A UN report details the increasingly serious consequences of AI as it relates to water, land, and carbon emissions

The report by the United Nations University – Institute for Water, Environment and Health provides new data on the water, land, carbon, and waste footprints associated with the rise of AI. While these are not new issues, analyzing and quantifying them brings clarity and confronts us with what is undoubtedly a massive problem.

The document presents scenarios in which resource consumption increases through 2030, including water and land resources, alongside rising CO2 emissions. It is important to note that the report uses actual figures to make projections through models; therefore, these projections depend on critical assumptions regarding AI’s share of data center consumption, advances in energy efficiency, the composition of regional electricity bills, and the cooling techniques employed, among other factors; and may vary depending on potential efficiency improvements, the adoption of models in devices (rather than data centers), or regulatory or political changes. Therefore, the figures should be interpreted within this context and not considered absolute truths. In any case, projections place electricity consumption at between hundreds and nearly a thousand TWh by 2030, equivalent to the consumption of 1.3 billion people in sub-Saharan Africa over five years.

Some perhaps lesser-known data points are relevant to users. For example, inference—the day-to-day responses to users—appears to consume as much or more energy than model training. Simply keeping prompts to the minimum necessary information and requesting short, concise responses would save an enormous amount of resources.

Although the information provided by the report is interesting, I find myself missing a more detailed analysis of the impact of different uses—for example, large-scale models versus on-device models, or scientific applications versus recreational use—comparing high-value social uses (research, medical applications) with uses such as the mass creation of entertainment content.

Given its technical nature, the report speaks in general terms about how data centers should be integrated into energy, water, and land-use planning, applying cumulative impact assessments that take other local demands into account. In this regard, the report seems to fall short in its assessment of the geopolitical situation and the real problem posed by the concentration of decision-making power in a few companies; personally, I find that the recent encyclical [by Pope Leo XIV] offers a much more comprehensive and interesting perspective.

In my opinion, the press release accurately reflects the report, which is a solid piece of work that addresses several issues of great relevance to the field of green AI. It does not merely focus on the carbon emissions aspect of AI models, but also on other factors that are often overlooked, such as the fact that a particular data centre’s emissions may be lower, but at the cost of requiring more water. It argues that all these factors should be taken into account when evaluating different models or the efficiency of data centres.

It is a very interesting article for the general public, as it can help raise awareness of an issue that people are generally unaware of, and does so with clear data and evidence; it also invites critical reflection on the use we are making of new technologies.

This is a very comprehensive report that takes a holistic approach to the environmental cost associated with data centre energy consumption, taking into account not only CO₂ emissions but also the water consumption and land use footprints linked to data centres.

To summarise its impact, I would highlight the following statement: ‘If the electricity consumption of data centres were considered that of a country, it would rank eleventh globally in terms of electricity consumption’ (my own translation). The report clearly reflects an economic reality in which the technology sector occupies a dominant position. It is worth noting that nine of the ten companies with the highest market capitalisation in the world belong to this sector, with Nvidia having the highest market value. This company, a manufacturer of the processors used to train and operate AI systems, has a market capitalisation exceeding the GDP of every country in the world, with the exception of the United States and China. In fact, this market capitalisation is equivalent to more than 13% of US GDP.

The report also highlights the growing share of artificial intelligence in data centre electricity consumption, rising from around 20% of the electricity consumed by these facilities in 2025 to a projected 40% in 2030. However, only 16% of countries have specialised AI infrastructure, and within this group, 90% of installed capacity is concentrated in two countries: the United States and China. Consequently, a very significant proportion of the increase in the environmental impact associated with artificial intelligence is also concentrated in these two economies.

Among the principles proposed by the report to move towards a responsible AI ecosystem, I find the principle of efficiency by design particularly relevant. The document highlights that a typical text query to a system such as ChatGPT can require approximately 200 times more energy than text classification tasks, such as spam filtering (the comparison is even greater when considering images or videos). In this context, the development and deployment of smaller, specialised models appears to be a promising way to reduce energy consumption without sacrificing the utility of these technologies.

Another aspect that caught my attention is the comparison of the environmental footprints of data centres in different countries. For example, France, which hosts one of the largest numbers of data centres in Europe, has a significantly lower carbon, water and land-use footprint per kWh than other countries with a high concentration of such infrastructure, such as the United Kingdom, Italy or Germany. Such comparisons highlight that environmental impact depends not only on the scale of installed capacity, but also on the characteristics of the energy system that powers it.