Your AI habit is draining rivers in Aragon, Spain and Texas, US

Data centers cool processors through evaporation - water transforms into vapor and enters the atmosphere. Cooling towers evaporate roughly 80%; chillers have similar rates. This doesn't return to local rivers or reservoirs. In 2024, Google's thirstiest Iowa data center consumed 2.7 million gallons (10.2 million liters) daily. Meta restored over 1.59 billion gallons to watersheds, but facilities still consumed massive volumes initially.

The type of AI task matters significantly. Simple text queries use relatively little. Google's Gemini averages 0.26 milliliters per prompt, about five drops. Image generation is the most water-intensive, requiring substantially more cooling power. Advanced "reasoning" models can use 50-100 times more resources than basic queries. When you ask AI to generate a detailed image or process complex reasoning tasks, you're triggering far more computational work and water consumption than a simple question about the weather.

Amazon's Aragon facility needs 500 million liters (132M gallons) yearly - amid severe drought. The U.S. firm requested 48% more as Spain sought EU drought aid. 'Tu Nube Seca Mi Río' demands moratoriums on new facilities. Europe's use projected to double: 0.82 trillion liters (2025) to 1.58 trillion by 2030. Frankfurt, London, Amsterdam, Paris, Dublin took 85% of 2024's new capacity. Spain and Greece face highest water stress exposure.

Major companies are implementing solutions. Microsoft's next-generation data centers use closed-loop liquid cooling that recycles water continuously, avoiding evaporation of 125 million liters per year per facility. Google reports improving water efficiency by 80% since early 2000s. Liquid immersion cooling can reduce water consumption by up to 91% compared to traditional air cooling. Direct-to-chip cooling targets processors at the source, minimizing overheating and excessive water use. These aren't future concepts - they're being deployed now.

Your choices matter more than you think. Use AI intentionally - ask yourself if the task actually needs it or if a regular search would work. Batch your questions instead of multiple separate queries. Choose providers reporting efficiency metrics: Google's Gemini uses about 0.24 watt-hours per prompt; ChatGPT uses 0.34. That difference scales. When generating images, consider if you need AI or if stock photos or other tools work. Think of AI like driving - sometimes necessary, sometimes optional, always using resources.

Data centers measure efficiency through Water Usage Effectiveness (WUE) - liters consumed per kilowatt-hour of energy. The average across facilities is 1.9 liters per kWh. The goal is zero for air-cooled centers, but climate conditions often prevent this. Recent research discovered different cooling methods can vary efficiency dramatically. Hybrid systems combining liquid and air cooling optimize based on specific hardware needs. The question isn't whether to use water, but how efficiently facilities use it for the computing power delivered.