INTRODUCTION
Recent research has begun to uncover the hidden environmental costs of artificial intelligence with growing attention on its significant water footprint. Studies conducted by environmental scientists and computer engineers have used a combination of data audits, model training logs and cooling system analyses to estimate water usage across AI operations. Findings reveal that training large scale AI models such as Chat GPT can consume millions of litres of fresh water primarily for cooling the massive servers required to process complex calculations. Training GPT in Microsoft US data centers is estimated to evaporate about 700,000 litres of water directly for cooling. A 2025 study (University of California, Riverside) estimates 10–50 Chat GPT responses consume about 500 mL of fresh water, so roughly 0.5 L per 10–50 queries, or around 10–50 mL per response. The water demand is often localised placing additional stress on regions already facing water scarcity. For example, data centres in arid areas draw heavily on municipal supplies affecting agriculture, eco systems and local communities. Researchers also noted that the timing of AI model training frequently coincides with peak periods of water stress, heightening the severity of the impact. While often AI is applauded for its efficiency and innovation these findings challenge the notion of its sustainability. The environmental trade offs, particularly the strain on freshwater resources call for urgent improvements in data centre infrastructure and energy water efficiency. Without intervention, the pursuit of smarter technology may come at the cost of one of our most vital natural resources.
LITERATURE REVIEW
As AI technologies like ChatGPT, Google Bard, and Microsoft Copilot continue to scale rapidly, the environmental impact of the data centers that power them is becoming increasingly visible. While energy consumption and carbon emissions often dominate public discourse, a growing concern lies in the massive freshwater usage for server cooling, an issue particularly troubling in drought prone regions.
A report by News.com.au, titled “AI’s next crisis isn’t electricity, it’s water scarcity,” warns that by 2027, AI could consume between 4.2 to 6.6 billion cubic meters of freshwater annually, more than the total yearly usage of Denmark. This alarming projection is already influencing policy: cities like Singapore and Dublin have paused approvals for new AI data centers due to stretched water and energy resources.
The high consumption of water outlined in the News.com report is not only supported but expanded upon by Investopedia’s article, “The Hidden Cost of AI: How Data Centers Are Draining Water Resources,” which reveals that ChatGPT and similar AI tools can consume up to 0.5 liters of water per user session, and that model training can require up to 185,000 gallons. This issue is set to worsen with a projected 165% growth in data center capacity by 2030, especially since 20% of these facilities are in water stressed regions highlighting the compounding nature of the crisis.
Business Insider’s June 2025 investigation, “How data centers are deepening the water crisis,” further reinforces the urgency, revealing that 40% of U.S. data centers supporting AI operations are located in areas with high or extreme water scarcity. While major players like Amazon and Microsoft are experimenting with air based cooling and water stewardship programs, these efforts are still sporadic and inconsistently adopted across the industry.
Despite the mounting evidence and rising awareness, a critical literature gap remains, there is a lack of in depth, localized studies that examine how diverting water for AI data centers affects surrounding communities, particularly in regions already experiencing water scarcity. While macro level consumption data is available, very little is known about the social, economic, and health impacts on communities where water resources are being redirected for AI infrastructure. This oversight limits policymakers’ ability to balance technological growth with human and ecological needs, and it leaves vulnerable populations without a voice in critical resource allocation decisions.
Written by: Zymal Tahir and Anaya
Comments
Post a Comment