims exceed the disclosed evidence. The immediate evidence gap is operator-level metering and public reporting.

Core sources

Cooling, energy, and U.S. modeling

• Lawrence Berkeley National Laboratory, 2024 United States Data Center Energy Usage Report
• U.S. Department of Energy, Cooling Tower Management
• U.S. Department of Energy, Cooling Water Efficiency Opportunities for Federal Data Centers
• Microsoft data-center PUE and WUE disclosures
• Microsoft non-evaporative AI cooling design
• Meta ESG data index
• NVIDIA GB200 NVL72 hardware documentation

Planned AI infrastructure

• OpenAI and SB Energy: 1.2 GW lease
• OpenAI: five additional Stargate sites

Kenya

• iXAfrica NBOX1 campus specifications
• iXAfrica sustainability disclosures
• Airtel Africa data-center projects
• Kenyan ICT Ministry: Nxtra groundbreaking
• Africa Data Centres: Nairobi expansion cooling statement
• Kenya Demographic and Health Survey 2022
• Nairobi County Integrated Development Plan 2023–2027
• WASREB Impact Report 16

Human-use comparisons

• U.S. EPA WaterSense household-use estimate
• England 2023–2024 water-resource performance
• Nairobi City Water, August 2024 Water News Update