The figures now circulate with the certainty of facts. The world's data centres consumed approximately 415 terawatt-hours in 2024, nearly 1.5% of all electricity on the planet, and could approach 945 by 2030. In the United States, the figure was 176 terawatt-hours in 2023, 4.4% of national consumption. Training GPT-3 may have evaporated some seven hundred thousand litres of fresh water. I have cited these figures myself, and they are the best we have.
It is worth examining what kind of figures these are. The first is an aggregate estimate from an international agency, and the 2030 figure is, furthermore, a scenario projection. The second models the consumption of an entire country's infrastructure. The water figure is an academic calculation, and its own authors present it as such, not as an audited measurement. None of the three says anything about a specific installation. They are not designed for that purpose: they serve to establish the order of magnitude of a global phenomenon, which is not insignificant, rather than to determine what will occur in the building they intend to construct ten kilometres from your home.
For years, there was almost nothing on this subject, and the little we knew about specific cases came from other sources. It was reported that Meta's project in Talavera de la Reina would consume around eight per cent of the water allocated to the city, after reducing its initial forecast; in Aragon, groups that organised against various projects summarised the matter in a slogan that is difficult to improve: “your cloud dries up my river”. That information came from journalism and the environmental documentation that projects must submit, not from the operator publishing its meter readings.
The situation has changed very recently, and almost no one has reported it. Since 2024, a European regulation has required all data centres of a certain size to declare, one by one, their total energy consumption, water consumption, potable water usage, refrigerants, heat reuse, and percentage of renewables. Not in aggregate form: by installation, with a separate set of data for each physical location. The first submission took place in September 2024 and is repeated annually. The meter that we claimed for years did not exist, does exist, is read, and someone records the figure.
What is happening, and it is more interesting than the absence I had assumed, is that this figure per facility goes into a European database, and the same regulation that mandates its production orders it to be kept confidential. It states it clearly: the database will be public only in an aggregated form, and the Commission and Member States will keep the data and indicators for each individual centre confidential. It is measured with precision, delivered to the regulator, and what reaches the public is the sum. The data exists and is reserved.
Thus, the problem is not that no one is measuring; it is that the individual measurement is taken, verified in its format, and stored, and the law decides that the citizen, the local council, and the journalist see the country's total but not the building next door. And this is not an oversight: it is a regulatory choice, justified by the commercial secrecy of the operators. The aggregated figure protects one form of privacy—that of the company—and sacrifices another form of transparency: that of the territory hosting the facility and bearing its water bill.
It is appropriate to be fair about what does exist, as not everything is reserved. Some companies publish more than the law requires them to make public. Meta offers electricity and water consumption per facility in its environmental index, self-declared and with limited review by an auditor for selected metrics. Google publishes water usage by location of its centres and efficiency by campus, although not the absolute electricity consumption of each one. Microsoft and Amazon, however, remain at the global or regional average. The picture is uneven: there are those who show the meter of their own volition and those who show only the average for their fleet, and between the two there is no common criterion or third party to guarantee that what is shown is comparable.
This explains, incidentally, why the arithmetic of consumption per query remains a dead end, albeit now for a more precise reason. It is not just that it shifts the responsibility to the user, who is the only point in the chain without decision-making capacity. It is that the fine calculation per facility—the one that would allow one to state the true cost of an operation in a specific centre, at a specific time, with the electricity mix of that grid—is made on data that exists but is not public. Whoever publishes the prompt figure is estimating based on an average, because the exact data is locked in a drawer to which they have no access.
One final point remains, and I note it here as it pertains to a different domain. The impact of a data centre is not an inherent property of the building. The same design, with the same servers, proves almost innocuous on a decarbonised grid and in a water-abundant basin, yet highly contentious on a saturated grid and an aquifer under stress in August. It is a relational magnitude, not a characteristic of the object. Consequently, the national aggregate data, which is the only information made public, is precisely what is least helpful to those tasked with deciding on a specific installation: it averages out the very variable that matters, which is the location.
In Spain, meanwhile, the royal decree that would transpose all of this remains in progress: the public consultation closed in September of last year, and the ministry states it is working on the matter. When it arrives, it will inherit the same European framework, and with it the same question, which is no longer whether it is measured. It is measured. The question is who is permitted to read what has been measured. And today, the answer is that the territory hosting the centre and paying for its water is provided with the national average, while the figure for its own building remains, measured and reserved, in a database that it cannot access.
On open conversation
This text revisits the material footprint of computing from an angle that verification compelled me to correct halfway through: I believed the problem was the absence of measurement, yet it turns out that since 2024, a European regulation has mandated measurement by installation. What I discovered is something else, and more uncomfortable: that individual measurement exists, the regulator possesses it, and the law mandates that it be kept confidential, with only aggregates published. The dispute is not whether it is measured, but who may read the figure per installation. If anyone wishes to intervene from the perspective of environmental impact assessment, administrative law, data centre engineering, or the administration that grants authorisation, this notebook remains open, and I am particularly interested in being corrected if those individual data points are more accessible than I have found.
Sources
International Energy Agency. Energy and AI. 2025.
Shehabi, A. et al. 2024 United States Data Center Energy Usage Report. Lawrence Berkeley National Laboratory, 2024. DOI: 10.71468/P1WC7Q.
Li, P.; Yang, J.; Islam, M. A.; Ren, S. Making AI Less “Thirsty”. arXiv:2304.03271, 2023.
Commission Delegated Regulation (EU) 2024/1364 on the first common Union scheme for rating the sustainability of data centres, developing Directive (EU) 2023/1791 on energy efficiency. Obligation to report by installation; aggregate publication; confidentiality of individual data.
Meta, Environmental Data Index (electricity and water consumption per facility, with limited assurance by EY). Google, Environmental Report (water by location, PUE by campus). Microsoft and Amazon Web Services, sustainability reports (global and regional PUE/WUE).
Ministry for the Ecological Transition and the Demographic Challenge. Draft Royal Decree on the efficiency and sustainability of data centres (public consultation closed 15 September 2025; in progress).
Journalistic coverage of the Talavera de la Reina case (eldiario.es, Toledodiario) and data centres in Aragon (RTVE, eldiario.es).