A systematic review synthesizing forty-six environmental investigations has mapped the comprehensive resource metabolism of artificial intelligence, demonstrating that the physical infrastructure supporting modern computational algorithms requires massive and interconnected inputs of electricity, water, minerals, and land. Conducted by researcher Sibongiseni Hlabisa and published in the journal Carbon Neutral Systems, the evaluation establishes that computational systems cannot be understood through isolated carbon metrics or narrow efficiency indicators. Instead, artificial intelligence operates as an open socio-metabolic subsystem whose operational and embodied footprints escalate alongside expanding computational workloads. Evaluating evidence from 2015 to 2025 across academic databases, the synthesis identifies that the physical apparatus supporting machine learning links digital processing directly to global extraction, regional utility strains, and landscape conversion. By framing computing within planetary boundaries and socio-metabolic theory, the analysis details how the expanding digital economy externalizes heavy biophysical burdens onto surrounding ecosystems while concentrating infrastructural benefits within limited geographic regions.

To evaluate the biophysical scale of computing infrastructure, the investigation conducted a structured literature search across Web of Science, Scopus, and Google Scholar databases for environmental assessments published between 2015 and 2025. An initial search query combining terms related to artificial intelligence, data centers, energy consumption, water use, semiconductor fabrication, and land occupation yielded approximately 3,130,000 preliminary records. Successive filtering restricted candidates to English-language peer-reviewed publications, reducing the total to roughly 18,300 records before keyword screening narrowed the focus to 312 explicitly relevant studies. Subsequent title and abstract evaluations excluded conceptual papers lacking empirical data, leaving 94 candidates for full-text eligibility appraisal. Ultimately, forty-six sources satisfied all inclusion criteria by providing quantitative or model-based estimates with sufficient methodological transparency. This final corpus incorporated twenty-two peer-reviewed journal articles and proceedings, seventeen technical and government reports, four preprints, and three scholarly reference volumes, providing a normalized cross-domain dataset.

Energy requirements represent the primary operational driver within this socio-metabolic nexus, marked by an acute divergence between per-unit computational efficiency and aggregate demand. The synthesized data establish that training a single large model, such as PaLM-2 or GPT-3, consumes between 1.3 and 1.5 gigawatt-hours of electricity, roughly matching the annual power consumption of a small town. Looking across future trajectories, international projections indicate that global artificial intelligence electricity demand could expand three- to five-fold between 2022 and 2028, reaching 150 to 300 terawatt-hours annually. Modeling by the International Energy Agency indicates that data center electricity demand under a standard base case will more than double from 415 terawatt-hours in 2024 to 945 terawatt-hours by 2030, approaching 1,200 terawatt-hours by 2035. Under an unconstrained lift-off scenario, demand could exceed 1,700 terawatt-hours by 2035, while restrained headwinds and high-efficiency pathways project trajectories of 700 and 970 terawatt-hours, respectively.

The geographical allocation of this expanding electricity requirement demonstrates intense concentration, creating localized grid hotspots that complicate regional utility management. Data center power growth is heavily anchored in the United States, which accounts for nearly half of global expansion with an anticipated consumption surge of approximately 240 terawatt-hours between 2024 and 2030. China forms the second-largest national contributor, adding an estimated 175 terawatt-hours over the same period, while European facilities exhibit more constrained growth under stricter regulatory frameworks. Although corporate reports highlight dramatic operational enhancements—such as a reported thirty-fold reduction in energy per inference between 2023 and 2024—aggregate power draw continues to escalate. The synthesis underscores that efficiency dividends are systematically reinvested into larger neural architectures and higher computational densities. This dynamic illustrates the Jevons paradox, wherein reduced marginal computational costs stimulate wider deployment, preventing efficiency gains alone from decoupling algorithmic advancement from rising gross energy consumption.

Water acts as a critical thermodynamic sink within artificial intelligence infrastructure, dissipating concentrated thermal energy from processors while enabling precision semiconductor manufacturing. Evaporative data center cooling architectures generally consume between 7,000 and 12,000 liters of water per megawatt-hour of electricity, generating acute competition with municipal and agricultural users in arid regions. Upstream manufacturing compounds this hydrological burden, as semiconductor foundries require 1.8 to 2.5 cubic meters of water, predominantly processed to ultrapure standards, for every fabricated silicon wafer. Future projections spanning 2024 to 2030 illustrate divergent water consumption trajectories dictated by technological adoption rates and cooling designs. Under a rapid adoption scenario featuring a fifteen-fold compute workload surge and minimal five percent efficiency gains, data center water consumption could escalate by more than 1,000 percent by 2030. Conversely, a sustainable transition scenario incorporating aggressive water-efficient cooling innovations yielding a sixty percent efficiency gain could restrict growth to approximately four times current levels.

Material extraction and chemical processing anchor the computational sector firmly within physical supply chains, revealing significant upstream environmental burdens embedded in specialized computing hardware. An analysis of hardware manufacturing reveals that producing a single NVIDIA HGX H100 baseboard generates approximately 1.3 metric tons of carbon dioxide equivalent, with more than ninety percent of lifecycle emissions originating directly from upstream material extraction and component manufacturing. Advanced processors depend on continuous supplies of refined silicon, copper, cobalt, tantalum, and rare-earth elements such as neodymium and terbium. Sourcing these critical minerals generates severe local externalities, including toxic tailings, groundwater contamination, and riverine pollution near extraction basins. The literature documents an entropic escalation across chip manufacturing: as lithographic nodes advance from 28 down to 2 nanometers, eco-efficiency gains plateau while absolute resource demands surge. Consequently, artificial intelligence hardware exhibits 1.8 to 4.5 times higher water and rare-earth material demands than conventional general-purpose silicon chips.

Physical land occupation represents a visible yet frequently neglected dimension of digital expansion, converting rural and peri-urban landscapes into specialized industrial zones. Hyperscale data center installations routinely occupy 30 to 100 hectares of land, while modern semiconductor fabrication complexes span between 20 and 60 hectares, requiring dedicated substations, transport links, and cooling equipment. Because these capital-intensive facilities possess multi-decade operational lifecycles, their construction locks vast contiguous land parcels into industrial functions, displacing agricultural acreage, buffer zones, and natural habitats. Multi-constraint analyses reveal that infrastructural hubs are colliding with environmental thresholds, provoking local regulatory pushback. In Ireland, grid saturation and public concerns over cumulative power draws prompted a formal moratorium on unconstrained data center connections. Similarly, Singapore imposed a construction freeze in 2019, transitioning to restrictive conditional permits that mandate strict power efficiency benchmarks, stringent water metrics, and active thermal heat recovery for new facility approvals.

The structural clustering of core computing assets into narrow geographical corridors creates profound systemic vulnerabilities, exposing global digital services to regional climate and infrastructure shocks. Today, the majority of advanced processing and high-performance manufacturing takes place within fewer than one hundred gigawatt-scale sites globally. In resilience theory, excessive centralization optimizes operational cost and throughput speed but severely reduces spatial redundancy, creating acute single points of failure. During the severe 2021 drought in Taiwan, water shortages forced authorities to divert supplies from agricultural irrigation to sustain continuous operations at Hsinchu Science Park fabrication facilities. Similar water allocation disputes have surfaced around semiconductor expansions in Arizona. Furthermore, this distribution reinforces global socio-ecological asymmetries: the high-margin economic and informational benefits of computational systems remain concentrated within the Global North, while extractive degradation, e-waste hazards, and localized environmental strains are externalized to resource frontiers across the Global South.

Addressing these interconnected resource pressures demands a fundamental transition in regulatory philosophy, shifting away from voluntary corporate efficiency metrics toward systemic sufficiency and absolute resource caps. Existing performance indicators, such as power usage effectiveness and water usage effectiveness, obscure aggregate growth by focusing strictly on marginal operational improvements. The synthesis argues that policymakers must implement integrated resource accounting that establishes binding regional quotas for cumulative electricity, water withdrawal, and land conversion linked directly to local ecological capacities. Practical pathways include adopting circular economic models that capture waste heat for urban district heating networks, recirculating ultrapure water within semiconductor foundries, and recovering critical metals from decommissioned hardware. Additionally, emerging operational strategies like temporal workload migration—termed follow-the-sun and follow-the-wind computing—can dynamically shift non-urgent computational tasks across regions to synchronize processing loads with local surpluses of renewable wind and solar generation, thereby mitigating peak grid stress and lowering fossil reliance.

Despite the breadth of the synthesized literature, the review emphasizes significant empirical limitations that constrain environmental governance across the digital technology sector. Of the forty-six evaluated sources, only approximately one-third presented direct empirical measurements or computational models specifically isolating artificial intelligence workloads, while the remaining two-thirds relied on broader communications infrastructure data or conceptual frameworks. Corporate sustainability disclosures remain largely self-reported, proprietary, and focused on operational emissions, while routinely omitting Scope 3 supply chain footprints and absolute material throughput metrics. To resolve these data deficits, the study calls for establishing standardized life-cycle assessment protocols, real-time infrastructural metering, and an independent Global Observatory for Digital Resource Use modeled after international environmental assessment bodies. Future research must bridge computer science and environmental systems analysis, developing empirical material flow assessments that couple computational complexity with thermodynamic boundaries. Ultimately, aligning artificial intelligence with planetary boundaries requires treating digital infrastructure as an industrial sector governed by biophysical limits.

Subject of Research: Systematic review of artificial intelligence resource consumption across energy, water, materials, and land systems

Article Title: The ecology of artificial intelligence: energy, water, materials, and land limits of digital systems

Article References: Hlabisa, S. (2025). The ecology of artificial intelligence: energy, water, materials, and land limits of digital systems. Carbon Neutral Systems, 1(1), Article 19. https://doi.org/10.1007/s44438-025-00018-8

Image Credits: AI Generated

DOI: 10.1007/s44438-025-00018-8

Keywords: artificial intelligence, resource nexus, data centers, energy consumption, water footprint, semiconductor manufacturing, socio-metabolic analysis, environmental governance, ecology, artificial, intelligence, energy