Navigating the Environmental Paradox of AI: A Decision Framework for Clean Technology Practitioners

Abstract

Artificial intelligence presents a critical paradox for clean technology: while enabling unprecedented environmental optimization, AI deployment demands massive resource inputs that threaten to offset benefits. As global AI infrastructure investment approaches $500 billion annually, data center electricity consumption is projected to exceed 1000 TWh by 2030. We conducted a systematic literature review of 73 peer-reviewed empirical studies (2021–2025) to develop an Environmental Asset-Cost Framework categorizing AI’s impacts across five asset categories (energy optimization, production enhancement, green innovation, resource conservation, precision applications) and five cost categories (energy consumption, water use, e-waste, infrastructure, supply chain extraction). Our analysis reveals three critical insights: First, AI’s environmental impact follows a synthesized S-curve heuristic—a pattern derived from convergent but methodologically diverse evidence strands—characterized by initial emission reductions (0–2 years), mid-term rebound effects (2–5 years), and conditionally projected long-term optimization (5+ years). Second, geographical context creates 10–60× variation in outcomes; regions with high renewable electricity and water abundance achieve net benefits within 2–3 years, while fossil fuel-heavy, water-stressed regions may never reach net positive outcomes. Third, the rebound effect is predictable and manageable through strategic interventions. Our framework provides actionable deployment guidance, demonstrating that achieving AI’s net environmental benefits requires renewable energy infrastructure development before AI deployment, alternative cooling technologies, and policy frameworks incorporating temporal dynamics.

1. Introduction

The clean technology sector faces a critical paradox: artificial intelligence offers unprecedented capabilities for environmental optimization yet demands massive resource inputs that threaten to offset its benefits. As global AI infrastructure investment approaches $500 billion annually [1,2], data center electricity consumption is projected to more than double from 460 TWh in 2022 to 620–1050 TWh by 2026 [3,4] (460 TWh baseline: International Energy Agency (IEA) 2022 global data center electricity consumption, reported in [4]; projection range: IEA scenario analysis reported in [4]; unit: TWh), equivalent to adding a country the size of Sweden or Germany to global electricity demand. This expansion raises fundamental questions about AI’s true net environmental impact. Simultaneously, AI enables 30–50% reductions in industrial energy consumption [4,5,6], optimizes renewable energy integration into power grids [7,8], and accelerates clean technology innovation [1]. This dual nature can create decision paralysis for clean technology practitioners deciding whether to deploy AI systems and risk environmental rebound effects or forego AI and sacrifice optimization benefits.

Current literature documents both AI’s environmental assets (e.g., optimization capabilities, efficiency gains, and innovation acceleration) [5,8,9,10,11,12], and costs (e.g., energy consumption, water use, and electronic waste) [13,14,15]. However, existing research lacks a systematic framework for evaluating these trade-offs across the technology lifecycle. This gap is consequential, as point-in-time assessments cannot capture the temporal dynamics we identify in this analysis. In this S-curve pattern, initial environmental benefits are followed by mid-term rebound effects before potential long-term optimization. Without frameworks to predict and navigate these dynamics, clean technology deployment strategies risk amplifying, rather than mitigating, AI’s environmental costs.

Three critical gaps motivate the development of this systematic framework. First, existing environmental impact assessments treat AI as a static technology with fixed environmental characteristics. The analysis in this paper reveals temporal variability: AI applications initially demonstrate carbon reductions, followed by rebound periods as infrastructure scales [8,10,15,16]. No existing framework systematically accounts for these lifecycle dynamics or provides practitioners with predictive guidance. Second, current metrics (i.e., Power Usage Effectiveness (PUE) for energy, Water Usage Effectiveness (WUE) for water, and lifecycle assessments for hardware) provide point-in-time snapshots but cannot guide deployment strategies across temporal phases [8,10,12,15,16,17]. Clean technology practitioners need integrated decision criteria that transform these isolated measurements into actionable guidance. Third, geographical context like grid carbon intensity, water availability, and renewable energy access fundamentally alters whether AI deployment yields net environmental benefits [1,16,18,19], yet no systematic framework maps these contextual factors to deployment recommendations.

These gaps prevent clean technology practitioners from answering fundamental questions: When should AI be deployed? Where should AI infrastructure be located? How can rebound effects be mitigated? What conditions enable net positive environmental outcomes? The absence of systematic frameworks forces practitioners to rely on incomplete assessments, which may lead to counterproductive deployment decisions.

We develop an Environmental Asset-Cost Framework synthesizing empirical evidence from 2021 to 2025. The framework provides clean technology practitioners with: (1) systematic categorization of AI’s multi-dimensional environmental impacts across five asset categories and five cost categories, (2) a temporal dynamics model characterizing the S-curve pattern and predicting rebound effect timing, (3) geographical context mapping identifying optimal deployment regions based on grid characteristics and resource availability, (4) decision criteria for evaluating when AI implementation supports clean technology objectives, and (5) mitigation strategies for managing the predictable rebound effect period. Our framework transforms fragmented evidence on environmental impact into actionable guidance for sustainable AI deployment in clean technology transitions.

We focus on the literature from January 2021 to October 2025 for three strategic reasons. First, this period marks the post-generative-AI-boom maturation phase when empirical data on infrastructure impacts became available following the surge in deployment. Although generative AI has existed for many years, earlier research primarily relied on projections and simulations. We observe that the last five years of literature reflect actual deployment outcomes from organizations that have implemented large language models and advanced AI systems at scale [3,4,19,20,21]. Second, this timeframe captures the first wave of comprehensive industry sustainability disclosures, prompted by increased regulatory pressure, including the implementation of the EU AI Act and expanded U.S. Environmental Protection Agency reporting standards. These disclosures offer unprecedented transparency into previously opaque environmental impacts, facilitating the development of an evidence-based framework [18,22].

Third, limiting the timeframe ensures consistency in a technological context. We examine literature reflecting current-generation AI systems rather than conflating projections across different technological generations with distinct environmental profiles. This five-year analysis allows for precise characterization of contemporary AI’s environmental impacts under comparable technological and regulatory conditions, providing decision-relevant evidence for clean technology practitioners rather than averaging across incompatible technological eras.

2. Theoretical Framework: The Environmental Asset-Cost Model

2.1. Framework Overview and Development

The Environmental Asset-Cost Framework was developed through iterative analysis of empirical literature, showing that AI’s environmental impacts cannot be accurately described using single-metric assessments or simple good/bad categories. An initial review identified common patterns: some AI applications consistently showed environmental benefits through optimization and efficiency improvements [5,6,11,18,23], while infrastructure needs created ongoing resource demands [3,19,24,25]. More importantly, temporal analysis uncovered that impact trajectories are dynamic rather than fixed, with deployment outcomes varying significantly across different geographical contexts and implementation timelines [3,16].

This complexity required a multi-dimensional evaluation structure capable of: (1) systematically categorizing various impact types across environmental domains, (2) incorporating temporal dynamics to track lifecycle trajectories, (3) including geographical mediators that significantly influence impact magnitude and direction, and (4) offering actionable decision criteria for practitioners. We developed the framework through four phases: initial categorization of impact themes from literature, pattern recognition to identify the S-curve temporal trajectory, mapping dimensions to structure findings into asset and cost categories, and validation to ensure the framework can accommodate all reviewed studies without forcing evidence into predetermined categories [26,27,28].

The framework is explicitly grounded in three established theoretical traditions. First, lifecycle assessment (LCA) theory provides the foundation for evaluating environmental impacts across the full system boundary—from supply chain extraction and hardware manufacturing through operational energy and water use to end-of-life disposal—rather than limiting assessment to operational emissions alone [29,30,31,32]. Second, the Environmental Kuznets Curve (EKC) literature informs the S-curve temporal argument: The EKC theory proposes that environmental degradation initially increases with technological and economic development before declining as efficiency, substitution, and governance mechanisms mature [8,16,33]. Our framework applies this logic specifically to AI deployment trajectories, treating the three-phase pattern as an AI-specific instantiation of EKC dynamics. Third, rebound effect theory—specifically the Jevons Paradox and digital rebound literature—explains the mechanism driving Phase 2: efficiency improvements lower the effective cost of AI services, stimulating increased consumption that partially or fully offsets initial gains [8,16,34,35]. Together these theoretical foundations position the Asset–Cost matrix as a decision-support complement to established environmental accounting frameworks such as the GHG Protocol’s Scope 1/2/3 boundary structure, which addresses operational emissions (Scope 1–2) and value chain impacts (Scope 3) but does not incorporate temporal dynamics or AI-specific rebound mechanisms. The 5 × 5 matrix structure is a purposeful analytical simplification; future extensions could incorporate dynamic feedback mechanisms between asset and cost dimensions to capture how optimization gains in one category influence resource demands in another.

The resulting framework organizes AI environmental impacts along two main dimensions, Assets and Costs, each containing five categories, intersected by two key moderating factors, temporal dynamics and geographical context.

Table 1. Environmental Asset-Cost Framework for AI.

		COSTS
		Energy Consumption	Water Use	E-Waste Generation	Infrastructure Impacts	Supply Chain
ASSETS	1. Energy Optimization	AI-driven smart grids and demand response reduce energy use but require computational energy for model training and inference [9,15,36,37].	Building Heating, Ventilation, and Air Conditioning (HVAC) optimization reduces energy and associated water for electricity generation, but AI infrastructure cooling consumes water [4,14,19,21,34].	Smart grid optimization extends equipment life through predictive maintenance while requiring specialized hardware with short lifecycles [7,9,12,15,38,39,40,41,42].	Distributed optimization reduces need for generation capacity but requires edge computing infrastructure deployment [1,5,7,12,17,25,37,38,43,44,45].	Energy optimization reduces overall demand on supply chains while extraction of rare earth metals for GPUs and AI chip manufacturing creates upstream extraction burdens [7,9,11,12,14,39,40,41,46].
	2. Production Enhancement	Increased computational demands in manufacturing optimization [2,6,29,35].	Water use in cooling intensive computation for production AI. Precision agriculture cuts irrigation yet AI infrastructure for farm optimization uses cooling water [4,14,15,19,21,23].	Process optimization extends equipment lifecycles through predictive maintenance while AI hardware faces rapid obsolescence [7,9,12,30,38].	Production efficiency reduces industrial facility footprint but requires data center infrastructure for optimization algorithms [1,36,46,47,48,49,50].	Manufacturing optimization reduces material extraction through waste reduction while AI chips require rare earth mining [9,11,14,15,23,39,46,51].
	3. Green Innovation	Energy consumed in Research and Development (R&D) and training of new AI models for sustainability. AI accelerates renewable energy tech development, but discovery process requires substantial computational energy [1,4,14,35,49,52].	Materials discovery for water-efficient technologies proceeds faster with AI while discovery infrastructure consumes cooling water [4,13,19,31].	AI accelerates circular economy solutions but creates waste from discarded experimental hardware [4,9,22,23,40,46].	Clean tech R&D acceleration enabled by AI reduces long-term infrastructure needs but requires immediate research computing capacity [3,13,16,31,37,49,53,54].	Supply chain impacts from new hardware fabrication [23,30,39,40,47].
	4. Resource Conservation	Energy for AI models optimizing natural resource use (e.g., water, minerals) [4,7,13,14,27,31,55].	AI-enabled smart irrigation and leak detection reduces water system losses while infrastructure consumes electricity and associated water [19,37,49].	AI optimizes refurbishment timing extending hardware life, yet AI systems themselves face rapid obsolescence [12,15,30,37,38,48,51,56,57].	Infrastructure such as IoT sensor networks for resource monitoring [11,14,45,58,59,60].	Supply chain optimization reduces transportation emissions while AI chip supply chains generate upstream burdens [4,9,37].
	5. Precision Applications	Medical AI reduces unnecessary procedures lowering healthcare energy footprint but requires diagnostic computing infrastructure [17,38,51].	Infrastructure monitoring prevents water losses through early leak detection while monitoring systems require electricity and cooling [4,12,13,17,19,23,42,45,46].	Precision medicine extends equipment life through optimal maintenance while medical AI systems face technology turnover [14,15,23,37,38,51,61].	Targeted interventions reduce overall facility needs but require edge computing deployment for real-time analysis [9,36,44,47].	Precision agriculture reduces fertilizer production demand through targeted application while agricultural AI requires chip manufacturing [9,23,24,46,62].

presents the complete Environmental Asset-Cost Framework matrix.

2.2. Asset Dimension: Categories of Environmental Benefits

The Asset dimension characterizes how AI systems create environmental benefits through five distinct mechanisms:

• Energy Optimization: AI-driven systems optimize energy consumption across industrial processes, building management systems, and electrical grid operations. Machine learning algorithms detect inefficiencies that traditional monitoring misses, enabling 20–40% reductions in industrial energy consumption and 15–30% improvements in building HVAC performance [4,9,13,17,37,49,52,63,64].
• Production Enhancement: Precision agriculture applications reduce fertilizer and water use by 20–30% [23] while maintaining or improving yields [49,65]. Manufacturing optimization reduces material waste and production energy through predictive maintenance and process refinement [11,46,49,66].
• Green Innovation Acceleration: AI speeds up clean technology research and development by enabling materials discovery, molecular simulation, and design optimization [1,46]. Machine learning-driven materials science has shortened discovery timelines for battery materials, catalysts, and carbon capture compounds [31,46,66]. Generative design algorithms develop energy-efficient product geometries and sustainable material mixes [49,66].
• Resource Conservation: Circular economy applications promote waste reduction, optimized reuse of components, and planning for extending product lifecycles [11]. Predictive models determine the best times for refurbishment, cutting down premature disposal [11,30,37]. Supply chain optimization reduces transportation emissions and inventory waste through demand forecasting and route optimization [37,46].
• Precision Applications: AI-driven targeted interventions facilitate precise and efficient resource allocation by analyzing high-resolution data from sensors, drones, and satellites to ensure that inputs like water and fertilizer are applied only when and where they are necessary [49,65]. In the healthcare sector, AI acts as an “organizational consciousness” that helps reduce unnecessary procedures and redundant diagnostic tests, reducing the industry’s environmental impact [38,39]. Infrastructure monitoring systems detect leaks, structural weaknesses, and inefficiencies early, preventing larger resource losses [4,37,49].

2.3. Cost Dimension: Categories of Environmental Burdens

The cost dimension characterizes AI’s environmental resource demands through five impact categories:

• Energy Consumption: Training large language models requires 1287 to 7200 MWh of electricity per model, equivalent to 500–1500 U.S. homes’ annual electricity consumption [2,4,15,19,21,42,49,67] (GPT-3 baseline of 1287 MWh from Patterson et al. 2021, reported as contextual background in [2,49,68]; GPT-4 estimate of 7200 MWh from industry modeling reported in [4,49]; unit: MWh per training run; these figures derive from external industry and academic sources cited through reviewed studies, not directly measured in this review’s empirical corpus). More advanced models like GPT-4 have significantly higher requirements. Inference operations at scale consume 40 to 260 MWh daily for major platforms [4,15,49,68]. Data centers supporting AI workloads demand three to five times more energy than traditional computing centers due to accelerated hardware requirements [15]. Global AI data center electricity consumption is projected to reach 620 to 1050 TWh annually in 2026, representing more than 2% of global electricity demand [3,4,36] (baseline: IEA 2022 estimate of 460 TWh; projection: IEA scenario range reported in [4]; unit: TWh annually; contextual background via reviewed study). One source predicted this share could surge to 10% [4] (industry analyst forecast reported as contextual background in [4]; unit: % of global electricity demand).
• Water Use: Data center cooling systems use 1.8 to 12 L of water per kWh depending on the technology and climate [4,19,48] (unit: liters per kWh of electricity consumed; lower bound from air-cooled systems; upper bound from evaporative cooling in warm climates; range derived from Water Use Efficiency (WUE) data in reviewed empirical study [19] and contextual background reported in [4] via the Organisation for Economic Co-operation and Development (OECD) AI estimates). Training a single large language model indirectly consumes over 700,000 L through electricity generation and direct cooling [14,19,21,23] (estimate based on GPT-3-scale model; includes indirect consumption through electricity generation and direct evaporative cooling; reported as contextual background in [19], citing Hao 2024 and Bhaskar & Seth 2024; unit: liters per training run). U.S. data centers are expected to use 660–1200 billion liters of water annually by 2030 [48], with AI workloads accounting for a significant portion of this demand. Globally, AI water withdrawal is expected to reach 4.2 to 6.6 billion cubic meters by 2027 [4,19,38]. Data centers are frequently located in water-stressed regions such as Arizona, Nevada, India, and Australia, primarily to take advantage of lower energy costs, exacerbating regional shortages [19,48].
• E-waste Generation: AI requires specialized hardware, such as Graphics Processing Units (GPUs) and Tensor Processing Units (TPUs), which are frequently retired before the end of their operational life due to constant demand for faster systems [38]. This has accelerated obsolescence cycles, with GPU lifespans averaging two to three years compared to four to six years for traditional servers [39]. One source states that processing just one ton of rare earth ores, as critical for AI capacitors and sensors, can generate 2000 tons of toxic waste [39] (unit: tons of toxic waste per ton of ore processed; contextual background reported in [39]; original industry/environmental assessment source cited through the reviewed study). Recycling is insufficient, with about 20% of global e-waste properly recycled due to the complexity of specialized components [13,15,39].
• Infrastructure Impacts: Data centers are housed in large structures requiring massive amounts of steel, aluminum, and concrete [15]. As these facilities are typically built horizontally to reduce costs, the footprint of one facility can span dozens of acres, transforming land and disrupting local ecosystems. The construction and resource extraction for these facilities are linked to the destruction of natural habitats and loss of biodiversity. The environmental burden includes not just the servers but also the mechanical cooling systems, transformers, and backup diesel generators required to maintain operations and supply backup power and ongoing maintenance [13,15,52].
• Supply Chain Extraction: Semiconductor supply chains require rare earth mining, contaminating water sources and degrading land. Manufacturing a single AI accelerator GPU requires 20,000 to 30,000 L of ultrapure water and releases CO₂ [19]. Mining for critical AI minerals like lithium and cobalt is linked to deforestation, soil contamination, and water pollution [7,48]. These extractive processes are often concentrated in the Global South under poor labor conditions [13,15]. The information and communications technology sector accounts for 3.9% of global greenhouse gas emissions, surpassing global air travel, with data center emissions projected to triple by 2030 [7,49].

2.4. Temporal Dynamics: The S-Curve Pattern

Our key organizing framework for interpreting the evidence is the identification of a synthesized conceptual heuristic—a pattern emerging from convergent but methodologically diverse evidence strands in AI’s environmental impact trajectory. This heuristic, derived from studies spanning multiple methodological approaches, sectors, and geographies, challenges static assessments and explains why point-in-time evaluations often fail to capture long-term dynamics—though the strength of evidence varies considerably across its three phases.

The S-curve manifests through three distinct phases:

• Phase 1—Initial Optimization (0–2 years): [Evidence Level: Moderate—primarily cross-sectional empirical and econometric studies] Research using a dynamic threshold model found that in countries or periods with low AI intensity, increases in AI activity were associated with a significant reduction in emissions [16]. Early deployment, leveraged to enhance data monitoring, environmental governance, and eco-efficiency, show net environmental benefits as efficiency improvements outweigh infrastructure costs [16]. Infrastructure needs stay limited during this stage, using existing data center capacity or needing modest expansion. The carbon savings from optimization far surpass the additional emissions from infrastructure, resulting in net positive environmental effects. Studies in manufacturing, transportation, agriculture, and energy sectors consistently show this early positive phase [16].
• Phase 2—Rebound Effect (2–5 years): [Evidence Level: Moderate—supported by observed longitudinal and econometric data across multiple sectors and geographies] As AI deployment expands, research uncovers the rebound effect as infrastructure growth outpaces efficiency improvements. As AI adoption broadens beyond initial applications, organizations deploy more models, increase inference frequency, and expand training datasets. One study explicitly identifies a “Digital Rebound Paradox” where theoretical efficiency gains from AI applications are offset by increased consumption, computational demands, and system expansion [16]. Data center capacity additions surpass efficiency gains, hardware refresh cycles speed up to maintain competitive performance, and water use rises due to cooling needs. Beyond a critical threshold of AI intensity, the relationship reverses, and further intensification in highly digitized contexts leads to deteriorated environmental performance and rising emissions [16]. The implementation of AI often triggers Jevons Paradox, a phenomenon where improvements in energy efficiency paradoxically stimulate increased energy consumption [8]. The energy rebound effect is identified as a critical mediating mechanism that amplifies emissions in the early-to-mid stages of AI technology development [8]. Chen et al. [3] project China’s AI carbon footprint will double between 2030 and 2035 despite efficiency advancements, illustrating this rebound pattern. Ren et al. [34] note that between 2020 and 2024, large language model (LLM) scaling increased overall emissions despite per-parameter efficiencies, another sign of rebound. The rebound period poses a key challenge for clean technology efforts. Efficiency improvements plateau while infrastructure demands keep growing, potentially causing negative environmental impacts.
• Phase 3—Mature Optimization (5+ years): [Evidence Level: Indicative—based primarily on modeling projections and conditional extrapolations; limited observed longitudinal data] Long-term maturation can lead to positive outcomes if infrastructure growth stabilizes and system-wide improvements build up. AI technology follows an inverted U-shaped pattern (similar to the Environmental Kuznets Curve), where it initially increases emissions but ultimately facilitates their reduction as the technology develops [8,33]. In the context of AI data centers, projections indicate a sharp rise in carbon footprints reaching a peak in 2038, followed by a significant decline through 2050 [3]. This long-term maturation leads to positive outcomes as data center efficiency improves and clean electricity sources become more prevalent [3]. In High-Income Countries (HICs), which represent a more mature stage of AI development, AI technology is already associated with a nonlinear decline in emissions [8]. Organizations that continue AI deployment through the rebound period can gain significant benefits as optimizations spread across operations. However, reaching this stage requires ongoing commitment during the tough rebound period and depends heavily on local context, renewable energy availability, and proactive mitigation plans. Evidence for Phase 3 results is limited, as most current AI deployments have not yet achieved this level of maturity.

Conditions Required for Phase 3 Materialization: The projected long-term net benefits of Phase 3 are conditional. Based on the modeling and econometric evidence in our corpus, three conditions must hold simultaneously for AI deployment to yield net environmental benefits in the mature phase:

• Grid decarbonization: The electricity supplying AI infrastructure must transition substantially toward renewable or low-carbon sources. Studies indicate this condition is the single most influential moderator of Phase 3 timing [3,8,33].
• Infrastructure growth stabilization: The rate of data center expansion and hardware refresh cycles must slow relative to efficiency gains. Absent this, rebound effects persist indefinitely regardless of per-unit efficiency improvements [3,16,34].
• Adoption of best-practice efficiency standards: Widespread deployment of energy-efficient hardware, software optimization, and alternative cooling technologies must occur across the sector, not only among early adopters [1,5,27,49].

Where these conditions are absent—particularly in fossil-fuel-heavy, water-stressed, or institutionally limited deployment contexts—Phase 3 may not materialize within any foreseeable time horizon [16,33].

Table 2. Evidence Basis for S-Curve Phases.

Phase	Study Type	Representative Studies	Time Horizon	Evidence Level
Phase 1: Initial Optimization	Cross-sectional empirical; econometric	[8,16,33,42,62]	1990–2023	Moderate
Phase 1: Initial Optimization	Observed sector-level	[6,63,65]	2021–2025	Moderate
Phase 2: Rebound Effect	Observed longitudinal	[3,34,35]	2020–2024	Moderate
Phase 2: Rebound Effect	Econometric projection	[8,16]	1995–2023	Moderate
Phase 3: Mature Optimization	Computational model/simulation	[3]	2025–2050	Indicative
Phase 3: Mature Optimization	Econometric projection (conditional)	[8,33]	1990–2020	Indicative

summarizes the evidentiary basis for each S-curve phase, categorizing supporting studies by study type and time horizon.

AI’s environmental trajectory resembles building a highway. Initially, efficient cars reduce travel time and emissions (Initial Optimization). As the highway expands, more drivers increase traffic and emissions (Rebound Effect). When it matures with electric transit and smarter traffic management, environmental costs decrease (Mature Optimization).

This S-curve pattern explains conflicting narratives about AI’s environmental impact. Studies focusing on early deployment or isolated applications show benefits, while analyses of overall infrastructure trends reveal increasing environmental burdens [15,16]. Both views are correct but limited to specific timeframes. The key message for clean technology practitioners is that deployment decisions must consider the full lifecycle, not just short-term snapshots. Short-term gains do not ensure long-term sustainability without deliberate rebound mitigation strategies.

2.5. Geographical Mediators

Geographical context fundamentally alters AI’s environmental impact trajectory by influencing both the scale of ecological costs and the effectiveness of AI as a sustainable asset. The environmental “cost” of AI, primarily energy and water consumption, is highly sensitive to where data centers are located and how they are powered. The potential for AI to serve as a “sustainability enabler” is often conditional on a country’s socio-economic and institutional maturity.

Our framework identifies four critical geographical mediators:

• Carbon Intensity of Local Grids. The carbon footprint of a data center is directly proportional to the carbon intensity of its location’s electricity mix [49]. Regional electrical grid composition determines whether AI energy consumption translates to high or low carbon emissions. Data centers in regions with 80%+ renewable electricity (e.g., Norway and Switzerland) generate less than 20 g CO₂/kWh, while those in coal-dependent regions (e.g., Australia and South Africa) generate over 800 g CO₂/kWh—a difference of 40 times those that rely primarily on renewable electricity [49]. This disparity means identical AI workloads produce vastly different carbon footprints based solely on location. Chen et al. [3] demonstrate that China’s AI carbon footprint projection varies widely depending on grid decarbonization speed. Even within the U.S., carbon emissions for the same AI workload can vary significantly between regions like California, West Nebraska, and the Rocky Mountains due to different levels of renewable energy integration [25]. Availability of additional renewable energy capacity determines whether AI deployment drives new fossil fuel generation or absorbs existing renewable surplus. Regions with renewable overcapacity during certain periods (e.g., high solar and wind production) can deploy AI workloads with minimal marginal emissions by utilizing otherwise-curtailed energy. Conversely, regions at grid capacity limits face the opposite scenario: new AI demand necessitates new generation, typically fossil fuel-based due to faster construction timelines.
• Water Availability. Water scarcity determines whether cooling systems create acute environmental stress or manageable impact. AI’s water footprint is exacerbated in water-stressed regions (e.g., Western U.S., India, Australia, parts of China, and the Middle East) where data centers compete with agriculture and domestic needs [19,38]. Estimates suggest global AI demand could account for 4.2 to 6.6 billion cubic meters of water withdrawal by 2027, a burden concentrated in these vulnerable watersheds [38]. Herrera et al. project that AI infrastructure water consumption in water-stressed regions could exceed sustainable withdrawal rates in these areas by 2030 without alternative cooling technologies. Conversely, deployment in water-abundant regions with access to seawater or ample freshwater (Nordic countries, Pacific Northwest) creates minimal scarcity pressure.
• Climate and Thermal Impacts. Geography influences power usage effectiveness (PUE) and cooling costs [3,28]. Ambient temperature and humidity determine the level of cooling energy intensity required. Data centers in cooler climates (e.g., Nordic countries, Canada) can save energy by using free ambient air for up to 60–80% of the year, reducing cooling energy compared to those in hot or humid regions, which will require more electricity and water [19,38,49]. This geographical advantage compounds over years, creating substantial cumulative energy and water savings.
• Socioeconomic Strength. Lower-income countries often adopt advanced AI technologies without the necessary regulatory capacity or digital infrastructure to mitigate environmental costs [16]. In these contexts, AI may follow a linear trajectory of rising emissions rather than the inverted U-shaped pattern (peaking and then declining) observed in mature economies [8]. Strong institutions (e.g., in Germany and Sweden) can use AI to meet decarbonization targets through strict environmental auditing [16]. Empirical panel data analysis across 35 OECD countries confirms that institutional quality, measured by governance effectiveness, regulatory quality, and rule of law, has a significant positive effect on environmental performance, and that without strong institutions and effective energy transition strategies, AI deployment increases rather than reduces carbon emissions [33]. Conversely, in regions with weak governance, AI adoption may occur in extractive or high-emission sectors without oversight, intensifying carbon output, and natural resource depletion [16]. Trade restrictions and export controls on advanced semiconductors can force certain regions to rely on less energy-efficient legacy hardware, which increases the energy footprint of their domestic AI sectors and limits their access to efficiency gains [1].

These geographical mediators interact multiplicatively rather than additively. A data center in Iceland benefits from renewable electricity, water abundance, and favorable cooling climate, achieving near-minimal environmental impact. Conversely, a center in a coal-dependent, water-stressed, hot climate faces compounded burdens across all dimensions. This geographical dependency means universal AI deployment recommendations are inappropriate. Context-sensitive strategies are essential.

China’s “East-West Computing Resources Transmission Project” aims to relocate data centers to western regions with abundant renewable energy [3]. However, this shift initially creates “new carbon hotspots” in the west because local grid factors remain high and energy infrastructure development lags behind computational expansion [3]. By 2030, states like Texas, Illinois, and Washington are projected to reduce data center emissions due to high renewable shares, while New York and California may see increases due to continued reliance on natural gas for baseload power [48].

3. Methodology

This systematic review follows established protocols for literature synthesis adapted for clean technology assessment. We used a multi-stage process of searching, screening, and analyzing to develop the Environmental Asset-Cost Framework.

3.1. Search Strategy and Data Sources

We conducted systematic searches in Web of Science and Scopus databases for peer-reviewed articles published between 1 January 2021, and 15 October 2025. Searches were completed in Web of Science and Scopus using the following terms.

Web of science search: AK = ((“artificial intelligence” OR “AI” OR “machine learning” OR “large language model”) AND (“environmental” OR “carbon” OR “greenhouse” OR “emissions” OR “climate” OR “ecological” OR “sustainability” OR “sustainable”)) AND TI = ((“artificial intelligence” OR “AI” OR “machine learning” OR “large language model”) AND (“environmental” OR “carbon” OR “greenhouse” OR “emissions” OR “climate” OR “ecological” OR “sustainability” OR “sustainable”)).

Scopus search: KEY (“artificial intelligence” OR “AI” OR “machine learning” OR “large language model”) AND (“environment” OR “carbon” OR “greenhouse” OR “emissions” OR “climate” OR “ecological”) AND (TITLE ((“artificial intelligence” OR “AI” OR “machine learning” OR “large language model”) AND (“environment” OR “carbon” OR “greenhouse” OR “emissions” OR “climate” OR “ecological”))).

A sensitivity check was conducted using a title-only search omitting the environmental keyword requirement for a subset of highly cited AI infrastructure publications. Results confirmed that the vast majority of empirical studies with substantive environmental data also contain environmental terms in titles or abstracts, suggesting the practical magnitude of any exclusion bias is limited within the defined scope. This check is reported as a methodological note and does not alter the included study set.

To complement the peer-reviewed corpus and address the recognized limitation of academic databases in capturing practitioner-relevant quantitative data, a targeted supplemental search of high-signal grey literature was conducted covering five source categories: IEA World Energy Outlook and Data Centre reports; IPCC Working Group III contribution to the Sixth Assessment Report (AR6); EU AI Act regulatory impact assessments; major hyperscaler sustainability disclosures (Google Environmental Report 2024, Microsoft Environmental Sustainability Report 2024, Amazon Sustainability Report 2023); and IEEE Standards Association sustainability documentation. Sources from these grey literature searches that provided quantitative environmental data not represented in the peer-reviewed corpus—primarily the IEA baseline electricity consumption figures and IEA projection scenarios—are incorporated as contextual background evidence throughout the manuscript and are clearly labeled as grey literature rather than systematic review findings (see Appendix B, “Source Type” column). Grey literature sources are not included in the 73-study evidence corpus and do not contribute to framework category assignments in Appendix C.

3.2. Inclusion and Exclusion Criteria

Inclusion criteria required articles published in English to: (1) address AI’s net environmental impact from a trade-off perspective, encompassing either (a) studies quantifying the direct environmental footprint of AI systems (e.g., energy consumption, water use, e-waste, supply chain burdens) or (b) studies measuring AI-enabled environmental benefits (e.g., efficiency gains in smart grids, HVAC systems, precision agriculture, or industrial optimization)—studies addressing only one side of this ledger are eligible provided they contribute measurable data to either the Asset or Cost dimension of the Environmental Asset-Cost Framework; (2) report empirical data, measurements, or systematic assessment frameworks; (3) address one or more environmental dimensions (e.g., energy, carbon, water, materials, waste, etc.); and (4) provide sufficient methodological detail for framework development.

Exclusion criteria eliminated articles that: (1) addressed neither the direct environmental footprint of AI systems nor AI-enabled environmental benefits with measurable data (e.g., purely technical performance studies with no environmental dimension); (2) presented purely conceptual arguments without empirical grounding; (3) reported only preliminary or incomplete results; or (4) focused exclusively on non-environmental aspects of AI sustainability (social equity, economic impacts, governance).

3.3. Article Selection Process

Initial searches identified 7827 records (Web of Science: 2908; Scopus: 4919) as shown in the PRISMA Flow diagram in Figure 1. After removing 2156 duplicates using reference management software, 5671 unique records underwent title and abstract screening. Two reviewers independently screened titles and abstracts using the inclusion/exclusion criteria, with disagreements resolved through structured discussion; cases where consensus could not be reached were adjudicated by a third reviewer. This process excluded 5543 records, primarily AI applications for environmental purposes (3800), irrelevant topics (1390), and insufficient methodological detail (353). Full-text review of 128 articles yielded 73 articles meeting all inclusion criteria after excluding AI-for-environment applications (40), duplicates missed in initial screening (10), and insufficient environmental impact focus (5). Inter-rater reliability for final inclusion decisions was κ = 0.89, indicating strong agreement. Framework coding—the assignment of included studies to Asset and Cost categories within the 5 × 5 matrix—was conducted collaboratively through structured consensus discussion between the two reviewers rather than independent parallel coding, with a third independent reviewer consulted to resolve cases where consensus could not be reached. We acknowledge this as a methodological limitation. The absence of independently calculated inter-rater reliability statistics for the framework coding stage reduces the replicability of category assignments. Future systematic reviews employing this framework should incorporate independent parallel coding at the framework stage to generate Kappa statistics for both the inclusion and coding phases.

3.4. Framework Development and Validation

Framework development progressed through four iterative phases. Initial categorization identified recurring impact themes via open coding of the first 20 articles, revealing distinct asset categories such as optimization, enhancement, innovation, conservation, and precision, and cost categories including energy, water, e-waste, infrastructure, and supply chain. Pattern recognition through chronological analysis of studies reporting temporal data revealed the S-curve trajectory, confirmed across multiple independent studies. Dimension mapping structured all reviewed evidence into the 5 × 5 Asset-Cost matrix, with geographical mediators identified through comparative analysis of regional studies. Validation confirmed that all 73 included studies were assignable to at least one framework cell without forcing evidence into predetermined categories; 41 studies were assigned to multiple cells (average 2.2 category assignments per study), reflecting the multi-dimensional nature of AI’s environmental impacts.

To illustrate the application of the decision rules and how the coding team resolved ambiguous assignments, three representative borderline cases are described below.

• Case 1: Dual Asset assignment (A1 and A2): Ajagekar et al. [6]. This study applies deep reinforcement learning to control greenhouse climate and supplementary lighting, simultaneously yielding measurable reductions in energy consumption (A1) and improvements in crop production outcomes (A2). Initial review considered assigning it solely to A1, as energy optimization is the primary intervention mechanism. However, the study explicitly reports crop yield outcomes alongside energy metrics, providing distinct empirical data for both dimensions. Applying Decision Rule 3 (multi-category assignment), the study was coded to both A1 and A2, with A1 as the primary category.
• Case 2: Asset–Cost intersection (A5 and C1/C2/C3): Kocak et al. [38]. This study examines radiology AI, which simultaneously generates environmental benefits through precision diagnostics that reduce unnecessary procedures and repeat imaging (A5) and imposes environmental costs through the energy demand, cooling water use, and hardware waste of the AI systems supporting those diagnostics (C1, C2, C3). The borderline question was whether the environmental benefit of precision medicine (A5) is sufficiently distinct from the cost burden of the AI system itself. Applying Decision Rule 4 (Asset–Cost intersection), the study was coded to A5, C1, C2, and C3, as both sides are empirically quantified and the study explicitly frames them as a sustainability paradox.
• Case 3: Scope boundary (borderline inclusion): Muhammad et al. [60]. This study investigates the strategic enablers of AI-driven public sector transformation in Nigeria using a Fuzzy DEMATEL model, with leadership commitment, institutional frameworks, and digital infrastructure identified as primary drivers. Because the paper does not directly quantify AI’s environmental footprint or AI-enabled environmental benefits, it represents a boundary case under the revised inclusion criteria. It was retained in the corpus because it contributes measurable data on digital infrastructure requirements in a lower-income country context, directly informing the C4 (Infrastructure Impacts) dimension, and because understanding institutional readiness in such contexts is necessary for the framework’s geographic moderator analysis. Studies of this type were flagged during coding for consensus review, and their primary category assignment (C4) was confirmed by the third reviewer.

3.5. Data Extraction and Synthesis

From each article, we extracted: (1) environmental impact type and magnitude, (2) temporal characteristics and reported timelines, (3) geographical context and location-specific factors, (4) methodological approach and data sources, and (5) sector-specific application context. Quantitative data were synthesized using narrative synthesis methods appropriate for heterogeneous study designs. Where multiple studies reported similar metrics, we present ranges reflecting variation across contexts rather than statistical aggregates, as heterogeneity in methods, locations, and AI applications precluded meta-analysis.

4. Results: Environmental Asset-Cost Analysis

Our systematic analysis of existing studies reveals complex, multi-dimensional environmental impacts that vary substantially across temporal phases and geographical contexts. We structured the results according to the Environmental Asset-Cost Framework (Table 1), emphasizing the critical S-curve temporal pattern and geographical dependencies.

4.1. The S-Curve Temporal Pattern

A key organizing finding from our systematic analysis is the identification of a synthesized S-curve heuristic characterizing AI’s net environmental impact over time. This pattern emerged independently across multiple studies examining different AI applications, geographical regions, and temporal scales, indicating a synthesized decision heuristic rather than context-specific variation.

Chen et al. [3] provide the most comprehensive quantitative evidence for this pattern through modeling China’s AI infrastructure expansion through 2050. Their analysis identifies three distinct phases: (1) From 2020 to 2025, net carbon reductions occurred as AI optimization in manufacturing and energy systems lowered emissions. (2) From 2025 to 2035, they project a rebound effect with AI infrastructure emissions rising significantly while optimization benefits level off. (3) Between 2035 and 2045, they project a potential return to net benefits if grid decarbonization speeds up and infrastructure growth stabilizes. Specifically, they cite emissions at 82 metric tons (Mt.) in 2022, up to 695 Mt. in 2038, then decreasing to 474 Mt. by 2050. In the long term, they highlight improvements in data center efficiency and the increased use of clean electricity sources as key factors that could reduce total emissions by 2050.

Alnafrah [16] supports the first two stages of the pattern using a “dynamic threshold model” applied to a decade of data (2013–2023). The study found that in “low AI intensity” regimes (representing early-stage adoption), increases in AI activity were significantly associated with reductions in CO₂ emissions. Beyond a critical threshold, the relationship reverses. The study identifies a “Digital Rebound Paradox” where efficiency gains are offset by increased consumption, system expansion, and escalating computational demands, leading to a significant rise in emissions.

Yu et al. [35] document the carbon footprint surge through a quantified analysis of the carbon emission associated with 79 AI systems released between 2020 and 2024. They found that carbon emissions of GPT-4 reached 21,660-t CO₂e, which is a 12-fold (1100%) increase over its pre-iteration model, GPT-3.5. They project that the demand for AI services will grow by 30–40% annually over the next decade, leading to ongoing updates of AI models, which will increase energy consumption and carbon emissions. However, they note that the higher the proportion of sustainable energy used, the lower AI-related carbon emissions will be. They propose that setting emission limits could encourage the industry to adopt more environmentally friendly practices, technologies, and sustainable energy sources, helping create a more sustainable future for AI. This exemplifies how efficiency gains can be overwhelmed by scale increases during the rebound phase.

Ren et al. [34] conducted life cycle assessments, comparing the environmental costs of a typical large language model (LLM) to a human, revealing that LLMs potentially outperform humans (using data for both U.S. residents and Indians) when evaluating energy consumption, carbon emissions, water consumption, and economic costs of each. As today’s comparatively lightweight LLMs offer potential efficiency advantages, they acknowledge the likelihood that the larger models of the future will have greater environmental impacts, which could reduce the benefits and make a case for human workers to perform these tasks until research efforts are strengthened to ensure the long-term sustainability of LLMs.

According to Leuthe et al. [53], machine learning (ML) projects often get stuck in an experimental pilot phase and fail to transition into productive, value-adding applications. Drawing on previous research, they note that many ML projects are terminated prior to completion because they fail to live up to their intended outcomes. This lack of success is often attributed to a failure to consider the end-to-end development process, focusing too narrowly on initial modeling and training rather than long-term deployment. Through expert interviews and real-world case studies, they showed that while teams initially focus on performance and “the bigger, the better,” their sustainable ML design pattern matrix stimulates behavioral intention to implement sustainable practices that eventually lead to long-term cost and energy savings.

The S-curve pattern has profound implications for clean technology decision-making. First, it explains why point-in-time assessments sometimes lead to conflicting conclusions. Studies conducted during Phase 1 emphasize benefits, while Phase 2 studies highlight burdens, and Phase 3 evidence remains limited due to temporal constraints. Second, it shows that short-term environmental gains do not ensure long-term sustainability without explicit rebound mitigation strategies. Third, it suggests that many current AI deployments may be in the most environmentally damaging phase (rebound period), considering typical two to four-year deployment timelines, which explains recent concerns about AI’s increasing environmental footprint. Fourth, it indicates that stopping AI during the rebound phase can prevent the realization of potential long-term benefits, emphasizing the importance of sustained effort along with mitigation strategies.

4.2. Geographical Disparities in Net Environmental Impact

Geographical context determines whether AI deployment achieves net environmental benefits and dramatically influences both the magnitude and timing of the S-curve pattern. Our analysis reveals 10–60x variation in environmental impacts based solely on deployment location, independent of AI application or system design. (Derivation: the lower bound of this range reflects carbon intensity differences between optimal and challenging grid contexts—less than 20 gCO₂e/kWh vs. over 800 gCO₂e/kWh, a 40x carbon variation per [49]; the upper bound incorporates compounding water use multipliers across cooling technology types of 1.8–12 L/kWh per [19], yielding combined lifecycle impact differentials approaching 60x in extreme comparisons; unit: ratio of lifecycle environmental impact per unit AI workload across deployment contexts)

Alnafrah [16] confirms that AI’s environmental role is conditional on national policy and productive capabilities. The trajectory is not uniform. High-income countries often see AI already contributing to emission reductions, while lower-middle-income countries experience a “technological leapfrogging paradox” where rapid AI adoption outpaces regulatory capacity, leading to a linear increase in emissions.

Zhang et al. [8] demonstrate that the magnitude and timing of the turning point where AI begins to yield net environmental benefits are heavily influenced by the local energy mix. Specifically, high renewable energy consumption shifts the turning point to the left, enabling earlier emissions reductions, while nuclear energy helps flatten the initial emissions curve by meeting the baseload demand of AI infrastructure.

Bolón-Canedo et al. [49] highlight “great variability between countries” in the carbon intensity of their electricity grids, ranging from less than 20 gCO₂ e/kWh in Norway and Switzerland to over 800 gCO₂ e/kWh in Australia and South Africa. This represents a 40x variation based on the energy provider and geographical location.

Jagannadharao et al. [25] note that the carbon intensity of the grid in different U.S. regions varies dramatically; for example, the WAPA Rocky Mountain Region (WACM) can reach intensities over 2000 lbs. CO₂/MWh, while the California ISO North grid can drop to near zero. This wide spectrum confirms that the environmental impact of the same AI workload is highly dependent on where and when it is executed.

Regional comparative analysis reveals three distinct geographical contexts with fundamentally different environmental trajectories:

• Optimal Context (High Renewables and Water Abundance). Nordic countries, the Pacific Northwest U.S., Scotland, and parts of Canada combine renewable electricity, water abundance, and cool climates, enabling free cooling. AI deployments in these regions demonstrate lower carbon footprints and lower water consumption than global averages [19,38]. The S-curve rebound phase is attenuated or eliminated, with some deployments maintaining net environmental benefits throughout the lifecycle [16]. Iceland’s data center sector exemplifies this optimal context: 100% renewable electricity (geothermal and hydro), negligible water stress, and ambient temperatures enabling year-round free cooling [35]. The carbon intensity of electricity is less than 20 gCO₂e/kWh in Norway and Switzerland, a carbon footprint roughly 40 times lower than coal-dependent counterparts [49]. Higher renewable energy consumption shifts the turning point of the AI emissions curve to the left, effectively eliminating or shortening the period where AI adoption increases net emissions [8].
• Challenging Context (Fossil-Heavy and Water-Stressed). Parts of the Middle East, Western U.S., Northern China, and India face compounded challenges: >70% fossil fuel electricity, severe water scarcity, and hot climates requiring intensive cooling. China’s data centers are concentrated in northern provinces that experience chronic water scarcity, while India faces high water stress across much of the country [19]. Both regions rely on coal-dominated energy structures [3,15]. Data center hotspots in Arizona and Virginia consume millions of liters of water daily, often in arid climates which make water-intensive evaporative cooling systems ecologically problematic [19]. AI deployments in these regions generate 800–1200 g CO₂/kWh and consume 5–9 L water per kWh, creating acute environmental burdens [4,19,25,32,49]. The S-curve rebound phase is amplified and extended, with some scenarios never achieving net environmental benefits even in long-term projections [8,16]. Jha et al. [48] project that data center CO₂ emissions will account for 3–14% of the U.S. power sector’s total emissions by 2030, representing a significant hurdle for national climate goals. Due to high water consumption, expected to reach 1006 billion liters annually by 2030, data centers in water-stressed areas like Arizona and Nevada are facing increased public scrutiny and likely will see regulatory constraints on water withdrawals which could lead to costly technology upgrades or facility relocations [48].
• Transition Context (Mixed Grids and Moderate Resources). Eastern U.S., Central Europe, Coastal China, and parts of East Asia occupy middle ground: 40–60% renewable electricity with ongoing transition, moderate water availability, and variable climates. AI deployments in transition regions demonstrate high variability in environmental outcomes depending on specific local conditions, renewable energy procurement strategies, and technology choices [25]. These regions offer potential for accelerated improvement as grid decarbonization progresses, but also risk extended rebound phases if infrastructure growth outpaces grid transformation [1]. Zhang et al. [8] demonstrate that regions coupling AI deployment with dedicated renewable energy procurement can achieve impacts comparable to optimal contexts, while those relying on existing grid mix face impacts approaching challenging contexts.

Table 3. Geographic Context and Net Environmental Outcome: Illustrative Comparison for AI-Optimized Data Center Workload.

Geographic Context	Grid Carbon Intensity	Water Stress	Cooling Climate	Net Carbon Impact	Net Water Impact	Phase 3 Likelihood	Representative Region
High-renewable/water-abundant	<20 gCO2e/kWh	Low	Cool (free cooling viable)	Strongly positive	Minimal burden	High	Iceland, Norway, Pacific NW U.S.
Transitioning grid/moderate resources	200–500 gCO2e/kWh	Moderate	Temperate	Variable—depends on procurement	Moderate	Moderate (conditional)	Eastern U.S., Central Europe
Fossil-heavy/water-stressed	>700 gCO2e/kWh	High	Hot (intensive cooling required)	Strongly negative	Acute burden	Low to none	Northern China, Arizona, parts of India

Sources: Grid carbon intensity data from [49]; water stress and cooling scenario analysis from [19]; regional deployment outcomes from [3,16].

illustrates how the same AI application generates fundamentally different net environmental outcomes across contrasting geographic contexts.

Case Vignette 1: Grid Carbon Intensity and HVAC Optimization. Bolón-Canedo et al. [49] document that the carbon intensity of electricity grids ranges from less than 20 gCO₂e/kWh in Norway and Switzerland to over 800 gCO₂e/kWh in Australia, South Africa, and some U.S. states (a 40-fold difference). For an AI-optimized HVAC system achieving a 25% reduction in building energy consumption, this geographic variation determines whether the system is a net carbon reducer or a net carbon contributor. In a Nordic deployment powered by near-zero-carbon hydroelectric or geothermal electricity, the HVAC optimization yields genuine emissions reductions with negligible offsetting infrastructure burden. The same system deployed in a coal-dependent region generates substantial operational emissions from its own computational infrastructure, potentially eliminating or reversing the carbon savings from the optimization it enables. Zhang et al. [8] confirm this mechanism econometrically, demonstrating that high renewable energy consumption shifts the AI emissions turning point earlier, while fossil-fuel dependence delays or prevents net environmental benefits.

Case Vignette 2: Water Stress and Cooling Infrastructure Choices. Herrera et al. [19] model AI infrastructure water consumption under uncertainty across global regions, finding that cooling technology choice interacts critically with local water availability to determine net water impact. In water-abundant regions with access to cool ambient air or seawater (e.g., Nordic countries, Pacific Northwest), AI data centers can operate with minimal freshwater withdrawal using free air cooling for the majority of the year. The same facility in a water-stressed region (e.g., Arizona, Northern China) relying on evaporative cooling systems consumes 5–9 L of water per kWh of electricity, creating direct competition with agricultural and municipal water supplies. Jha et al. [48] project that U.S. data center water consumption could reach 1006 billion liters annually by 2030 under high AI growth scenarios, with consumption concentrated in water-stressed geographies already experiencing groundwater depletion. These regional dynamics mean that cooling technology choice, not just energy source, is a critical deployment decision variable in stressed contexts.

4.3. Asset Categories: Quantified Environmental Benefits

Systematic analysis of asset categories reveals substantial environmental benefits when AI is strategically deployed, though benefits vary significantly across applications and contexts.

Energy Optimization applications demonstrate 15–50% reductions depending on baseline efficiency and application context [5,26,49]. Industrial AI systems achieve 20–40% energy savings in manufacturing through process optimization, predictive maintenance, and production scheduling [6,11,52]. Smart building systems can reduce HVAC energy consumption by 15–35% and overall building consumption by 10–25% through occupancy prediction, load balancing, and equipment optimization [4,17,57]. Grid management AI improves renewable integration by 10–25%, reducing curtailment and enabling higher renewable penetration without reliability compromises [7,69]. These optimization benefits represent genuine emission reductions only when avoided energy would otherwise be consumed—an important caveat for rebound analysis [16,38,52].

Production Enhancement in precision agriculture reduces fertilizer and water usage by 20–30% while maintaining yields [23]. AI enables the precise and targeted use of pesticides to reduce environmental impact [42,49]. Manufacturing optimization reduces material waste by 10–30% and production energy by 15–35% [11,29]. These benefits translate to direct environmental impact reduction through avoided resource extraction, reduced chemical pollution, and decreased water stress.

Green Innovation Acceleration shows promise but remains difficult to quantify. AI-driven materials discovery and quantum computing hold immense potential to accelerate materials research for clean energy, solar panels, and batteries [1,49]. However, translating accelerated discovery to realized environmental benefits depends on commercialization success and deployment scale, outcomes that remain uncertain for recent discoveries [21,23,38].

Resource Conservation through circular economy applications can extend product lifecycles through predictive refurbishment and optimal repair timing [11,30,38]. Supply chain optimization reduces transportation emissions up to 30% through route optimization and demand forecasting [11,45,46]. These benefits accumulate over time and may strengthen during Phase 3 of the S-curve as systems mature [8,16].

Precision Applications in healthcare reduce unnecessary procedures by up to 20%, lowering medical waste and energy consumption [38]. Infrastructure monitoring prevents resource losses through early detection of leaks, failures, and inefficiencies, with reported savings of 10–30% in water distribution systems and up to 30% in electrical grids [7,23,42].

Critical assessment reveals three important patterns. First, optimization benefits are largest in baseline inefficient systems. While modern cloud data centers are highly optimized, older, smaller data centers lack the scale for efficient cooling [49]. Specific examples of “maximal gains” in baseline systems include a 57% reduction in energy for greenhouse climate control and a 40% reduction in data center cooling costs [4,6]. However, the carbon footprint of AI is scaling rapidly and at the systems level may partly offset these efficiency gains [70]. Even small improvements in accuracy (marginal gains) in already high-performing models can lead to disproportionately high increases in energy consumption [31,71]. Second, emissions initially increase during the early stages of adoption (Phase 1) due to “capital deepening” and high energy intensity [8]. The benefits (net emission reductions) are generally realized only after a turning point as the technology matures and efficiency gains eventually outweigh the scale effect [8]. The sources suggest the relationship is “stage-dependent” [16]. Third, realized benefits depend critically on counterfactual assumptions: if efficiency gains enable production expansion rather than absolute reduction, rebound effects eliminate environmental benefits [1,16,25,72,73].

4.4. Cost Categories: Quantified Environmental Burdens

Systematic analysis of cost categories reveals substantial and growing environmental burdens that escalate during the S-curve rebound phase.

Energy Consumption for AI training and inference constitutes the dominant environmental burden [70]. GPT-3 training consumed approximately 1287 MWh, generating 550 metric tons CO₂ in average grid conditions, with leading models now far exceeding this range (e.g., GPT-4 at 7200 MWh) [4,23,49,68,70] (GPT-3: 1287 MWh/550 tCO₂eq—from Patterson et al. 2021 reported as contextual background through reviewed studies [49,68]; GPT-4: 7200 MWh—industry modeling estimate reported as contextual background in [4,49]; unit: MWh per training run/metric tons CO₂eq at average U.S. grid carbon intensity of ~429 gCO₂/kWh). Inference operations at scale consume 260–500 MWh daily for major platforms serving millions of users [4,49,68]. Data centers supporting AI workloads demand 300–500% more electricity than traditional computing due to GPU acceleration requirements [4,15]. Global projections estimate AI data center electricity consumption will reach 340–945 TWh annually by 2030, representing 2–5% of global electricity demand (equivalent to adding Japan’s entire electricity consumption to global demand) [3,48].

Water Use for data center cooling and electricity generation represents an often overlooked but critical burden [15,19,23]. Direct cooling water consumption averages 1.8–12 L per kWh depending on technology and climate [4,19,48]. Training a single large language model indirectly consumes approximately 700,000 L through electricity generation and direct cooling [21,23]. U.S. data centers are projected to consume over 1000 billion liters annually by 2030 [4,48]. Water stress is concentrated geographically, with a significant portion of U.S. data center water consumption occurring in water-stressed regions experiencing groundwater depletion and ecosystem degradation [19,48,70].

E-waste Generation accelerates due to AI hardware’s rapid obsolescence [15,23,36,38]. AI hardware, including servers and AI-specialized GPUs, is frequently decommissioned at the world’s approximately 11,800 data centers every 3–5 years [15,48]. Of the 62 million tons of e-waste discarded in 2022, representing an increase of 82% since 2010, the global recycling rate is less than 22% [13,15]. Rare earth element recovery from e-waste is technically feasible but economically uncompetitive with primary mining, creating a linear rather than circular material flow [15,30,39]. The circularity challenge is acute because refurbished or harvested chips often fail to meet the high-performance requirements of new AI applications, though they may be repurposed for lower-end, general-purpose tasks [15,30].

Infrastructure Impacts from AI data centers encompass significant material and environmental costs beyond operational electricity. Construction and hardware manufacturing contribute approximately 10% to 18% of the total lifecycle carbon footprint of AI systems [3]. These facilities are physical objects housed in buildings made of aluminum, steel, and concrete, which depend on complex mechanical systems, such as compressors, heat rejection fans, and pumps, and electrical systems, including transformers and backup generators [15]. Data centers often prioritize horizontal construction to reduce costs, a practice that requires vast tracts of land and leads to habitat fragmentation, ecosystem disruption, and land-use conflicts [13,15]. In the US, data center water consumption is projected to reach as high as 1006 billion liters annually by 2030, concentrated in regions already experiencing water stress and groundwater depletion [48]. These burdens on infrastructure are compounded especially in regions reliant on fossil fuel energy, where deployment of AI may exacerbate rather than reduce environmental impacts [48].

Supply Chain Extraction impacts are centered on the resource-intensive semiconductor industry. The production of a single AI accelerator GPU requires between 20,000 and 30,000 L of ultrapure water [19,21]. In 2021, global semiconductor manufacturing generated an estimated 76.5 million metric tons of CO₂ equivalent [30] (unit: million metric tons CO₂eq; baseline year: 2021; contextual background reported in [30] citing Pelcat 2023; scope 3 emissions represent approximately 79% of this total per [30]). The extraction of rare earth elements (CRMs) like lithium, cobalt, and gallium for AI hardware is linked to severe environmental degradation, including deforestation and water contamination [7]. Processing just one ton of rare earth ore can produce up to 2000 tons of toxic waste, which is often left to pollute the surrounding natural environment [21,39]. Manufacturing activities remain highly concentrated in regions such as Taiwan, South Korea, and Japan, where industrial grids show a significantly higher reliance on fossil fuel energy compared to parts of Europe and the U.S. [15,30].

Table 4. Number of Included Studies Contributing Evidence to Each Framework Category.

Dimension	Category	n	Key Supporting Studies (Selected)
Asset	A1: Energy Optimization	25	[1,4,5,6,7,8,9,10,16,18,22,27,33,37,42,44,49,52,53,56,58,62,63,64]
	A2: Production Enhancement	8	[6,9,23,24,29,58,63,65]
	A3: Green Innovation Acceleration	12	[1,7,9,10,22,29,32,45,50,52,62,66]
	A4: Resource Conservation	6	[9,11,12,30,65,66]
	A5: Precision Applications	11	[12,17,23,24,36,38,39,51,61,65,67]
Cost	C1: Energy Consumption	63	[1,2,3,4,5,6,7,8,10,11,13,14,15,16,17,18,20,21,22,23,24,25,26,27,28,35,36,38,39,40,41,42,43,44,45,46,47,48,49,51,52,53,54,55,56,57,59,61,62,64,66,67,68,69,70,71,72,73]
	C2: Water Use	14	[1,3,4,7,13,15,19,21,34,38,39,51,61,71]
	C3: E-Waste Generation	12	[4,7,10,13,15,21,34,35,38,39,51,61]
	C4: Infrastructure Impacts	7	[2,32,35,47,48,59,60]
	C5: Supply Chain Extraction	6	[1,4,7,30,39,40]
Total assignments		164	Average 2.2 categories per study

illustrates the number of studies reviewed that contribute evidence to each of framework category.

5. Discussion: Framework Applications for Clean Technology Decision-Making

The Environmental Asset-Cost Framework transforms our systematic findings into actionable guidance for clean technology practitioners. This section presents decision criteria, regional deployment strategies, rebound mitigation approaches, and policy implications derived from the framework.

5.1. Decision Criteria for Sustainable AI Deployment

Clean technology practitioners should evaluate AI deployment through three systematic assessments structured by our framework:

• Asset Optimization Potential Assessment. Evaluate: (1) What specific efficiency gains are achievable in this application context? (2) What is the baseline energy or resource consumption being optimized? (3) Are optimization opportunities substantial (>20% potential reduction) or marginal (<10%)? (4) What is the projected timeline to realize optimization benefits? Applications with large optimization potential in baseline inefficient systems offer the strongest asset benefits. Evidence for substantial optimization gains (>20%) in baseline-inefficient systems is well-established across sectors: manufacturing [6,9,49,52,64], building energy management [5,17,58], grid integration [7,49,63], and precision agriculture [23,65]. Evidence for marginal-gain scenarios in already-optimized systems is limited to a smaller number of studies and should be treated with caution [31,71]. [Evidence Level: Established for high-baseline systems; Emerging for near-efficiency-frontier systems]. Systems already operating near efficiency frontiers show limited optimization potential, weakening the asset case.
• Cost Burden Assessment. Evaluate: (1) What is the carbon intensity of the local electrical grid (<200 g CO₂/kWh favorable, >600 g CO₂/kWh challenging)? (2) Is the deployment region water-stressed or water-abundant? (3) What is the expected hardware refresh cycle (longer is better for e-waste reduction)? (4) Can dedicated renewable energy procurement accompany deployment? (5) Are alternative cooling technologies viable? Grid carbon intensity thresholds are supported by observed grid data reported across multiple reviewed studies [25,49], with the <200 gCO₂/kWh favorable threshold derived from grid intensity data in optimal deployment regions and the >600 gCO₂/kWh challenging threshold from fossil-heavy regional comparisons [49]. Water stress assessments are supported by scenario-based forecasting [19] and U.S. data center analyses [48]. Hardware refresh cycle evidence derives from industry lifecycle data reported in healthcare and general AI hardware studies [30,39]. [Evidence Level: Established for carbon intensity and water stress dimensions; Emerging for combined multi-factor threshold scoring]. High-renewable grids with water abundance and cool climates minimize cost burdens, while fossil-heavy, water-stressed, hot regions maximize burdens.
• Temporal Planning Assessment. Evaluate: (1) When will the rebound effect likely occur based on planned infrastructure scaling? (2) What mitigation strategies can bridge the rebound period? (3) What is the projected timeline to net positive outcomes? (4) Can the organization sustain commitment through the rebound phase? Deployments should include explicit rebound mitigation strategies rather than assuming initial benefits will persist. Phase 1 and Phase 2 timing evidence is supported by observed and econometric data across multiple studies [3,8,16,34,35]. Phase 3 timeline projections (5–7 years to net positive outcomes) are based primarily on modeled projections and are conditional on grid decarbonization and infrastructure stabilization [3,8,33]; organizations should treat these timelines as indicative rather than definitive. [Evidence Level: Moderate for Phase 1–2 timing; Indicative/Emerging for Phase 3 timelines—see Conditions for Phase 3 in Section 2.4.]. Organizations unable to commit to five to seven-year timelines should reconsider deployment or focus on applications with minimal infrastructure growth requirements.

5.2. Regional Deployment Strategies

Our framework reveals that deployment location is as important as system design in determining environmental outcomes. We identify three regional contexts requiring distinct strategies:

• Optimal Deployment Regions (Prioritize AI Infrastructure). Characteristics: >70% renewable electricity, water abundance, cool climates enabling free cooling. Examples: Iceland, Norway, Sweden, Finland, Scotland, Pacific Northwest U.S., parts of Canada. Strategy: Maximize AI deployment for optimization applications. Infrastructure investment yields minimal environmental burden while asset benefits remain high. These regions can absorb substantial AI infrastructure growth without environmental degradation. Recommendation: Governments should incentivize data center development and encourage transparent ecological reporting, organizations should prioritize these locations for infrastructure expansion, and clean technology sectors should leverage AI extensively for optimization applications. This recommendation is supported by grid carbon intensity data showing <20 gCO₂e/kWh in Nordic regions [49], water footprint scenario analysis confirming low scarcity risk in these geographies [19], and evidence that high renewable energy consumption shifts the AI emissions curve toward earlier net benefits [8,16,33]. [Evidence Level: Established].
• Challenging Deployment Regions (Delay or Constrain AI Infrastructure). Characteristics: >70% fossil fuel electricity, severe water stress, hot climates. Examples: Parts of Middle East, Western U.S. (Nevada, Arizona), Northern China, parts of India. Strategy: Delay AI infrastructure deployment until grid decarbonization substantially progresses or restrict deployment to low-intensity, AI applications with proven optimization benefits exceeding infrastructure costs (as opposed to training massive new models). Recommendation: Organizations should avoid building new AI infrastructure in these regions, governments should implement strict efficiency standards and water use limits, and clean technology priorities should focus on grid decarbonization before AI expansion. This recommendation is supported by China-specific carbon trajectory modeling projecting rebound amplification under fossil-heavy grids [3], U.S. data center emissions projections showing 3–14% of power sector CO₂ under high-growth scenarios [48], water footprint forecasting in stressed regions [19], and multi-country empirical evidence that fossil-heavy, institutionally limited contexts may never achieve Phase 3 net benefits [16,33]. [Evidence Level: Established].
• Transition Regions (Strategic Deployment with Renewable Coupling). Characteristics: 40–60% renewable electricity with ongoing transition, moderate water availability. Examples: Eastern U.S., Central Europe, Coastal China, parts of East Asia. Strategy: Deploy AI infrastructure strategically with mandatory renewable energy procurement agreements and advanced cooling technologies. These regions offer variable outcomes depending on specific implementation choices. Recommendation: Require renewable energy procurement for new AI infrastructure, incentivize zero-water cooling technologies, and accelerate grid decarbonization to improve deployment conditions over time. This recommendation draws on evidence that AI workload carbon impact varies significantly by grid mix within the same country [25], that renewable energy coupling demonstrably shifts emission turning points earlier [8], and that regions at transitional grid composition show high variability in outcomes [1,16]. Direct empirical evidence for transition-region outcomes is more limited than for optimal or challenging contexts. [Evidence Level: Emerging—primarily modeled projections and cross-sectional studies; fewer observed longitudinal case studies].

5.3. Mitigating the Rebound Effect

The S-curve rebound effect is predictable and thus manageable through proactive strategies. We identify five evidence-based mitigation approaches:

• Renewable Energy Coupling. Ensure AI infrastructure deployment coincides with new renewable energy capacity additions rather than consuming existing renewable supply that could displace fossil generation elsewhere. Temporal matching through battery storage or demand flexibility can align AI workloads with renewable availability periods. Evidence base: multi-country econometric analysis confirming renewable energy consumption as the dominant moderator of the AI-emissions relationship [8,33]; China trajectory modeling showing grid decarbonization as the primary Phase 3 enabler [3]; global assessment linking high-income country emission reductions to renewable energy maturity [16]; AI infrastructure expansion review confirming renewable integration as the critical alignment variable [1]. [Evidence Level: Established for the principle; Emerging for specific implementation mechanisms such as temporal matching and battery storage].
• Energy Consumption Caps. Implement absolute energy consumption limits that prevent infrastructure scaling from overwhelming efficiency gains. Set caps at levels preserving net environmental benefits even during infrastructure expansion. Caps should adjust for validated optimization benefits but maintain absolute limits on infrastructure growth. Evidence base: rebound analysis showing infrastructure scaling consistently outpaces per-unit efficiency gains in the absence of absolute limits [3,16,35]; sustainability budget frameworks proposing analogous cap mechanisms for AI development [71]; digital rebound paradox documentation confirming that efficiency alone does not constrain total consumption [16]. [Evidence Level: Emerging—the principle of caps is supported, but direct empirical evidence of cap effectiveness as an implemented policy is limited in the reviewed corpus].
• Hardware Lifecycle Extension. Design for hardware longevity to delay e-waste impacts and reduce embodied emissions from manufacturing. Target 4–6-year lifecycles rather than 2–3 years through modular upgrades, robust cooling, and conservative utilization. Refurbishment and secondary markets can further extend useful life, though specialized AI hardware presents challenges. Evidence base: reported GPU lifecycle data of 2–3 years versus 4–6 years for traditional servers [39]; semiconductor circularity analysis showing lifecycle extension as the highest-leverage supply chain intervention [30]; circular economy AI architecture frameworks demonstrating resource optimization through extended lifecycle design [11]. [Evidence Level: Emerging—supported by 3 independent studies; direct evidence of lifecycle extension as an implemented organizational strategy is limited].
• Alternative Cooling Technologies. Deploy zero-water cooling (air cooling, immersion cooling in dielectric fluids) or closed-loop systems eliminating evaporative loss in water-stressed regions. Evidence base: scenario-based water footprint forecasting identifying cooling technology choice as the primary lever for reducing AI water consumption [19]; U.S. data center analysis projecting water consumption trajectories under different cooling scenarios [48]. [Evidence Level: Emerging—supported by 2 directly relevant reviewed studies; technical feasibility is established but large-scale deployment evidence is limited].
• Circular Economy Integration. Plan for component reuse and refurbishment from deployment onset rather than as end-of-life afterthought. Design infrastructure for disassembly, establish take-back programs, and create secondary markets for functional components. Evidence base: semiconductor value chain circularity analysis identifying design-for-disassembly and secondary markets as viable strategies [30]; green AI architecture research demonstrating resource reuse integration in AI infrastructure design [11]; Industry 5.0 framework analysis confirming circular economy as a core AI sustainability pathway [66]. [Evidence Level: Emerging—supported by 3 studies; implementation evidence is primarily conceptual/framework-level rather than from observed longitudinal deployments].

Practitioner Monitoring Template

Proactive rebound management requires ongoing measurement. The following template in

Table 5. Practitioner Monitoring Template for AI Deployment Environmental Tracking.

Metric	Unit	Monitoring Frequency	Measurement Approach	Public Data Sources	Alert Threshold
Energy consumption per inference unit	kWh per 1000 inferences	Monthly	Measure total data center electricity consumption; divide by inference volume from application logs	Utility smart meter data; cloud provider dashboards (AWS, Azure, GCP energy reports)	>10% increase month-over-month without corresponding inference volume growth signals rebound onset
Water use effectiveness (WUE)	Liters per kWh of IT load	Quarterly	Measure total cooling water withdrawal; divide by IT equipment energy load	Facility water utility bills; The Green Grid WUE benchmarks; EPA WaterSense data for regional baseline	WUE >2.0 L/kWh in water-stressed regions warrants cooling technology review
E-waste per replacement cycle	kg hardware retired per active server	Annually	Track hardware decommissioned vs. hardware deployed; calculate ratio	Asset management records; EPA Electronics Challenge data; manufacturer take-back program records	Hardware retirement cycles shorter than 48 months signal accelerating obsolescence
Grid carbon intensity at deployment site	gCO2eq per kWh	Monthly (or real-time where available)	Query regional grid operator for current carbon intensity of electricity supply	U.S.: EPA eGRID database; EU: Electricity Maps (electricitymaps.com); Global: IEA Electricity Data Explorer	>400 gCO2e/kWh warrants renewable procurement review; >700 gCO2e/kWh signals misalignment with net benefit deployment conditions
Efficiency gain per operational unit	% reduction in target resource (energy, water, fertilizer, etc.) vs. pre-AI baseline	Quarterly	Compare target resource consumption pre- and post-AI deployment; normalize for output volume changes	Application-specific operational records; industry benchmarks from sector associations	If efficiency gains fall below 10% while infrastructure energy/water costs rise, deployment has entered rebound territory

below provides a ready-to-use tracking structure built around five core metrics that practitioners can apply with currently available data sources. Monitoring frequency and data sources are specified for each metric.

These five metrics, tracked together, allow practitioners to detect the transition from Phase 1 (efficiency gains outpacing infrastructure costs) to Phase 2 (rebound onset) before it becomes entrenched. When two or more metrics simultaneously show adverse trends—rising infrastructure consumption alongside stagnating or declining efficiency gains—this signals that rebound mitigation strategies (Section 5.3) should be escalated from planning to active implementation.

5.4. Measurement and Standardization Needs

Our framework reveals critical gaps in current environmental measurement approaches for AI systems. We propose three standardization initiatives:

• AI Lifecycle Environmental Score (ALES). Develop integrated metric accounting for energy, water, e-waste, material extraction, and infrastructure impacts across the full lifecycle. Current metrics measure individual dimensions in isolation, preventing comprehensive assessment. ALES would weight dimensions by regional scarcity (higher water weight in water-stressed regions, higher carbon weight in fossil-heavy grids) and incorporate temporal dynamics through multi-year projections. This enables comparison across deployment contexts and identification of optimization priorities. The need for integrated lifecycle metrics is supported by evidence that existing single-dimension metrics (PUE, WUE, carbon per inference) consistently fail to capture multi-dimensional and temporal impacts [22,27,28,53,55]. The sustainability budget framework in [71] provides a directly analogous governance model. Carbon tracking tool development [26] and hybrid environmental impact assessment methods [55] demonstrate technical feasibility. [Evidence Level: Established for the need; Emerging for the specific ALES construct, which is proposed here as a synthesis contribution].
• Temporal Impact Disclosure Requirements. Require AI developers and deployers to project S-curve impacts rather than reporting point-in-time snapshots. Disclosures should include (1) Phase 1 optimization benefits with evidence of additionality, (2) Phase 2 infrastructure scaling plans and projected peak impact, (3) Phase 3 mitigation strategies and timeline to net positive outcomes. Phase 3 evidence is more limited and conditional than Phases 1 and 2. This transforms environmental reporting from backward-looking accounting to forward-looking planning, enabling stakeholders to evaluate long-term sustainability rather than current status. Supported by evidence that point-in-time assessments produce systematically misleading conclusions depending on deployment phase [3,16,35] and that practitioners lack forward-looking projection tools [1,53]. Phase 3 projections within any disclosure framework should be explicitly labeled as conditional (see Section 2.4 Conditions for Phase 3). [Evidence Level: Emerging—the principle is well-motivated but no reviewed study directly evaluates the effectiveness of temporal disclosure requirements].
• Geographical Context Reporting Standards. Mandate transparent reporting of grid carbon intensity, water stress indices, and renewable energy procurement for all AI infrastructure. Current voluntary reporting allows selective disclosure that obscures unfavorable conditions. Standardized reporting would enable stakeholders to assess deployment context appropriateness and compare environmental performance across organizations. Reporting should include marginal grid impact (not average grid mix) to capture true environmental consequences of additional demand. Supported by documented carbon intensity variation across deployment regions [25,49], water footprint geographic sensitivity analysis [19], and multi-country empirical evidence that identical AI workloads produce dramatically different outcomes based on location alone [8,16,49]. [Evidence Level: Moderate—well-supported empirically but policy implementation evidence is limited to frameworks rather than observed outcomes].

5.5. Policy Implications and Recommendations

Our framework informs context-sensitive policy design, recognizing that uniform regulations are inappropriate given geographical variation:

• Location-Based Deployment Incentives. Provide tax incentives, expedited permitting, and infrastructure support for AI facilities in optimal regions (high renewables, water abundance) while implementing disincentives (carbon taxes, water use fees, efficiency mandates) for challenging regions. This aligns private incentives with environmental outcomes and channels infrastructure investment toward sustainable locations. Supported by econometric evidence that institutional quality and energy transition policy are the dominant moderators of whether AI deployment reduces or increases emissions [33], multi-country evidence that high-income countries with strong institutions achieve AI-enabled emission reductions [16], and grid carbon intensity data confirming 40× variation in deployment impact based solely on location [49]. [Evidence Level: Emerging—principle is strongly evidenced; specific incentive mechanism effectiveness has not been empirically tested in the reviewed corpus].
• Renewable Energy Procurement Mandates. Require new AI data centers to procure renewable energy with verified additionality rather than relying on existing grid mix. Procurement should include temporal matching requirements, ensuring AI workloads align with renewable generation periods, not just annual volume matching that allows fossil fuel consumption during renewable scarcity periods. Supported by evidence that renewable energy share is the primary determinant of Phase 3 timing [8,33] and that infrastructure expansion without accompanying renewable capacity additions extends and amplifies the rebound phase [1,3]. [Evidence Level: Emerging—causal mechanism is empirically supported; policy effectiveness evidence is not available in the reviewed corpus].
• Water Use Regulations in Stressed Regions. Implement strict water withdrawal limits and technology requirements for data centers in water-stressed regions. Regulations should mandate zero-water cooling technologies or closed-loop systems, prohibit evaporative cooling expansion, and require water accounting in environmental impact assessments. Water pricing should reflect scarcity value rather than historical cost-recovery rates that subsidize excessive consumption. Supported by scenario-based forecasting projecting AI water demand exceeding sustainable withdrawal rates in stressed regions by 2030 without intervention [19] and U.S. data center analyses showing concentrated consumption in already water-stressed geographies [48]. [Evidence Level: Emerging—supported by 2 directly relevant reviewed studies; limited to forecasting rather than observed regulatory outcomes].
• Extended Producer Responsibility for E-waste. Establish take-back requirements for AI hardware manufacturers and infrastructure operators. Policies should incentivize design for longevity, component reuse, and material recovery rather than the current linear disposal model. Producer responsibility creates economic incentives for lifecycle extension and circular practices currently lacking competitive markets prioritizing performance over sustainability. Supported by semiconductor supply chain circularity analysis demonstrating that current linear disposal economics are the primary barrier to circular material flows [30] and healthcare AI lifecycle analysis documenting accelerated obsolescence cycles as the dominant e-waste driver [39]. [Evidence Level: Emerging—supported by 2 studies; policy mechanism has not been empirically evaluated in the reviewed corpus].
• Temporal Impact Assessment Requirements. Require environmental impact assessments for major AI deployments to project S-curve dynamics rather than static snapshots. Assessments should identify rebound timing, quantify mitigation strategies, and demonstrate a pathway to net environmental benefits. This prevents approval of deployments that appear sustainable based on Phase 1 analysis but create long-term environmental burdens through Phase 2 rebound. Supported by evidence that Phase 1 analyses systematically underestimate long-term environmental costs [3,16,35] and that rebound timing is predictable given infrastructure scaling data [3,16]. Phase 3 outcomes in any assessment should be labeled as conditional projections rather than forecasts. [Evidence Level: Emerging—rationale is empirically supported; regulatory implementation evidence is absent from the reviewed corpus].

Table 6. Summary Recommendation Table with Evidence Basis.

Section	Recommendation	Supporting Studies	Evidence Level
5.1	Deploy AI where asset optimization potential exceeds 20% in baseline-inefficient systems	[5,6,9,23,49,52,64,65]	Established
5.1	Evaluate grid carbon intensity (<200 g CO2/kWh favorable; >600 g CO2/kWh challenging) before deployment	[25,49]	Established
5.1	Evaluate water stress index and cooling technology options before deployment	[19,48]	Established
5.1	Plan for 5–7 year commitment including explicit rebound mitigation	[3,8,16]	Indicative (Phase 3 conditional)
5.2	Prioritize AI infrastructure in optimal regions (>70% renewable, water-abundant, cool climates)	[8,16,19,33,49]	Established
5.2	Delay AI infrastructure in challenging regions until grid decarbonization advances	[3,16,19,33,48]	Established
5.2	In transition regions, mandate renewable procurement and advanced cooling with deployment	[1,8,25]	Emerging
5.3	Couple AI infrastructure with new renewable capacity additions	[1,3,8,16,33]	Established (principle); Emerging (mechanism)
5.3	Implement absolute energy consumption caps to prevent rebound scaling	[3,16,35,71]	Emerging
5.3	Extend hardware lifecycles to 4–6 years through modular design	[11,30,39]	Emerging
5.3	Deploy zero-water or closed-loop cooling in water-stressed regions	[19,48]	Emerging
5.3	Integrate circular economy design from deployment onset	[11,30,66]	Emerging
5.4	Develop integrated AI Lifecycle Environmental Score (ALES)	[22,26,27,28,53,55,71]	Established (need); Emerging (construct)
5.4	Require temporal S-curve disclosure rather than point-in-time reporting	[3,16,35,53]	Emerging
5.4	Mandate geographic context reporting (grid intensity, water stress, renewable procurement)	[8,19,25,49]	Moderate
5.5	Location-based deployment incentives aligned with renewable grid availability	[16,33,49]	Emerging
5.5	Renewable procurement mandates with verified additionality	[1,3,8,33]	Emerging
5.5	Water withdrawal limits and zero-water technology requirements in stressed regions	[19,48]	Emerging
5.5	Extended producer responsibility for AI hardware e-waste	[30,39]	Emerging
5.5	Temporal impact assessment requirements for major AI deployments	[3,16,35]	Emerging

provides an at-a-glance summary of all practitioner recommendations, their supporting evidence basis, and evidence level classification.

5.6. Cross-Category Interactions and Trade-Offs

The 5 × 5 Asset-Cost matrix represents an analytical decomposition of AI’s environmental impacts, not an assumption of additive independence between categories. In practice, optimizing one dimension frequently generates or amplifies burdens in another, and practitioners should consider these interdependencies when aggregating cell-level assessments into deployment decisions.

Several cross-category interactions are directly documented in the reviewed corpus. Energy optimization benefits can be offset by water consumption costs: AI-driven improvements in data center energy efficiency (A1 → reduced C1) frequently rely on evaporative cooling systems that substantially increase water withdrawal (C2), particularly in warm climates [4,19,48]. A facility that improves its Power Usage Effectiveness (PUE) through more intensive cooling may simultaneously worsen its Water Usage Effectiveness (WUE), producing a trade-off between carbon and water impact that single-dimension assessments would miss entirely.

Precision agriculture gains can be offset by semiconductor supply chain burdens: AI applications in precision agriculture (A2, A5) reduce fertilizer and water use at the field level [23,65], but the sensors, edge computing devices, and connectivity infrastructure enabling precision delivery require semiconductor hardware with significant upstream extraction and manufacturing impacts (C5) [30]. The net environmental benefit of precision agriculture AI depends on whether field-level input reductions over the hardware lifecycle outweigh the supply chain burden of the enabling infrastructure—an assessment that requires cross-category analysis rather than evaluation of the A2/A5 benefit alone.

Green innovation acceleration can generate e-waste through experimental hardware cycles: AI-accelerated materials discovery and clean technology R&D (A3) drives rapid cycles of computational experimentation that accelerate hardware refresh rates and generate e-waste (C3) from decommissioned research infrastructure [4,35,39]. The environmental benefit of discovering a more efficient battery material may be partially offset by the hardware waste generated during the discovery process, particularly if experimental AI systems have short operational lifespans.

These interactions reinforce the framework’s value as an integrated tool: practitioners who evaluate only the most prominent impact dimension risk systematically underestimating total environmental burden. Where included studies provide direct evidence of cross-sector spillover effects, these are flagged in the matrix discussion above. Future research should develop quantitative methods for aggregating cross-category net impacts, moving beyond the qualitative trade-off identification this framework currently enables.

5.7. Governance and Equity Dimensions: Decision-Relevant Considerations Beyond Scope

This review focuses on empirically quantifiable environmental impacts of AI systems. Two additional dimensions—regulatory governance and environmental equity—fall outside the defined scope but interact directly with the environmental outcomes the framework measures, and practitioners should consider them alongside the Asset–Cost analysis.

Regulatory governance shapes whether AI deployment achieves the framework’s environmental conditions. The EU AI Act (2024) establishes transparency and risk classification requirements that indirectly incentivize energy-efficient model design, though it does not yet mandate direct environmental disclosure. The U.S. Executive Order on Safe, Secure, and Trustworthy AI (2023) directed federal agencies to assess AI’s environmental impacts, creating nascent reporting infrastructure. Carbon pricing mechanisms—where implemented, such as the EU Emissions Trading System—alter the economic calculus of high-carbon AI deployment in ways that parallel the framework’s geographic cost assessments [21,71]. Strong institutional quality measurably moderates whether AI deployment achieves net environmental benefits, as documented across 35 OECD countries [33] and in global multi-country analysis [16]. For practitioners, this means that the framework’s deployment recommendations are more reliably achievable in institutional contexts with robust environmental regulation—governance quality is a moderating variable that the 5 × 5 matrix does not explicitly capture but that practitioners should assess alongside the geographic and temporal factors it does address.

Environmental equity concerns the distribution of AI’s environmental costs across communities. Data center siting decisions that optimize for low energy cost frequently concentrate water withdrawal and land use impacts in communities with limited political capacity to resist—a pattern documented in data center water stress analyses [19,48] and supply chain extraction literature [29,30]. The AI infrastructure expansion driving the rebound phase is not geographically neutral: it tends to concentrate in regions where regulatory capacity is weakest or land and water costs are lowest, often environmental justice communities [15]. This dynamic means that aggregate environmental assessments—including the framework’s geographic moderator analysis—can obscure distributional impacts that fall disproportionately on specific populations. Practitioners making deployment location decisions should supplement the framework’s efficiency-focused geographic analysis with environmental justice screening tools such as the U.S. EPA EJScreen or analogous regional instruments. Key references for governance and equity dimensions as they intersect with AI environmental outcomes include [13,16,21,33,71].

5.8. Worked Lifecycle Scoring Example: AI-Optimized HVAC in a Mid-Sized Commercial Building

To demonstrate how practitioners can apply the Environmental Asset–Cost Framework using currently available data, this section walks through a deployment decision for AI-optimized HVAC control in a 50,000 square-foot commercial office building. The example uses publicly available data throughout.

Deployment scenario: A building manager is considering deploying an AI-based HVAC optimization system (similar to the deep reinforcement learning controller studied in [6]) that is projected to reduce building HVAC energy consumption by 25%. The system requires dedicated edge computing hardware (two GPU-enabled servers) with an expected 4-year lifecycle before replacement.

Step 1 Asset Score: Quantify the optimization benefit. The building’s current HVAC energy consumption is 400 MWh/year (a typical figure for a building of this size in a temperate U.S. climate, per DOE Commercial Buildings Energy Consumption Survey). A 25% reduction yields 100 MWh/year in energy savings. This maps to Asset category A1 (Energy Optimization). The evidence basis for a 25% reduction is Moderate, consistent with the range documented across building HVAC studies in the corpus [5,6,9,58].

Step 2 Cost Score: Quantify the AI system’s own environmental burden.

• Energy cost (C1): Two GPU servers running inference continuously consume approximately 10–15 MWh/year (based on typical server power draw of 1.2–1.7 kW per unit at continuous load). Net energy balance: 100 MWh saved minus ~12 MWh consumed = 88 MWh net annual energy reduction.
• Carbon cost (C1): Depends on grid carbon intensity at deployment site. Using EPA eGRID data for two regional scenarios:

  ◦

  Scenario A (New England ISO, ~170 gCO₂e/kWh): Net 88 MWh × 0.170 = ~15 tCO₂/year net reduction. AI hardware consumption adds ~2 tCO₂/year. Net: ~13 tCO₂/year reduction [Evidence Level: Established].

  ◦

  Scenario B (SERC Midwest, ~610 gCO₂e/kWh): Gross savings 100 MWh × 0.610 = 61 tCO₂/year, hardware consumption adds ~7.3 tCO₂/year. Net: ~53.7 tCO₂/year reduction—substantially larger in absolute terms but the same AI system represents a higher emissions cost in this context [Evidence Level: Established].

• Water cost (C2): GPU server cooling adds minimal direct water use if air-cooled (~0 L of direct withdrawal). The electricity consumed (12 MWh/year) generates indirect water consumption through power generation: at the U.S. average of ~1.7 L/kWh [19], this adds approximately 20,400 L/year of indirect water use—an order of magnitude less than the direct cooling needs of a data center and unlikely to constitute a meaningful burden at this scale.
• E-waste cost (C3): Two GPU servers at ~15 kg each, replaced every 4 years = 7.5 kg e-waste/year amortized. At current recycling rates of ~20% [13,15], approximately 6 kg enters landfill annually. This is a minor burden at this deployment scale.
• Infrastructure cost (C4): Edge deployment within an existing building adds negligible new construction footprint. No meaningful C4 burden at this scale.
• Supply chain cost (C5): Two GPU servers require an estimated 40,000–60,000 L of ultrapure water in fabrication [19,21] and generate upstream supply chain emissions. Amortized over the 4-year lifecycle, this represents approximately 12,500 L/year in upstream water use—smaller than the direct water savings in many building contexts.

Step 3 Net Assessment by Scenario:

• Scenario A (Low-Carbon Grid—e.g., New England ISO, ~170 gCO₂e/kWh):

  ◦

  Net energy saved: 88 MWh/year

  ◦

  Net CO₂ impact: approximately 13 tCO₂/year reduction

  ◦

  Net water impact: minor indirect addition (~20,400 L/year through electricity generation)

  ◦

  E-waste: minor (~7.5 kg/year amortized)

  ◦

  Supply chain water: minor (~12,500 L/year amortized)

  ◦

  Recommendation: Deploy—net positive across all dimensions

• Scenario B (Mid-Carbon Grid—e.g., SERC Midwest, ~610 gCO₂e/kWh):

  ◦

  Net energy saved: 88 MWh/year

  ◦

  Net CO₂ impact: approximately 54 tCO₂/year reduction—larger in absolute terms, though the AI hardware’s own carbon cost is proportionally higher in this grid context

  ◦

  Net water, e-waste, and supply chain impacts: identical to Scenario A (minor)

  ◦

  Recommendation: Deploy—net positive across all dimensions in both scenarios; the mid-carbon grid actually produces a larger absolute carbon benefit from the same optimization, underscoring that building-scale HVAC AI is low-risk even outside optimal grid contexts

Step 4 Temporal and Phase Assessment: At this small deployment scale (2 servers), the Phase 2 rebound risk is low—infrastructure is not scaling, and the optimization benefit does not incentivize further AI expansion. This deployment is unlikely to trigger a meaningful rebound phase, making it suitable for organizations that cannot commit to multi-year mitigation strategies. [Evidence Level: Emerging—based on extrapolation from Ajagekar et al. [6] and the framework’s phase logic; direct longitudinal evidence for building-scale HVAC AI over 5+ years is limited in the corpus.]

Key takeaway for practitioners: This example illustrates that small-scale, edge-deployed AI optimization applications in existing buildings represent the framework’s lowest-risk deployment profile: meaningful efficiency gains, minimal infrastructure burden, no meaningful rebound risk, and positive net outcomes across both carbon and water dimensions in all but the highest-carbon grid contexts. The same analysis applied to a hyperscale data center training large language models would produce radically different results—which is precisely the framework’s purpose.

6. Conclusions

This research addresses AI’s sustainability paradox through the development of an Environmental Asset-Cost Framework that provides clean technology practitioners with systematic guidance for sustainable deployment. Our analysis of empirical studies from 2021 to 2025 reveals that AI’s environmental impact cannot be accurately characterized through simple binary assessments or point-in-time measurements. Instead, impacts follow predictable temporal dynamics and vary dramatically across geographical contexts, creating opportunities for strategic optimization.

6.1. Core Findings and Their Significance

The synthesis of the S-curve temporal heuristic represents a core conceptual contribution of this review. This pattern, characterized by initial optimization benefits, mid-term rebound effects, and conditionally projected long-term sustainability, emerged from convergent but methodologically diverse evidence strands spanning multiple AI applications, geographical regions, and temporal scales. It is presented as a synthesized decision heuristic rather than a universally validated empirical regularity; evidence strength varies by phase, with Phase 1 and Phase 2 supported by observed and econometric data, while Phase 3 rests primarily on modeled projections. Critically, the rebound phase is predictable, typically occurring 2–5 years post-deployment as infrastructure scaling outpaces efficiency improvements. This predictability enables proactive mitigation rather than reactive damage control.

Geographical context proves equally consequential. Our analysis documents 10–60x variation in environmental impacts based solely on deployment location, independent of AI system design or application. Regions combining high renewable electricity penetration (>70%), water abundance, and favorable cooling climates can achieve and maintain net environmental benefits throughout the deployment lifecycle. Conversely, regions with fossil fuel-heavy grids (>70%), severe water stress, and hot climates face amplified and extended rebound phases, with some scenarios never achieving net positive environmental outcomes even in long-term projections. This geographical dependency demonstrates that universal AI deployment recommendations are inappropriate—context-sensitive strategies are essential.

6.2. Actionable Pathways Forward

Our framework translates these findings into concrete decision criteria. Clean technology practitioners can assess deployment viability through three systematic evaluations: asset optimization potential (targeting applications with >20% efficiency improvement opportunities), cost burden assessment (prioritizing locations with <200 g CO₂/kWh grid intensity and water abundance), and temporal planning (incorporating explicit rebound mitigation strategies). These assessments enable practitioners to identify deployment contexts where AI genuinely supports clean technology objectives versus contexts where deployment would create net environmental harm.

The rebound effect, while challenging, is manageable through evidence-based strategies. Renewable energy coupling ensures AI infrastructure deployment coincides with new renewable capacity additions rather than consuming existing supply. Hardware lifecycle extension from current 2–3 year cycles to 4–6 years substantially reduces embodied emissions and e-waste generation. Alternative cooling technologies, including zero-water systems and closed-loop configurations, eliminate freshwater consumption in water-stressed regions. Energy consumption caps prevent infrastructure scaling from overwhelming efficiency gains during critical transition periods. Organizations implementing these strategies can navigate the rebound phase while maintaining progress toward net environmental benefits.

6.3. Policy and Standardization Imperatives

Achieving sustainable AI deployment at scale requires coordinated policy frameworks and measurement standardization. Location-based deployment incentives can channel infrastructure investment toward optimal regions (high renewables, water abundance) while discouraging expansion in challenging contexts (fossil-heavy grids, water stress). Renewable energy procurement mandates with verified additionality prevent AI expansion from triggering new fossil fuel generation. Water use regulations in stressed regions protect critical freshwater resources through technology requirements and withdrawal limits. Extended producer responsibility for e-waste creates economic incentives for design longevity and circular material flows.

Current environmental reporting approaches fail to capture temporal dynamics and geographical dependencies central to our framework. We propose the development of integrated metrics, including the AI Lifecycle Environmental Score (ALES), that accounts for energy, water, e-waste, and material impacts across full lifecycles, weighted by regional resource scarcity. Temporal impact disclosure requirements should mandate projection of S-curve dynamics rather than point-in-time snapshots, enabling stakeholders to evaluate long-term sustainability. Geographical context reporting standards must transparently disclose grid carbon intensity, water stress indices, and renewable energy procurement, allowing meaningful performance comparisons across deployments.

6.4. Future Research Directions

A boundary constraint of the current search strategy should be acknowledged: requiring environmental terms in both titles and keywords may have excluded studies with primarily technical AI framing but embedded environmental data (e.g., papers focused on algorithmic efficiency or hardware benchmarking that report energy figures as secondary outputs rather than primary contributions). Studies of this type are likely to report energy consumption data rather than broader multi-dimensional environmental impacts, and their primary contribution would be to the C1 (Energy Consumption) category. The likely direction of this bias is toward underrepresentation of energy efficiency gains in highly technical AI literature, which would if anything strengthen rather than weaken the Cost-side findings. Future reviews should extend to full-text searches or snowball sampling to capture indirectly relevant contributions.

While our framework synthesizes substantial empirical evidence, important knowledge gaps remain. Longitudinal studies tracking specific AI deployments through all three S-curve phases would validate temporal projections and refine mitigation strategy effectiveness. Economic analysis comparing mitigation strategy costs against environmental benefits would inform resource allocation decisions. Policy effectiveness evaluations assessing real-world outcomes of regulatory interventions would guide optimal policy design. Research examining AI’s environmental impacts in the Global South, currently underrepresented in the literature, would ensure equitable and globally applicable frameworks.

Evidence for Phase 3 (mature optimization, 5+ years) remains limited as most current AI deployments have not yet achieved this temporal stage. As deployments mature over the next 3–5 years, systematic monitoring and documentation will either validate or challenge our framework’s Phase 3 projections. This represents both a research opportunity and a practical imperative, as understanding long-term trajectories informs current deployment decisions with multi-year consequences.

6.5. Concluding Perspective

The path to sustainable AI is neither predetermined toward environmental catastrophe nor guaranteed toward technological salvation. Our framework demonstrates that outcomes depend fundamentally on strategic choices regarding deployment location, temporal planning, and mitigation implementation. The evidence clearly indicates that AI can serve as a powerful tool for clean technology transitions—but only when deployed with rigorous attention to geographical context, explicit planning for temporal dynamics, and systematic implementation of rebound mitigation strategies.

The critical window for establishing sustainable AI deployment practices remains open but is narrowing. Current infrastructure expansion trajectories, if continued without strategic intervention, risk locking in unsustainable patterns that persist for decades given infrastructure lifespans. However, the tools for sustainable deployment exist: our framework provides decision criteria, geographical guidance, and mitigation strategies grounded in empirical evidence rather than speculative projections.

Clean technology practitioners, policymakers, and AI developers share responsibility for translating these findings into practice. The question is not whether AI will play a role in clean technology transitions; that role is already established and expanding. Rather, the question is whether that role will be characterized by strategic sustainability or inadvertent environmental burden. Our framework provides the analytical foundation for choosing the former. The implementation choices made over the next 2–3 years, during the critical transition from Phase 1 optimization to Phase 2 rebound across many current deployments, will largely determine whether AI becomes a net contributor to or detractor from global sustainability objectives. We have the knowledge and tools to navigate this transition successfully. What remains is the collective commitment to do so.