When you save a file to the cloud, watch a video, or ask an artificial intelligence a question, the impression is that everything happens in an abstract, distant space with almost no physical impact. But the cloud has an address. It relies on buildings filled with servers, cooling systems, communication networks, and massive amounts of electricity.

And there is a resource in this equation that usually appears less in discussions about the environmental impact of technology: water.

A study published by Cambridge Prisms: Water, from Cambridge University Press, estimated that the artificial intelligence infrastructure cluster in the São Paulo metropolitan area, with an operational information technology load of approximately 550 MW, has a water footprint of 16.1 million cubic meters per year. This volume is equivalent to the consumption needs of more than 100,000 households.

The most curious fact, however, lies elsewhere: more than 46% of this water is not used directly inside the data centers. It is an indirect water footprint related to electricity generation by hydroelectric power plants.

This changes the way we must look at the environmental cost of computing.

The cloud needs water to work

Behind every server lies a simple physical problem: heat. The more processing the equipment performs, the greater the amount of heat generated, and in high-computational-density structures, such as those dedicated to artificial intelligence, keeping this temperature under control is fundamental to preserving system performance and availability.

This is where water enters the operation. Depending on the architecture used, it can participate directly in the cooling system, either in circuits that remove heat from the equipment or in cooling towers that dissipate this heat through evaporation. The Lawrence Berkeley National Laboratory highlights that direct water consumption in data centers is generally associated with cooling needs.

But this does not mean that every data center consumes the same amount of water. The chosen cooling system makes a huge difference. Liquid cooling technologies, hybrid systems, and solutions that use outdoor air can significantly alter this demand. An analysis by Berkeley Lab itself found differences of over 10,000 times in water use per workload, influenced by factors such as server efficiency, cooling type, infrastructure utilization, climate, and electricity grid characteristics.

Therefore, when people say that the cloud "consumes water," the question is not just how much water enters a data center, but also how that infrastructure was designed to transform energy into processing without letting heat interrupt operations.

And it is precisely when we expand this calculation beyond the walls of the data center that the most surprising part of the story emerges: the water used to generate the electricity that keeps these servers running.

São Paulo reaches 16.1 million m³ per year

This is when the relationship between computing and water ceases to be just an engineering question and acquires a concrete dimension. A study published in April 2026 in Cambridge Prisms: Water estimated that the artificial intelligence infrastructure cluster in the São Paulo Metropolitan Area, with an operational IT load of approximately 550 MW, has an annual water footprint of 16.1 million m³.

The volume is equivalent to about 16.1 billion liters of water per year. To provide a dimension that is easier to visualize, the researchers themselves estimate that this quantity corresponds to the needs of more than 100,000 households. It is important, however, to make a caveat: this number represents a water footprint, and does not mean that 16.1 billion liters are withdrawn directly from São Paulo's supply system. The estimate combines both the consumption associated with cooling the data centers and the indirect consumption related to generating the electricity they use.

And it is precisely this second portion that makes the number even more interesting: more than 46% of the estimated footprint is indirect, linked to the so-called virtual water of hydroelectric generation. In other words, a considerable part of the water associated with the operation of this infrastructure is not in the servers, but in the energy chain that keeps them functioning.

Almost half of the water is "hidden" in the electricity

The water associated with data centers is not just in the systems used to cool servers. A significant portion is in the very electricity that keeps this infrastructure running. In the case analyzed in São Paulo, of the 16.1 million m³ of water footprint estimated for the AI cluster, 7.45 million m³, or 46.3%, correspond to indirect consumption associated with power generation.

This happens mainly because of the share of hydroelectric plants in the Brazilian electricity matrix. Hydroelectric generation depends on reservoirs that lose water through evaporation. In the study's methodology, a portion of the evaporated water is attributed to the electricity consumed by the data centers, forming what researchers call "virtual water." Thus, even if the water does not physically pass through the servers, it is part of the water footprint required to keep the equipment running.

The concept helps reveal a connection that normally stays out of the discussion on digital infrastructure: the more computing is required, the greater the demand for electricity and, depending on the source of that energy, the greater the pressure on water resources can also be. In periods of drought, this relationship becomes even more sensitive, since changes in the share of hydroelectric power can alter the amount of water associated with energy generation.

In other words, the water that sustains the cloud is not always where we imagine. Part of it is hidden in the chain that produces the energy needed to keep the cloud running.

The problem is not just in the data center

Looking only at the water used inside data centers leaves an important part of the calculation out. In São Paulo, 46.3% of the estimated water footprint is indirectly related to the generation of the electricity used by the AI cluster.

This water can be miles away from the servers, as in the case of evaporation from hydroelectric reservoirs that help supply the electrical system. And this relationship can change depending on weather conditions: in the 2021 drought scenario analyzed by the study, indirect water consumption for the same data center load increased by 27.2%.

Therefore, assessing the environmental impact of the cloud requires looking beyond the physical structure of the servers. Digital infrastructure is connected to a water and energy chain that extends beyond the limits of the data center itself.

Artificial intelligence is increasing the scale of infrastructure

Artificial intelligence is not just adding new applications to existing digital infrastructure. It is changing the scale of processing required. AI models require large volumes of computing for training and operation, which drives the expansion of data centers and, especially, the installation of more powerful and dense servers. The International Energy Agency (IEA) points out that data center electricity consumption grew by 17% in 2025, while that of AI-focused data centers advanced by about 50% in the same period.

This expansion also changes the physical structure of these facilities themselves. According to the IEA, the power density of AI-focused servers increased 11-fold between 2020 and 2025 and could grow another four-fold by 2027. More capacity concentrated in less space also means a greater need for systems capable of removing the heat produced by this equipment.

And growth does not seem to be close to stabilizing. The IEA's current estimate is that global data center electricity consumption will practically double, going from 485 TWh in 2025 to about 950 TWh in 2030, while consumption by data centers dedicated to AI is expected to triple in this period.

For São Paulo, this is especially relevant because the study on the water footprint of AI already starts from an estimated cluster of 550 MW of operational IT load in the Metropolitan Region. If computational capacity continues to grow, the discussion about water stops being just a snapshot of current consumption and starts to involve how this infrastructure will be expanded and what resources will be needed to sustain it.

São Paulo cannot treat water and energy as separate problems

When a data center increases its demand for electricity, it is not just putting pressure on the energy system. In a region whose generation depends significantly on water, this demand can also translate into pressure on water resources. It is precisely this connection that the study on São Paulo calls the water-energy nexus.

The problem becomes more evident in periods of drought. In the study's analysis, when the 2021 scenario is applied, with a lower hydroelectric share and higher use of thermoelectric generation, the indirect water consumption associated with the same data center load rises to 9.48 million m³, a 27.2% increase. In other words, the same digital infrastructure can have a different water footprint depending on the conditions of the electrical system that supplies it.

There is also a geographical issue. The analyzed cluster is concentrated in the Metropolitan Region of São Paulo, within the Alto Tietê basin, a region that the study classifies as having high water stress. This means that it is not enough to assess how much a new facility consumes: it is also necessary to consider where it will be installed, from which electrical system it will receive energy, and what the water situation is in that region.

This is why the expansion of data centers should be planned in an integrated manner. Water, energy, and digital infrastructure are part of the same equation, and ignoring one of these variables can cause an apparently efficient solution to transfer the impact to another point in the system.

The next data center metric could be water

For years, data center efficiency was measured mainly by energy consumption. Now, water is beginning to gain ground in this discussion. WUE (Water Usage Effectiveness), standardized by ISO/IEC 30134-9:2022, already allows measuring the water efficiency of these facilities.

With the advancement of AI and the increased demand for processing and cooling, this indicator tends to gain even more importance. For São Paulo, this means looking not only at how much energy a data center consumes, but also at how much water is needed to keep it running.

The cloud may seem invisible, but its resources are not. As artificial intelligence expands the demand for processing, understanding the water cost of technology will be just as important as measuring its energy consumption. After all, the digital future also depends on resources that are not digital.