A new report from the United Nations University warns that the growth of AI is creating a vast environmental footprint: data centers, cooling, power generation, mineral mining, e-waste, and inequality between countries that deploy the technology and those that bear some of its costs.
Artificial intelligence is often portrayed as a “virtual” technology, almost weightless. But a new report from the United Nations University Institute for Water, Environment and Health, UNU-INWEH, warns that behind every query, image or video generated by AI is a vast physical infrastructure: data centers, chips, cooling systems, electricity consumption, mineral mining, land use, water pumping and e-waste. According to the report, the public and regulatory debate on AI has so far focused mainly on electricity and carbon emissions, but the true environmental footprint of AI is much broader and more complex.
The report, titled Environmental Cost of AI's Energy Use: Carbon, Water and Land Footprints , argues that examining the sustainability of AI through just one metric can be misleading. “Low-carbon” electricity is not necessarily “low-water” or “low-land” electricity. Switching from one energy source to another may reduce emissions, but at the same time greatly increase water consumption or land use. Therefore, the report’s authors warn, decisions about building data centers and operating AI systems should be examined according to the full range of environmental impacts, not just carbon emissions.
Data centers already consume as much electricity as a large country
According to the report, global investment in AI is expected to surpass $2.5 trillion this year, and the global market could grow from about $189 billion in 2023 to almost $5 trillion in 2033. This growth is supported by a huge energy infrastructure. If data centers were a country, their electricity consumption in 2025, estimated at 448 terawatt-hours, would place them in 11th place in the world, about the same level as France.
According to the report, AI workloads accounted for about 20% of data center electricity consumption in 2025. If their share increases to about 40% by 2030, AI electricity consumption alone could reach hundreds of terawatt-hours per year, and in some scenarios approach about 3% of projected global electricity consumption. Depending on the electricity generation mix, the associated emissions could reach about 400 million tonnes of carbon dioxide equivalent, a similar order of magnitude to the UK’s annual emissions from all sectors.
But electricity is only part of the picture. The report estimates that the land footprint of AI power generation by 2030 could exceed 14 square kilometers, an area similar to Northern Ireland. Data centers’ estimated water consumption is 9.3 trillion liters, enough to meet the minimum drinking needs of the entire world’s population for about a year and a half. Even when some of the water is returned to the system, the report’s authors warn, large-scale pumping could strain aquifers and river systems, especially in dry areas or those already suffering from groundwater depletion.
Not just model training: daily usage is accumulating to huge proportions
Training large AI models is energy-intensive, but the report highlights that their ongoing use could become even more significant. According to one estimate in the report, training a large model on the scale of GPT-5 could require about 100 gigawatt-hours of electricity, along with an estimated water footprint of about a billion liters and a land footprint of about 1.5 square kilometers. However, once such models are integrated into mass products, the cumulative footprint of billions of daily uses becomes much larger.
The report estimates that ChatGPT alone processes about 2.5 billion queries per day. While a single text query consumes relatively little energy, multiplying that consumption by billions of operations creates a significant annual power consumption. An even greater impact is created when AI is integrated by default into mass platforms, such as search, image creation, and video creation.
One area of particular concern is generative video. According to the report, creating a single high-quality AI clip can require more than 415 watt-hours, far more than creating hundreds of AI images. As resolution and frame rate increase, the energy demand increases rapidly. So if generative video is to become mainstream, it will no longer be a niche use case but a broad infrastructure issue.
The report also points to Jevons’ paradox: Technological efficiency does not guarantee a decrease in overall resource consumption. As usage becomes cheaper and more efficient, the volume of usage may increase until it erases all savings. That’s why the report’s authors call not only for better hardware, but also for resource budgets: limits on tokens, GPU hours, or kilowatt-hours, depending on the context.
An environmental and social gap between those who operate AI and those who pay the price
One of the report’s key claims is that AI is also exacerbating environmental inequality. Advanced AI infrastructure is concentrated in a small number of countries. According to the report, only 32 countries host dedicated AI cloud infrastructure, and about 90% of this computing power is concentrated in the US and China. More than 150 countries do not have any significant sovereign AI infrastructure at all. This means that many countries are dependent on external providers, without sufficient control over access, price, data governance, or environmental impacts.
At the same time, some of the environmental burden falls on other regions. Mining of minerals critical to chips and servers is often done in the Global South and in places where environmental oversight is weaker. E-waste, if not properly managed, can expose local communities to hazardous materials. According to the report, by 2030, AI infrastructure could generate up to 2.5 million tons of e-waste per year, the equivalent of throwing away about 250 Eiffel Towers each year.
The report presents Ireland as a practical warning. Data centers there already account for about 21% of measured electricity consumption, up from just 5% in 2015, and their consumption exceeds that of all urban households in the country. The Irish grid operator has even halted new permits in the Dublin region until 2028. For the report’s authors, this is a sign of what can happen when the growth of AI infrastructure preempts energy, water and land planning.
A roadmap for more responsible AI
The report does not call for a halt to artificial intelligence. On the contrary, it recognizes its great potential in areas such as medicine, education, scientific discovery and climate resilience. But it emphasizes that innovation without responsibility could deepen inequality and increase pressure on already stressed environmental systems.
The report’s authors therefore propose a roadmap based on six principles: transparency, efficiency as a design choice, environmental justice, life-cycle responsibility, global collaboration, and sustainable use. Governments are called upon to integrate AI infrastructure into energy, water, and land planning, and to mandate uniform environmental reporting. AI companies are required to treat model selection, defaults, and output configuration as environmental decisions. Users and organizations are urged to choose the easiest model and most cost-effective format that meets the task.
Even the wording of the query becomes, according to the report, an environmental question. A concise answer mode can reduce the number of tokens produced by an AI system, and on a scale of billions of users, this translates into real savings in electricity, water, and land. The report’s central message is that the environmental footprint of AI is not a fixed one. It depends on how systems are designed, where they operate, the types of tasks they perform, and how humans use them.
Short FAQ:
why AI It takes so much energy.?
AI systems require large data centers, advanced processors, and continuous cooling. Model training and daily use of billions of queries combine to create high power consumption.
Why the report also emphasizes water and land, not just carbon?
Because every kilowatt-hour of electricity also involves the use of water, land, and infrastructure. A low-carbon energy source can still require a lot of water or space.
What can be done to reduce the footprint of AI?
You can choose lighter models, reduce unnecessary outputs, use AI only when there is a real need, require environmental transparency, and integrate data centers into energy, water, and land planning.
More of the topic in Hayadan:
3 תגובות
I deleted my ChatGPT account.
I deleted my chatGPT account.
What you wrote here also consumes electricity and water and wastes the reader's time.
Artificial intelligence brings progress, medicine, and science to all corners of the world. That's worth something too...