The Hidden Cost of AI: Energy, Water, and Data Sovereignty

Introduction: From the Abstract Cloud to the Physical Infrastructure of AI

The Expansion of Generative Models and the Growth of Their Environmental Impact

The expansion of generative artificial intelligence has transformed computing into a large-scale physical infrastructure. These models require specialized servers, a continuous supply of electricity, cooling systems, water, semiconductors, industrial land, and transmission networks. The “cloud” metaphor obscures these material flows and shifts their costs onto specific regions, local communities, and international supply chains (Hlabisa, 2025; Monstadt & Saltzman, 2025). The future scale depends on server expansion, usage intensity, and efficiency improvements, so estimates have wide ranges (Herrera et al., 2025; Xiao et al., 2025).

Scope of the Review: Life Cycle, Data Centers, and Infrastructure Governance

This review examines the AI footprint throughout its life cycle, focusing on training, inference, hardware manufacturing, and data center operations. It also analyzes cooling, geographic localization, corporate concentration, and sovereignty over data and resources. The central argument is that technical efficiency is not enough if the absolute expansion of computing outweighs its environmental benefits.

Energy and Carbon: AI’s Growing Footprint Throughout Its Lifecycle

Electricity Demand, Training, Inference, and Data Center Expansion

AI’s electricity demand stems from several phases that must be accounted for collectively. Training state-of-the-art models involves intensive energy consumption, but inference can match or exceed this consumption when systems are integrated into widely used services (Oliveira et al., 2026). Added to this are storage, networks, auxiliary equipment, and cooling. Therefore, attributing the footprint exclusively to training results in an incomplete representation of the life cycle.

Recent estimates show the potential scale of this growth. In the United States, the expansion of AI servers between 2024 and 2030 could generate between 24 and 44 Mt of annual CO₂ equivalent emissions, depending on the pace of deployment and grid conditions (Xiao et al., 2025). Another estimate places carbon emissions specifically associated with AI in 2025 between 32.6 and 79.7 million metric tons, although it notes that distinguishing between AI workloads and other workloads remains challenging (de Vries-Gao, 2025). These figures are not directly comparable: the first projects U.S. servers under expansion scenarios, while the second estimates the global AI footprint using approximations based on data center data.

Electricity consumption also alters the operation of energy systems. AI loads combine high power density, temporal variability, and demanding thermal requirements—conditions that affect grid planning, stability, and electricity markets (Sheng et al., 2026). This article was subsequently retracted; therefore, its claims should be treated as an unvalidated reference rather than a basis for definitive conclusions. In contrast, a framework for coordinated response between data centers and electric systems proposes shifting tasks based on the carbon intensity of electricity, incorporating solar power and storage without exchanging sensitive information (Ruan et al., 2024). Time-shifting can reduce operational emissions, but it does not eliminate the material demand associated with building new capacity.

The relationship between efficiency and total consumption depends on how improvements are implemented. Thermal optimization, load balancing, and equipment control can reduce electricity consumption per unit of service (Rostami et al., 2024; Zhu et al., 2024). At a data center in Guangzhou, a predictive control strategy reduced annual cooling consumption by up to 22.15% compared to a baseline strategy under a 55% computing load (Zhu et al., 2024). This result applies to a specific installation and configuration, not to a general reduction in the AI footprint. The expansion of services may offset—or even exceed—unit-level gains.

From Operational Efficiency to Systemic Climate Footprint

The climate footprint depends on where the computing takes place, the time of consumption, and the electricity mix. Grid decarbonization can reduce operational emissions, but the construction of servers, semiconductors, and buildings entails embedded emissions that electricity-focused metrics often exclude (Oliveira et al., 2026). Life-cycle analysis must also incorporate hardware replacement and supply chains, since manufacturing is part of the infrastructure that supports each workload (Herrera et al., 2025; Hlabisa, 2025).

Location introduces environmental and policy differences. In the United States, the spatial distribution of servers, the pace of grid decarbonization, and efficiency measures account for much of the uncertainty in projections (Xiao et al., 2025). Best practices could reduce emissions by up to 73% in certain scenarios, but their effectiveness is limited by existing energy infrastructure (Xiao et al., 2025). This finding calls into question promises of climate neutrality based solely on internal corporate improvements or offsets.

The appropriate response requires a shift from isolated indicators to systemic constraints. The review literature concludes that hardware efficiency, software optimization, and cooling must be coordinated with decisions regarding scale, location, intensity of use, and timing of operation (Oliveira et al., 2026). The demand-side response can align part of computing with less carbon-intensive electricity, but it requires rules on transparency and data protection (Ruan et al., 2024). Therefore, the energy issue leads directly to the issue of water: every kilowatt consumed has a water component that depends on both cooling and electricity generation.

Water and Cooling: The Invisible Resource Behind Intensive Computing

Factors Determining Water Consumption per Workload

AI’s water consumption includes water used directly on-site, water associated with electricity generation, and water incorporated into manufacturing and supply chains (Herrera et al., 2025). This classification avoids reducing the problem to just the water evaporated in cooling towers. It also allows for a distinction between local consumption—which can affect aquifers and municipal water systems—and consumption displaced to regions that produce electricity or hardware.

Available estimates show substantial uncertainty. The expansion of AI servers in the United States could require between 731 and 1,125 million cubic meters of water per year during 2024–2030 (Xiao et al., 2025). For global AI, another estimate places the 2025 water footprint between 312.5 and 764.6 billion liters (de Vries-Gao, 2025). These figures should not be added together or compared as if they measured the same system: they differ in geography, footprint definition, time horizon, and attribution method.

Water use per workload varies by more than 10,000-fold in the review by Lei et al. (Lei et al., 2025). This variation stems from differences of more than 1,000-fold in water consumption per kilowatt-hour of server electricity and differences of approximately ten-fold in server efficiency (Lei et al., 2025). The determinants are ranked by server efficiency, the water factor of the power grid, utilization, cooling type, infrastructure efficiency, climate, proportion of idle servers, and hardware replacement (Lei et al., 2025). This wide range makes it impossible to assign a universal value in liters per query or per model.

Future scale also depends on the growth of computing. A probabilistic framework projects that, without mitigation measures, global water consumption associated with data centers could increase more than sevenfold by mid-century, with operational cooling consumption accounting for the majority (Herrera et al., 2025). The projection combines limited empirical data with expert-informed assumptions, so it presents scenarios rather than a specific prediction. Its value lies in showing which variables—particularly computing growth and efficiency trends—have the greatest impact on the outcome.

Air, Water, and Liquid Cooling: Savings, Trade-offs, and Local Limitations

Air cooling remains widespread but becomes less effective as the thermal density of processors increases. Reviews describe alternatives that include free cooling, rack-scale systems, cold plates, immersion cooling, and thermofluids (Alkrush et al., 2024; Li et al., 2024). Liquid cooling performs better under high thermal loads, although it introduces operational complexity, control requirements, and potential environmental impacts from the fluids used (Azarifar et al., 2024). In two-phase immersion systems, some engineered fluids pose global warming potential as well as long-term health and reliability issues (Azarifar et al., 2024).

Free cooling can simultaneously reduce electricity and water consumption under suitable climatic conditions. Passive radiative cooling, simulated for data centers in Singapore, achieved an average annual energy savings of approximately 20% when directly cooling the return air, while another configuration reduced water consumption by about 84% (Aili et al., 2024). These simulation results are dependent on the configuration and climate and are not guarantees that can be applied to any location. The choice of technology must consider temperature, humidity, water availability, and compatibility with existing equipment.

Operational improvements also involve trade-offs. In a facility with direct liquid cooling, raising the water supply temperature from 17 to 25 °C reduced the cooling system’s energy consumption by 63.3% at an outdoor wet-bulb temperature of 19 °C, but it nearly doubled the heat transfer to the room air and increased IT equipment consumption by 3.15% (Gheni et al., 2025). The result demonstrates that the control system must jointly model liquid cooling, air, IT equipment, and heat reuse. Optimizing one subsystem can shift energy consumption to another.

Rack-scale cooling and flow control offer additional benefits. A review identifies energy savings of up to 67.2% compared to traditional methods, although it points out shortcomings in reliability, control of heterogeneous loads, and facility retrofitting (Alkrush et al., 2024; Li et al., 2024). The predictive control strategy with chilled water storage reduced cooling consumption by between 4.04% and 22.15%, depending on the operating mode considered (Zhu et al., 2024). These results support specific technical interventions but do not address the territorial distribution of water. An efficient facility can still exacerbate water scarcity if it is located in a stressed watershed, an issue that calls for planning and accountability.

Infrastructure, Corporate Concentration, and Data Sovereignty

The industrialization of AI and dependence on Big Tech, the cloud, and territory

The industrialization of AI integrates models, applications, data centers, cloud services, financing, and developer markets into a single business chain. An analysis of Amazon, Microsoft, and Google shows that these companies control different levels of the technology stack and that numerous firms depend on their platforms (van der Vlist et al., 2024). This concentration determines who can train models, where computing takes place, and what environmental information falls under corporate control.

The infrastructure also reshapes the landscape. Data centers occupy land, consume electricity and water, generate logistics flows, and can produce waste heat, although urban policies have paid limited attention to these physical aspects (Monstadt & Saltzman, 2025). The case of the Amsterdam metropolitan area illustrates the shift from nearly unconditional political support toward reforms aimed at limiting growth, without resolving conflicts over land, electricity, and resources (Monstadt & Saltzman, 2025). Siting, therefore, is not a neutral technical decision: it distributes benefits and costs among actors with unequal capacities.

Environmental transparency and control over data and resource flows

Data sovereignty implies self-determination over the use of data, underpinned by trust, infrastructure, and contractual agreements throughout its lifecycle (von Scherenberg et al., 2024). When applied to AI, it requires establishing a link between who controls the data and who controls computing power, energy, and environmental records. Without this link, sovereignty may be reduced to privacy, while the physical resources that enable processing remain opaque.

Current transparency does not allow for the reliable separation of AI workloads from other data center workloads, nor, in many cases, does it allow for an estimation of the overall performance of technology companies (de Vries-Gao, 2025). The lack of information on direct water use, electricity, power generation, and supply chains limits regulation and community planning (Privette et al., 2026). Disclosure should include comparable metrics by facility and by workload type, along with regional information on water sources, carbon intensity, and grid capacity. Without public access to this data, digital sovereignty remains subordinate to private decisions regarding shared infrastructure.

Critical Synthesis: Uncertainties, Controversies, and Governance Gaps

The difficulty of attributing energy, carbon, and water specifically to AI loads

Environmental attribution faces a central methodological limitation: operators typically report on aggregated data centers, without distinguishing between AI, storage, networks, or other services (de Vries-Gao, 2025). Estimates must approximate these loads using general performance metrics, which widens the ranges of uncertainty. Furthermore, the water footprint may account for cooling, electricity, and manufacturing, whereas corporate indicators typically prioritize a single phase (Herrera et al., 2025). Comparisons across studies are valid only when the system boundary, geography, time period, and functional unit align.

Uncertainty is not solely statistical either. Server location, electricity mix, climate, and usage materially alter the results (Lei et al., 2025; Xiao et al., 2025). The scarcity of local data prevents an accurate assessment of distributional effects, especially in areas experiencing drought or water stress, where low national consumption can produce intense regional impacts (Privette et al., 2026).

Technological Efficiency Versus Absolute Demand Growth

The evidence supports improvements in cooling, load control, and server efficiency, but does not demonstrate that these automatically reduce total demand. The life-cycle review concludes that large-scale inference can match or exceed training and that unit gains are subject to systemic feedback effects (Oliveira et al., 2026). Therefore, efficiency must be accompanied by deployment limits, siting criteria, and an assessment of computational needs. Increased capacity can negate savings if demand for services grows more rapidly.

Gaps in comparable data, accountability, and environmental sovereignty

Governance requires disclosure obligations, AI-specific environmental assessments, and coordination among companies, network operators, water authorities, and communities. Existing frameworks do not yet guarantee comparable data on water, carbon, hardware, and land (Hlabisa, 2025; Privette et al., 2026). Environmental sovereignty raises a distributional question: who decides which uses of water, land, and electricity are acceptable when economic benefits are concentrated and costs are shifted elsewhere? Without public participation and verifiable accountability mechanisms, the expansion of AI will continue to rely on infrastructure whose impact remains partially invisible.

Conclusion: Toward a Digital Sovereignty Compatible with Ecological Limits

Implications for energy, water, and land-use planning

Ecological digital sovereignty requires planning for AI as physical infrastructure. Decisions regarding data centers must integrate network capacity, carbon intensity, water availability, climate, land use, and opportunities for heat reuse (Herrera et al., 2025; Monstadt & Saltzman, 2025). Expansion should not be approved based on aggregated national metrics that obscure local impacts. Energy and water planning must incorporate growth scenarios and disclose the assumptions underlying their results.

Priorities for Transparent, Accountable, and Materially Sustainable AI

The priorities are to establish mandatory and comparable metrics, separate AI workloads, report on direct and indirect water use, and account for hardware manufacturing and replacement. Efficiency, demand flexibility, climate-adapted cooling, and territorial limits must also be combined, because no single intervention can control the entire footprint (Lei et al., 2025; Oliveira et al., 2026). An AI system compatible with ecological limits will depend less on corporate promises of neutrality and more on public regulations that link data, infrastructure, resources, and accountability.

References

• Aili, A., Long, W.-J., Cao, Z., & Wen, Y. (2024). Radiative free cooling for energy and water saving in data centers. Applied Energy, 359, 122672. https://doi.org/10.1016/j.apenergy.2024.122672
• Alkrush, A. A., Salem, M. S. A., Abdelrehim, O., & Hegazi, A. A. (2024). Data center cooling: A critical review of techniques, challenges, and energy-saving solutions. International Journal of Refrigeration, 160, 246–262. https://doi.org/10.1016/j.ijrefrig.2024.02.007
• Azarifar, M., Arık, M., & Chang, J.-Y. (2024). Liquid cooling of data centers: A necessity facing challenges. Applied Thermal Engineering, 247, 123112. https://doi.org/10.1016/j.applthermaleng.2024.123112
• de Vries-Gao, A. (2025). The carbon and water footprints of data centers and what this could mean for artificial intelligence. Patterns, 7(1), 101430. https://doi.org/10.1016/j.patter.2025.101430
• Gheni, M., Kerskes, H., & Stergiaropoulos, K. (2025). Operational analysis of the cooling system in a direct liquid-cooled data center: a measurement and simulation study on the impact of supply water temperature. Applied Energy, 403, 127061. https://doi.org/10.1016/j.apenergy.2025.127061
• Herrera, M., Xie, X., Menapace, A., Zanfei, A., & Brentan, B. (2025). Sustainable AI infrastructure: A scenario-based forecast of water footprint under uncertainty. Journal of Cleaner Production, 526, 146528. https://doi.org/10.1016/j.jclepro.2025.146528
• Hlabisa, S. B. (2025). The ecology of artificial intelligence: energy, water, materials, and land limits of digital systems. Carbon Neutral Systems, 1(1). https://doi.org/10.1007/s44438-025-00018-8
• Lei, N., Lu, J., Shehabi, A., & Masanet, E. R. (2025). The water use of data center workloads: A review and assessment of key determinants. Resources Conservation and Recycling, 219, 108310. https://doi.org/10.1016/j.resconrec.2025.108310
• Li, X., Li, M., Zhang, Y., Han, Z., & Wang, S. (2024). Rack-level cooling technologies for data centers – a comprehensive review. Journal of Building Engineering, 90, 109535. https://doi.org/10.1016/j.jobe.2024.109535
• Monstadt, J., & Saltzman, K. (2025). HOW DATA CENTERS HAVE COME TO MATTER: Governing the spatial and environmental footprint of the ‘digital gateway to Europe.’ International Journal of Urban and Regional Research, 49(4), 757–778. https://doi.org/10.1111/1468-2427.13316
• Oliveira, A. P., Carraquico, T., & Martínez-Pérez, C. (2026). Beyond efficiency: A systematic review of energy consumption and carbon footprint across the AI lifecycle. Sustainability, 18(3), 1359. https://doi.org/10.3390/su18031359
• Privette, A. P., Barros, A. P., & Cai, X. (2026). Data centers’ water footprint: The need for more transparency. AGU Advances, 7(2). https://doi.org/10.1029/2025av002140
• Rostami, S., Down, D. G., & Karakostas, G. (2024). Linearized data center workload and cooling management. IEEE Transactions on Automation Science and Engineering, 22, 3502–3514. https://doi.org/10.1109/tase.2024.3395471
• Ruan, J., Zhu, Y., Cao, Y., Sun, X., Lei, S., Liang, G., Qiu, J., & Xu, Z. (2024). Privacy-Preserving bi-level optimization of internet data centers for electricity-carbon collaborative demand response. IEEE Internet of Things Journal, 11(14), 24948–24959. https://doi.org/10.1109/jiot.2024.3391762
• Sheng, Y., Chenxuan, Z., Zhu, Z., Xu, H., Wen, J., Wang, R., Yang, J., Wang, Q., & Bu, S. (2026). RETRACTED: Power for AI data centers: Energy demand, grid impacts, challenges, and perspectives. *Energies*, 19(3), 722. https://doi.org/10.3390/en19030722
• van der Vlist, F., Helmond, A., & Ferrari, F. (2024). Big AI: Cloud infrastructure dependence and the industrialization of artificial intelligence. Big Data & Society, 11(1). https://doi.org/10.1177/20539517241232630
• von Scherenberg, F., Hellmeier, M., & Otto, B. (2024). Data sovereignty in information systems. Electronic Markets, 34(1). https://doi.org/10.1007/s12525-024-00693-4
• Xiao, T., Nerini, F. F., Matthews, H. D., Tavoni, M., & You, F. (2025). Environmental impact and net-zero pathways for sustainable artificial intelligence servers in the USA. Nature Sustainability, 8(12), 1541–1553. https://doi.org/10.1038/s41893-025-01681-y
• Zhu, Y., Zhang, Q., Zeng, L., Wang, J., & Zou, S. (2024). An advanced control strategy for a hybrid cooling system with a cold water storage system in a data center. Energy, 291, 130304. https://doi.org/10.1016/j.energy.2024.130304