AI is thirsty - and the world is drying up
Every AI prompt has an environmental cost: servers consume vast amounts of water and electricity. Increased system efficiency only drives higher usage, following the Jevons paradox.
The realization that every prompt you type into your computer or mobile phone has an environmental price tag is not particularly intuitive, but it is the truth: behind the scenes, in a remote server farm, your request is translated into actual water consumption. The water consumption of AI is derived directly from the complexity of the task. UC Riverside researchers estimate that writing a 100-word email consumes about half a liter. Creating a video consumes about 4.1 liters — twice the recommended daily drinking amount for a person.
Almost a billion people use ChatGPT every week, sending the service about 2.5 billion requests per day. And that is before counting the rest of the giant models operating in parallel. Multiplying this data by the scale of global usage presents a worrying picture: artificial intelligence is becoming a huge and demanding consumer of our natural resources.
The Jevons Paradox: Efficiency as a Trap
Facing these numbers, the same reassuring answer is usually pulled out: technology will solve the problem. It is assumed that as we become more efficient, the crisis will disappear. However, this approach ignores a 150-year-old economic mechanism. In the winter of 2025, the Chinese startup DeepSeek presented an advanced model for which it was claimed that the computing resources required were 10 times lower than those invested in a similar model by Meta. In response, the market value of Nvidia plummeted by about 600 billion dollars in one day, as investors believed efficiency would reduce the need for more chips. However, Microsoft CEO Satya Nadella responded by citing the "Jevons paradox."
Formulated by William Stanley Jevons in 1865, the theory explains why new steam engines, which required less coal, did not lead to a decrease in national coal consumption in England but rather to an increase. When technology becomes more efficient and cheaper, its use expands significantly. Eventually, the sharp increase in demand swallows the individual savings, and total consumption jumps. This is exactly what is happening with AI today: 80% to 90% of energy consumption comes from ongoing activity. Once an action becomes cheap, usage expands, leading to an aggressive expansion of data center infrastructure.
The Environmental Bill in the Present
Data centers are huge industrial hangars housing hundreds of thousands of chips operating around the clock. Forecasts suggest that by 2030, global electricity consumption by data centers will double, equaling the annual electricity consumption of all of Japan. Alongside electricity, cooling systems draw millions of liters of water. According to a Microsoft report, about 42% of the tech giant's water consumption came from areas suffering from water scarcity.
Tech giants promise a "green" solution, yet their past commitments to reach a carbon-negative balance by 2030 are failing: Microsoft's carbon footprint jumped by 25% and Google's by 18% due to data center construction. The irony of the revolution is that the more virtual artificial intelligence becomes, the more land, water, and energy it needs. Nothing really floats in the cloud — this bill is always paid here, on the ground.
