Understanding the water footprint of artificial intelligence infrastructure
Artificial intelligence is rapidly expanding the world’s digital infrastructure. New data centers are being built to support AI training, cloud services, and large-scale inference systems. Alongside concerns about electricity demand, another environmental question is gaining attention: AI data center water usage.
Why do AI data centers use so much water?
At ShrinkThatFootprint, we have previously examined the energy and carbon footprint of digital technologies, including the climate impacts of AI training and the electricity use of modern data centers. Water consumption is a related but less visible part of the same infrastructure story.
The short answer is that AI data centers concentrate enormous amounts of electricity into small physical spaces, and nearly all of that electricity becomes heat. Removing that heat efficiently often involves cooling technologies that rely on water.
Understanding this water footprint requires looking at three things:
- The physics of computing and heat
- The cooling systems used in data centers
- The broader water–energy–carbon tradeoffs in digital infrastructure.
AI infrastructure is fundamentally a heat problem
Every computation produces heat.
When a processor runs calculations, electrical energy flows through billions of transistors. Almost all of that electrical energy is ultimately converted into heat. If that heat is not removed continuously, the hardware will overheat and fail.
For decades, data centers have therefore been designed primarily as heat removal systems.
Traditional enterprise data centers typically operate racks of servers consuming roughly 5–15 kilowatts of power per rack. AI infrastructure is much more energy dense.
Modern AI systems built around GPUs and high-speed networking can exceed 100 kilowatts per rack, several times the density of earlier systems. This concentration of power produces extremely high heat flux that must be removed reliably.
A typical heat flow pathway inside an AI facility looks like this:
- heat generated in GPU chips
- transferred into server heat sinks or liquid cooling plates
- carried into facility cooling systems
- rejected to the outside environment
The final step — rejecting heat into the atmosphere — is where water is often used.
Why water is used in data-center cooling
The key reason for AI data center water usage is simple physics.
When water evaporates, it absorbs large amounts of heat. This makes evaporative cooling one of the most energy-efficient ways to remove heat from industrial systems.
Many data centers therefore use cooling towers or evaporative cooling systems, sometimes they’re also called evaporative heat-rejection systems. Warm water carrying heat from servers is sprayed through air, where a portion evaporates. The evaporation removes heat from the remaining water, which is then recirculated to continue cooling equipment.
This process consumes water continuously because the evaporated water must be replaced. Additional water is periodically discharged to control mineral buildup in the system.
In other words, water is not required for computing itself — it is used because it allows data centers to remove heat while minimizing electricity use for cooling.
That tradeoff is central to the environmental footprint of AI infrastructure.
The hidden water footprint of electricity
Water used directly at the facility is only part of the story.
Electricity generation can also consume large quantities of water, particularly in thermoelectric power plants that use steam cycles and cooling systems.
As a result, researchers often separate two components of data-center water use:
Direct water use
Water consumed on site for cooling and facility operations.
Indirect water use
Water used elsewhere to produce the electricity consumed by the data center.
Recent estimates from Lawrence Berkeley National Laboratory’s Data Center Energy Use Report suggest that the indirect component can be much larger than the direct one. For example, researchers estimate that U.S. data centers consumed about 66 billion liters of water directly in 2023, but the electricity powering them carried an indirect water footprint approaching 800 billion liters.
This distinction matters because cooling choices affect both water and energy consumption. Eliminating water use on site may require more electricity, which can increase water use upstream at power plants.
From a footprint perspective, water and energy are tightly coupled.
Cooling technologies used in AI data centers
Data centers rely on several major cooling approaches, each with different environmental tradeoffs.
Air-cooled systems
Traditional facilities often rely on refrigeration systems similar to industrial air conditioning. Heat is removed from servers using chilled air and rejected outdoors.
These systems use little or no water but require more electricity to achieve the same cooling effect.
Water-cooled systems with cooling towers
Many hyperscale facilities use water-cooled chillers combined with cooling towers.
These systems evaporate water to remove heat and are generally more energy efficient than air-cooled alternatives, especially in warm climates.
The tradeoff is continuous water consumption.
Evaporative cooling
Some modern facilities use evaporative cooling directly in airflow streams. These systems can achieve very low energy consumption but rely heavily on water when evaporation occurs.
Liquid cooling for AI hardware
High-density AI clusters increasingly use liquid cooling inside servers.
Instead of relying entirely on air and fans, coolant flows through plates attached to processors, carrying heat away more efficiently. This approach allows higher power densities but does not automatically eliminate water use. The facility must still reject heat outdoors, which may still involve cooling towers.
In practice, liquid cooling shifts where heat is removed rather than eliminating the cooling challenge.
Measuring water efficiency in data centers
To compare water performance across facilities, the industry uses a metric called Water Usage Effectiveness (WUE).
WUE measures the amount of water consumed relative to the electricity used by IT equipment. It is typically expressed in liters of water per kilowatt-hour of computing.
Lower numbers indicate less water consumed per unit of computing.
Reported values vary widely depending on climate and design. In cooler regions, some hyperscale facilities report values around 0.04 liters per kWh. In hotter regions the numbers can exceed 1 liter per kWh.
Even these small numbers can translate into large total volumes when multiplied by the enormous electricity use of modern AI clusters.
This is a recurring theme in footprint accounting: small intensity values multiplied by massive scale produce significant environmental impacts.
Why AI is increasing the pressure on water resources
Several trends in AI infrastructure are amplifying the cooling challenge.
Higher hardware power density
AI accelerators consume far more power than conventional CPUs. Dense racks of GPUs can exceed 100 kilowatts, dramatically increasing heat loads.
Sustained computing workloads
AI training and large inference clusters often operate at near-constant utilization for long periods. This creates continuous cooling demand rather than intermittent spikes.
Rapid growth in computing infrastructure
AI is driving a global expansion of data-center capacity. Even if efficiency improves, total water use may grow simply because more infrastructure is being deployed.
This dynamic mirrors trends seen in other technologies: improvements in efficiency are often offset by increases in scale.
Water stress and the geography of data centers
The environmental impact of data-center water use depends heavily on location.
Some regions have abundant freshwater resources, while others experience chronic water scarcity. Analysts often evaluate “water stress,” which measures how much water is withdrawn relative to available renewable supply.
Technology companies increasingly consider water stress when selecting data-center locations.
In practice, however, data centers are often located based on a combination of factors including electricity availability, fiber connectivity, tax policy, and proximity to major markets. As AI infrastructure expands, some communities have raised concerns about water use in regions already facing drought risk.
This illustrates a broader point emphasized in footprint analysis: environmental impacts are highly context-dependent.
Strategies to reduce the water footprint of AI
Technology companies and researchers are exploring several ways to reduce water use.
Eliminating evaporative cooling
Some new designs attempt to remove water from cooling systems entirely by using dry coolers or air-cooled chillers.
This can reduce direct water consumption but may increase electricity use.
Using reclaimed water
Another strategy is to use treated wastewater rather than freshwater. Several data centers already use reclaimed water for cooling towers, reducing pressure on drinking-water supplies.
Improving AI efficiency
Reducing the energy required for AI workloads also reduces cooling demand. More efficient models, hardware improvements, and optimized training methods all lower the energy required per unit of computation.
Transitioning to low-water electricity sources
Because indirect water use from electricity can dominate total water footprints, shifting data centers to renewable electricity sources such as wind and solar can significantly reduce water consumption.
The water–energy–carbon tradeoff
The key insight is that water use cannot be evaluated in isolation.
Cooling systems that consume water may reduce electricity use. Systems that eliminate water may require more electricity.
The environmental outcome therefore depends on multiple factors:
- local climate
- water availability
- electricity grid mix
- facility design
- workload efficiency.
For sustainability analysis, this becomes a three-way optimization problem between water, energy, and carbon emissions.
Conclusion
AI data centers use water primarily because it is one of the most efficient ways to remove heat from extremely dense computing systems.
As artificial intelligence expands, the environmental footprint of digital infrastructure is becoming more visible. Water consumption from cooling is significant, but the larger hidden component often lies in the electricity used to power these systems.
Understanding the water footprint of AI therefore requires looking at the entire system — computing efficiency, cooling technology, electricity sources, and location.
For readers interested in reducing environmental impacts, the most important takeaway is that the footprint of AI infrastructure is not determined by a single metric. It emerges from the interaction between water, energy, and carbon across the entire digital ecosystem.
