The rapid proliferation of Artificial Intelligence (AI) has not only brought about a digital revolution but is also imposing increasingly severe, quantifiable physical environmental burdens on the planet. The latest landmark report by UNU-INWEH (Institute for Water, Environment and Health), an institute under the United Nations University, highlights that current analyses focusing solely on carbon emissions fundamentally miscalculate AI’s true ecological footprint by neglecting critical water and land use costs.
Technology sector representatives frequently argue that as algorithms and networks evolve, the energy efficiency of AI systems improves drastically, which will eventually mitigate their environmental impact. However, the comprehensive study titled “Environmental Cost of Artificial Intelligence: Carbon, Water, and Land Footprints,” published by UN scientists, firmly refutes this optimism. The report points to the manifestation of the Jevons paradox: as a technology becomes more efficient and less expensive, it generates a substantially larger volume of consumption globally, ultimately neutralizing any efficiency gains. AI is not merely virtual, cloud-based software, but a tangible material system whose operation consumes direct physical resources.
Drastic Growth: Data Centers with Country-Sized Consumption
According to researchers at the United Nations University, the electricity demand of global data centers supporting AI infrastructure will reach critical levels in the near future. While the electricity consumption of global data centers is estimated at approximately 448 terawatt-hours (TWh) in 2025, this metric is projected to more than double by 2030, reaching around 945 terawatt-hours. This immense amount of energy equals or exceeds the total annual consumption of developing nations like Pakistan, Bangladesh, and Nigeria combined, and closely approaches the annual electricity requirements of highly industrialized Japan.
The central finding of the report is that AI’s environmental impact does not depend exclusively on the number of kilowatt-hours (kWh) consumed, but heavily on where the electricity is generated and the specific regional energy mix. A carbon-centric approach is misleading because low-carbon power sources do not automatically translate to a low water or land use footprint. Every kilowatt-hour consumed simultaneously carries carbon, water, and land-use implications, and these factors often move in opposite directions: reducing the burden on one environmental element (for instance, transitioning to certain clean energies) frequently triggers an increase in another factor.
The Invisible Cost: Drinking Water for 1.3 Billion People
Data center cooling systems, along with the generation of electricity itself, require highly intensive freshwater consumption. The study reveals that by 2030, the annual water footprint attributable to the AI sector will reach 9.3 trillion (9,300,000,000,000) liters. To put this abstract number into perspective, this volume corresponds to the total annual basic survival and domestic water needs of 1.3 billion people living in Sub-Saharan Africa. Concurrently, the direct and indirect land requirements (land footprint) of these facilities will exceed 14,500 square kilometers, which is roughly twice the size of the Jakarta metropolitan region.
Model Deployment is More Expensive Than Training
In the public consciousness and previous analyses, the environmental impacts of AI were almost exclusively linked to the initial training phase of Large Language Models (LLMs). Scientists at UNU-INWEH point out that this perspective is fundamentally flawed. In reality, model deployment—meaning everyday operations and serving user queries (inference)—consumes 80 to 90 percent of the systems’ total life-cycle energy demand. Once a model goes live, billions of daily user interactions create a continuous and cumulative load. A prime example is ChatGPT, which is estimated to process approximately 2.5 billion prompts (user instructions) every day.
The report also highlights that the energy requirements of individual queries vary by orders of magnitude depending on the type of task. Taking the power consumption of a basic text classification task as a baseline, a typical conversational chat query requires about 200 times more energy. Generating a single AI image consumes 1,450 times the baseline value, while producing a single short AI video uses as much electricity as the automated classification of 200,000 spam emails.
This becomes tangible even at the level of daily consumer habits:
-
Generating a single AI image requires the same amount of energy as running a 10-watt LED bulb continuously for 17 minutes. In terms of water consumption, this translates to two tablespoons (29 ml) of water due to the water footprint of electricity generation.
-
Generating a single complex AI video can power that same 10-watt LED bulb for 42 hours. In this case, the indirect water consumption jumps to 4.1 liters, which nearly equals a human’s total drinking water requirement for two full days.
Global Inequalities and Digital Colonialism
The study sharply criticizes the asymmetric global distribution of the benefits and burdens associated with AI. While the economic, defense, and sovereignty benefits of AI, along with control over computational capacities, are concentrated almost entirely in wealthy, developed nations, the environmental damages of the physical infrastructure hit local communities. Currently, only 32 countries in the world host dedicated AI data centers, and 90 percent of the total global capacity is concentrated in just two countries.
Conversely, the extraction of critical minerals needed for hardware manufacturing often takes place in vulnerable regions with weak environmental regulations. Furthermore, a significant portion of the electronic waste (e-waste) generated by the rapid depreciation of AI infrastructure—estimated to reach 2.5 million tonnes annually by 2030—ends up in low-income countries, where processing occurs under unsafe and unregulated conditions.
UN researchers emphasize that integrating AI into energy management, climate action, water, and land-use planning is inevitable. Only through this can it be ensured that technological innovation does not shift the true environmental cost of digital growth onto the world’s most vulnerable communities.
Official Institutional Reference:
The original study and announcement from the official United Nations University structure (UNU-INWEH) can be accessed at the following links:
