When the world’s power grid was compared to a national consumption, data centers powering AI stepped into the spotlight. In 2025, the infrastructure that feeds artificial intelligence drew 448 terawatt‑hours (TWh) of electricity—enough to rank them eleventh on the global list, roughly equal to France’s entire annual usage.

The United Nations University Institute for Water, Environment and Health (UNU‑INWEH) released a report on 3 June 2026 that details the carbon, water and land costs of this expanding sector. Edited by Kaveh Madani, the study projects that AI‑related data‑centre spending will exceed $2.5 trillion by 2026, with the bulk of the bill directed toward energy.

Looking ahead, the report warns of a 40 % rise in electricity use by 2040, pushing AI’s share of global consumption to 3 %. That figure is comparable to the power needed by 1.3 billion people in sub‑Saharan Africa for five years. In terms of climate impact, AI’s annual CO₂e emissions are projected to match those of the United Kingdom in 2025.

Water use is equally staggering. The study estimates that AI data centres will withdraw 9.3 trillion litres by the decade’s end—a volume that could supply the drinking‑water needs of the world’s 8.1 billion inhabitants for about a year and a half. Even with circular water‑management practices, large‑scale withdrawals could strain aquifers and rivers, especially in arid regions.

Land occupation is expected to exceed 14,000 km² by 2030, roughly the size of Northern Ireland, as servers, cooling infrastructure and associated facilities expand.

Hardware waste adds another layer of concern. By 2030, e‑waste from AI infrastructure could reach 2.5 million tonnes per year, the equivalent of discarding 250 Eiffel Towers annually.

Despite these massive infrastructure costs, most AI use today is relatively low‑value. For instance, ChatGPT processes an estimated 2.5 billion requests per day. Using a conservative estimate of 0.42 Wh per request, the model consumes about 383 GWh of electricity annually. The corresponding water footprint would meet the minimum domestic needs of roughly 500,000 people in sub‑Saharan Africa, and the land footprint would cover more than 800 football pitches.

Generative AI services also consume more energy than traditional search. A typical AI‑powered search request can use up to 3 Wh, compared with 0.3 Wh for a conventional search. High‑resolution AI‑generated videos can require over 415 Wh per clip, a figure that rises quadratically with resolution and frame rate.

"The future of artificial intelligence should not be assessed solely on what machines can do but also on humanity’s ability to employ those powers within the planet’s limits," Kaveh Madani said. "The findings underscore the need for AI developers and users to consider environmental limits."

The report calls on the United States and China—countries that together host about 90 % of AI infrastructure—to take action. It urges policymakers to incorporate AI’s environmental costs into energy, water and land‑use planning, and to enforce stricter reporting and efficiency standards for data centres.

Part of a growing body of research that highlights the hidden environmental costs of AI, the study feeds into recent discussions within the United Nations and other international bodies about balancing AI innovation with sustainability.

As AI continues to expand, the report’s figures suggest that the sector’s environmental impact will become increasingly significant. The challenge for governments, industry and civil society is to align AI development with global climate, water‑security and land‑management goals.

The report is available through UNU‑INWEH and is cited by several United Nations agencies and research institutions, representing the most comprehensive assessment to date of AI’s carbon, water and land footprints.