The Global Data Center Water Crisis

# The Global Data Center Water Crisis

Every time you run a large language model query, train a neural network, or stream high-
definition video, a data center somewhere consumes electricity. That electricity generates
heat. Cooling that heat, at the scale of modern data centers, requires water — enormous
quantities of water. As AI workloads drive data center capacity to unprecedented levels,
the water consumption implications are coming under increasing scrutiny from regulators,
local governments, and environmental researchers.

## How Data Centers Use Water

Data centers use water primarily for cooling. The dominant approach for large facilities
is evaporative cooling: hot water or air from server rooms passes through cooling towers,
where a fraction of the water evaporates, carrying heat away. The remaining water is
recirculated. The water that evaporates is consumptively used — it does not return to
the local watershed.

A second mode of water use is for on-site generation of chilled water and for humidity
control. Some facilities also use water-cooled server racks — direct liquid cooling —
where coolant circulates through cold plates attached to processors, then transfers
heat to a central cooling loop.

## The Scale of Consumption

Google disclosed in its 2023 environmental report that its global data centers consumed
approximately 5.6 billion liters of water. Microsoft disclosed roughly 7 billion liters.
Meta and Amazon have disclosed similar magnitudes. These numbers represent direct
operational water withdrawal; they do not include the water consumed by the power plants
that generate electricity for the data centers, which adds significantly to the total
water footprint.

Training a large language model is particularly water-intensive. Researchers from UC
Riverside estimated that training GPT-3 consumed roughly 700,000 liters of fresh water.
Larger models trained in 2024-2025 consumed more. Inference — running the model after
training — also consumes water continuously at scale.

## Geographic Concentration and Local Impact

The problem is not just the total volume of water consumed globally, but where it is
consumed. Data center construction has concentrated in specific geographic clusters:
Northern Virginia (the world's largest data center market), central Arizona, the
Netherlands, Singapore, and a growing set of secondary markets.

Northern Virginia has adequate water supply from the Potomac River, but the pace of
data center construction — hundreds of facilities approved and under construction —
is straining local infrastructure. Arizona presents a more acute challenge. The Phoenix
metropolitan area is in an arid climate dependent on the Colorado River, which has been
at historically low levels. Data centers in Arizona consume water that is already
under allocation pressure from agriculture, municipalities, and tribal nations.

The Netherlands has restricted new data center construction in certain areas due to
energy grid capacity and land use concerns. Singapore placed a moratorium on new data
centers from 2019 to 2022 and continues to impose strict water and energy efficiency
requirements on approved projects.

## Technical Responses: Air Cooling and Immersion Cooling

Several technical approaches can reduce or eliminate water consumption in data centers.
Air-side economization — using outside air to cool servers when ambient temperatures
are low enough — reduces dependence on evaporative cooling. This works well in cooler
climates but has limited effectiveness in hot, arid locations.

Direct liquid cooling and immersion cooling eliminate the need for large evaporative
cooling towers by transferring heat directly from chips to coolant. Immersion cooling
submerges entire server boards in a dielectric fluid. The heat removed can potentially
be recaptured and used for district heating, industrial processes, or greenhouse
agriculture — turning a waste heat problem into a resource recovery opportunity.

Microsoft experimented with an underwater data center (Project Natick) specifically to
test cooling by proximity to seawater. The project demonstrated that sealed submarine
data center vessels could operate reliably, though the economic and maintenance
arguments for large-scale adoption remain unresolved.

## Regulatory and Disclosure Trends

Water disclosure requirements for data centers are tightening. The European Union's
Energy Efficiency Directive has established requirements for data centers above certain
thresholds to report water consumption efficiency metrics. The US has no federal
disclosure mandate, but several states with significant data center clusters are
considering requirements.

The Green Grid consortium has developed the Water Usage Effectiveness (WUE) metric,
analogous to PUE for energy. A WUE of 1.0 means one liter of water is consumed per
kilowatt-hour of IT equipment energy. Leading facilities report WUE below 0.5;
less efficient facilities can exceed 2.0.

## The Path Forward

The intersection of AI-driven data center growth and water stress is a solvable engineering
problem, but only if water consumption is treated as a binding design constraint from
the start — not an afterthought. Facilities choosing locations and cooling technology
in 2026 will operate for decades. Decisions made now about water consumption embed
infrastructure that will either relieve or exacerbate local water stress through the
mid-century period when climate change projections indicate water stress will increase
in many existing data center markets.

The Global Data Center Water Crisis

// COMMENTS

ON THIS PAGE