AI and Water Usage: Challenges and Solutions
Artificial intelligence is reshaping many domains, but one often overlooked cost is its water footprint. As data centers grow both in number and scale, the water required to keep AI systems cool — directly via cooling and indirectly via electricity generation — is becoming a material environmental concern. IEEE Spectrum’s recent article “The Real Story on AI’s Water Use — and How to Tackle It” outlines how we use water today for AI, why it matters, and what paths forward exist. Below is a synthesis of those issues and possible strategies.
What is the Water Footprint of AI?
AI’s water usage falls into two principal categories:
- Direct water use (on-site cooling)
Servers generate heat; to prevent overheating, data centers use cooling systems. Evaporative cooling is common, where water is evaporated to carry away heat. Some portion of the withdrawn water is consumed (i.e. lost to the atmosphere). In many systems, 45-60% of the water withdrawn is consumed. Examples:- Data centers in water-stress areas tend to place higher load during hot days, so water demand can spike when it is least convenient.
- In the U.S., it’s estimated that data centers’ direct water consumption in 2023 was around 17.5 billion gallons; assuming about half of what’s withdrawn is consumed, withdrawal is ~35 billion gallons.
- Indirect water use (via electricity generation)
A lot of the water impact comes not from the data center’s cooling systems per se, but from how the electricity that powers these data centers is produced. Many power plants require cooling water (e.g. from rivers, lakes, or aquifers), which may be withdrawn and often consumed. The energy required to run servers, air conditioners, auxiliary systems, etc., means indirect water use often outweighs the direct cooling water usage.
The article estimates that in many U.S. data centers, indirect use (from power production) makes up a large share of total water use. For example, a single GPT-3 text generation output (150-300 words) consumed ~16.9 ml of water in total, of which ~14.7 ml was due to electricity generation and only ~2.2 ml for onsite cooling.
Why It Matters: Stress on Local Resources & Trade-Offs
- Local water stress: Even if data centers globally use a modest fraction of water, their concentration in water-scarce or drought-prone regions makes them problematic. Areas of high water stress see data centers pulling large volumes from municipal water supplies, potentially competing with residential, agricultural, and ecological needs.
- Temporal peaks: On hot days or during dry seasons, water demand from data centers spikes precisely when local systems are under maximum stress. This compounds existing scarcity problems.
- Indirect environmental impact: Water bodies used for cooling power plants can be stressed; aquifers and rivers can be depleted. Additionally, water usage interacts with carbon emissions through the energy source of electricity. If grids rely on fossil fuels, water use for generation also corresponds to higher carbon footprints.
Solutions: How to Cut AI’s Water Footprint
The IEEE Spectrum article, along with supporting research, outlines several strategies. They often involve trade-offs (e.g. more electricity vs. less water), but with wise design choices, the footprint can be reduced significantly.
- Cooling system improvements
- Alternatives to evaporative cooling: Use air-based cooling, liquid-immersion cooling, or cooling fluids that don’t require evaporation. These can drastically reduce direct water consumption.
- Closed-loop or “zero-water” designs: Systems that recycle cooling water or otherwise avoid evaporative loss. Some designs eliminate potable water withdrawal for cooling. But note: they may demand more energy or more robust infrastructure.
- Recycled water usage: Use greywater, treated wastewater, or non-potable sources for cooling, avoiding using fresh potable water.
- Onsite water storage, thermal energy storage: These can buffer peaks in demand when water is most scarce or expensive.
- Grid improvements and cleaner electricity Since much of the water footprint comes via electricity generation, shifting electricity supply to sources that use less water (renewables, solar, wind) helps reduce indirect water use. Also, making electricity grids more efficient reduces wastage, thereby indirectly reducing water needed to generate extra electricity.
- Location planning Built-in geographic considerations: avoid building large AI/data center installations in regions already under high water stress. If near water but high stress, explicitly plan for low-water cooling. Also, during siting, consider climate and access to clean water.
- Demand management
- Efficiency of operations: better thermal design in data centers, more efficient chips, optimizing workload scheduling (e.g. shifting heavy loads to cooler parts of day).
- Disclosures and metrics: More transparency by companies in how much water they withdraw vs. consume; standardized reporting to monitor progress. IEEE Spectrum notes many do disclose direct water use, but data are still incomplete.
- Holistic trade-offs and balancing Important to recognize that reducing one kind of resource usage may increase another. For example, moving to non-evaporative cooling (liquid immersion) might increase energy consumption, which (if powered by water-intensive sources) could increase indirect water usage or carbon emissions. Thus, decisions should consider both water and energy footprints together. IEEE Spectrum emphasizes that until electricity grids are dominated by renewables, many “water-saving” technologies may shift burden elsewhere.
Recent Research & Data Points
Some additional data and research help illustrate and quantify the issue:
- A paper “Making AI Less ‘Thirsty’: Uncovering and Addressing the Secret Water Footprint of AI Models” estimates that globally AI could withdraw 4.2-6.6 billion cubic meters of water in 2027.
- They also estimate the water consumption for training large models: for instance, training GPT-3 in U.S. data centers had direct evaporation of ~700,000 liters of freshwater.
- Another dataset for African data centers shows that water usage per AI task (e.g. writing a report, sending emails) differs significantly between countries depending largely on how electricity is generated locally (i.e. water intensity of that electricity) and local climate. In some African countries, the same task can use far less water than in the U.S., simply because their grid is less water-intensive or climate is cooler.
Challenges & Open Questions
- Data gaps and transparency: Many operators do not fully report water withdrawal vs. water consumption, or direct vs. indirect water use. Without good data, planning is hard.
- Trade-offs between water vs. energy vs. carbon: As above, minimizing water often means more energy, which may or may not come from clean sources. Must balance these metrics.
- Policy & regulation: In many regions there are no binding standards for water usage for data centers, or for how to measure water use. Also, zoning and permitting may not take into account water stress.
- Technological cost and maturity: Some cooling alternatives are still expensive or less proven at large scale. Also retrofitting existing data centers is more challenging than designing new ones with water-efficiency in mind.
Conclusion
Artificial intelligence is not just a digital phenomenon — its operations manifest physical costs, among which water usage is increasingly significant. The compounding of direct cooling demands with the hidden water costs of electricity generation makes AI’s water footprint non-trivial. However, as IEEE Spectrum and others show, there are clear strategies for mitigation: smarter cooling technology, choosing better locations, improving grid cleanliness, and using recycled/non-potable water.
To move forward, action is needed from multiple stakeholders: companies building and operating data centers, power utilities, local governments, and the research community. Only by taking a holistic view — one that weighs both water and energy (and carbon) — can we ensure AI’s growth is sustainable and doesn’t exacerbate water stress in vulnerable regions.