AI data centers are on track to drink as much water as 1.3 billion people by 2030. That’s the headline finding from a new United Nations University report, covered this week by Futurism AI, and it reframes how we should think about the cost of running these systems. The report comes from the UN University Institute for Water, Environment and Health (UNU-INWEH), and its core argument is blunt: we’ve been measuring the environmental cost of AI wrong.
Most assessments focus on the carbon emissions from training large language models. According to Futurism AI, the report says this approach “systematically mismeasures” the real impact, because it ignores AI’s water and land footprints. Water gets consumed cooling and powering data centers. Land gets eaten up by energy infrastructure and supply chains. Neither shows up cleanly on a carbon ledger.
Inference is the real cost
Here’s the part practitioners should sit with. Training is not where most of the energy goes. Running the models to answer everyday prompts, what researchers call inference, makes up 80 to 90 percent of total AI energy use.
The numbers tell the story:
- Training GPT-4 consumed up to 70 gigawatt-hours of electricity.
- Running ChatGPT to answer billions of daily prompts burns an estimated 383 GWh.
So the one-time training cost is dwarfed by the daily grind of serving answers. That flips the usual narrative. The expensive part of AI isn’t building it, it’s using it at scale.
The 2030 projections
Factor in inference, and the forecast gets heavy. By 2030, the report estimates AI data centers will pull 945 terawatt-hours of electricity. To put that in scale, that’s triple the combined electricity use of Pakistan, Bangladesh, and Nigeria, three countries home to more than 650 million people.
The water figure is the one that stings. By the same year, AI is projected to consume 9.3 trillion liters of water. That matches the basic annual water needs of all 1.3 billion people in Sub-Saharan Africa.
Why “go green” isn’t a clean fix
What stands out here is that the obvious solution backfires. You can’t just swap in renewables and call it solved. The report found that ditching coal for bioenergy would cut electricity-related carbon emissions by 70 percent, but the water footprint would surge 30 times over and the land footprint 100 times.
“What surprised us most is how often the choices that look greenest from a carbon perspective end up worse for water or for land,” lead author Miriam Aczel, a UNU-INWEH researcher, said in a statement reported by Futurism AI. She warned that judging AI by carbon alone risks “solving one problem while creating other problems, often in places that didn’t ask for it.”
Then there’s the efficiency trap. You’d think better, leaner models shrink the footprint. The report says the opposite tends to happen. “More efficient and affordable AI and energy mean more consumption of AI, making the overall footprint far bigger than what we save through efficiency gains,” said coauthor Kaveh Madani, director of UNU-INWEH. It’s the Jevons paradox in action: make something cheaper, people use far more of it.
What this means for you
A few practical takeaways:
- Account for inference, not just training. If you’re building or buying AI, the running cost is where the real environmental and dollar weight sits. Heavy usage at scale is the line item to watch.
- Don’t trust single-metric sustainability claims. A vendor touting “carbon-neutral AI” may be quietly shifting the burden to water or land. Ask about all three.
- Location matters. Where a data center sits decides whose water and land it draws on. That’s becoming a real procurement and reputational question.
The report’s limitation is worth noting: these are projections built on current trends, not certainties, and 2030 figures depend on how usage and infrastructure actually evolve. Still, the direction is clear. Efficiency gains alone won’t save us, and carbon math alone won’t show us the full bill. Expect water and land to move from footnotes to front-page metrics in the AI sustainability debate. You can read the full breakdown at the original Futurism AI report.