Ask an AI chatbot a question and receive a response within seconds. When you request a photo, it appears at about the same speed. It feels like it takes almost no effort.
The reality behind that screen looks very different. All prompts are sent to a humming machine warehouse powered by electricity that also utilizes water and land.
A new UN report measures all three costs together and finds that the total is much greater than the carbon numbers alone had previously suggested.
The cost behind the prompt
These warehouses are data centers and expand faster than the power grid can comfortably absorb them.
By 2030, dedicated AI could consume approximately 945 terawatt-hours per year. This is nearly three times the amount of electricity used in some of the world’s most populous countries.
Its footprint is already huge. In 2025, data centers around the world will consume approximately 448 terawatt-hours of electricity, enough to rank a country as the 11th largest electricity consumer in the world.
External forecasts predict that demand will nearly double by 2030.
Dr Miriam Akzel of the United Nations University Institute for Water, Environment and Health (UNU-INWEH) is the report’s lead author.
He noted that most assessments to date have only tracked carbon dioxide from training large models.
Water charges and land charges are also charged for each unit of electricity.
Reducing carbon can quietly raise two other costs.
Switching from coal to bioenergy (electricity produced by burning plants) can reduce the carbon footprint of electricity by about 70%.
Hidden costs are high. The same exchange could increase the water footprint by more than 30 times and the land footprint by more than 100 times.
Choices that look clean on carbon charts can quietly drain rivers and fields elsewhere.
“What surprised us most is that the options that seem to be the most environmentally friendly from a carbon perspective very often have worse outcomes for water and land,” Akzel said.
The mismatch caught the team off guard. Quietly misleading green label.
where does the energy go
Public concern centers on training, the costly phase in which a model first learns.
But researchers have found that inference, or the mundane task of answering prompts, consumes an estimated 80 to 90 percent of an AI system’s energy.
Not all prompts cost the same, and the difference can be staggering. A typical chat reply consumes about 200 times more energy than a simple text sorting task, while a single AI image can consume about 1,450 times more energy than its baseline.
One study of dozens of models confirmed that pattern.
To put it simply, one AI image lights up a small LED bulb for about 17 minutes. 42 hours is sufficient for complex videos.
ChatGPT alone processes an estimated 2.5 billion prompts per day, and it would take a tree sapling covering an area the size of Manhattan to offset that carbon.
Efficiency that backfires
A comforting assumption runs through much of the coverage of AI. It is believed that as technology becomes more efficient, environmental costs should naturally shrink.
The report strongly pushes back against that idea. Economists call this trap the rebound effect, and it is sometimes named the Jevons paradox, after the 19th century thinker.
When something becomes cheaper and easier, people use it much more.
Professor Kaveh Madani, director of the institute and leader of the team, declares: More efficient and cheaper AI tends to mean more AI until the total footprint exceeds the efficiency saved.
Local costs, distant benefits
The burdens of AI are rarely seen where there are benefits. In Ireland, data centers will consume around 21 per cent of all metered electricity in 2023, which is more than all urban households combined nationally.
The electricity grid operator has frozen new licenses in the Dublin area until 2028.
In drought-stricken areas of Mexico, new computing sites are competing for dwindling water supplies.
Uruguay planned to build a water-dry data center in 2023 after a drought depleted the capital’s reserves and made tap water unsafe to drink.
Hardware leaves its own mark. AI devices could generate up to 2.5 million tonnes of e-waste annually by 2030, much of which will end up in poor countries with few safeguards.
Only 32 countries host these data centers, and most computing power is concentrated in just two countries.
What does disclosure change?
Until this report, the environmental costs of AI were primarily measured in carbon, with a single number representing a larger picture.
Set carbon, water, and land side by side in a single calculation and show how often they are separated.
The findings already have political weight. A few weeks after the report, the United Nations called on AI companies to disclose their carbon, water and land footprints.
The companies were also required to run their data centers on renewable electricity by 2030.
Comparable public figures allow regulators and customers to weigh one company against another.
From here, your to-do list becomes more concrete. Governments can incorporate data centers into water and land plans, and companies can treat default settings as environmental decisions.
The report concludes on a hopeful note, arguing that competency and care can grow together when real costs are made visible.
The report is published as follows: United Nations University Inway Research Report.
—–
Like what you read? Subscribe to our newsletter for fascinating articles, exclusive content and the latest updates.
Check us out on EarthSnap, the free app from Eric Ralls and Earth.com.
—–
