India, June 11 -- Is Artificial Intelligence Draining Our Freshwater Supplies? A study says that training a single large AI model (GPT-4) consumes an estimated 5.4 million liters of freshwater, which is enough to fill over two Olympic swimming pools. And ever 10-50 medium length chat "drink" a 500 ml water bottle (1).
Artificial Intelligence has delivered breakthroughs in healthcare, from early cancer detection to drug discovery. But according to a new analysis published in Communications of the ACM, the same large language models (LLMs) powering tools like ChatGPT are consuming freshwater at rates that could rival major beverage companies-and threaten already stressed community water supplies.
The Thirst behind the Screen: How AI Uses Water
Unlike human thirst, AI's water use is industrial and invisible. The study explains that AI models run on thousands of energy- hungry servers inside warehouse-sized data centers. These servers generate massive heat, which should be removed using cooling system, often water intensive towers or evaporative air conditioning.
On an average the researchers have given a breakdown of water consumption by AI:
Scope 1 (Onsite) : Water evaporated directly inside data centers for cooling. In hot regions like Arizona, this can reach 9 liters per kWh of server energy.
Scope 2 (Offsite) : Water consumed at power plants generating electricity for AI. The U.S. average is about 3.1 liters per kWh
Scope 3 (Supply chain) : Water used to manufacture AI chips-often overlooked but potentially accounting for up to 99% of a tech company's total water footprint (1)
. The researchers have warned that, "AI's water foot print has largely remained under the radar and if not addressed it can become a major roadblock to sustainability and create social conflicts"(1).
Why This Is a Public Health Emergency
Freshwater scarcity is only an environmental issue is not just an environmental issue, it also a direct driver of diarrheal diseases, heat stroke, sanitation failures and food insecurity. By 2028, AI serves alone could consume 150-280 billion liters of U.S. freshwater annually (1) .
Critically, data centers are often located in water-stressed regions (Arizona, Texas, India) where groundwater is already over-extracted. When a data center evaporates millions of liters of potable water, that water is permanently removed from the local watershed-competing directly with farms, hospitals, and homes.
The study says that: reducing AI's carbon footprint (by "following the sun" for solar energy) may actually increase water use during hot daytime hours. Carbon efficiency and water efficiency are not the same, and optimizing one can worsen the other.
Can AI Become Less "Thirsty"? Recommendations from the Study
The study supported by the U.S.National Science Foundation, offer a prescription for sustainable AI:
Transparency: Every AI "model card" (like a nutrition label for algorithms) must disclose water footprint alongside carbon emissions. Currently, water data is almost missing.
Timing & Location: AI training should be scheduled during cooler nighttime hours or moved to regions with better water efficiency. Users could be shown real-time "water efficiency" ratings for their AI queries.
Holistic Standard: The world's first international standard on sustainable AI (ISO/IEC) must include water foot print as a key metric, not an afterthought.
Healthy living tips
"Water and carbon footprints are complementary to, not substitutable for, each other," the researchers emphasize. "Judiciously deciding when and where to run a large AI model can significantly cut the water footprint."(1)