The ecology of artificial intelligence: energy, water, materials, and land limits of digital systems
Open accessThe DOI matches its Crossref record; 31 of the 31 references it cites were found in a registry or an index; 3 of the 80 sources its claims cite were checked for support.
Abstract
Artificial intelligence (AI) is promoted as a tool for ecological sustainability, yet the ecological effects of AI itself remain poorly understood. Existing studies document aspects of AI’s footprint, including energy demand, water withdrawals, mineral extraction, and land occupation, but treat these impacts in isolation. This article develops an integrated resource nexus framework that situates AI within global environmental systems’ energy–water–material–land metabolism. Drawing on recent evidence, we synthesize quantitative estimates of AI’s resource intensity, from thousands of megawatt-hours and millions of liters required to train frontier models, to embodied mineral and land footprints embedded in semiconductor production and hyperscale data centers. We show that these demands are systemic, entangled, and spatially uneven, with benefits concentrated in the Global North and ecological costs externalized to resource frontiers in the Global South. The contribution of this article is twofold. First, it presents a novel conceptual model of AI as a socio-metabolic infrastructure, highlighting the entropic trade-offs associated with digital expansion. Second, it derives governance pathways, including AI-specific environmental impact assessments, multi-dimensional ecological metrics, and spatially responsible siting strategies that embed AI within planetary boundary frameworks. By reframing AI as an ecological actor with biophysical limits, we advance environmental systems research beyond narrow carbon accounting toward a holistic understanding of digital infrastructures in the Anthropocene.
Overview
CoChat Summary
TL;DR
This review argues that AI is not “weightless” but a materially intensive socio-metabolic system whose energy, water, materials, and land demands are tightly coupled. Synthesizing 46 sources, it concludes that efficiency gains alone do not reduce aggregate impacts and that governance must move toward sufficiency, integrated accounting, and spatially responsible siting.
Key findings
- The synthesis of 46 selected sources finds that AI infrastructures have substantial and interlinked demands for energy, water, materials, and land, with these domains forming a mutually reinforcing resource nexus.
- The paper reports that training a single large model such as GPT-3 or PaLM-2 can consume 1.3–1.5 GWh of electricity, and that global AI-related electricity demand could reach 150–300 TWh per year by 2028, a three- to five-fold increase from 2022.
- Reported water intensities include 7,000–12,000 L per MWh for data-center cooling and 1.8–2.5 m³ of ultrapure water per wafer in semiconductor fabrication; the paper also states that a Rapid AI Adoption scenario could produce more than a 1,000% increase in water consumption, while a Sustainable Transition scenario could limit growth to approximately 300% despite a fourfold expansion in AI workloads.
- For materials and land, the paper states that an NVIDIA HGX H100 baseboard emits approximately 1.3 t CO₂-equivalent, with more than 90% of emissions from materials and components, and that hyperscale data centers occupy 30–100 ha while semiconductor fab complexes extend over 20–60 ha.
Methods
- The study is a review combining systematic literature synthesis with a socio-metabolic conceptual framework and the planetary boundaries framework.
- It used systematic searches in Web of Science, Scopus, and Google Scholar for studies published between 2015 and 2025; the initial search returned approximately 3,130,000 records, narrowed to 18,300 after filters, then 312 after keyword refinement, 94 full texts screened, and 46 sources retained for analysis.
- The selected literature was analyzed through a qualitative-quantitative synthesis, extracting and normalizing metrics such as electricity demand per model training run, water withdrawal per MWh, embodied emissions per semiconductor unit, and land footprints of data centers and fabs.
Limitations
- The paper states that it is constrained by the scarcity and variability of available data; many environmental estimates rely on modeled or proprietary values rather than empirical reporting, and few sources disaggregate AI-specific resource use from broader ICT infrastructure.
- It says quantitative synthesis therefore emphasizes representative orders of magnitude rather than precise totals.
- Evident from the text: the review relies on secondary sources and includes a mix of empirical studies, reports, preprints, and conceptual works, so the evidence base is heterogeneous and some figures are drawn from illustrative or scenario-based estimates rather than direct measurement.
Model-written summary. Check it against the paper before relying on it.
Integrity signals
Registry
Cites
Statistics
Verification
Claims the paper makes
3 Valid · 18 Not valid · 63 Not checked · 135 Own results and interpretations
The cited sources support only part of this claim — read each source’s verdict below.
In the paper
“Advanced processors and memory chips require refined silicon, copper, cobalt, tantalum, and rare-earth elements such as neodymium and terbium.”
What the sources say
- Partly supportedcites Ali et al. 2017 matched to “Mineral supply for sustainable development requires resource governance” (2017) sourceExperimental — accuracy not yet measured
Global Resource Governance Needed (p. 5) says high-technology applications require minerals including copper and rare earths. It does not identify processors or memory chips, say they require refined materials, or name silicon, cobalt, tantalum, neodymium, or terbium. (LLM, conf=0.83)
- Not checked yetcites Wübbeke 2013
plain- Asserted: passes — The authors put it forward as true.
- Checkable: passes — It can be checked against evidence.
- Load-bearing: passes — The paper’s conclusions depend on it.
- Based: passes — The paper gives a reason why it holds.
- Partly supported
The cited sources support only part of this claim — read each source’s verdict below.
In the paper
“Their extraction generates extensive ecological externalities, including toxic waste and riverine contamination (Ali etal. 2017; Wübbeke 2013).”
What the sources say
- Partly supportedcites Ali et al. 2017 matched to “Mineral supply for sustainable development requires resource governance” (2017) sourceExperimental — accuracy not yet measured
Our copy of Ali et al. (2017) does treat mining's ecological costs in general terms — the Introduction (p. 3) cites the Samarco tailings dam collapse as illustrating 'the consequences when mining and mineral processing is not done in an environmentally and socially responsible manner', and Actions for sustainable mineral sourcing (pp. 7-8) calls for standards assuring 'minimal negative impact on the environment', flags 'environmental alarms - are the risks to local ecosystems and populations too high?', and notes valuable metals otherwise 'ending up in the tailings dam' alongside waste and water consumption. That supports the broad premise that extraction imposes ecological externalities and generates tailings/waste. But no passage states that extraction generates 'extensive ecological externalities' specifically 'including toxic waste and riverine contamination' — the paper never characterises mine wastes as toxic or describes river/water contamination from extraction, so the two named externality types are not supported by this source (they likely rest on the co-cited Wübbeke 2013). (LLM, conf=0.70)
- Not checked yetcites Wübbeke 2013
plainstates a cause- Asserted: passes — The authors put it forward as true.
- Checkable: passes — It can be checked against evidence.
- Load-bearing: passes — The paper’s conclusions depend on it.
- Based: passes — The paper gives a reason why it holds.
- Partly supported
The cited sources support only part of this claim — read each source’s verdict below.
In the paper
“Semiconductor manufacturing depends on high-purity silicon, copper, cobalt, and rare earth elements such as neodymium and terbium (Ali et al. 2017; Wübbeke 2013).”
What the sources say
- Partly supportedcites Ali et al. 2017 matched to “Mineral supply for sustainable development requires resource governance” (2017) sourceExperimental — accuracy not yet measured
Our copy supports only the broad background part: 'Global Resource Governance Needed' (p. 5) says high-technology applications will consume 'not only "critical" metals such as the rare earths, but also vast amounts of common commodities such as copper, steel and cement', and the Introduction (pp. 3-4) treats copper and specialty/by-product metals (germanium, indium, gallium) as supply concerns. Nothing in our copy mentions semiconductor manufacturing, high-purity silicon, cobalt, or the specific rare earths neodymium and terbium (searches: 'semiconductor manufacturing high-purity silicon', 'silicon cobalt supply electronics', 'cobalt neodymium terbium', 'neodymium magnets'); the 'high-purity silicon, copper, cobalt, neodymium, terbium' specification and the semiconductor-sector framing are therefore not supported by this source. (LLM, conf=0.65)
- Not checked yetcites Wübbeke 2013
plainstates a cause- Asserted: passes — The authors put it forward as true.
- Checkable: passes — It can be checked against evidence.
- Load-bearing: unclear — It is unclear whether the paper’s conclusions depend on it.
- Based: passes — The paper gives a reason why it holds.
- Partly supported
In the paper
“Although successive generations of processors have achieved efficiency gains, Google (2024) reports a more than 30-fold improvement in energy per inference between 2023 and 2024.”
What the sources say
- Not checked yetcites Google 2024
plainstates an association- Asserted: passes — The authors put it forward as true.
- Checkable: passes — It can be checked against evidence.
- Load-bearing: unclear — It is unclear whether the paper’s conclusions depend on it.
- Based: passes — The paper gives a reason why it holds.
In the paper
“As shown in Table1, training a single large model, such as GPT-3 or PaLM-2, can consume between 1.3 and 1.5 gigawatt-hours (GWh) of electricity, roughly equivalent to the annual energy usage of a small town (Patterson etal. 2021).”
What the sources say
- Not checked yetcites Patterson et al. 2021
plain- Asserted: passes — The authors put it forward as true.
- Checkable: passes — It can be checked against evidence.
- Load-bearing: unclear — It is unclear whether the paper’s conclusions depend on it.
- Based: passes — The paper gives a reason why it holds.
108 more claims
Create a free account to see all 219 claims and to verify the 77 references still open.
- It may state two separate claims as one. 76
- Probably not a substantial claim. 52
Self-check
Percentages, sums and reported differences re-computed from the paper’s own numbers.
Each contradiction between two passages listed here was confirmed by an independent check.
Rated minor: it would not touch a result — a value no result rests on, a p-value that stays on the same side of the significance level, or a conflict that is clerical, resolved elsewhere in the paper, or outside its own findings.
In the paper
§ 3.2 Energy
“AI-training operational water footprint”
§ 3.2 Energy
“2-5 L per 1,000 inferences (approx.)”
Rated minor: it would not touch a result — a value no result rests on, a p-value that stays on the same side of the significance level, or a conflict that is clerical, resolved elsewhere in the paper, or outside its own findings.
In the paper
§ 2 Methods
“The remaining 94 papers were assessed in full text, of which 56 met all inclusion criteria”
§ 2 Methods
“After successive screening and eligibility checks, 46 sources were retained for analysis (see Fig.1).”
Citations
Cites
A safe operating space for humanity
Rockström, J., Steffen, W., Noone, K., Persson, Å., Chapin, F. S., Lambin, E. F., Lenton, T. M., Scheffer, M., Folke, C., Schellnhuber, H. J., et al. (2009) · Nature · doi:10.1038/461472a · Cited by 13,333 · cited 3×
“…systems research, ranging from socio-metabolic theory (Fischer-Kowalski & Haberl 2007; Krausmann et al. 2008) and planetary boundaries (Rockström et al. 2009; Stef-fen et al. 2015) to contemporary analyses of data center energy and water consumption (LBNL 2024; Li et al. 2023; Mytton…”
Planetary boundaries: Guiding human development on a changing planet
Steffen, W., Richardson, K., Rockström, J., Cornell, S. E., Fetzer, I., Bennett, E. M., Biggs, R., Carpenter, S. R., de Vries, W., de Wit, C. A., et al. (2015) · Science · doi:10.1126/science.1259855 · Cited by 12,527 · cited 1×
“Such frameworks could be aligned with planetary boundary thinking (Rockström et al. 2009; Steffen et al. 2015), establishing quantitative safe operating limits for AI’s metabolic flows.”
Energy and Policy Considerations for Deep Learning in NLP
Strubell, E., Ganesh, A., & McCallum, A. (2019) · Proceedings of the 57th Annual Meeting of the Association for Computational Linguistics · doi:10.18653/v1/p19-1355 · Cited by 3,684
IPBES (2019) · Socio-Environmental Systems Modeling · doi:10.5281/zenodo.3831673 · Cited by 3,354
A Unified Framework of Five Principles for AI in Society
Floridi, L., & Cowls, J. (2019) · Harvard Data Science Review · doi:10.1162/99608f92.8cd550d1 · Cited by 1,631 · cited 1×
“As Floridi and Cowls (2019) argue, ethical AI requires fair algorithms and just infrastructures, systems whose material foundations reflect distributive equity and ecological care.”
United States Data Center Energy Usage Report
Lawrence Berkeley National Laboratory (LBNL), B., Shehabi, A., USDOE, Smith, S., Sartor, D., Brown, R., Herrlin, M., Koomey, J., Masanet, E., Horner, N., et al. (2016) · doi:10.2172/1372902 · Cited by 1,040 · cited 2×
“Conversely, water-free cooling solutions such as closed-loop chillers can increase energy demand by 10–15 percent (Shehabi et al. 2024).”
Urban land teleconnections and sustainability
Seto, K. C., Reenberg, A., Boone, C. G., Fragkias, M., Haase, D., Langanke, T., Marcotullio, P., Munroe, D. K., Olah, B., & Simon, D. (2012) · Proceedings of the National Academy of Sciences · doi:10.1073/pnas.1117622109 · Cited by 905 · cited 1×
“As Seto et al. (2012) emphasize, the sustainability of urban and industrial systems depends on recognizing land tele-connections: the distant ecological linkages that connect production, consumption, and environmental change across scales.”
Mineral supply for sustainable development requires resource governance
Ali, S. H., Giurco, D., Arndt, N., Nickless, E., Brown, G., Demetriades, A., Durrheim, R., Enriquez, M. A., Kinnaird, J., Littleboy, A., et al. (2017) · Nature · doi:10.1038/nature21359 · Cited by 803 · cited 3×
“Their extraction generates extensive ecological exter-nalities, including toxic waste and riverine contamination (Ali et al. 2017; Wübbeke 2013).”
Jevons’ Paradox revisited: The evidence for backfire from improved energy efficiency
Sorrell, S. (2009) · Energy Policy · doi:10.1016/j.enpol.2008.12.003 · Cited by 801 · cited 4×
“As Sorrell (2009) warns, governance through efficiency metrics may reproduce rebound effects if not coupled with absolute limits on consumption.”
InfluentialAligning artificial intelligence with climate change mitigation
Kaack, L. H., Donti, P. L., Strubell, E., Kamiya, G., Creutzig, F., & Rolnick, D. (2022) · Nature Climate Change · doi:10.1038/s41558-022-01377-7 · Cited by 570 · cited 2×
“From climate modeling and biodiversity monitoring to energy optimization, AI is often portrayed as a cornerstone of sustainability innovation (Kaack et al. 2022; Lhoumeau, et al. 2025; Rolnick et al. 2022).”
Research works printed but not linked: 5
- Hornborg A (2001) The power of the machine: global inequalities of economy, technology, and environment. AltaMira Press, Walnut Creek
- International Energy Agency (IEA) (2025) Energy and AI. IEA, Paris
- Lawrence Berkeley National Laboratory (LBNL). (2024). U.S. data center energy outlook. Berkeley: LBNL. https://doi.org/10.71468/P1WC7Q · DOI
- Pisano, U. (2012). Resilience and Sustainable Development. ESDN Quarterly Report No. 26. European Sustainable Development Network. Available at: https://www.esdn.eu/fileadmin/ESDN_Reports/2012-September-Resilience_and_Sustainable_Development.pdf
- Wiener, N. (1948). Cybernetics: Or Control and Communication in the Animal and the Machine. MIT Press.
Cited by
The AI Act and its green blind spots: Hidden environmental risks in the AI lifecycle
Ibrahim, I. A., Zaidan, E., Truby, J., & Hoppe, T. (2026) · Technology in Society · doi:10.1016/j.techsoc.2026.103284 · Cited by 11
Ejiyi, C. J., Djukem, D. L. W., Diokpo, C. N., Eze, F. O., Muoka, G. W., Sam, F., Zhang, Q., Cai, D., Atwereboannah, A. A., Dossa, J. V., et al. (2026) · Archives of Computational Methods in Engineering · doi:10.1007/s11831-026-10559-x · Cited by 6
“Clean technology applications impose boundary conditions governed by physical laws and safety requirements, necessitating hybrid architectures that integrate data-driven learning with mechanistic consistency [21].”
Environmental Policy and Risk Regulatory Framework for Sustainable Tidal Current Energy in China
Guo, R., Liu, Z., Hanbing, W., Wang, G. F., Zhao, S., & Deng, Y. (2026) · Sustainability · doi:10.3390/su18073224 · Cited by 1
Wang, R., & Lin, Q. (2026) · ACM International Conference on Interaction Design and Children (IDC) · doi:10.1145/3773077.3813775
“Training and operating machine learning models require large amounts of electricity [5], while data centers consume substantial water for cooling processes [11].”
AI Infrastructure Expansion and Water-Energy Sustainability Challenges
Manojkumar, R. (2026) · Advances in computational intelligence and robotics book series · doi:10.4018/979-8-3373-8434-4.ch001
Metadata
- Authors
- Hlabisa, S.
- Affiliations
- University of South Africa (ZA)
- Corresponding author
- Hlabisa, Sibongiseni
- ORCID
- Hlabisa, Sibongiseni
- Venue
- Carbon Neutral Systems
- Published
- 2025-12-02
- Volume
- 1
- Issue
- 1
- ISSN
- 3059-3875 (ISSN-L)
- OpenAlex
- W4416899445
- Cited by
- 15
- References
- 29
- Citations by year
- 2026: 15
- FWCI
- 3.38
- Type
- journal article
- Language
- en
- Open access
- hybrid
- doi.org
- Licence
- cc-by
- creativecommons.org/licenses/by/4.0 (tdm)
- creativecommons.org/licenses/by/4.0 (vor)
- Locations
- Carbon Neutral Systems (publishedVersion, cc-by)
- Topics
- Physical Sciences › Engineering › Electrical and Electronic Engineering › Green IT and Sustainability
- Physical Sciences › Environmental Science › Water Science and Technology › Water-Energy-Food Nexus Studies
- Physical Sciences › Environmental Science › Environmental Engineering › Environmental Impact and Sustainability
- Keywords
- artificial intelligence, environmental impact of AI, energy demand, water consumption, mining, land use, resource nexus, digital infrastructure, hyperscale data centers, semiconductor manufacturing, planetary boundaries, environmental impact assessment, North-South inequality, Anthropocene
About these checks
Registry, retraction, Cites and statistics results are computed, not judged, with one exception: a printed reference that no index lists is read once by a model before it counts as a citation error. Claims are extracted by a model and kept only when their quote appears verbatim in the text; each cited source is then checked, with a model’s help, for whether it says what the claim relies on. Those verdicts can be wrong — read the claim’s quote and the cited source before relying on one.
The claim verdicts this report counts are experimental: their accuracy is not yet measured.
A language model drew the claims from the paper on 2026-10-11, under version 3 of the claim definition (model deepseek-v4.1-flash, prompt extract-v3).
A second model labelled each claim on 2026-10-11 (model jev-1.13-20260917, definition revision gold-guideline-v0.2).
Cited sources are checked with a language model’s help (gpt-6-sol and qwen3.8-omni-flash).
Statistics and arithmetic are recomputed, not judged.
Confidence bands are shown only for fields calibrated on our annotated test set. None is calibrated yet, so no band is shown.
What counts as a claim ›How we measure confidence ›How we check this ›
Are you an author, or do you see a problem with this page? Request a review
How to cite
Cite CoChat for the checks on this page, and the paper for its own findings.
- This page
- CoChat. (2026). Citation check: The ecology of artificial intelligence: energy, water, materials, and land limits of digital systems. https://app.cochat.ai/papers/10.1007/s44438-025-00018-8
- The paper
- Hlabisa, S. (2025). The ecology of artificial intelligence: energy, water, materials, and land limits of digital systems. Carbon Neutral Systems. https://doi.org/10.1007/s44438-025-00018-8
Data sources
Bibliographic data: Semantic Scholar, OpenAlex, Crossref.
Contains information from the Semantic Scholar Academic Graph Datasets, made available under the ODC Attribution License (ODC-BY 1.0).
Semantic Scholar’s data is described in: Kinney, R., et al. (2023). The Semantic Scholar Open Data Platform. arXiv:2301.10140.