The Overlooked Water Challenge in the Age of AI
In this blog
Introduction
The high energy consumption of data centers has been well known, monitored and optimized for years. However, data centers' water consumption is usually overlooked. As AI models continue to be trained at increasingly higher capacities and companies across the globe embed AI in their everyday practices, innovation in cooling technologies and regulations around water consumption emerge.
Why is water consumption critical to data centers?
Building-level cooling
Building cooling is primarily done through air cooling or evaporative cooling. Air cooling is more common, using a chiller to cool air that enters the facility. Evaporative cooling involves a separate water source in a cooling tower that is evaporated by the heat rejection. While it consumes significantly less energy than air cooling, it also consumes significantly more water and requires continuous freshwater to prevent mineral and salt buildup. Data center designers will typically choose between the cooling methods based on the unique conditions of each location[i], such as the local climate, available power and water stress.
Server-level cooling
Data centers traditionally use air cooling in servers, but liquid cooling is growing increasingly popular. To combat server density (because of more compute-intensive workloads), liquid cooling circulates water or dielectric liquid through pipes to quickly absorb heat from dense equipment stacks. The two types of liquid cooling that are most common today are direct-to-chip and immersion. Direct-to-chip is sometimes referred to as "cold plate cooling," as it uses cold plates placed directly on the CPU or GPU. Immersion cooling submerges servers and equipment into large tanks filled with dielectric liquid. Other types of liquid cooling include single-phase immersion cooling, dual-phase immersion cooling, two-stage evaporative cooling and closed-couple cooling[ii].
AI is driving the shift in cooling technologies
As AI providers continue to invest hundreds of billions into AI infrastructure[iii], water usage continues to rise. In Virginia's "Data Center Alley," data center water usage surged nearly 66% from 2019 to 2023[iv]. AI requires intensive cooling, both for the general usage of AI and for training AI models. Using the GPT-3 model with 175 billion parameters as an example, researchers at UC Riverside found that training the GPT-3 model in Microsoft's US data centers used approximately 5.4 million liters of water[v], enough to power 5,400 households for one day. By 2028, hyperscale data centers (ones that deploy internet services and platforms at massive scale) alone are expected to directly consume between 60 and 124 billion liters of water.
Figure 1: Direct water consumption by data center type [vi].
Innovation & alternative methods in cooling
AI is accelerating the adoption of liquid cooling since it runs on power-intensive equipment that requires intensive cooling. Projections in the valuation of the liquid cooling market estimate growth from $3.2 billion in 2023 to $16.4 billion by 2032[vii]. The need for innovation is driving more efficient high-density data center designs - one example is Microsoft's "sidekick" system, which circulates chilled liquid through cold plates in a closed loop to remove heat without large-scale chillers[viii].
Alternative cooling methods are also gaining popularity, such as geothermal cooling practices like aquifer thermal energy storage (ATES) or deep lake water cooling. Rather than exporting heat into the atmosphere via evaporation, geothermal cooling returns water to its original source. Another example is data center heat export, which allows a portion of the heat in a data center to be transferred to a third-party heat network[ix].
Rethinking water as a finite resource
Community impact
Due to their need for clean freshwater, data centers compete with communities for potable water. In 2024, 78% of Google's data center water withdrawal was potable water - around 7,700 million gallons.[x] Increasingly more communities are feeling the direct impacts of these large-scale potable water withdrawals. In Newton County, Georgia, Meta's data center was permitted to use over 500,000 gallons of water daily, leading residents to report lower water pressure and supply issues[xi]. This impact can also be seen in the lowered groundwater levels in Minnesota and in Texas suburbs, where residents face drought restrictions while nearby data centers continue high-volume water consumption.
Monitoring & regulations
Currently, less than one-third of data center operators report water usage. Widespread tracking and reporting of Water Usage Effectiveness (WUE) can optimize cooling systems and help companies reduce costs. Firms that rely on water-intensive data centers are subject to financial risk due to droughts and regulatory restrictions when they do not consider alternatives, dependency tracking, and mitigation options. General Motors, for example, incurred $2.1 million in extra water costs during a drought in Brazil in 2015[xii].
Policies around data center water consumption are emerging. California introduced SB 58 in January 2025, which provides a tax credit for data centers that adopt "sustainable practices", such as using recycled water-cooling systems[xiii]. In Minnesota, new regulations require developers to partner with the Department of Natural Resources to ensure new data centers have an adequate water supply.[xiv] In the EU, the European Commission revised its Energy Efficiency Directive to include mandated data center energy reporting[xv], which can be used to benchmark, rate and enforce improvements in water consumption.
Looking ahead
Green AI
A new emerging approach and methodology is green AI. Focused on green-by AI and green-in AI, this approach illustrates how AI can both enable sustainability and embed sustainability into its everyday operations. Green-by AI is the concept of AI playing a pivotal role in decreasing GHG emissions and enhancing sustainability across sectors. Examples include AI in smart grids, forecasting of solar and wind power, precision farming, IoT sensors, and contributions to mitigating climate change[xvi].
Green-in AI tackles algorithm optimization to reduce the computational resources that produce the extreme environmental impacts of AI. Prioritizing optimized ML model architecture and efficient hardware use (such as parallelization and edge computing) will have the most direct impact. Green-in AI also includes data center optimization, such as algorithms and frameworks that dynamically manage server loads and adjust cooling systems. New research even suggests that the time of day can play a role in water efficiency[xviii] - thus, choosing both when and how to train AI models could significantly improve its environmental impact.
Conclusion
Water must be treated with strategic importance. As the world evolves around AI and digital innovation, the water footprint of data centers will continue to grow. Data center operators and hyperscale companies can mitigate risk by considering their local environment when making decisions on cooling, exploring alternative cooling methods and innovations in liquid cooling, and prioritizing operational efficiencies in their model architecture and hardware.
Sources
[i] Setmajer, A. (2024, September 19). How data centers use water, and how we're working to use water responsibly. Equinix. https://blog.equinix.com/blog/2024/09/19/how-data-centers-use-water-and-how-were-working-to-use-water-responsibly/
[ii] Mathur, A. (2024, October 15). Liquid cooling: Enhancing sustainable data center operations. HCLTech. https://www.hcltech.com/blogs/liquid-cooling-enhancing-sustainable-data-center-operations
[iii] Capoot, A. (2025, July 14). Meta's Zuckerberg: First AI supercluster will come online in 2026. CNBC. https://www.cnbc.com/2025/07/14/meta-zuckerberg-ai.html
[iv] Wiggers, K. (2024, August 19). Demand for AI is driving data center water consumption sky high. TechCrunch. https://techcrunch.com/2024/08/19/demand-for-ai-is-driving-data-center-water-consumption-sky-high/
[v] Li, P., Yang, J., Islam, M. A., & Ren, S. (2025). Making AI less "thirsty": Uncovering and addressing the secret water footprint of AI models (Version 5). arXiv. https://doi.org/10.48550/arXiv.2304.03271
[vi] Shehabi, A., Newkirk, A., Smith, S. J., Hubbard, A., Lei, N., Siddik, M. A. B., Holecek, B., Koomey, J., Masanet, E., & Sartor, D. (2024). 2024 United States data center energy usage report. Lawrence Berkeley National Laboratory. https://escholarship.org/uc/item/32d6m0d1
[vii] Global Market Insights - Data Center Liquid Cooling Market Share, Outlook 2024-2032. https://www.gminsights.com/industry-analysis/data-center-liquid-cooling-market
[viii] Microsoft. (2025, April). Liquid cooling infographic: Heat exchanger units. https://datacenters.microsoft.com/wp-content/uploads/2025/04/Liquid_Cooling_Infographic_FINAL-3.pdf
[ix] Nkonge, N. (2024, June 5). What is data center heat export and how does it work? Equinix. https://blog.equinix.com/blog/2024/06/05/what-is-data-center-heat-export-and-how-does-it-work/
[x] Google LLC. (2025, June). 2025 environmental report. https://www.gstatic.com/gumdrop/sustainability/google-2025-environmental-report.pdf
[xi] Rijo, L. (2025, July 21). Meta data center impacts local water supply in Newton County. PPC Land. https://ppc.land/meta-data-center-impacts-local-water-supply-in-newton-county/
[xii] Gupta, H. (2024, January 11). A thirst for change: Navigating businesses through water risk and resilience. Forbes Technology Council. https://www.forbes.com/councils/forbestechcouncil/2024/01/11/a-thirst-for-change-navigating-businesses-through-water-risk-and-resilience
[xiii] California State Legislature. (2025). Senate Bill No. 58: Sales and Use Tax Law—Exemptions for certified data center facilities. LegiScan. https://legiscan.com/CA/text/SB58/id/3153196
[xiv] Marohn, K. (2025, June 18). Data centers face new regulations in Minnesota. MPR News. https://www.mprnews.org/story/2025/06/18/data-centers-face-new-regulations-in-minnesota
[xv] Datacenters.com. (2025, July 8). Inside the EU's crackdown on data center emissions: What it means globally. https://www.datacenters.com/news/inside-the-eu-s-crackdown-on-data-center-emissions-what-it-means-globally
[xvi] Zhang, Y., Li, H., & Chen, X. (2024). Optimizing water usage in AI-driven data centers: A hybrid cooling approach. Neurocomputing. Advance online publication. https://www.sciencedirect.com/science/article/pii/S0925231224008671
[xvii] Zhang, Y., Li, H., & Chen, X. (2024). Optimizing water usage in AI-driven data centers: A hybrid cooling approach. Neurocomputing. Advance online publication. https://www.sciencedirect.com/science/article/pii/S0925231224008671
[xviii] Li, P., Yang, J., Islam, M. A., & Ren, S. (2025). Making AI less "thirsty": Uncovering and addressing the secret water footprint of AI models (Version 5). arXiv. https://doi.org/10.48550/arXiv.2304.03271