The Cloud Is Not Ethereal: AI, the Electric Grid, and Other People’s Water

Introduction

In the current artificial intelligence- discussion , the management and use of natural resources have taken on considerable importance—not only because of the quantities consumed, but also because of their effects on the communities living near this infrastructure and on the environment. The sources analyzed for this essay reflect a complex reality in the fields of energy and water, in which genuine technical progress coexists with structural deficiencies that affect electrical grids, river basins, and those same communities. All of this reveals, with increasing clarity, the capacity—or lack thereof—of states and regions to govern artificial intelligence, a technology whose expansion appears difficult to reverse. This essay interprets this as part of a transformation comparable to the Industrial Revolution, because it is reshaping how energy, water, and land are allocated among states, companies, and communities.[1] The comparison is not merely rhetorical. The first Industrial Revolution was also, in this interpretation, a story of coal and rivers, and its costs were felt first in the towns that hosted the mills.

The electricity consumption of artificial intelligence is one of the central questions of this study. Although it is true that artificial intelligence accounts for a small share of global electricity consumption, the reports consulted warn of tensions that undermine the ability of public authorities to set the terms for its deployment, and these tensions call for structural decisions that will be analyzed below. Data centers consumed about 1.5 percent of the world’s electricity in 2024, according to the International Energy Agency.[2] In 2025, they consumed 23 percent of Ireland’s metered electricity, according to the Irish Central Statistics Office and press reports based on its data.[3] Although it may seem contradictory, both figures are correct. This essay examines that apparent contradiction, because the choice between the two figures determines almost everything that follows.

With regard to the structure of the article, it first examines the general situation of energy consumption and how it is measured. It then turns to the specific challenges in the areas of electricity and grid capacity, water and land use, and the access models that different countries have adopted. Finally, it proposes the conditions a country should meet in order to play a strategic role in this field and avoid falling behind—which, as will be argued, does not mean having the most capacity.

Before beginning, a methodological note is in order. This is a documentary analysis. It draws on official statistics, reports from international agencies, company disclosures, court and regulatory decisions, and specialized press coverage, using data from 2023 to 2026, all of which were consulted in October 2026. Since these sources combine measured data, modeled estimates, and self-reported information, the notes specify which type of data is involved in each case and indicate when a figure is based on a single secondary source.

Overview of Consumption and How It Is Measured

The global figures can, in principle, be considered reliable, but at the same time they are subject to a significant framing issue, which is evident in the public’ discussion s regarding the consumption of natural resources. The IEA estimates that data centers consumed 415 TWh in 2024 and 485 TWh in 2025—a 17 percent increase—and projects consumption of approximately 950 TWh by 2030, representing roughly 3 percent of global demand.[4] In contrast, electric motor systems accounted for 53 percent of global electricity consumption in 2023—some 13,462 TWh—and space cooling for about 10 percent, according to a 2018 IEA assessment.[5] Data centers also account for less than one-tenth of the growth in global electricity demand through 2030.

Artificial intelligence in the strict sense accounts for an even smaller fraction, although the definition used determines the result. Gartner estimates that AI-optimized servers will consume 95 TWh in 2025, representing 21 percent of its estimate of 447 TWh for data centers as a whole, which accounts for roughly 0.3 percent of global demand.[6] In terms of individual use, the figures are very small, since Google reports a median of 0.24 Wh for a text prompt and Epoch AI estimates about 0.3 Wh.[7]

These per-query figures, however, tend to circulate in the public discussion as if they described the total cost of artificial intelligence, and when interpreted that way, they are misleading. The figure published by Google covers a median text prompt and explicitly leaves the measurement of training to future work, while the figure that OpenAI’s CEO provided for ChatGPT in 2025 was presented in a blog post without a published methodology.[8]Among the sources reviewed, Mistral AI is the only provider that publishes a lifecycle analysis that includes training.[9] Table 1 outlines what each disclosure covers and what it omits. The energy consumption of artificial intelligence does not, therefore, stem from simple conversational use, but rather from training, from computationally intensive tasks such as reasoning, image processing, and video processing, and from the sheer volume of deployment.

Perceptions of the problem vary, and two commonly cited comparisons do not hold up when checked against the sources consulted. Electric vehicles, often said to consume more energy, used about 250 TWh in 2025—roughly half as much as data centers.[10] Comparisons with steel, cement, aviation, or streaming typically lack a primary source and pit activities that rely largely on fossil fuels against electricity alone; therefore, this article does not use them.

Taken together, the sources consulted suggest that the main problem with current discussion is not the lack of figures, but the choice of the unit in which those figures are expressed. The familiar argument compares stocks when the discussion concerns flows and locations. A use case that represents 1.5 percent of the total but grows 17 percent a year—while facilities focused on artificial intelligence grew about 50 percent in 2025—carries considerably more weight at the margin than on average, given that total demand grows about 3 percent.[11] In the United States, data centers account for about half of the expected increase in demand through 2030, and in advanced economies for more than one-fifth.[12] Nor does efficiency settle the matter. Google reports a thirty-three-fold reduction in energy consumption per prompt in a single year, yet its total electricity consumption rose 37 percent in 2025.[13] This is the dynamic that the specialized literature describes as a rebound effect, although the available evidence does not allow for its magnitude to be measured.[14]

This diagnosis becomes particularly evident in the specific areas examined in the next section.

Analysis of Specific Areas

i. Electricity and the power grid: the gap between the global average and the local share

In terms of geographical distribution, the proportion of electricity consumption by data centers varies by a factor of fifteen or more depending on the scale considered. Figure 1 shows the local shares represented by the sources, along with the type of figure each one represents, since mixing metered consumption with requested capacity is a common error in this discussion.

 
Ireland is a striking extreme case. Data centers accounted for 5 percent of metered electricity in 2015, rising to 22 percent in 2024—a period during which total metered consumption grew by 30 percent and their consumption grew by 462 percent. By 2024, their share exceeded that of urban households (18 percent) and rural households (10 percent), taken separately.[15] In 2025, they consumed 7,663 GWh, while residential consumption rose by only 1 percent. In Virginia, the share exceeded one-quarter in 2023, and commercial demand shifted from an annual growth rate of 1.4 percent between 2009 and 2019 to 5.8 percent between 2019 and 2023.[16] In Frankfurt, the figures range from 20 to 40 percent, likely because they are based on different areas or years; therefore, no single figure should be cited.[17]

Aragon is revealing for a different reason. It is an autonomous community of Spain that is, to a large extent, a demographic desert, since some 85 percent of its territory has fewer than 12.5 inhabitants per square kilometer.[18] There, data centers accounted for roughly 14 percent of the total in 2025, according to a press estimate, with some 108 MW of computing capacity installed.[19] An sparsely populated area is attractive to this industry precisely because land is available, but its power grid was not built to handle loads of this magnitude.

The question of who pays for this expansion further complicates the discussion. The independent market monitor of PJM—the largest electricity market in the United States—attributes 45 percent of the $47.2 billion in capacity costs across three auctions to data center load forecasts.[20] By contrast, an analysis by the Lawrence Berkeley National Laboratory, cited by the Congressional Research Service, concludes that grid investment—and not data centers—was the main driver of retail price increases between 2019 and 2025.[21]Since the discrepancy concerns attribution and not measurement, both perspectives deserve to be presented.

To preserve the objectivity and clarity of this analysis, it is also necessary to warn of the risk of treating announcements as actual consumption. The 233 GW of large-load requests recorded in Texas in March 2026 represent “requested capacity,” a category that includes applications which specialists from the University of Texas and the Regulatory Assistance Project describe as not all being genuine.[22] The same caution applies to the 38 GW of access requests reported in Brazil and to the figures granted in Spain. The latter are examined below.

ii. Water and territory: the national average masks the river basin

The treatment of water in this discussion has evolved from a stage of near-invisibility toward a framework of litigation and local resistance, although measurement remains inadequate. The most reliable direct figure is modeled rather than measured, since the Lawrence Berkeley National Laboratory estimates approximately 17.1 billion gallons—about 65 million cubic meters—of on-site consumption by data centers in the United States in 2023, while acknowledging that few direct measurements exist.[23] Indirect water—consumed by the power plants that supply electricity—changes the order of magnitude, and the team led by Shaolei Ren projects global water withdrawal for artificial intelligence of between 4.2 and 6.6 billion cubic meters in 2027, a projection based on assumptions rather than actual measurements.[24] These figures cannot be placed on a single scale or interpreted as a trend, because each one measures something different. Table 2 shows why.

 
Despite the prevalence of reassuring comparisons, the most frequently cited one—that between data centers and golf courses—does not hold up to scrutiny. It conflates direct and indirect water use, equates consumption with withdrawal, and overlooks the fact that a golf course is located within a watershed shared by its own users, whereas a data center may provide services from a stressed watershed to a much broader user base, often located far away.[25] The national figure obscures the basin, which is precisely where the conflict takes place.

Documented cases illustrate this mechanism. In February 2024, a Chilean environmental court partially annulled the permit for Google’s data center in Cerrillos, Santiago, and required that the assessment of the aquifer take into account the effects of climate change following some fifteen years of drought; the company subsequently redesigned the facility to use air cooling.[26] In August 2025, the Tucson City Council in Arizona unanimously rejected Project Blue, a development linked to Amazon, amid disputes over the confidentiality agreements that had surrounded the negotiations.[27] In March 2026, a Spanish regional high court accepted the first administrative appeal against a plan by Amazon Web Services in Aragon, with arguments citing a lack of transparency regarding water use.[28] Finally, in Georgia, four families sued goal in September 2026 over alleged damage to an aquifer and contamination of wells—allegations that the company disputes and that remain unproven.[29]

In the most clearly resolved case—that of Cerrillos—the dispute with one of the Big Tech companies resulted in a redesign rather than outright cancellation, and that outcome has implications that deserve to be made explicit. A common solution for reducing water use is air cooling. However, air cooling simply shifts the burden to the electricity grid, because it requires more power; thus, the water problem is merely transferred rather than resolved. The conclusion is that water and electricity cannot be treated as separate policy issues.

Not every conflict over this infrastructure involves water, however. The Irish dispute is a matter of grid capacity, and goal’s project in Zeewolde, in the Netherlands, was halted due to concerns about renewable energy and land use.[30] Presenting them as water-related cases would weaken the argument rather than strengthen it.

The Gulf exhibits the structural manifestation of this same tension. Annual rainfall there is less than 100 millimeters, and desalination provides between 60 and 90 percent of the water supply—a process powered mainly by natural gas—so cooling with water competes directly with a resource produced from fossil fuels.[31]

iii. Models of access: capital, saturated grid, regulated access, and water-stressed periphery

The evidence reviewed allows us to outline a typology of four models— provisional —which is proposed here as an analytical tool rather than as a classification established by the sources themselves. Table 3 summarizes them.

The first model is capital. It combines abundant capital and energy with borrowed chips, and corresponds to the Gulf. Stargate UAE plans a 5-GW campus , with 200 MW expected to be in operation by the end of 2026, while Humain, in Saudi Arabia, aims for 1.9 GW in 2030 and 6.6 GW in 2034.[32] These are plans, not operational capacity. What this model offers is capital, land, and state-directed energy; what it lacks is water and advanced chips, since in May 2025 Washington was considering the export of large volumes of semiconductors to Saudi Arabia and the United Arab Emirates—a practice that has been restricted since the end of 2023—which makes deployment dependent on foreign authorization.[33] The lesson of this model is that capital can buy capacity, but not control over the inputs on which that capacity depends.

The second model is that of the saturated grid, exemplified by Ireland and Virginia, where the challenge lies not in generation but in the grid itself. Dublin has been operating under a de facto moratorium since 2021. Subsequently, in December 2025, the Irish regulator ruled that large energy users must provide dispatchable generation or storage and obtain 80 percent of their electricity from new renewable projects within six years.[34] In the United States, regulation operates at the state level, and Texas approved Senate Bill 6 in 2025 for loads of 75 MW or more, while Oregon created a separate tariff class for loads above 20 MW.[35]Virginia adds a fiscal dimension that the Irish case lacks. In Loudoun County, more than 40 percent of tax revenue—some 1.2 billion dollars in fiscal year 2026—comes from data centers, which leaves the county vulnerable if demand falls.[36] Local officials suspended new approvals in September 2026 amid resident backlash, established a $120 million revenue stability fund in 2024, and are drafting a requirement for operators to post decommissioning bonds.[37] The lesson to be learned from this model is that a state may lose control of its grid—and its budget—by granting approvals too quickly.

The third model imposes conditions on grid access and is represented by Spain and the European Union in general. At the EU level, only transparency obligations are in force: specifically, the requirement for operators of facilities with a capacity of at least 500 kW to submit annual reports under the Energy Efficiency Directive, and the requirement under the Artificial Intelligence Act that providers of general-purpose models document their energy consumption.[38] Spain, for its part, has a draft royal decree currently under public consultation—and therefore not yet in force—applicable to facilities exceeding 1 MW, which would make grid access contingent on at least 80 percent renewable energy consumption, with additionality and hourly matching requirements.[39] The Spanish case also highlights the gap between capacity allocation and actual consumption, since more than 12 GW of access rights have reportedly been granted against a sector target of 2.5 GW of computing power by 2030, while Aragon has approximately 108 MW installed.[40] Here too, the limiting factor is the grid, given that only 141 of 542 interconnection nodes have agreed-upon reference values.[41] The lesson from this model is that transparency informs the government but does not yet bind the operator.

The fourth model corresponds to the water-stressed periphery, as seen in Chile, Mexico, and Brazil, where the requested capacity is considerable, the installed capacity is small, and water is the main source of conflict. Brazil reports 38 GW in access requests, which are applications rather than actual consumption.[42] Mexico has approximately 279 MW in operation and at least thirty facilities in Querétaro, a state subject to severe water stress.[43]A Latin American study documents the absence of informed consent in Querétaro, in Canelones, Uruguay, and in Patagonia.[44] In addition, it should be noted that the sources consulted do not provide the share of national electricity that data centers represent in any of these countries, so no comparison with the Irish figures can be made. The lesson from this model is that capacity can be brought to a region that never chose it, with the watershed bearing the cost.

 
One limitation affects the entire analysis. Epoch AI's open database tracks ninety-three large facilities, but it is skewed toward the United States and covers an estimated 43 percent of global capacity, which may not be sufficient to draw general conclusions.[45] Satellite imagery tracks construction but not consumption, and the lack of transparency at individual sites is established only indirectly, through disputes such as those in Aragon and Tucson.[46] Open sources therefore provide an adequate description of the United States but a poor one of the Gulf and the periphery—an asymmetry that is a finding in itself, since transparency is distributed just as unevenly as the infrastructure.

iv. Conditions for a strategic country: from hosting capacity to setting terms

Regardless of the pace at which artificial intelligence advances, the question for a country is no longer whether to host its infrastructure, but on what terms. A country’s strategic importance in this field stems not from the computing capacity it hosts, but from the terms it is able to set. Five conditions emerge from the cases examined, and they encapsulate the lessons drawn above.

The first is a reliable electricity supply and a grid capable of carrying it. The IEA warns that about 20 percent of planned projects could be delayed by grid bottlenecks, that transmission lines require between four and eight years to build in advanced economies, and that lead times for transformers and cables have doubled in three years.[47] A commitment to gigawatts, therefore, does not equate to a grid.

The second is water and territory, which requires an explicit rule on priority when a watershed faces a shortage. Access to water is recognized as a human right, and General Comment 15 of the Committee on Economic, Social, and Cultural Rights gives priority to staff and domestic uses.[48]

The third is access to advanced chips, which remains subject to export authorization, as the specific case of the Gulf demonstrates. Emma Strubell and her co-authors warned in 2019 that the cost of cutting-edge computing excludes a large part of the research community, and the AI Now Institute maintains that computing capacity is highly concentrated, so that access to it functions as a form of power.[49] This dependence begins even before the data center. Taiwan produces just under 90 percent of the most sophisticated chips, and during the 2021 drought—the worst on the island since 1964—manufacturers such as TSMC had to bring in water by truck.[50] The water footprint of artificial intelligence therefore begins in a basin that is neither that of the data center nor that of its users.

The fourth is rules and transparency, since reporting by site and by model is a prerequisite for everything else. Europe has begun implementing reporting requirements, and Spain and Ireland have gone a step further by linking compliance to access to additional clean energy generation. Some requirements are already law. Since the 2023 recast of the Energy Efficiency Directive, member states must ensure that data centers with a rated energy input above 1 MW use their waste heat—for example, in district heating—unless they can demonstrate that this is not technically or economically feasible.[51] A rule of this kind transforms a byproduct into a benefit for the region that hosts the facility. A country that cannot measure cannot negotiate.

The fifth is social legitimacy. Projects announced under confidentiality agreements, as happened in Tucson, are met with resistance. Participation should be understood here as an ethical requirement based on subsidiarity, and not as a legal obligation currently in force, since no general legal standard of informed consent applies to this type of project.

The Spanish case—of which Aragon is the clearest example—illustrates why the fourth and fifth conditions are the most demanding. A state that grants access to twelve gigawatts despite a plan for two and a half has attracted capacity without having set terms. A strategic country, by contrast, is one that can refuse and that can grant consent on conditions it has stated in advance, such as the decommissioning bonds currently being drafted in Loudoun County.

Conclusion

The energy and water consumption of artificial intelligence presents a clear duality. While it accounts for a modest share of global electricity consumption and has a genuinely low cost per query, it also creates deficits that affect power grids, river basins, and the accountability of those who authorize this infrastructure.

The data show that the main challenges stem not so much from the volume consumed as from the location, the pace, and the visibility of that consumption. The figure that best summarizes this analysis is not the 1.5 percent of global electricity, but the gap between that figure and Ireland’s 23 percent. Perceived lack of transparency, the inadequacy of site-specific measurement, the treatment of requested capacity as if it were demand, and the difficulty of determining who bears the cost of the grid all underscore the need to strengthen institutional controls before—rather than after—the authorization of each project.

Three moral references allow this diagnosis to be framed beyond mere statistics. *Antiqua et Nova*, the Holy See’s 2025 note on artificial intelligence, recalls that “‘the cloud’ is not an ethereal domain separate from the physical world” and that current models and their hardware “consume vast amounts of energy and water,” while warning of “the concentration of power over mainstream AI applications in the hands of a few powerful companies.”[52] Laudato si’ provides the vocabulary of water, stating that “access to safe drinking water is a basic and universal human right,” and speaks of a true “ecological debt” between the global North and South.[53] Finally, *Magnifica humanitas*, Leo XIV’s encyclical on artificial intelligence, acknowledges that current systems “require enormous amounts of energy and water,” rejects development that shifts “costs and burdens onto others,” and calls for local communities to have a voice in decisions regarding data and digital environments, since “we cannot allow a handful of actors to dictate these processes on their own.”[54] None of these texts constitutes a policy, but together they provide a criterion that the figures cannot offer—namely, whether the burden of a project falls on those who did not choose it and will not share in its benefits. That criterion is not exclusive to Catholic social thought. Elinor Ostrom demonstrated that shared watersheds and aquifers—including the irrigation huertas of Valencia and Murcia—can be managed sustainably when those who depend on them participate in establishing the rules.[55]

Ultimately, the consolidation of a strategic position in this field will depend on the ability of states to assess the situation before granting authorization, to set their conditions before signing, and to prioritize people over machines when a watershed runs low, integrating energy policy, water governance, and the common good within a single framework of legitimacy. The fundamental question is not how much energy artificial intelligence consumes. It is who makes the decisions, based on what information, and for whose benefit.