The AI boom is producing ever more powerful AI models, and more and more people are using the technology. According to a new report, not only is the power consumption of AI data centers increasing, water and space are also being required in ever larger quantities – including for “green” AI data centers. The problem of electronic waste is also underestimated, according to the researchers. However, the report also received criticism, among other things because of the partly outdated database and because particularly efficient AI technologies were hardly taken into account.
Artificial intelligence is making rapid progress, AI models and AI agents can take on more and more tasks and process them independently. GPT, Gemini, Claude and Co can no longer only generate texts, images, videos or computer code, but can also plan processes, optimize structures and even solve more complex problems. However, this comes at a price: the more powerful the large AI models become, the more computing capacity they require. High-performance processors are required for both training and ongoing operations.
“AI is not just code”
“AI is not just code, it also includes physical infrastructure and supply chains, including data centers, chips, electricity, cooling systems, water, land use, critical raw materials and finally e-waste,” state Miriam Aczel from the Institute for Water, Environment and Health at the United Nations University (UNU-INWEH) and her colleagues. When it comes to the potential environmental consequences of the AI boom, the focus has so far been on electricity consumption and CO2 emissions.
In their new report, the research team now also looks at other consequences of increasing AI use, including the water footprint, land use, the unequal distribution of AI resources and the problem of electronic waste disposal. “This report is not a plea against artificial intelligence – a technological transformation that is improving the lives of billions of people around the world,” emphasizes project leader Kaveh Madani, Director of UNU-INWEH.
“It is a call to use this technology responsibly and proactively address its unintended consequences to make it sustainable and equitable,” said Madani. Given the rapid growth of the AI industry, only a narrow window of time remains to ensure sustainable development that respects planetary boundaries.
Even “green” AI data centers do not solve all problems
Specifically, the report states that AI technologies were responsible for around 20 percent of global data center electricity consumption in 2025. This was around 448 terawatt hours. However, the researchers predict a rapid increase in AI capacities by 2030, which could double the electricity consumption of data centers to 945 terawatt hours. This would correspond to around three percent of global electricity consumption, according to the report. He estimates the water demand due to AI to be around 9.3 trillion liters in 2030 and the land consumption to be 14,500 square kilometers.
What’s interesting is that switching to a “greener” energy mix in order to reduce CO2 emissions through AI data centers does not always benefit the environment: “What surprised us most is how often supposedly ‘greener’ alternatives from a CO2 perspective ultimately have a negative impact on water or land,” says Aczel. Switching from coal to bioenergy reduces the CO2 footprint by 70 percent, but the water requirement increases 30 times and land use increases 100 times.
“Low CO2 does not automatically mean water or space saving,” says the report. “These asymmetries can therefore exacerbate the environmental problems of local communities, while the strategic advantages of AI tend to manifest themselves elsewhere.”
Prompts account for 80 to 90 percent of the electricity demand
The report also corrects a common assumption about at which stage an artificial intelligence requires the most resources. So far, the focus has been on training the AI models. It is estimated that training GPT-4 alone required between 50 and 70 gigawatt hours. “However, in order to understand the true environmental costs of AI, one must also take into account the consequences of its daily operation,” explain the researchers. Because even if a single prompt doesn’t make a big difference, it adds up to billions of prompts every day worldwide.
As a result, normal AI operation has long since accounted for the lion’s share of electricity consumption: 80 to 90 percent of total energy consumption goes to processing prompts and tasks, according to the report. When the AI generates an image, it requires around 1,450 times as much energy as simply classifying a text. For a video clip it is 415 watt hours, around 200,000 times more. Factors such as the length of the prompt also influence how many tokens and therefore how much computing power the AI model requires.
“Many people think that as the technology gets better and more efficient, the environmental footprint of AI will shrink,” says Madani. “But that’s only partly true.” Because the more efficient and cheaper AI technology becomes, the more people will use it. However, this increases the number of users and inquiries – and with it the computing and power requirements.
The problem of disposal
And there is another problem that is often underestimated: the amount of electronic waste generated. Because the high-performance processors often only have a short lifespan, they are replaced after just a few months to years – which means that a lot of worn-out microchips and circuit boards are produced. “By 2030, the AI infrastructure could produce up to 2.5 million tons of electronic waste per year – that is roughly the mass of 250 Eiffel Towers per year,” write the researchers.
Overall, the report comes to the conclusion: “AI has remarkable potential, but in order to realize this responsibly, system change is required,” state Aczel and her colleagues. Real progress requires the anchoring of sustainability at all levels – from hardware and model design through supply chains, implementation and use to the disposal of worn-out components. “Responsible AI is possible when performance and responsible use within planetary boundaries go hand in hand,” say the researchers.
Report draws criticism from experts
However, there is also criticism of the UNU’s AI report: “It is currently very important to publish reliable and reliable figures on the resource consumption of artificial intelligence in order to correctly classify the wealth of information. Unfortunately, the report does not live up to this claim: it is sometimes difficult to understand, is based on old data or does not present it in the appropriate context,” comments AI researcher David Kappel from Bielefeld University, for example. The figures given in the report on the electricity mix and electricity consumption of data centers worldwide are based on data from 2015.
Another point of criticism from several AI experts is the fact that particularly efficient AI models such as DeepSeek are not taken into account and advances and new developments in hardware are not taken into account. “The technical efficiency gains in recent years are remarkable,” says Wolfgang Maaß from the German Research Center for Artificial Intelligence (DFKI). “Quantization, knowledge distillation, MoE architectures and hardware-specific optimization, for example through Flash Attention, already enable significant increases in efficiency today.”
Another criticism of the water footprint information is that the report does not distinguish between consumption and use. The former refers to water that can no longer be used because, for example, it evaporates, seeps into the ground or is contaminated. However, water that is only used for cooling or generating electricity can be reused. “If the water is also available for other things after energy production, then this is far less problematic,” explains hydrologist Thorsten Wagener from the University of Potsdam.
Source: United Nations University – Institute for Water, Environment and Health, 2026, Report: Environmental Cost of AI’s Energy Use