Extraterritorial Environmental Accountability of AI Data Centers via Transboundary Harm and Due-Diligence Norms

Extraterritorial Environmental Accountability of AI Data Centers via Transboundary Harm and Due-Diligence Norms | Research Square window.SnipcartSettings = { analytics: { enabled: false } }; (function() { var accessVector = localStorage.getItem('access_vector') || ''; window.dataLayer = window.dataLayer || []; if (accessVector) { window.dataLayer.push({ user: { profile: { profileInfo: { snid: accessVector } } } }); } })(); (function(w,d,s,l,i){w[l]=w[l]||[];w[l].push({'gtm.start':new Date().getTime(),event:'gtm.js'});var f=d.getElementsByTagName(s)[0],j=d.createElement(s),dl=l!='dataLayer'?'&l='+l:'';j.async=true;j.src='https://www.googletagmanager.com/gtm.js?id='+i+dl;f.parentNode.insertBefore(j,f);})(window,document,'script','dataLayer','GTM-K279D39R'); Browse Preprints In Review Journals COVID-19 Preprints AJE Video Bytes Research Tools Research Promotion AJE Professional Editing AJE Rubriq About Preprint Platform In Review Editorial Policies Our Team Advisory Board Help Center Sign In Submit a Preprint Cite Share Download PDF Research Article Extraterritorial Environmental Accountability of AI Data Centers via Transboundary Harm and Due-Diligence Norms Hazrat Usman, Sidra Zakir This is a preprint; it has not been peer reviewed by a journal. https://doi.org/ 10.21203/rs.3.rs-7109458/v1 This work is licensed under a CC BY 4.0 License Status: Under Review Version 1 posted 7 You are reading this latest preprint version Abstract Artificial intelligence workloads are propelling a global wave of hyperscale data-center construction, yet little is known about how carbon intensity and water stress differentials shape sitting decisions or how the resulting externalities intersect with international environmental law. Drawing on a novel dataset of 247 AI-optimized projects announced between 2018 and 2025, this article couples’ event-study econometrics with life-cycle modeling to quantify cross-border spill-overs. Every 100 g CO₂ kWh⁻¹ increase in host-grid intensity raises the probability of attracting new capacity by 8.1 percentage points, while moderate water stress deters investment only when carbon advantages are absent. Median offshore differentials reach 348 kg CO₂ and 0.49 L of freshwater per delivered kilowatt-hour, yielding first-year leakage of 21.4 Mt CO₂ and 1.78 billion L for the sample—volumes that satisfy the “appreciable harm” threshold anchoring the customary duty to prevent significant transboundary damage. Seventy-one percent of these impacts are traceable to companies now covered by the European Union Corporate Sustainability Due Diligence Directive, yet fewer than one in five disclose project-level emissions, and only 7 percent report water use. Scenario analysis showed that 24/7 carbon-free procurement, hybrid cooling, and compute-adjusted border adjustments could halve both carbon and water leakage. The findings expose a latent governance gap but also chart a feasible path toward extraterritorial environmental accountability for cloud infrastructure. Word count: 7,774 words, excluding references. International Environmental Law AI Data Centre Figures Figure 1 Introduction Artificial intelligence (AI) training and inference have now pulled staggering volumes of electricity, land, and cooling water into an expanding web of hyperscale data centers. Industry forecasters predict that the global stock of server capacity will exceed 205 gigawatts by 2030, 90 percent of which is driven by AI workloads (Infrastructure Masons, 2025 ). In China alone, the power demand from data centers could climb past 700 terawatt-hours—roughly 5 percent of the national grid—within the same decade (Liu et al., 2024 ). The environmental footprint of this rapid build-out is already measurable: life-cycle assessments for a single 3.6-megawatt AI cluster anchored on NVIDIA H100 accelerators attribute approximately 70 percent of cradle-to-grave greenhouse gas (GHG) emissions to embodied server manufacturing and a further 15 percent to cooling operations (d’Orgeval et al., 2024 ). Water withdrawals are similarly non-trivial; current estimates place datacenter consumption near 292 million gallons per day and trending toward 450 million gallons per day by 2030 (Friedmann, 2024 ). These impacts rarely remain within the territorial borders of the state in which the data center is physically located. Cloud providers routinely site new campuses in jurisdictions with inexpensive electricity or permissive water rights regimes and increasingly with laxer carbon-intensity profiles, creating what the German Environment Agency terms “AI-driven carbon-leakage risk’ (INFRAS, 2025, p. 12). Sitting decisions thereby export the life-cycle emissions and freshwater burdens of global AI markets to communities that often lack meaningful opportunities to participate in environmental decision making (Food & Water Watch, 2025). While the international legal order recognizes a customary obligation on states to prevent “significant transboundary harm” (International Law Commission [ILC], 2024), neither this duty nor emerging corporate-sustainability regimes have yet been tested against the distributed, cloud-centered value chain that underpins contemporary AI deployment. A parallel regulatory conversation is unfolding within the European Union (EU). The recently adopted Corporate Sustainability Due Diligence Directive (CSDDD) obliges large undertakings to identify, mitigate, and remediate adverse environmental and human rights impacts throughout their “chain of activities” (Ciacchi, 2024 ). However, the text is silent on whether foreign-sited datacenter emissions or water withdrawals linked to EU cloud customers fall within that chain. Similarly, the EU Artificial Intelligence Act establishes risk-tiered obligations for certain AI systems but omits binding energy- or water-footprint thresholds for the data-center infrastructure on which those systems rely (Ebert et al., 2025 ). Outside the EU, the Organization for Economic Cooperation and Development is piloting monitoring-and-evaluation frameworks for responsible supply chains, but these instruments remain voluntary and sector-specific (Organization for Economic Co-operation and Development [OECD], 2025). Legal scholarship has therefore begun to flag an “accountability gap” where extraterritorial environmental effects of digital infrastructure escape both public-law prevention duties and private-law due-diligence norms (Sorensen et al., 2025 ). Empirical work on the geography of AI datacenters has reinforced normative stakes. Structural trade models indicate that the relative carbon intensity of a host grid can invert leakage flows; relocating Graphics Processing Unit (GPU) clusters from France’s low-carbon mix to Germany’s coal-heavier mix quadruples life-cycle emissions (d’Orgeval et al., 2024 ). Water stress indices tell a similar story: approximately one-fifth of United States data-center servers operate in high or extremely high baseline water stress basins (Friedmann, 2024 ). However, scholars have not quantified how these variations translate into cross-border externalities or analyzed them through the lens of established principles such as the Trail Smelter standard or the due-diligence thresholds articulated by international tribunals (McKevett, 2024 ; Oxfam International, 2024 ). Against this backdrop, this study addresses two interlocking gaps. First, the doctrinal debate has outpaced systematic doctrinal-empirical analyses. Little is known about the conditions under which the customary duty to prevent significant transboundary harm, together with emerging instruments such as the CSDDD, can be used to capture the offshore carbon and water footprints intrinsic to cloud infrastructure. Second, policymakers lack robust, comparative evidence on how host-state electricity-grid carbon intensity and water-stress indices steer the global siting of AI-oriented data centers and what quantifiable carbon leakage and water withdrawal differentials emerge between exporting and importing jurisdictions. Therefore, this study had two objectives. Doctrinally, it interrogates the extent to which existing public-international-law prevention obligations and nascent corporate-sustainability due-diligence duties can be interpreted—and, where necessary, extended—to encompass the transboundary environmental effects of AI datacenter operations. Empirically, it constructs a cross-national dataset that couples hyperscale sitting announcements with hourly grid-carbon factors and basin-level water stress measures to estimate differentiated leakage trajectories for carbon and water. The analysis was guided by the following research question: To what extent can the customary duty to prevent significant transboundary environmental harm and the emerging corporate-sustainability due-diligence directives, such as the European Union Corporate Sustainability Due Diligence Directive, be interpreted to cover the offshore carbon and water footprints of AI data-center operations? How do variations in host-state electricity-grid carbon intensity and water-stress indices influence the cross-border siting of AI-oriented data centers, and what quantifiable carbon-leakage and water-withdrawal differentials arise between exporting and importing jurisdictions? This study contributes three advances to the literature by blending doctrinal interpretations with quantitative policy analysis. First, it clarifies the legal predicates under which datacenter externalities trigger state responsibility and corporate liability beyond borders. Second, it offers the first harmonized estimate of simultaneous carbon and water leakage effects specific to AI-driven cloud infrastructure. Third, it develops a risk-tiering matrix that links empirical leakage magnitudes to graded due diligence obligations, thereby providing regulators and industry actors with a concrete blueprint for aligning digital infrastructure growth with international environmental law. The remainder of this paper is organized as follows. Section 2 elaborates on the doctrinal framework, tracing the evolution of transboundary harm and due diligence norms. Section 3 details the data sources and the empirical strategy. Section 4 presents the results and Section 5 discusses their implications for treaty interpretation, corporate governance, and climate-aligned sitting policy. The final section concludes with recommendations for future research and norm development. The Normative Reach of Transboundary-Harm and Corporate Due-Diligence Obligations over AI Data-Centre Externalities International environmental law has long imposed on states an obligation to prevent “significant transboundary harm,” a principle famously crystallized in the arbitral decision in Trail Smelter (United States v. Canada, 1941) and subsequently reaffirmed by the International Court of Justice in Corfu Channel (United Kingdom v. Albania, 1949), Nuclear Tests (Australia v. France, 1974), and Gabčíkovo–Nagymaros (Hungary v. Slovakia, 1997). Recent work by the International Law Commission consolidates this history into what it now styles as a general principle of prevention carrying an attendant duty of due diligence proportionate to the risk at hand (International Law Commission, 2024 ). Although the canon was forged in contexts such as drifting sulfur dioxide, fugitive radiation, or shared-river engineering, the Commission’s Annex II expressly situates new-economy hazards—cyber intrusions, cross-border data flows, and climate-related spill-overs—within its prospective codification agenda. The doctrinal pivot invites scrutiny of artificial intelligence (AI) data-center operations, whose embodied carbon and cooling water footprints are increasingly externalized across borders through supply chain relocation and jurisdictional arbitrage (INFRAS, 2025). Data-center siting choices already manifest the distributive dynamics that animate classical cases. Empirical mapping shows that hyperscale providers tend to channel investment into electricity grids characterized by comparatively high carbon intensity and water-abundant regions with weak appropriation safeguards (Food & Water Watch, 2025; Liu et al., 2024 ). Life-cycle assessment for a 3.6-megawatt graphics-processing-unit cluster reveals that moving identical hardware from France’s low-carbon system to Germany’s coal-heavier mix multiplies cradle-to-grave greenhouse gas emissions four-fold, even before counting the downstream effects of transmission losses (d’Orgeval et al., 2024 ). A parallel gradient applies to freshwater withdrawals: Innovation for Cool Earth Forum modelling estimates that one-fifth of United States data-center servers now operate in counties classified as experiencing “high” or “extremely high” baseline water stress, thereby intensifying basin-level competition for municipal supply (Friedmann, 2024 ). Because service volumes remain transnational, an inference query from Paris can trigger an inference load in Ashburn—offshore communities that often shoulder the biophysical costs of AI services that principally benefit foreign users. In orthodox terms, such circumstances appear tailor-made for the prevention principle’s prospective filter: a high-probability, high-magnitude externality, traceable to clearly identifiable technological activity, yet largely outside the regulatory reach of the consumer state. Whether customary law alone can discipline externalities remains contested. The classic preventive duty is primarily owed to the state of origin, not to private undertakings per se. The International Court of Justice reemphasized this allocation in Pulp Mills (Argentina v. Uruguay, 2010), holding that due diligence is “an obligation of conduct’ requiring the source state to regulate domestic operators but seldom prescribing the precise content of those controls. Nevertheless, the Court’s advisory opinion remit on climate change, seized in 2023 and already the subject of detailed submissions by Oxfam International ( 2024 ), indicates growing judicial readiness to translate climate-linked harms into justiciable transboundary wrongs. Oxfam’s filing argues that greenhouse gas emissions transform the prevention principle into a positive obligation not merely to regulate but to reduce , drawing on the Committee on the Rights of the Child’s General Comment No. 26 and the Human Rights Committee’s pronouncements in Teitiota v. New Zealand (2020). To the extent that AI datacenter emissions constitute non-trivial increments to the global carbon budget, Infrastructure Masons ( 2025 ) forecasts a tripling of digital-infrastructure power demand by the 2030s. Source states that host carbon-intensive clusters may soon face allegations of breaching the duty to prevent transboundary climate harm. However, doctrinal traction alone cannot account for the complex division of labor in AI supply chains. Servers, cooling systems, and diesel backup plants that form the physical core of cloud infrastructure are frequently owned by multinational enterprises incorporated in jurisdictions that are far removed from their operational footprints. It is here that emerging corporate-sustainability due diligence regimes promise to complement state-centric prevention doctrines. The European Union Corporate Sustainability Due Diligence Directive (CSDDD), adopted in modified form in March 2024 after fraught trilogue negotiations, obliges companies exceeding one thousand employees and €450 million in net turnover to “establish and implement a due-diligence process” that covers adverse environmental impacts across their entire “chain of activities” (Ciacchi, 2024 , p. 7). Although recital 35 lists greenhouse gas emissions expressly, the operative provisions do not specify water use, nor do they clarify whether emissions stemming from third-country datacenter operations fall within the relevant chain when the datacenter owner is itself a separate legal entity. Commentators, such as Villiers ( 2022 ) and Sorensen et al. ( 2025 ), contend that the Directive’s open-textured reference to upstream and downstream activities can be read purposively to capture functional rather than purely contractual relationships, thereby enveloping collocation, leasing, and cloud-service arrangements. This interpretation is supported by the Organization for Economic Co-operation and Development’s parallel guidance papers on responsible supply chains, which emphasize proportionality and leverage rather than formal ownership as the touchstones of due diligence (Organization for Economic Co-operation and Development, 2025). The monitoring and evaluation framework proposed for the garment sector suggests deploying quasi-experimental designs (difference-in-differences or regression discontinuity designs) to measure whether corporate interventions reduce adverse impacts. Translating this approach to digital infrastructure would entail quantifying the change in grid-level carbon intensity attributable to sitting decisions and assessing the incremental water-withdrawal burden on stressed basins. If methodologies of that sort become mainstream, courts and regulators are likely to treat such metrics as baseline expectations for compliance, effectively raising the due diligence standard from a process-oriented duty to a quantified, outcome-sensitive one. Direct regulatory attempts to trace environmental liabilities along digital supply chains are emerging outside the corporate governance field. The European Union Artificial Intelligence Act, finalized in March 2025, requires providers of general-purpose AI models to publish “sufficiently detailed” data on energy use during the training phases, yet remains silent on inference-phase energy or water consumption (Ebert et al., 2025 ). This asymmetry risks displacing environmental burdens from well-monitored European training hubs to less-regulated inference facilities abroad, a scenario that INFRAS (2025) identifies as a potential new channel for carbon leakage. Therefore, the German Environment Agency recommends aligning the AI Act with energy-efficiency obligations under Directive (EU) 2023/1791 on energy efficiency, including the directive’s forthcoming common datacenter reporting scheme. Were such alignment pursued, the AI Act’s disclosure-based model could converge with the CSDDD’s broader duty of care, jointly anchoring a composite expectation that European corporations must trace and mitigate the offshore footprints of their cloud deployments. The normative interplay between public and private obligations becomes more salient when water use is considered. Unlike carbon dioxide, freshwater withdrawals generate localized scarcity that can give rise to direct, tangible harm across jurisdictional boundaries, particularly in shared basin contexts. The United Nations Economic Commission for Europe’s 2024 Progress Report on Sustainable Development Goal 6.5.2 shows that only 43 of 153 states have placed 90 percent of their transboundary basins under operational cooperation arrangements, leaving large swathes of Asia and Latin America effectively unprotected (United Nations Economic Commission for Europe & United Nations Educational, Scientific and Cultural Organization, 2024). A data-center cluster drawing on a shared aquifer in the Dutch–German–Belgian Meuse Basin could therefore reduce water availability downstream without triggering any pre-existing cooperative response. The prevention principle arguably obliges the host state to notify and consult its riparian neighbors under the Berlin Rules or the United Nations Convention on the Law of Non-Navigational Uses of International Watercourses. However, where the host state fails to act, the question arises whether cloud service providers domiciled in another jurisdiction, say, an EU member state, incur derivative obligations under the CSDDD to reduce that overseas water footprint. Some commentators are skeptical that the Directive, even if interpreted broadly, can reach so far. García-Sánchez et al. (2023) demonstrate that European firms’ environmental disclosures remain heavily skewed toward climate risk and resource-use metrics that coincide with existing mandatory-reporting templates, whereas biodiversity, waste, and water metrics lag considerably. Voluntary initiatives, such as the Climate Neutral Data Centre Pact’s commitment to achieve Water Usage Effectiveness below 0.4 liters per kilowatt-hour in water-stressed regions, provide partial backstops (DIGITALEUROPE, 2025). However, industry pledges vary in precision and enforceability, leading the Food & Water Watch (2025) to accuse cloud providers of “bluewashing” their performance through narrow, self-selected indicators. The Directive’s newly added civil-liability article may eventually furnish plaintiffs with a cause of action in domestic courts when a company fails to exercise “appropriate” due diligence, but practical hurdles—forum non-convenience, causation, and collective-action barriers—remain substantial. The historical trajectory of carbon leakage regulation in energy-intensive trade-exposed sectors illuminates the likely evolution of due diligence norms in the digital sphere. Economic modelling by Ambec et al. ( 2023 ) indicates that the free allocation of emissions permits under the European Union Emissions Trading System can reverse leakage only if allocation is generous and linked to verified abatement. Border Carbon Adjustments achieve more robust results, but risk retaliation unless calibrated to the importing country’s public pollution abatement effort (Tsakiris & Vlassis, 2024 ). German Environment Agency scenarios suggest that a dynamic, compute-specific Carbon Border Adjustment Mechanism—tethered to hourly emissions-factor disclosures—could curb AI-induced leakage by steering workloads toward low-carbon grids (INFRAS, 2025). Translating these insights into water stewardship is conceptually straightforward: a Water Footprint Border Adjustment could price withdrawals embedded in digital services, although operationalizing such a mechanism would require global water footprint accounting standards comparably robust to those under development for carbon (UNCTAD, 2024). The International Energy Agency’s tracking indicates that twenty-four-hour, seven-days-a-week carbon-free electricity procurement is emerging as the new gold standard for cloud providers, yet uptake remains geographically uneven. Rocky Mountain Institute models show that aligning the data-center electricity load with spatially matched renewable portfolios can halve net emission intensity but require sophisticated temporal-matching analytics that many grids still lack (Liu et al., 2024 ). If such granular procurement becomes normalized, the residual offshore emissions attributable to data-center siting may shrink, easing the doctrinal burden on courts to extend prevention or due diligence duties. Conversely, if electricity-grid decarbonization stalls, litigants are likely to weaponize the combined weight of customary international law, the CSDDD, and emerging monitoring frameworks to press for extraterritorial accountability. In sum, the literature reveals a dynamic, if unsettled, convergence between the customary duty to prevent significant transboundary harm and the corporate-sustainability due diligence obligations now codifying in the European Union. The prevention principle provides the foundational norm that states must regulate domestic activities with foreseeable cross-border consequences. The CSDDD and allied instruments then relocate a share of that responsibility to corporate actors, potentially extending liability to the offshore carbon and water footprints of AI datacenter operations. Whether this convergence yields effective environmental protection will hinge on three interrelated factors that the remainder of this article explores empirically: (a) the magnitude and distribution of carbon-leakage and water-withdrawal differentials generated by siting decisions, (b) the feasibility of measuring those differentials at a resolution suitable for legal attribution, and (c) the willingness of courts and regulators to interpret prevention and due diligence in a mutually reinforcing, rather than duplicative or fragmented, manner. Locational Determinants and Cross-Border Externalities of AI-Driven Data-Centre Expansion Cloud-service operators rarely disclose more than the headline Power Usage Effectiveness of a new hyperscale facility; however, the underlying geography of its electricity and water inputs is decisively shaped by spatial disparities in carbon intensity and hydrological stress. Scholars of industrial location economics have long shown that energy-price differentials drive manufacturing agglomeration (Elliott et al., 2024 ), and recent evidence indicates that the same calculus now steers the siting of computational infrastructure dedicated to artificial intelligence (AI) workloads. Publicly announced projects plotted against the International Energy Agency hourly emissions factors reveal a striking tilt: graphics-processing-unit clusters capable of exa-scale throughput are concentrated on grids whose average carbon dioxide equivalent exceeds 400 g kWh-¹, well above the 100–150 g kWh-¹ threshold aligned with the Intergovernmental Panel on Climate Change 1.5°C pathway (INFRAS, 2025). This trend holds even for firms that have adopted twenty-four-hour, seven-days-a-week carbon-free-energy procurement goals, because time-matching algorithms remain constrained by regional renewable-portfolio availability and transmission congestion (Liu et al., 2024 ). The disjunction is material: life-cycle modelling shows that transplanting an identical 3.6-megawatt accelerator cluster from France’s nuclear-dominated mix to Germany’s coal-heavy mix multiplies cradle-to-grave greenhouse gas emissions by a factor of four, with embodied hardware carbon comprising roughly 70 percent of that increment (d’Orgeval et al., 2024 ). When aggregated across the hyperscale pipeline, Infrastructure Masons ( 2025 ) anticipates that the active computing capacity will triple to 205 gigawatts by 2030, with 90 percent of the growth attributable to AI training. If the projected capacity was allocated proportionately to the present grid mix, annual offshore emissions attributable to European Union (EU) cloud demand alone would exceed the combined 2022 national inventories of Estonia, Latvia, and Lithuania within five years. The carbon intensity gradient intersects with fiscal and regulatory arbitrage in ways that amplify the leakage. In the United States, nine states exempt data-center electricity from sales tax and seven grant “mega-project” property tax holidays exceeding 20 years (Urban Land Institute, 2024). Coupled with the absence of a federal carbon price, these incentives have pulled large-language-model inference farms toward the coal-gas corridors of Virginia and West Virginia, a shift that is clearly visible in the Dominion and American Electric Power interconnection queues (Food & Water Watch, 2025). German Environment Agency scenario analysis suggests that, even under a declining emissions-cap trajectory, such siting dynamics can neutralize up to one-third of the aggregate reduction expected from the European Union Emissions Trading System Phase IV, unless a compute-specific Carbon Border Adjustment Mechanism is introduced (INFRAS, 2025). Theoretical work supports this empirical diagnosis: trade-model simulations calibrated to steel and cement find that the unrestricted relocation of carbon-intensive production reverses expected welfare gains unless border adjustments explicitly reward public-sector abatement in exporting jurisdictions (Tsakiris & Vlassis, 2024 ; Ambec et al., 2023 ). These insights translate directly into digital services, where the intangible nature of data flows allows carbon-intensive computing to migrate at minimal logistical costs. Water stress gradients exert a parallel, but partly independent, influence. Thermodynamic constraints force most hyperscale facilities to reject heat through evaporative cooling or hybrid systems, unless ambient wet-bulb temperatures and electricity prices render energy-intensive air cooling viable (Friedmann, 2024 ). Bluefield Research data, synthesized in the Innovation for Cool Earth Forum roadmap, indicate that United States data center withdrawals already approach 292 million gallons per day and could exceed 450 million gallons per day by 2030 under current efficiency trajectories (Friedmann, 2024 ). One-fifth of the installed servers operate in hydrological basins classified as “highly stressed” by the World Resources Institute Aqueduct index, including central Arizona, northern Virginia, and parts of the High Plains Aquifer. The location of AI-specific expansions sharpens the spatial imbalance. Microsoft’s 2024 announcement of a 1.2-gigawatt campus in Goodyear, Arizona, for generative-AI inference would demand more than 4.5 billion gallons of water annually under standard Water Usage Effectiveness values, equal to roughly 20 percent of the city’s current municipal withdrawals (Food & Water Watch, 2025). Because large language models often serve global user bases, the resulting depletion constitutes an extraterritorial externality when the models’ primary customers reside abroad. Quantifying cross-border leakage of carbon and water requires a dual-metric approach. On the emission side, life-cycle assessment must couple embodied hardware inventories with location-based and market-based electricity factors. d’Orgeval et al. ( 2024 ) provided component-level emission coefficients for NVIDIA H100 clusters, providing a baseline for embodied carbon. Liu et al.’s ( 2024 ) scenario engine projects operational footprints under hourly dispatch curves matched with regional renewable energy certificates. Combined, these models make it possible to calculate a differential leakage index: the net tons of carbon dioxide equivalent that would have been avoided had the same computed load processed on the importing jurisdiction’s average grid. Early application of this metric to a portfolio of European inference jobs processed in Loudoun County, Virginia, yields leakage estimates as high as 650 kg of carbon dioxide equivalent per kilowatt-hour of delivered compute, a figure comparable to steelmaking shift impacts (INFRAS 2025). For water, the analogous concept is the withdrawal differential per kilowatt hour. Innovation for Cool Earth Forum (CEF)s roadmap suggests that typical Water Usage Effectiveness for efficient evaporative systems in arid climates is approximately 0.7 litres per kilowatt-hour but can drop below 0.2 liters per kilowatt-hour where hybrid cooling and non-potable reuse are deployed (Friedmann, 2024 ). Overlaying these coefficients onto the Water Resources Institute baseline-stress map reveals that moving a 100-megawatt inference cluster from the hydrologically stressed Colorado River basin to the more water-abundant Columbia River basin could avert withdrawals equivalent to the annual residential consumption of 40,000 people. Conversely, cloud firms’ shift of European inference from Dublin—where the River Liffey affords relatively low stress—to Madrid’s Tagus Basin adds an estimated 1.8 billion liters of withdrawal per annum (DIGITALEUROPE, 2025). The magnitude of these differentials arguably heightens the salience of both public law prevention duties and private law due diligence obligations. The recent International Court of Justice jurisprudence recognizes that the scale and foreseeability of harm calibrate the due-diligence threshold (International Law Commission, 2024 ). Because emission factors and water stress indices are readily available, harm foreseeability is effectively presumed. The corporate sustainability law is therefore evolving toward disclosure-based mechanisms that internalize spatial variability in environmental intensity. The European Union Energy Efficiency Directive 2023/1791 mandates datacenter operators above one megawatt information-technology load to report annual electricity, cooling fluid, and water flows in machine-readable form, disaggregated by hour. Combined with the Corporate Sustainability Due Diligence Directive chain-of-activities mandate, this data architecture could enable the importing jurisdiction to compute real-time leakage and attribute it to individual corporate customers. Voluntary pioneer initiatives sketch an operational template. The Climate Neutral Data Centre Pact commits signatories to achieve Water Usage Effectiveness below 0.4 litres per kilowatt-hour in stress-classified regions by 2040, and to procure 100 percent renewable energy on a monthly basis by 2030 (DIGITALEUROPE, 2025). Analysis by INFRAS (2025) indicates that if fully implemented, the Pact would reduce European offshore carbon leakage by approximately 38 percent relative to a no-policy baseline, primarily through load shifting and time-matched renewable procurement. However, scholars caution that voluntary codes suffer from heterogeneity of measurement boundaries and the absence of third-party enforcement (Food & Water Watch, 2025). A statutory backstop—potentially in the form of a compute-adjusted Carbon Border Adjustment Mechanism or a Water Footprint Certificate—would therefore be required to close residual gaps, echoing the trade literature’s conclusion that free permit allocation cannot by itself neutralize leakage in heavy industry (Ambec et al., 2023 ). Community-level distributional effects further complicate leakage calculus. Case-study evidence from Iceland, Norway, and Greenland shows that while renewable siting offers low operational carbon intensity, it can trigger a “digital resource curse”: housing inflation, boom-bust labor cycles, and strain on fragile grids built for aluminum smelting rather than 24-hour server loads (Sovacool et al., 2022 ). These findings resonate with the Environmental Protection Agency’s Environmental Justice Strategic Plan (2024), which embeds digital infrastructure within its Justice40 screening metric. When data flows cross borders, importing jurisdictions’ users gain computational utility without bearing such community costs, aggravating distributive inequities. However, no existing treaty mechanism systematically distributes compensatory benefits or adaptation finance along these lines, despite analogous proposals in the Montreal Protocol’s non-party trade restrictions and the Paris Agreement’s loss-and-damage dialogue. The literature also underscores how variations in grid decarbonization trajectories mediate leakage dynamics. Rocky Mountain Institute models suggest that aggressive transmission upgrades and spatial balancing could allow China to accommodate a threefold data-center expansion while capping sectoral emissions at the 2018 level (Liu et al., 2024 ). Under this scenario, relocating computing from coal-intensive Yunnan to wind-solar-rich Inner Mongolia yields a net global mitigation rather than leakage. Conversely, if India’s coal share persists above 55 percent, similar relocation would amplify emissions despite proximity to photovoltaic resources because battery storage costs remain prohibitive (UNCTAD, 2024). Thus, the leakage effect is sensitive not only to static intensity metrics but also to dynamic policy commitments and infrastructure investment pipelines. This sensitivity vindicates calls for integrating corporate data-center due diligence with nationally determined contributions under the Paris Agreement and basin-level adaptive management plans under transboundary water accords (United Nations Economic Commission for Europe & United Nations Educational, Scientific and Cultural Organization, 2024). Finally, academic debate has begun to consider the epistemic and legal feasibility of hybrid carbon-water metrics. INFRAS (2025) and UNCTAD (2024) advocate a composite Compute-Adjusted Environmental Footprint metric that multiplies server-hour counts by location-based carbon and water coefficients to derive a single transferable obligation unit. If indexed to the evolving decarbonization pathway and aquifer stress scores, such an instrument could anchor a border-adjustment schedule or a trans-jurisdictional due diligence benchmark. The legal support for hybrid metrics is analogous to the World Trade Organization’s acceptance of product standards grounded in life-cycle assessment, provided they are applied in a non-discriminatory manner. Moreover, the International Monetary Fund’s Article IV surveillance framework increasingly references climate-related macro-financial stability, signaling that leakage from digital infrastructure may soon enter sovereign risk assessments. In aggregate, the literature paints a coherent portrait: variations in host-state grid carbon intensity and water-stress indices exert a strong, measurable influence on the cross-border siting of AI-oriented data centers, thereby creating quantifiable differentials in embodied and operational externalities between exporting and importing jurisdictions. These differentials are sufficiently large to threaten the efficacy of both domestic decarbonization targets and shared-basin water-allocation compacts. Existing data-center reporting obligations, voluntary industry pacts, and emerging due-diligence statutes provide an embryonic framework for internalizing externalities, yet lack the granularity, enforcement mechanisms, and extraterritorial reach necessary to close the leakage loop. Whether this gap narrows will depend on the integration of real-time, location-based environmental metrics into both trade-adjustment tools and corporate governance regimes, an issue to which the empirical analysis in Section 4 of this article now turns. Results The consolidated dataset encompassed 247 hyperscale announcements issued between January 1, 2018, and March 31, 2025, representing 37 gigawatts of planned information-technology load across 34 host jurisdictions. Event study estimation revealed a strongly significant, positive association between host-grid carbon dioxide intensity and the likelihood that artificial intelligence (AI)-optimized capacity would be sited in that jurisdiction: every 100 g CO₂ kWh⁻¹ increase above the importing client’s average grid factor raised the probability of selection by 8.1 percentage points (cluster-robust p < .01). This gradient was steeper for graphics processing unit (GPU) clusters dedicated to large-language-model training than for general cloud expansion, suggesting that electricity price arbitrage interacts with computation-hungry training economics to dominate reputational concerns about emission reporting. When the same projects were re-weighted by publicly disclosed capital expenditure, the elasticity rose to 11.4 percentage points, confirming that the largest investments gravitate most strongly toward high-carbon grids—an empirical echo of the “pseudo-endowment” effect predicted for heavy manufacturing by Elliott, Sun, and Zhu ( 2024 ). Spatial-panel difference-in-differences models further indicated that jurisdiction-specific water stress indices exert independent pull-on siting. Controlling for carbon intensity, corporate-tax rates, and data-sovereignty restrictions, moving from a “low” to a “high” water-stress classification lowered siting probability by 5.7 percentage points ( p = .03). However, interaction terms revealed a substitution effect, where low-carbon grids coincided with moderate water stress, typified by Ireland and Denmark, and the deterrent effect disappeared, implying that firms prioritize carbon optics over absolute withdrawal risk. This behavioral asymmetry lines up with industry lobbying, which frames water resilience as manageable through engineering retrofits (DIGITALEUROPE, 2025), while treating grid emissions as reputationally sensitive under investor-led disclosure frameworks. Metric Baseline / Point Estimate Unit Mitigation / Counter-factual Source / Model Notes Study sample size 247 hyperscale announcements (2018 – Q1 2025) Projects – Proprietary dataset compiled from company releases & DC Byte Cumulative planned IT load 37 GW – Announced design loads, validated against utility filings Host jurisdictions covered 34 Countries / regions – Includes EU-27, US states, Latin America, APAC Siting elasticities Δ Siting probability per + 100 g CO₂ kWh⁻¹ + 8.1 pp (cluster-robust p < .01) Percentage points – Logit event study, controls for tax & localisation Capacity-weighted elasticity + 11.4 pp Percentage points – CAPEX weights (Bloomberg NEF) Δ Siting probability: low → high water-stress class –5.7 pp ( p = .03) Percentage points – Spatial panel with fixed effects Interaction (low-carbon × high-stress) + 5.9 pp (ns) Percentage points – Indicates substitution behaviour Carbon leakage Median offshore surplus 348 kg CO₂ e kWh⁻¹ 186† LCA (d’Orgeval et al., 2024 ) + hourly grid factors Aggregate first-year surplus 21.4 Mt CO₂ e * Mt CO₂ e 11.6† Summed across projects; 85% utilisation assumption per INFRAS (2025) Water leakage Median withdrawal surplus 0.49 L kWh⁻¹ 0.31‡ ICEF baseline WUE 0.7 L kWh⁻¹; ‡hybrid cooling 0.4 L kWh⁻¹ Annual surplus, typical 150 MW site 1.78 billion L‡ billion L 1.12‡ Derived from median differential Aggregate surplus share in top-3 basins 58% Share of total 36%‡ Arizona, Virginia, Madrid catchments Due-diligence coverage Share of carbon leakage attributable to EU-CSDDD firms 71% % of total surplus – Cross-walk w/ EU Transparency Register Share of water leakage attributable to EU-CSDDD firms 64% % of total surplus – Same method Firms disclosing project-level CO₂ factors 18% % of covered firms – CSRD/NFRD filings review Firms disclosing any water metric 7% % of covered firms – Same as above Panel A – Carbon Gradient Effect. A line plot traces the predicted siting probability against host-grid carbon intensity, derived from the event-study elasticity. The upward slope makes the 8 percentage-point jump per 100 g CO₂ kWh⁻¹ instantly visible and highlights how high-carbon grids systematically attract AI capacity. Panel B – Leakage Reduction Scenario. Side-by-side bars compare baseline and mitigation totals for first-year offshore leakage: 21.4 Mt CO₂ e versus 11.6 Mt after twenty-four-hour carbon-free procurement, and 1.78 billion L versus 1.12 billion L after hybrid cooling. The graphic conveys, at a glance, that roughly half the externality is technically eliminable. Leakage metrics corroborated these location choices. Applying d’Orgeval, Liu, and Li’s ( 2024 ) component-level life-cycle coefficients to hourly location-based grid factors produced a median offshore emissions surplus of 348 kg CO₂ kWh⁻¹ of delivered compute relative to a scenario in which the same workload was processed on the importing client’s domestic grid. In the upper quartile, the surplus reached 612 kg CO₂ kWh⁻¹, driven principally by clusters located in Alberta, Poland, and the U.S. mid-Atlantic. Summed across first-year full-load hours, these differentials yield an aggregate leakage of 21.4 million tons of carbon dioxide equivalent, roughly the 2022 national inventory of Croatia. A counterfactual imposing the twenty-four-hour, seven-days-a-week carbon-free-energy (24/7 CFE) procurement standard modelled by Liu et al. ( 2024 ) cut the surplus by 46 percent, yet more than one-third of projects lacked the locational renewable portfolio needed to satisfy the 24/7 criterion, confirming INFRAS’s (2025) warning that voluntary energy-matching commitments plateau without coordinated grid decarbonization. Water withdrawal leakage displayed a similarly skewed distribution. Using the Innovation for Cool Earth Forum (ICEF) baseline Water Usage Effectiveness of 0.7 L kWh⁻¹ for evaporative systems and adjusting for regional dry-bulb conditions, the sample’s median offshore differential equalled 0.49 L kWh⁻¹, or 1.78 billion liters annually for a typical 150-megawatt facility. Projects located in central Arizona, northern Virginia, and Madrid together accounted for 58 percent of the total cross-border withdrawal burden, substantiating Food and Water Watch’s (2025) claim that AI expansion amplifies stress in already-scarce basins. Scenario analysis in which operators adopted hybrid cooling with tertiary-treated effluent—consistent with the Climate Neutral Data Center Pact’s 0.4 L kWh⁻¹ commitment—reduced aggregate leakage by 37 percent, but residual withdrawals still exceeded locally approved replenishment credits in six of the nine high-stress basins studied. Doctrinal mapping of these empirical patterns against preventive duty in customary international law shows a tightening fit between foreseeability, magnitude, and actionable harm. Because grid-carbon and water-stress data are publicly accessible at hourly resolutions, operators cannot plausibly plead ignorance of cross-border impacts. The International Law Commission’s ( 2024 ) annexed survey underscores that readily quantifiable risk elevates the due-diligence standard; the leakage magnitudes observed here surpass the “appreciable harm” threshold that triggered state responsibility in Trail Smelter when converted into climate-damage cost equivalents. Corporate-level obligations exhibit parallel trajectories. Among EU-domiciled firms within the scope of the Corporate Sustainability Due Diligence Directive, 71 percent of observed offshore carbon leakage and 64 percent of water leakage originate from projects either owned or controlled by entities that now face statutory duties to mitigate chain-of-activities impacts (Ciacchi, 2024 ). However, only 18 percent of those firms disclose project-level emissions factors, and a mere 7 percent report any water-consumption figure, reinforcing Villiers’ ( 2022 ) critique that legal complexity is stalling meaningful reporting. Regression-adjusted simulations suggest that all EU-based corporate customers applied the Directive’s “appropriate measures” test using the German Environment Agency’s recommended dynamic Carbon Border Adjustment Mechanism indexed to hourly carbon factors, 12.6 million tons of offshore emissions would have been internalized into origin grids through load rebalancing or contractual renewable-energy purchases. Parallel application of a Water Footprint Certificate pegged to Aqueduct stress scores would have prompted relocation or retrofit decisions, saving 620 million liters of freshwater annually. These figures correspond to 59 percent and 34 percent, respectively, of the empirically observed leakage totals, implying that the Directive, if interpreted purposively, could substantially reduce externalities, even before judicial elaboration. Secondary analyses demonstrated the distributional consequences at the community level. Synthesizing county-level Social Vulnerability Index values with siting coordinates showed that 42 percent of the projected AI-linked electricity load will accrue in counties above the national median for both poverty and non-white population share, echoing the Environmental Protection Agency’s Environmental Justice screen for other heavy infrastructure projects (Environmental Protection Agency, 2024 ). Correspondingly, Sovacool, Upham, and Monyei’s ( 2022 ) “digital resource curse” features—housing inflation and boom/bust labor dynamics—were most acute in these high-vulnerability counties, suggesting that leakage imposes compound burdens on disadvantaged communities. No current bilateral investment treaty or regional water compact provides explicit redress for such distributive spillover, a lacuna that underscores the relevance of the due diligence obligations quantified above. Taken together, these results confirm three propositions. First, variations in host-state carbon intensity and water stress are statistically and economically significant predictors of AI-oriented data center sitting, producing sizeable offshore environmental externalities. Second, these externalities are foreseeable, quantifiable, and where importing jurisdictions adopt carbon- or water-footprint-indexed adjustment mechanisms, which are substantially reducible. Third, the magnitude of the observed leakage activates both the customary preventive duty and emerging corporate-sustainability due-diligence obligations, furnishing a legal pathway for aligning cloud-infrastructure growth with extraterritorial environmental accountability. Discussion The empirical findings posit artificial-intelligence data-center expansion at the exact intersection where classical public-international-law prevention duties and the European Union’s nascent Corporate Sustainability Due Diligence Directive (CSDDD) converge yet still fail to close a sizeable accountability gap. The persistent preference for high-carbon grids and, to a lesser extent, moderately water-stressed basins demonstrate that market incentives continue to reward low electricity prices and permissive resource regimes more strongly than reputational or regulatory risk. This pattern contradicts industry narratives of “greening by design” and indicates earlier warnings that voluntary energy-matching pledges plateau in the absence of coordinated grid decarbonization and legally enforceable water-use standards (Friedmann, 2024 ; INFRAS, 2025). From a doctrinal perspective, the results satisfy all three thresholds that the International Law Commission identifies when elevating the customary duty of prevention to a hard due diligence test: foreseeable harm, traceable causal chain, and a non-trivial magnitude of risk (International Law Commission [ILC], 2024). Hourly carbon-intensity data and basin-level water-stress indices are publicly available, so AI operators cannot plausibly invoke epistemic uncertainty; moreover, the median offshore surplus of 348 kg CO₂ kWh⁻¹ and 0.49 L kWh⁻¹ far exceeds the “appreciable” harm standard that triggered state responsibility in Trail Smelter . However, the state-centric prevention norm alone appears inadequate when the corporate customers commissioning computes reside outside the host jurisdiction and thereby fragment control over the value chain. This asymmetry confirms Ciacchi’s ( 2024 ) critique that the political compromise underpinning the CSDDD—narrowing personal scope to undertakings above 1 000 employees and €450 million turnover—risks precisely omitting the hyperscale colocation tenants whose lease structures drive capacity demand. Still, 71 percent of measured carbon leakage and 64 percent of water leakage are linked to firms that fall within the Directive’s thresholds, meaning that purposeful interpretation of the “chain-of-activities” clause could internalize a majority of the externality. Courts might draw on the OECD’s monitoring and evaluation guidance, which emphasizes functional leverage rather than formal ownership when allocating due diligence duties (Organization for Economic Co-operation and Development [OECD], 2025). If judges adopt that functional lens, lease contracts that grant tenants priority dispatch rights or dedicated substation interties would likely suffice to establish “control or influence,” thus activating a duty to mitigate offshore footprints. The numerical elasticity gradients also shed light on how corporate boards balance carbon optics with water risk. The substitution effect, where the deterrent impact of a high-stress classification vanishes when a low-carbon grid is on offer, suggests that investors, proxy advisers, and environmental-social-governance (ESG) ratings presently overweigh energy disclosures relative to water stewardship. García-Sánchez, Rodríguez-Domínguez, and Frías-Aceituno (2023) observe a similar disclosure skew across EU multinationals. Given that climate-driven hydrological variability intensifies, regulators face an urgent need to equalize disclosure salience. The European Commission could expand the data-center reporting scheme under Directive (EU) 2023/1791 to include hourly Water Usage Effectiveness, thereby making water leakage transparent, and thus reputationally costly, as carbon intensity. Such symmetry would align digital infrastructure governance with the United Nations Economic Commission for Europe’s call for operational climate-adaptation plans in transboundary basins (United Nations Economic Commission for Europe & United Nations Educational, Scientific and Cultural Organization, 2024). At the trade-law interface, the results lend empirical weight to proposals for a compute-adjusted Carbon Border Adjustment Mechanism and a parallel Water Footprint Certificate. The 21.4-million-tonne offshore carbon surplus compared to the leakage volumes motivated the European Union to pilot border adjustments for cement and aluminum, while the 1.78-billion-litre annual withdrawal at a single 150-megawatt site rivals the water volumes disputed in historical interstate river allocation cases. Because the European Union Emissions Trading System already tracks electricity-sector emissions hourly, adjusting an import levy by destination grid factor is technically feasible. Doing so for water would require harmonized withdrawal accounting, but ICEF’s framework and the Aqueduct index provide a credible starting point. Implementing these instruments would not only reduce the surplus; scenario modelling shows a 59-percent cut for carbon and a 34-percent cut for water but also satisfy the World Trade Organization’s non-discrimination test provided that both domestic and foreign computing are assessed against identical metrics (Tsakiris & Vlassis, 2024 ). Distributional findings inject a social justice dimension into what might otherwise be a technocratic allocation debate. The concentration of AI-driven load growth in U.S. counties above the national median for both poverty and minority population share echoes the “digital resource curse” observed in Nordic Arctic towns hosting export-oriented server farms (Sovacool, Upham, & Monyei, 2022 ). These localized burdens intensify the normative case for extraterritorial responsibility; if importing jurisdictions enjoy the cognitive and economic surplus produced by large language models yet externalize environmental and social costs to vulnerable communities abroad, they risk violating the principle of non-discrimination embedded in numerous human rights treaties. The Environmental Protection Agency’s Environmental Justice strategic plan, which now lists data centers among projects subject to Justice40 screening, offers a domestic analog (Environmental Protection Agency, 2024 ), but legal symmetry requires corporate customers to share distributive vigilance across borders. Courts adjudicating CSDDD claims might therefore treat social-vulnerability mapping as part of the “appropriate measures” calculus, thereby operationalizing the Directive’s reference to “severe and irreparable harm.” The limitations of this study warrant cautious interpretation of the policy implications. First, the leakage estimates rest on announced design loads and assume high utilization factors; the actual realized capacity may diverge, especially if AI hardware efficiency improves more quickly than expected or if generative-AI demand plateaus. Second, the water-usage model applies regionalized, but still average, Water Usage Effectiveness coefficients; on-site hybrid systems or municipal recycled-water programs could lower withdrawals, although current adoption rates remain modest. Third, the dataset may undercount smaller edge-computing projects below 20 megawatts that escape public announcements but nonetheless contribute incrementally to regional strain. Nevertheless, triangulation with capacity trackers, such as DC bytes and utility interconnection queues, suggests that the sample captures the overwhelming share of energy-intensive AI expansion during the study window. Future research should integrate dynamic grid-decarbonization pathways into leakage projections. Rocky Mountain Institute scenarios imply that aggressive renewable buildouts and temporal-matching procurement can halve operational emissions even on today’s grids (Liu et al., 2024 ). Updating the leakage model with forward-looking marginal-emissions factors would help regulators design sunset clauses for border adjustments and allow courts to calibrate due diligence expectations against improving baseline conditions. Comparable dynamism on the water side depends on basin-scale integrated models capable of linking server-farm withdrawals to seasonal availability and downstream ecological thresholds. In conclusion, the study’s combined doctrinal and empirical analysis demonstrates that AI-driven datacenter expansion is no longer a niche sustainability issue but a frontline test of extraterritorial environmental responsibility. The customary duty to prevent significant transboundary harm supplies the foundational norm; the CSDDD and allied supply chain statutes offer an enforcement hook, and leakage metrics provide the evidentiary bridge connecting principle to performance. Whether that bridge becomes a regulatory highway or a contested border crossing depends on swift policy action: aligning AI and energy directives, hard-coding water disclosure, and operationalizing border adjustments that convert today’s externalities into tomorrow’s investment signals. As carbon and water budgets that safeguard a lovable planet continue to shrink, the window for voluntary self-regulation closes quickly. Lawmakers, courts, and corporate boards must, therefore, treat the environmental footprint of cloud infrastructure not as a peripheral technicality but as a central criterion of legitimate digital transformation. Conclusion This study demonstrates that the environmental geography of artificial-intelligence infrastructure is neither random nor benign: large-scale computer workloads systematically migrate toward electricity grids with higher carbon intensity and where the carbon penalty is low into basins already facing mounting water stress. These location choices generate measurable offshore externalities—tens of millions of tons of greenhouse gas emissions and billions of litres of freshwater withdrawals— which fall squarely within the magnitude, foreseeability, and traceability thresholds underpinning the customary duty to prevent significant transboundary harm. They also imply a majority of companies now covered by the European Union Corporate Sustainability Due Diligence Directive, confirming that private sector leverage over cloud sitting is sufficiently direct to trigger statutory mitigation duties. Empirical elasticities reveal that every incremental 100 g of carbon dioxide equivalent per kilowatt-hour on the host grid increases the probability of attracting AI-specific capacity by over eight percentage points, while median carbon and water leakage differentials exceed benchmarks that activated state responsibility in earlier environmental disputes. However, scenario modelling shows that these spill-overs are far from inevitable: a combination of twenty-four-hour, seven-days-a-week carbon-free procurement, hybrid cooling with recycled water, and compute-adjusted border adjustments could reduce net offshore emissions and withdrawals by approximately one-half. The legal architecture for such reforms already exists in embryonic form—through hourly reporting mandates in the European Union Energy Efficiency Directive, the chain-of-activities clause in the Corporate Sustainability Due Diligence Directive, and emerging guidelines for responsible supply chains—suggesting that rapid regulatory integration, rather than entirely new treaty law, is the most practical path forward. These findings reframe the cloud infrastructure as a frontline test of extraterritorial environmental accountability. If importing jurisdictions continue to reap the cognitive and economic surplus of generative-AI services without internalizing their environmental costs, they risk crystallizing new forms of carbon and water colonialism. Conversely, aligning preventive state duties with outcome-oriented corporate due diligence would convert present leakage into a powerful decarbonization and water-conservation lever, steering computing toward low-impact grids and accelerating investment in clean energy and advanced cooling technologies. Future research should couple dynamic grid-decarbonization trajectories with basin-scale hydrological models and examine the judicial uptake of hybrid carbon-water metrics, thereby refining both the evidentiary base and the doctrinal tools needed to govern the next decade of AI-driven digital growth. Declarations Competing Interests All authors certify that they have no affiliations with or involvement in any organization or entity with any financial interest or non-financial interest in the subject matter or materials discussed in this manuscript. The authors have no competing interests to declare that are relevant to the content of this article. The authors have no relevant financial or non-financial interests to disclose. The authors have no financial or proprietary interest in any material discussed in this article. Funding The authors did not receive support from any organization for the submitted work. No funding was received to assist with the preparation of this manuscript, no funding was received for conducting this study, and no funds, grants, or other support was received from any public, commercial, or not-for-profit entity. Author Contribution (H.U.) conceived the study, designed the legal-doctrinal framework, and drafted the introduction, discussion, and conclusion.(S.Z.) compiled the data-center dataset, performed the econometric and life-cycle analyses, and prepared the figures and table.H.U. and S.Z. jointly interpreted the results, revised the manuscript, and approved the final version. Acknowledgments The authors completed every stage of this study independently—conceptualization, research, analysis, and writing—without external assistance. No individuals, institutions, or funding organizations contributed directly or indirectly to the conception, design, or preparation of this manuscript. References Ambec, S., Esposito, F., & Pacelli, A. (2023). The economics of carbon leakage mitigation policies (TSE Working Paper No. 23‑1408). Toulouse School of Economics. Ciacchi, S. (2024). The newly adopted Corporate Sustainability Due Diligence Directive: An overview of the law‑making process and analysis of the final text. ERA Forum, 25 (1), 29–48. https://doi.org/10.1007/s12027-024-00791-y d’Orgeval, A., Liu, Y., & Li, J. (2024). Carbon footprint of AI data centres: A life‑cycle approach. In Proceedings of the 16th International Conference on Applied Energy (pp. 112–123). ICAE2024. DIGITALEUROPE. (2025, July). Enhancing water resilience in the data‑centre industry. DIGITALEUROPE. https://www.digitaleurope.org/resources/enhancing-water-resilience-in-the-data-centre-industry/ Ebert, K., Alder, N., & Patel, R. (2025). AI, climate, and regulation: From data centers to the AI Act (Version 2) [Preprint]. arXiv. https://arxiv.org/abs/2410.06681 DBLP Elliott, R. J. R., Sun, P., & Zhu, T. (2024). Energy abundance, the geographical distribution of manufacturing, and international trade. Review of World Economics, 160 (4), 1361–1391. https://doi.org/10.1007/s10290-024-00544-6 Environmental Protection Agency. (2024, December). Environmental justice strategic plan 2024–2028 . U.S. Environmental Protection Agency. https://www.epa.gov/system/files/documents/2024-12/environmental-justice-strategic-plan-december-2024.pdf Food & Water Watch. (2025, March). A no-brainer: How AI’s energy and water footprints threaten climate progress . Food & Water Watch. https://www.foodandwaterwatch.org/fsw_0325_ai_water_energy/ Friedmann, J. (2024). Data‑centre water use. In Artificial Intelligence for Climate Change Mitigation Roadmap 2.0 (Box 15.5). Innovation for Cool Earth Forum. Retrieved from https://www.icef.go.jp/roadmap/ García‑Sánchez, I.-M., Rodríguez‑Domínguez, L., & Frías‑Aceituno, J. V. (2023). How does the European Green Deal affect the disclosure of environmental information? Corporate Social Responsibility and Environmental Management, 30 (4), 2766–2782. https://doi.org/10.1002/csr.2140 INFRAS, Schmid, N., Coroamă, V. C., Dumbravă, O., Eichler, M., Reisser, M., Kaack, L. H., … Füssler, J. (2025). Carbon leakage in AI‑driven data‑centre growth? An assessment of drivers and barriers to the localization of data centre operations and investments with respect to carbon pricing policies (TEXTE 68/2025). Umweltbundesamt. https://doi.org/10.60810/openumwelt‑7756 Infrastructure Masons. (2025). State of the digital infrastructure industry 2025 [Annual report]. https://imasons.org/publications/ International Law Commission. (2024). Report of the International Law Commission on the work of its seventy-fifth session (A/79/10) . United Nations. Retrieved from https://legal.un.org/ilc/reports/2024/english/a_79_10_advance.pdf efchina.org Liu, Y., Qi, Y., & Long, Y. (2024). Powering the data‑center boom with low‑carbon solutions: China’s perspective and global insights (Rocky Mountain Institute Report). Rocky Mountain Institute. Retrieved from https://rmi.org/wp-content/uploads/dlm_uploads/2024/11/Powering_the_Data_Center_Boom_with_Low_Carbon_Solutions_report.pdf McKevett, S. E. (2024). Between sky and space: National Environmental Policy Act’s extraterritorial application to the stratosphere and Starlink. Georgetown Environmental Law Review, 36 (3), 375–444. Retrieved from https://www.law.georgetown.edu/environmental-law-review/wp-content/uploads/sites/18/2024/12/GT-GELR240034.pdf Organisation for Economic Co-operation and Development. (2025). Supporting businesses in trade‑partner countries to meet social and environmental due‑diligence standards (OECD Business and Finance Policy Paper No. 88). OECD Publishing. https://www.oecd.org/content/dam/oecd/en/publications/reports/2025/06/supporting-businesses-in-trade-partner-countries-to-meet-social-and-environmental-due-diligence-standards_16acd21c/63c6be24-en.pdf Oxfam International. (2024, March). Written statement submitted to the International Court of Justice in the matter of the advisory opinion request on obligations of states in respect of climate change . Retrieved from https://climatecasechart.com/wp-content/uploads/non-us-case-documents/2024/20240322_18913_na.pdf Sorensen, L. S., Jacobsen, P., & Kvamsdal, S. F. (2025). A review of challenges and strategies towards integrating sustainability and due diligence in the buyer–supplier relationship. In Proceedings of the International Conference on Sustainable Supply Chains (pp. 45–60). Sovacool, B. K., Upham, P., & Monyei, C. G. (2022). The ‘whole systems’ energy sustainability of digitalisation: Humanising the community risks and benefits of Nordic datacentre development. Energy Research & Social Science, 88 , Article 102493. https://doi.org/10.1016/j.erss.2022.102493 Tsakiris, N., & Vlassis, N. (2024). Border carbon adjustments and leakage in the presence of public pollution‑abatement activities. Environmental and Resource Economics, 87 (9), 2231–2258. https://doi.org/10.1007/s10640-024-00882-x United Nations Conference on Trade and Development. (2024). Digital economy report 2024: Environmental impacts in the use phase of digitalisation (Chapter III). UNCTAD. Retrieved from https://unctad.org/system/files/official-document/der2024_ch03_en.pdf United Nations Economic Commission for Europe & United Nations Educational, Scientific and Cultural Organization. (2024). Progress on transboundary water cooperation: SDG 6.5.2 mid‑term status 2024 (with special focus on climate change) . UNECE & UNESCO. Retrieved from https://www.unece.org/environment-policy/publications/progress-transboundary-water-cooperation-mid-term-status-sdg Urban Land Institute. (2024). Local guidelines for data‑centre development . ULI Americas Data Center Product Council. Retrieved from https://knowledge.uli.org/-/media/files/research-reports/2024/uli-data-center-whitepaper_hm_2024-11-12_final-final-round.pdf Villiers, C. L. (2022). New directions in the European Union’s regulatory framework for corporate reporting, due diligence and accountability: The challenge of complexity. European Journal of Risk Regulation, 13 (4), 548–566. https://doi.org/10.1017/err.2022.25 Additional Declarations No competing interests reported. Cite Share Download PDF Status: Under Review Version 1 posted Editorial decision: Revision requested 18 Oct, 2025 Reviews received at journal 05 Sep, 2025 Reviewers agreed at journal 01 Sep, 2025 Reviewers invited by journal 24 Jul, 2025 Editor assigned by journal 17 Jul, 2025 Submission checks completed at journal 16 Jul, 2025 First submitted to journal 12 Jul, 2025 You are reading this latest preprint version Research Square lets you share your work early, gain feedback from the community, and start making changes to your manuscript prior to peer review in a journal. As a division of Research Square Company, we’re committed to making research communication faster, fairer, and more useful. We do this by developing innovative software and high quality services for the global research community. Our growing team is made up of researchers and industry professionals working together to solve the most critical problems facing scientific publishing. Also discoverable on Platform About Our Team In Review Editorial Policies Advisory Board Help Center Resources Author Services Accessibility API Access RSS feed Manage Cookie Preferences © Research Square 2026 | ISSN 2693-5015 (online) Privacy Policy Terms of Service Do Not Sell My Personal Information {"props":{"pageProps":{"initialData":{"identity":"rs-7109458","acceptedTermsAndConditions":true,"allowDirectSubmit":false,"archivedVersions":[],"articleType":"Research Article","associatedPublications":[],"authors":[{"id":491560763,"identity":"852c1a7b-a3db-4ff1-960c-60850eb88d3c","order_by":0,"name":"Hazrat Usman","email":"data:image/png;base64,iVBORw0KGgoAAAANSUhEUgAAAZAAAAAyAQMAAABI0h/eAAAABlBMVEX///8AAABVwtN+AAAACXBIWXMAAA7EAAAOxAGVKw4bAAABJUlEQVRIiWNgGAWjYBACAxQeD4RifABi8+HVkoCshY2BGSTKw0aKFjYJCI0dmLMff/bh54/D9rrtZw9+eNtmI88/v/lY5dccOxmgbQ8f3cDUYtmTkDyzJ+Fw4rYzecmSc9vSDGccY0u7LbstGegwNmPjHCwOO5BwmIEn4XCC2YEcA2meM4cTGI7xmN2W3MYM1MLDJo1Ny/mHzYx/Eg7bm51/Y/wbpEX+GP+3Yslt9bi13EhmZgbawrjtRo6ZNE/F4QSDYzxsjB+3Hcaj5Rkzs0xaeuK2G2/MLOdUpBluPJZmLM247TgPGzMOv5xPf8z4xsYa6LAc4xtvDGzk5Q4ffvjx57Zqe3725oePsWjBDpjBEcRMrHIQYPxBiupRMApGwSgY7gAA6ctjJR6iQYYAAAAASUVORK5CYII=","orcid":"","institution":"Punjab Bar Council","correspondingAuthor":true,"prefix":"","firstName":"Hazrat","middleName":"","lastName":"Usman","suffix":""},{"id":491560764,"identity":"a88b5afb-63a5-4a70-89ec-e11b155a9832","order_by":1,"name":"Sidra Zakir","email":"","orcid":"","institution":"Mohi-ud-Din Islamic University","correspondingAuthor":false,"prefix":"","firstName":"Sidra","middleName":"","lastName":"Zakir","suffix":""}],"badges":[],"createdAt":"2025-07-12 17:08:13","currentVersionCode":1,"declarations":"","doi":"10.21203/rs.3.rs-7109458/v1","doiUrl":"https://doi.org/10.21203/rs.3.rs-7109458/v1","draftVersion":[],"editorialEvents":[],"editorialNote":"","failedWorkflow":false,"files":[{"id":87746145,"identity":"bdc354aa-48fd-42f8-b738-940f14fbd08f","added_by":"auto","created_at":"2025-07-28 14:13:57","extension":"png","order_by":1,"title":"Figure 1","display":"","copyAsset":false,"role":"figure","size":89997,"visible":true,"origin":"","legend":"\u003cp\u003eEnvironmental externalities of AI data-center siting.\u003c/p\u003e","description":"","filename":"floatimage1.png","url":"https://assets-eu.researchsquare.com/files/rs-7109458/v1/1c4d1d3c441648ae98fdb985.png"},{"id":87747982,"identity":"05c2e04f-0a13-431f-81fc-739393644051","added_by":"auto","created_at":"2025-07-28 14:29:57","extension":"pdf","order_by":0,"title":"","display":"","copyAsset":false,"role":"manuscript-pdf","size":744597,"visible":true,"origin":"","legend":"","description":"","filename":"manuscript.pdf","url":"https://assets-eu.researchsquare.com/files/rs-7109458/v1/7fd8009a-c081-4ad7-99f8-7a2e93759da4.pdf"}],"financialInterests":"No competing interests reported.","formattedTitle":"Extraterritorial Environmental Accountability of AI Data Centers via Transboundary Harm and Due-Diligence Norms","fulltext":[{"header":"Introduction","content":"\u003cp\u003eArtificial intelligence (AI) training and inference have now pulled staggering volumes of electricity, land, and cooling water into an expanding web of hyperscale data centers. Industry forecasters predict that the global stock of server capacity will exceed 205 gigawatts by 2030, 90 percent of which is driven by AI workloads (Infrastructure Masons, \u003cspan citationid=\"CR12\" class=\"CitationRef\"\u003e2025\u003c/span\u003e). In China alone, the power demand from data centers could climb past 700 terawatt-hours\u0026mdash;roughly 5 percent of the national grid\u0026mdash;within the same decade (Liu et al., \u003cspan citationid=\"CR14\" class=\"CitationRef\"\u003e2024\u003c/span\u003e). The environmental footprint of this rapid build-out is already measurable: life-cycle assessments for a single 3.6-megawatt AI cluster anchored on NVIDIA H100 accelerators attribute approximately 70 percent of cradle-to-grave greenhouse gas (GHG) emissions to embodied server manufacturing and a further 15 percent to cooling operations (d\u0026rsquo;Orgeval et al., \u003cspan citationid=\"CR3\" class=\"CitationRef\"\u003e2024\u003c/span\u003e). Water withdrawals are similarly non-trivial; current estimates place datacenter consumption near 292\u0026nbsp;million gallons per day and trending toward 450\u0026nbsp;million gallons per day by 2030 (Friedmann, \u003cspan citationid=\"CR9\" class=\"CitationRef\"\u003e2024\u003c/span\u003e).\u003c/p\u003e\u003cp\u003eThese impacts rarely remain within the territorial borders of the state in which the data center is physically located. Cloud providers routinely site new campuses in jurisdictions with inexpensive electricity or permissive water rights regimes and increasingly with laxer carbon-intensity profiles, creating what the German Environment Agency terms \u0026ldquo;AI-driven carbon-leakage risk\u0026rsquo; (INFRAS, 2025, p. 12). Sitting decisions thereby export the life-cycle emissions and freshwater burdens of global AI markets to communities that often lack meaningful opportunities to participate in environmental decision making (Food \u0026amp; Water Watch, 2025). While the international legal order recognizes a customary obligation on states to prevent \u0026ldquo;significant transboundary harm\u0026rdquo; (International Law Commission [ILC], 2024), neither this duty nor emerging corporate-sustainability regimes have yet been tested against the distributed, cloud-centered value chain that underpins contemporary AI deployment.\u003c/p\u003e\u003cp\u003eA parallel regulatory conversation is unfolding within the European Union (EU). The recently adopted Corporate Sustainability Due Diligence Directive (CSDDD) obliges large undertakings to identify, mitigate, and remediate adverse environmental and human rights impacts throughout their \u0026ldquo;chain of activities\u0026rdquo; (Ciacchi, \u003cspan citationid=\"CR2\" class=\"CitationRef\"\u003e2024\u003c/span\u003e). However, the text is silent on whether foreign-sited datacenter emissions or water withdrawals linked to EU cloud customers fall within that chain. Similarly, the EU Artificial Intelligence Act establishes risk-tiered obligations for certain AI systems but omits binding energy- or water-footprint thresholds for the data-center infrastructure on which those systems rely (Ebert et al., \u003cspan citationid=\"CR5\" class=\"CitationRef\"\u003e2025\u003c/span\u003e). Outside the EU, the Organization for Economic Cooperation and Development is piloting monitoring-and-evaluation frameworks for responsible supply chains, but these instruments remain voluntary and sector-specific (Organization for Economic Co-operation and Development [OECD], 2025). Legal scholarship has therefore begun to flag an \u0026ldquo;accountability gap\u0026rdquo; where extraterritorial environmental effects of digital infrastructure escape both public-law prevention duties and private-law due-diligence norms (Sorensen et al., \u003cspan citationid=\"CR18\" class=\"CitationRef\"\u003e2025\u003c/span\u003e).\u003c/p\u003e\u003cp\u003eEmpirical work on the geography of AI datacenters has reinforced normative stakes. Structural trade models indicate that the relative carbon intensity of a host grid can invert leakage flows; relocating Graphics Processing Unit (GPU) clusters from France\u0026rsquo;s low-carbon mix to Germany\u0026rsquo;s coal-heavier mix quadruples life-cycle emissions (d\u0026rsquo;Orgeval et al., \u003cspan citationid=\"CR3\" class=\"CitationRef\"\u003e2024\u003c/span\u003e). Water stress indices tell a similar story: approximately one-fifth of United States data-center servers operate in high or extremely high baseline water stress basins (Friedmann, \u003cspan citationid=\"CR9\" class=\"CitationRef\"\u003e2024\u003c/span\u003e). However, scholars have not quantified how these variations translate into cross-border externalities or analyzed them through the lens of established principles such as the Trail Smelter standard or the due-diligence thresholds articulated by international tribunals (McKevett, \u003cspan citationid=\"CR15\" class=\"CitationRef\"\u003e2024\u003c/span\u003e; Oxfam International, \u003cspan citationid=\"CR17\" class=\"CitationRef\"\u003e2024\u003c/span\u003e).\u003c/p\u003e\u003cp\u003eAgainst this backdrop, this study addresses two interlocking gaps. First, the doctrinal debate has outpaced systematic doctrinal-empirical analyses. Little is known about the conditions under which the customary duty to prevent significant transboundary harm, together with emerging instruments such as the CSDDD, can be used to capture the offshore carbon and water footprints intrinsic to cloud infrastructure. Second, policymakers lack robust, comparative evidence on how host-state electricity-grid carbon intensity and water-stress indices steer the global siting of AI-oriented data centers and what quantifiable carbon leakage and water withdrawal differentials emerge between exporting and importing jurisdictions.\u003c/p\u003e\u003cp\u003eTherefore, this study had two objectives. Doctrinally, it interrogates the extent to which existing public-international-law prevention obligations and nascent corporate-sustainability due-diligence duties can be interpreted\u0026mdash;and, where necessary, extended\u0026mdash;to encompass the transboundary environmental effects of AI datacenter operations. Empirically, it constructs a cross-national dataset that couples hyperscale sitting announcements with hourly grid-carbon factors and basin-level water stress measures to estimate differentiated leakage trajectories for carbon and water. The analysis was guided by the following research question: \u003cem\u003eTo what extent can the customary duty to prevent significant transboundary environmental harm and the emerging corporate-sustainability due-diligence directives, such as the European Union Corporate Sustainability Due Diligence Directive, be interpreted to cover the offshore carbon and water footprints of AI data-center operations? How do variations in host-state electricity-grid carbon intensity and water-stress indices influence the cross-border siting of AI-oriented data centers, and what quantifiable carbon-leakage and water-withdrawal differentials arise between exporting and importing jurisdictions?\u003c/em\u003e\u003c/p\u003e\u003cp\u003eThis study contributes three advances to the literature by blending doctrinal interpretations with quantitative policy analysis. First, it clarifies the legal predicates under which datacenter externalities trigger state responsibility and corporate liability beyond borders. Second, it offers the first harmonized estimate of simultaneous carbon and water leakage effects specific to AI-driven cloud infrastructure. Third, it develops a risk-tiering matrix that links empirical leakage magnitudes to graded due diligence obligations, thereby providing regulators and industry actors with a concrete blueprint for aligning digital infrastructure growth with international environmental law.\u003c/p\u003e\u003cp\u003eThe remainder of this paper is organized as follows. Section 2 elaborates on the doctrinal framework, tracing the evolution of transboundary harm and due diligence norms. Section 3 details the data sources and the empirical strategy. Section 4 presents the results and Section 5 discusses their implications for treaty interpretation, corporate governance, and climate-aligned sitting policy. The final section concludes with recommendations for future research and norm development.\u003c/p\u003e\u003cp\u003e\u003cb\u003eThe Normative Reach of Transboundary-Harm and Corporate Due-Diligence Obligations over AI Data-Centre Externalities\u003c/b\u003e\u003c/p\u003e\u003cp\u003eInternational environmental law has long imposed on states an obligation to prevent \u0026ldquo;significant transboundary harm,\u0026rdquo; a principle famously crystallized in the arbitral decision in \u003cem\u003eTrail Smelter\u003c/em\u003e (United States v. Canada, 1941) and subsequently reaffirmed by the International Court of Justice in \u003cem\u003eCorfu Channel\u003c/em\u003e (United Kingdom v. Albania, 1949), \u003cem\u003eNuclear Tests\u003c/em\u003e (Australia v. France, 1974), and \u003cem\u003eGabč\u0026iacute;kovo\u0026ndash;Nagymaros\u003c/em\u003e (Hungary v. Slovakia, 1997). Recent work by the International Law Commission consolidates this history into what it now styles as a \u003cem\u003egeneral principle of prevention\u003c/em\u003e carrying an attendant duty of due diligence proportionate to the risk at hand (International Law Commission, \u003cspan citationid=\"CR13\" class=\"CitationRef\"\u003e2024\u003c/span\u003e). Although the canon was forged in contexts such as drifting sulfur dioxide, fugitive radiation, or shared-river engineering, the Commission\u0026rsquo;s Annex II expressly situates new-economy hazards\u0026mdash;cyber intrusions, cross-border data flows, and climate-related spill-overs\u0026mdash;within its prospective codification agenda. The doctrinal pivot invites scrutiny of artificial intelligence (AI) data-center operations, whose embodied carbon and cooling water footprints are increasingly externalized across borders through supply chain relocation and jurisdictional arbitrage (INFRAS, 2025).\u003c/p\u003e\u003cp\u003eData-center siting choices already manifest the distributive dynamics that animate classical cases. Empirical mapping shows that hyperscale providers tend to channel investment into electricity grids characterized by comparatively high carbon intensity and water-abundant regions with weak appropriation safeguards (Food \u0026amp; Water Watch, 2025; Liu et al., \u003cspan citationid=\"CR14\" class=\"CitationRef\"\u003e2024\u003c/span\u003e). Life-cycle assessment for a 3.6-megawatt graphics-processing-unit cluster reveals that moving identical hardware from France\u0026rsquo;s low-carbon system to Germany\u0026rsquo;s coal-heavier mix multiplies cradle-to-grave greenhouse gas emissions four-fold, even before counting the downstream effects of transmission losses (d\u0026rsquo;Orgeval et al., \u003cspan citationid=\"CR3\" class=\"CitationRef\"\u003e2024\u003c/span\u003e). A parallel gradient applies to freshwater withdrawals: Innovation for Cool Earth Forum modelling estimates that one-fifth of United States data-center servers now operate in counties classified as experiencing \u0026ldquo;high\u0026rdquo; or \u0026ldquo;extremely high\u0026rdquo; baseline water stress, thereby intensifying basin-level competition for municipal supply (Friedmann, \u003cspan citationid=\"CR9\" class=\"CitationRef\"\u003e2024\u003c/span\u003e). Because service volumes remain transnational, an inference query from Paris can trigger an inference load in Ashburn\u0026mdash;offshore communities that often shoulder the biophysical costs of AI services that principally benefit foreign users. In orthodox terms, such circumstances appear tailor-made for the prevention principle\u0026rsquo;s prospective filter: a high-probability, high-magnitude externality, traceable to clearly identifiable technological activity, yet largely outside the regulatory reach of the consumer state.\u003c/p\u003e\u003cp\u003eWhether customary law alone can discipline externalities remains contested. The classic preventive duty is primarily owed to the state of origin, not to private undertakings per se. The International Court of Justice reemphasized this allocation in \u003cem\u003ePulp Mills\u003c/em\u003e (Argentina v. Uruguay, 2010), holding that due diligence is \u0026ldquo;an obligation of conduct\u0026rsquo; requiring the source state to regulate domestic operators but seldom prescribing the precise content of those controls. Nevertheless, the Court\u0026rsquo;s advisory opinion remit on climate change, seized in 2023 and already the subject of detailed submissions by Oxfam International (\u003cspan citationid=\"CR17\" class=\"CitationRef\"\u003e2024\u003c/span\u003e), indicates growing judicial readiness to translate climate-linked harms into justiciable transboundary wrongs. Oxfam\u0026rsquo;s filing argues that greenhouse gas emissions transform the prevention principle into a \u003cem\u003epositive obligation not merely to regulate but to reduce\u003c/em\u003e, drawing on the Committee on the Rights of the Child\u0026rsquo;s General Comment No. 26 and the Human Rights Committee\u0026rsquo;s pronouncements in \u003cem\u003eTeitiota\u003c/em\u003e v. \u003cem\u003eNew Zealand\u003c/em\u003e (2020). To the extent that AI datacenter emissions constitute non-trivial increments to the global carbon budget, Infrastructure Masons (\u003cspan citationid=\"CR12\" class=\"CitationRef\"\u003e2025\u003c/span\u003e) forecasts a tripling of digital-infrastructure power demand by the 2030s. Source states that host carbon-intensive clusters may soon face allegations of breaching the duty to prevent transboundary climate harm.\u003c/p\u003e\u003cp\u003eHowever, doctrinal traction alone cannot account for the complex division of labor in AI supply chains. Servers, cooling systems, and diesel backup plants that form the physical core of cloud infrastructure are frequently owned by multinational enterprises incorporated in jurisdictions that are far removed from their operational footprints. It is here that emerging corporate-sustainability due diligence regimes promise to complement state-centric prevention doctrines. The European Union Corporate Sustainability Due Diligence Directive (CSDDD), adopted in modified form in March 2024 after fraught trilogue negotiations, obliges companies exceeding one thousand employees and \u0026euro;450\u0026nbsp;million in net turnover to \u0026ldquo;establish and implement a due-diligence process\u0026rdquo; that covers adverse environmental impacts across their entire \u0026ldquo;chain of activities\u0026rdquo; (Ciacchi, \u003cspan citationid=\"CR2\" class=\"CitationRef\"\u003e2024\u003c/span\u003e, p. 7). Although recital 35 lists greenhouse gas emissions expressly, the operative provisions do not specify water use, nor do they clarify whether emissions stemming from third-country datacenter operations fall within the relevant chain when the datacenter owner is itself a separate legal entity. Commentators, such as Villiers (\u003cspan citationid=\"CR24\" class=\"CitationRef\"\u003e2022\u003c/span\u003e) and Sorensen et al. (\u003cspan citationid=\"CR18\" class=\"CitationRef\"\u003e2025\u003c/span\u003e), contend that the Directive\u0026rsquo;s open-textured reference to upstream and downstream activities can be read purposively to capture \u003cem\u003efunctional\u003c/em\u003e rather than purely contractual relationships, thereby enveloping collocation, leasing, and cloud-service arrangements.\u003c/p\u003e\u003cp\u003eThis interpretation is supported by the Organization for Economic Co-operation and Development\u0026rsquo;s parallel guidance papers on responsible supply chains, which emphasize proportionality and leverage rather than formal ownership as the touchstones of due diligence (Organization for Economic Co-operation and Development, 2025). The monitoring and evaluation framework proposed for the garment sector suggests deploying quasi-experimental designs (difference-in-differences or regression discontinuity designs) to measure whether corporate interventions reduce adverse impacts. Translating this approach to digital infrastructure would entail quantifying the change in grid-level carbon intensity attributable to sitting decisions and assessing the incremental water-withdrawal burden on stressed basins. If methodologies of that sort become mainstream, courts and regulators are likely to treat such metrics as baseline expectations for compliance, effectively raising the due diligence standard from a process-oriented duty to a quantified, outcome-sensitive one.\u003c/p\u003e\u003cp\u003eDirect regulatory attempts to trace environmental liabilities along digital supply chains are emerging outside the corporate governance field. The European Union Artificial Intelligence Act, finalized in March 2025, requires providers of general-purpose AI models to publish \u0026ldquo;sufficiently detailed\u0026rdquo; data on energy use during the training phases, yet remains silent on inference-phase energy or water consumption (Ebert et al., \u003cspan citationid=\"CR5\" class=\"CitationRef\"\u003e2025\u003c/span\u003e). This asymmetry risks displacing environmental burdens from well-monitored European training hubs to less-regulated inference facilities abroad, a scenario that INFRAS (2025) identifies as a potential new channel for carbon leakage. Therefore, the German Environment Agency recommends aligning the AI Act with energy-efficiency obligations under Directive (EU) 2023/1791 on energy efficiency, including the directive\u0026rsquo;s forthcoming common datacenter reporting scheme. Were such alignment pursued, the AI Act\u0026rsquo;s disclosure-based model could converge with the CSDDD\u0026rsquo;s broader duty of care, jointly anchoring a composite expectation that European corporations must trace and mitigate the offshore footprints of their cloud deployments.\u003c/p\u003e\u003cp\u003eThe normative interplay between public and private obligations becomes more salient when water use is considered. Unlike carbon dioxide, freshwater withdrawals generate localized scarcity that can give rise to direct, tangible harm across jurisdictional boundaries, particularly in shared basin contexts. The United Nations Economic Commission for Europe\u0026rsquo;s 2024 Progress Report on Sustainable Development Goal 6.5.2 shows that only 43 of 153 states have placed 90 percent of their transboundary basins under operational cooperation arrangements, leaving large swathes of Asia and Latin America effectively unprotected (United Nations Economic Commission for Europe \u0026amp; United Nations Educational, Scientific and Cultural Organization, 2024). A data-center cluster drawing on a shared aquifer in the Dutch\u0026ndash;German\u0026ndash;Belgian Meuse Basin could therefore reduce water availability downstream without triggering any pre-existing cooperative response. The prevention principle arguably obliges the host state to notify and consult its riparian neighbors under the Berlin Rules or the United Nations Convention on the Law of Non-Navigational Uses of International Watercourses. However, where the host state fails to act, the question arises whether cloud service providers domiciled in another jurisdiction, say, an EU member state, incur derivative obligations under the CSDDD to reduce that overseas water footprint.\u003c/p\u003e\u003cp\u003eSome commentators are skeptical that the Directive, even if interpreted broadly, can reach so far. Garc\u0026iacute;a-S\u0026aacute;nchez et al. (2023) demonstrate that European firms\u0026rsquo; environmental disclosures remain heavily skewed toward climate risk and resource-use metrics that coincide with existing mandatory-reporting templates, whereas biodiversity, waste, and water metrics lag considerably. Voluntary initiatives, such as the Climate Neutral Data Centre Pact\u0026rsquo;s commitment to achieve Water Usage Effectiveness below 0.4 liters per kilowatt-hour in water-stressed regions, provide partial backstops (DIGITALEUROPE, 2025). However, industry pledges vary in precision and enforceability, leading the Food \u0026amp; Water Watch (2025) to accuse cloud providers of \u0026ldquo;bluewashing\u0026rdquo; their performance through narrow, self-selected indicators. The Directive\u0026rsquo;s newly added civil-liability article may eventually furnish plaintiffs with a cause of action in domestic courts when a company fails to exercise \u0026ldquo;appropriate\u0026rdquo; due diligence, but practical hurdles\u0026mdash;forum non-convenience, causation, and collective-action barriers\u0026mdash;remain substantial.\u003c/p\u003e\u003cp\u003eThe historical trajectory of carbon leakage regulation in energy-intensive trade-exposed sectors illuminates the likely evolution of due diligence norms in the digital sphere. Economic modelling by Ambec et al. (\u003cspan citationid=\"CR1\" class=\"CitationRef\"\u003e2023\u003c/span\u003e) indicates that the free allocation of emissions permits under the European Union Emissions Trading System can reverse leakage only if allocation is generous and linked to verified abatement. Border Carbon Adjustments achieve more robust results, but risk retaliation unless calibrated to the importing country\u0026rsquo;s public pollution abatement effort (Tsakiris \u0026amp; Vlassis, \u003cspan citationid=\"CR20\" class=\"CitationRef\"\u003e2024\u003c/span\u003e). German Environment Agency scenarios suggest that a dynamic, compute-specific Carbon Border Adjustment Mechanism\u0026mdash;tethered to hourly emissions-factor disclosures\u0026mdash;could curb AI-induced leakage by steering workloads toward low-carbon grids (INFRAS, 2025). Translating these insights into water stewardship is conceptually straightforward: a Water Footprint Border Adjustment could price withdrawals embedded in digital services, although operationalizing such a mechanism would require global water footprint accounting standards comparably robust to those under development for carbon (UNCTAD, 2024).\u003c/p\u003e\u003cp\u003eThe International Energy Agency\u0026rsquo;s tracking indicates that twenty-four-hour, seven-days-a-week carbon-free electricity procurement is emerging as the new gold standard for cloud providers, yet uptake remains geographically uneven. Rocky Mountain Institute models show that aligning the data-center electricity load with spatially matched renewable portfolios can halve net emission intensity but require sophisticated temporal-matching analytics that many grids still lack (Liu et al., \u003cspan citationid=\"CR14\" class=\"CitationRef\"\u003e2024\u003c/span\u003e). If such granular procurement becomes normalized, the residual offshore emissions attributable to data-center siting may shrink, easing the doctrinal burden on courts to extend prevention or due diligence duties. Conversely, if electricity-grid decarbonization stalls, litigants are likely to weaponize the combined weight of customary international law, the CSDDD, and emerging monitoring frameworks to press for extraterritorial accountability.\u003c/p\u003e\u003cp\u003eIn sum, the literature reveals a dynamic, if unsettled, convergence between the customary duty to prevent significant transboundary harm and the corporate-sustainability due diligence obligations now codifying in the European Union. The prevention principle provides the foundational norm that states must regulate domestic activities with foreseeable cross-border consequences. The CSDDD and allied instruments then relocate a share of that responsibility to corporate actors, potentially extending liability to the offshore carbon and water footprints of AI datacenter operations. Whether this convergence yields effective environmental protection will hinge on three interrelated factors that the remainder of this article explores empirically: (a) the magnitude and distribution of carbon-leakage and water-withdrawal differentials generated by siting decisions, (b) the feasibility of measuring those differentials at a resolution suitable for legal attribution, and (c) the willingness of courts and regulators to interpret prevention and due diligence in a mutually reinforcing, rather than duplicative or fragmented, manner.\u003c/p\u003e\u003cp\u003e\u003cb\u003eLocational Determinants and Cross-Border Externalities of AI-Driven Data-Centre Expansion\u003c/b\u003e\u003c/p\u003e\u003cp\u003eCloud-service operators rarely disclose more than the headline Power Usage Effectiveness of a new hyperscale facility; however, the underlying geography of its electricity and water inputs is decisively shaped by spatial disparities in carbon intensity and hydrological stress. Scholars of industrial location economics have long shown that energy-price differentials drive manufacturing agglomeration (Elliott et al., \u003cspan citationid=\"CR6\" class=\"CitationRef\"\u003e2024\u003c/span\u003e), and recent evidence indicates that the same calculus now steers the siting of computational infrastructure dedicated to artificial intelligence (AI) workloads. Publicly announced projects plotted against the International Energy Agency hourly emissions factors reveal a striking tilt: graphics-processing-unit clusters capable of exa-scale throughput are concentrated on grids whose average carbon dioxide equivalent exceeds 400 g kWh-\u0026sup1;, well above the 100\u0026ndash;150 g kWh-\u0026sup1; threshold aligned with the Intergovernmental Panel on Climate Change 1.5\u0026deg;C pathway (INFRAS, 2025). This trend holds even for firms that have adopted twenty-four-hour, seven-days-a-week carbon-free-energy procurement goals, because time-matching algorithms remain constrained by regional renewable-portfolio availability and transmission congestion (Liu et al., \u003cspan citationid=\"CR14\" class=\"CitationRef\"\u003e2024\u003c/span\u003e). The disjunction is material: life-cycle modelling shows that transplanting an identical 3.6-megawatt accelerator cluster from France\u0026rsquo;s nuclear-dominated mix to Germany\u0026rsquo;s coal-heavy mix multiplies cradle-to-grave greenhouse gas emissions by a factor of four, with embodied hardware carbon comprising roughly 70 percent of that increment (d\u0026rsquo;Orgeval et al., \u003cspan citationid=\"CR3\" class=\"CitationRef\"\u003e2024\u003c/span\u003e). When aggregated across the hyperscale pipeline, Infrastructure Masons (\u003cspan citationid=\"CR12\" class=\"CitationRef\"\u003e2025\u003c/span\u003e) anticipates that the active computing capacity will triple to 205 gigawatts by 2030, with 90 percent of the growth attributable to AI training. If the projected capacity was allocated proportionately to the present grid mix, annual offshore emissions attributable to European Union (EU) cloud demand alone would exceed the combined 2022 national inventories of Estonia, Latvia, and Lithuania within five years.\u003c/p\u003e\u003cp\u003eThe carbon intensity gradient intersects with fiscal and regulatory arbitrage in ways that amplify the leakage. In the United States, nine states exempt data-center electricity from sales tax and seven grant \u0026ldquo;mega-project\u0026rdquo; property tax holidays exceeding 20 years (Urban Land Institute, 2024). Coupled with the absence of a federal carbon price, these incentives have pulled large-language-model inference farms toward the coal-gas corridors of Virginia and West Virginia, a shift that is clearly visible in the Dominion and American Electric Power interconnection queues (Food \u0026amp; Water Watch, 2025). German Environment Agency scenario analysis suggests that, even under a declining emissions-cap trajectory, such siting dynamics can neutralize up to one-third of the aggregate reduction expected from the European Union Emissions Trading System Phase IV, unless a compute-specific Carbon Border Adjustment Mechanism is introduced (INFRAS, 2025). Theoretical work supports this empirical diagnosis: trade-model simulations calibrated to steel and cement find that the unrestricted relocation of carbon-intensive production reverses expected welfare gains unless border adjustments explicitly reward public-sector abatement in exporting jurisdictions (Tsakiris \u0026amp; Vlassis, \u003cspan citationid=\"CR20\" class=\"CitationRef\"\u003e2024\u003c/span\u003e; Ambec et al., \u003cspan citationid=\"CR1\" class=\"CitationRef\"\u003e2023\u003c/span\u003e). These insights translate directly into digital services, where the intangible nature of data flows allows carbon-intensive computing to migrate at minimal logistical costs.\u003c/p\u003e\u003cp\u003eWater stress gradients exert a parallel, but partly independent, influence. Thermodynamic constraints force most hyperscale facilities to reject heat through evaporative cooling or hybrid systems, unless ambient wet-bulb temperatures and electricity prices render energy-intensive air cooling viable (Friedmann, \u003cspan citationid=\"CR9\" class=\"CitationRef\"\u003e2024\u003c/span\u003e). Bluefield Research data, synthesized in the Innovation for Cool Earth Forum roadmap, indicate that United States data center withdrawals already approach 292\u0026nbsp;million gallons per day and could exceed 450\u0026nbsp;million gallons per day by 2030 under current efficiency trajectories (Friedmann, \u003cspan citationid=\"CR9\" class=\"CitationRef\"\u003e2024\u003c/span\u003e). One-fifth of the installed servers operate in hydrological basins classified as \u0026ldquo;highly stressed\u0026rdquo; by the World Resources Institute Aqueduct index, including central Arizona, northern Virginia, and parts of the High Plains Aquifer. The location of AI-specific expansions sharpens the spatial imbalance. Microsoft\u0026rsquo;s 2024 announcement of a 1.2-gigawatt campus in Goodyear, Arizona, for generative-AI inference would demand more than 4.5\u0026nbsp;billion gallons of water annually under standard Water Usage Effectiveness values, equal to roughly 20 percent of the city\u0026rsquo;s current municipal withdrawals (Food \u0026amp; Water Watch, 2025). Because large language models often serve global user bases, the resulting depletion constitutes an extraterritorial externality when the models\u0026rsquo; primary customers reside abroad.\u003c/p\u003e\u003cp\u003eQuantifying cross-border leakage of carbon and water requires a dual-metric approach. On the emission side, life-cycle assessment must couple embodied hardware inventories with location-based and market-based electricity factors. d\u0026rsquo;Orgeval et al. (\u003cspan citationid=\"CR3\" class=\"CitationRef\"\u003e2024\u003c/span\u003e) provided component-level emission coefficients for NVIDIA H100 clusters, providing a baseline for embodied carbon. Liu et al.\u0026rsquo;s (\u003cspan citationid=\"CR14\" class=\"CitationRef\"\u003e2024\u003c/span\u003e) scenario engine projects operational footprints under hourly dispatch curves matched with regional renewable energy certificates. Combined, these models make it possible to calculate a differential leakage index: the net tons of carbon dioxide equivalent that would have been avoided had the same computed load processed on the importing jurisdiction\u0026rsquo;s average grid. Early application of this metric to a portfolio of European inference jobs processed in Loudoun County, Virginia, yields leakage estimates as high as 650 kg of carbon dioxide equivalent per kilowatt-hour of delivered compute, a figure comparable to steelmaking shift impacts (INFRAS 2025).\u003c/p\u003e\u003cp\u003eFor water, the analogous concept is the withdrawal differential per kilowatt hour. Innovation for Cool Earth Forum (CEF)s roadmap suggests that typical Water Usage Effectiveness for efficient evaporative systems in arid climates is approximately 0.7 litres per kilowatt-hour but can drop below 0.2 liters per kilowatt-hour where hybrid cooling and non-potable reuse are deployed (Friedmann, \u003cspan citationid=\"CR9\" class=\"CitationRef\"\u003e2024\u003c/span\u003e). Overlaying these coefficients onto the Water Resources Institute baseline-stress map reveals that moving a 100-megawatt inference cluster from the hydrologically stressed Colorado River basin to the more water-abundant Columbia River basin could avert withdrawals equivalent to the annual residential consumption of 40,000 people. Conversely, cloud firms\u0026rsquo; shift of European inference from Dublin\u0026mdash;where the River Liffey affords relatively low stress\u0026mdash;to Madrid\u0026rsquo;s Tagus Basin adds an estimated 1.8\u0026nbsp;billion liters of withdrawal per annum (DIGITALEUROPE, 2025).\u003c/p\u003e\u003cp\u003eThe magnitude of these differentials arguably heightens the salience of both public law prevention duties and private law due diligence obligations. The recent International Court of Justice jurisprudence recognizes that the scale and foreseeability of harm calibrate the due-diligence threshold (International Law Commission, \u003cspan citationid=\"CR13\" class=\"CitationRef\"\u003e2024\u003c/span\u003e). Because emission factors and water stress indices are readily available, harm foreseeability is effectively presumed. The corporate sustainability law is therefore evolving toward disclosure-based mechanisms that internalize spatial variability in environmental intensity. The European Union Energy Efficiency Directive 2023/1791 mandates datacenter operators above one megawatt information-technology load to report annual electricity, cooling fluid, and water flows in machine-readable form, disaggregated by hour. Combined with the Corporate Sustainability Due Diligence Directive chain-of-activities mandate, this data architecture could enable the importing jurisdiction to compute real-time leakage and attribute it to individual corporate customers.\u003c/p\u003e\u003cp\u003eVoluntary pioneer initiatives sketch an operational template. The Climate Neutral Data Centre Pact commits signatories to achieve Water Usage Effectiveness below 0.4 litres per kilowatt-hour in stress-classified regions by 2040, and to procure 100 percent renewable energy on a monthly basis by 2030 (DIGITALEUROPE, 2025). Analysis by INFRAS (2025) indicates that if fully implemented, the Pact would reduce European offshore carbon leakage by approximately 38 percent relative to a no-policy baseline, primarily through load shifting and time-matched renewable procurement. However, scholars caution that voluntary codes suffer from heterogeneity of measurement boundaries and the absence of third-party enforcement (Food \u0026amp; Water Watch, 2025). A statutory backstop\u0026mdash;potentially in the form of a compute-adjusted Carbon Border Adjustment Mechanism or a Water Footprint Certificate\u0026mdash;would therefore be required to close residual gaps, echoing the trade literature\u0026rsquo;s conclusion that free permit allocation cannot by itself neutralize leakage in heavy industry (Ambec et al., \u003cspan citationid=\"CR1\" class=\"CitationRef\"\u003e2023\u003c/span\u003e).\u003c/p\u003e\u003cp\u003eCommunity-level distributional effects further complicate leakage calculus. Case-study evidence from Iceland, Norway, and Greenland shows that while renewable siting offers low operational carbon intensity, it can trigger a \u0026ldquo;digital resource curse\u0026rdquo;: housing inflation, boom-bust labor cycles, and strain on fragile grids built for aluminum smelting rather than 24-hour server loads (Sovacool et al., \u003cspan citationid=\"CR19\" class=\"CitationRef\"\u003e2022\u003c/span\u003e). These findings resonate with the Environmental Protection Agency\u0026rsquo;s Environmental Justice Strategic Plan (2024), which embeds digital infrastructure within its Justice40 screening metric. When data flows cross borders, importing jurisdictions\u0026rsquo; users gain computational utility without bearing such community costs, aggravating distributive inequities. However, no existing treaty mechanism systematically distributes compensatory benefits or adaptation finance along these lines, despite analogous proposals in the Montreal Protocol\u0026rsquo;s non-party trade restrictions and the Paris Agreement\u0026rsquo;s loss-and-damage dialogue.\u003c/p\u003e\u003cp\u003eThe literature also underscores how variations in grid decarbonization trajectories mediate leakage dynamics. Rocky Mountain Institute models suggest that aggressive transmission upgrades and spatial balancing could allow China to accommodate a threefold data-center expansion while capping sectoral emissions at the 2018 level (Liu et al., \u003cspan citationid=\"CR14\" class=\"CitationRef\"\u003e2024\u003c/span\u003e). Under this scenario, relocating computing from coal-intensive Yunnan to wind-solar-rich Inner Mongolia yields a net global mitigation rather than leakage. Conversely, if India\u0026rsquo;s coal share persists above 55 percent, similar relocation would amplify emissions despite proximity to photovoltaic resources because battery storage costs remain prohibitive (UNCTAD, 2024). Thus, the leakage effect is sensitive not only to static intensity metrics but also to dynamic policy commitments and infrastructure investment pipelines. This sensitivity vindicates calls for integrating corporate data-center due diligence with nationally determined contributions under the Paris Agreement and basin-level adaptive management plans under transboundary water accords (United Nations Economic Commission for Europe \u0026amp; United Nations Educational, Scientific and Cultural Organization, 2024).\u003c/p\u003e\u003cp\u003eFinally, academic debate has begun to consider the epistemic and legal feasibility of hybrid carbon-water metrics. INFRAS (2025) and UNCTAD (2024) advocate a composite Compute-Adjusted Environmental Footprint metric that multiplies server-hour counts by location-based carbon and water coefficients to derive a single transferable obligation unit. If indexed to the evolving decarbonization pathway and aquifer stress scores, such an instrument could anchor a border-adjustment schedule or a trans-jurisdictional due diligence benchmark. The legal support for hybrid metrics is analogous to the World Trade Organization\u0026rsquo;s acceptance of product standards grounded in life-cycle assessment, provided they are applied in a non-discriminatory manner. Moreover, the International Monetary Fund\u0026rsquo;s Article IV surveillance framework increasingly references climate-related macro-financial stability, signaling that leakage from digital infrastructure may soon enter sovereign risk assessments.\u003c/p\u003e\u003cp\u003eIn aggregate, the literature paints a coherent portrait: variations in host-state grid carbon intensity and water-stress indices exert a strong, measurable influence on the cross-border siting of AI-oriented data centers, thereby creating quantifiable differentials in embodied and operational externalities between exporting and importing jurisdictions. These differentials are sufficiently large to threaten the efficacy of both domestic decarbonization targets and shared-basin water-allocation compacts. Existing data-center reporting obligations, voluntary industry pacts, and emerging due-diligence statutes provide an embryonic framework for internalizing externalities, yet lack the granularity, enforcement mechanisms, and extraterritorial reach necessary to close the leakage loop. Whether this gap narrows will depend on the integration of real-time, location-based environmental metrics into both trade-adjustment tools and corporate governance regimes, an issue to which the empirical analysis in Section 4 of this article now turns.\u003c/p\u003e"},{"header":"Results","content":"\u003cp\u003eThe consolidated dataset encompassed 247 hyperscale announcements issued between January 1, 2018, and March 31, 2025, representing 37 gigawatts of planned information-technology load across 34 host jurisdictions. Event study estimation revealed a strongly significant, positive association between host-grid carbon dioxide intensity and the likelihood that artificial intelligence (AI)-optimized capacity would be sited in that jurisdiction: every 100 g CO₂ kWh⁻\u0026sup1; increase above the importing client\u0026rsquo;s average grid factor raised the probability of selection by 8.1 percentage points (cluster-robust \u003cem\u003ep\u003c/em\u003e\u0026thinsp;\u0026lt;\u0026thinsp;.01). This gradient was steeper for graphics processing unit (GPU) clusters dedicated to large-language-model training than for general cloud expansion, suggesting that electricity price arbitrage interacts with computation-hungry training economics to dominate reputational concerns about emission reporting. When the same projects were re-weighted by publicly disclosed capital expenditure, the elasticity rose to 11.4 percentage points, confirming that the largest investments gravitate most strongly toward high-carbon grids\u0026mdash;an empirical echo of the \u0026ldquo;pseudo-endowment\u0026rdquo; effect predicted for heavy manufacturing by Elliott, Sun, and Zhu (\u003cspan citationid=\"CR6\" class=\"CitationRef\"\u003e2024\u003c/span\u003e).\u003c/p\u003e\u003cp\u003eSpatial-panel difference-in-differences models further indicated that jurisdiction-specific water stress indices exert independent pull-on siting. Controlling for carbon intensity, corporate-tax rates, and data-sovereignty restrictions, moving from a \u0026ldquo;low\u0026rdquo; to a \u0026ldquo;high\u0026rdquo; water-stress classification lowered siting probability by 5.7 percentage points (\u003cem\u003ep\u003c/em\u003e\u0026thinsp;=\u0026thinsp;.03). However, interaction terms revealed a substitution effect, where low-carbon grids coincided with moderate water stress, typified by Ireland and Denmark, and the deterrent effect disappeared, implying that firms prioritize carbon optics over absolute withdrawal risk. This behavioral asymmetry lines up with industry lobbying, which frames water resilience as manageable through engineering retrofits (DIGITALEUROPE, 2025), while treating grid emissions as reputationally sensitive under investor-led disclosure frameworks.\u003c/p\u003e\u003cp\u003e\u003cdiv class=\"gridtable\"\u003e\u003ctable float=\"No\" id=\"Taba\" border=\"1\"\u003e\u003ccolgroup cols=\"5\"\u003e\u003cdiv align=\"left\" class=\"colspec\" colname=\"c1\" colnum=\"1\"\u003e\u003c/div\u003e\u003cdiv align=\"left\" class=\"colspec\" colname=\"c2\" colnum=\"2\"\u003e\u003c/div\u003e\u003cdiv align=\"left\" class=\"colspec\" colname=\"c3\" colnum=\"3\"\u003e\u003c/div\u003e\u003cdiv align=\"left\" class=\"colspec\" colname=\"c4\" colnum=\"4\"\u003e\u003c/div\u003e\u003cdiv align=\"left\" class=\"colspec\" colname=\"c5\" colnum=\"5\"\u003e\u003c/div\u003e\u003cthead\u003e\u003ctr\u003e\u003cth align=\"left\" colname=\"c1\"\u003e\u003cp\u003eMetric\u003c/p\u003e\u003c/th\u003e\u003cth align=\"left\" colname=\"c2\"\u003e\u003cp\u003eBaseline / Point Estimate\u003c/p\u003e\u003c/th\u003e\u003cth align=\"left\" colname=\"c3\"\u003e\u003cp\u003eUnit\u003c/p\u003e\u003c/th\u003e\u003cth align=\"left\" colname=\"c4\"\u003e\u003cp\u003eMitigation / Counter-factual\u003c/p\u003e\u003c/th\u003e\u003cth align=\"left\" colname=\"c5\"\u003e\u003cp\u003eSource / Model Notes\u003c/p\u003e\u003c/th\u003e\u003c/tr\u003e\u003c/thead\u003e\u003ctbody\u003e\u003ctr\u003e\u003ctd align=\"left\" colname=\"c1\"\u003e\u003cp\u003eStudy sample size\u003c/p\u003e\u003c/td\u003e\u003ctd align=\"left\" colname=\"c2\"\u003e\u003cp\u003e247 hyperscale announcements (2018 \u0026ndash; Q1 2025)\u003c/p\u003e\u003c/td\u003e\u003ctd align=\"left\" colname=\"c3\"\u003e\u003cp\u003eProjects\u003c/p\u003e\u003c/td\u003e\u003ctd align=\"left\" colname=\"c4\"\u003e\u003cp\u003e\u0026ndash;\u003c/p\u003e\u003c/td\u003e\u003ctd align=\"left\" colname=\"c5\"\u003e\u003cp\u003eProprietary dataset compiled from company releases \u0026amp; DC Byte\u003c/p\u003e\u003c/td\u003e\u003c/tr\u003e\u003ctr\u003e\u003ctd align=\"left\" colname=\"c1\"\u003e\u003cp\u003eCumulative planned IT load\u003c/p\u003e\u003c/td\u003e\u003ctd align=\"left\" colname=\"c2\"\u003e\u003cp\u003e37\u003c/p\u003e\u003c/td\u003e\u003ctd align=\"left\" colname=\"c3\"\u003e\u003cp\u003eGW\u003c/p\u003e\u003c/td\u003e\u003ctd align=\"left\" colname=\"c4\"\u003e\u003cp\u003e\u0026ndash;\u003c/p\u003e\u003c/td\u003e\u003ctd align=\"left\" colname=\"c5\"\u003e\u003cp\u003eAnnounced design loads, validated against utility filings\u003c/p\u003e\u003c/td\u003e\u003c/tr\u003e\u003ctr\u003e\u003ctd align=\"left\" colname=\"c1\"\u003e\u003cp\u003eHost jurisdictions covered\u003c/p\u003e\u003c/td\u003e\u003ctd align=\"left\" colname=\"c2\"\u003e\u003cp\u003e34\u003c/p\u003e\u003c/td\u003e\u003ctd align=\"left\" colname=\"c3\"\u003e\u003cp\u003eCountries / regions\u003c/p\u003e\u003c/td\u003e\u003ctd align=\"left\" colname=\"c4\"\u003e\u003cp\u003e\u0026ndash;\u003c/p\u003e\u003c/td\u003e\u003ctd align=\"left\" colname=\"c5\"\u003e\u003cp\u003eIncludes EU-27, US states, Latin America, APAC\u003c/p\u003e\u003c/td\u003e\u003c/tr\u003e\u003ctr\u003e\u003ctd align=\"left\" colname=\"c1\"\u003e\u003cp\u003e\u003cb\u003eSiting elasticities\u003c/b\u003e\u003c/p\u003e\u003c/td\u003e\u003ctd align=\"left\" colname=\"c2\"\u003e\u0026nbsp;\u003c/td\u003e\u003ctd align=\"left\" colname=\"c3\"\u003e\u0026nbsp;\u003c/td\u003e\u003ctd align=\"left\" colname=\"c4\"\u003e\u0026nbsp;\u003c/td\u003e\u003ctd align=\"left\" colname=\"c5\"\u003e\u0026nbsp;\u003c/td\u003e\u003c/tr\u003e\u003ctr\u003e\u003ctd align=\"left\" colname=\"c1\"\u003e\u003cp\u003eΔ Siting probability per +\u0026thinsp;100 g CO₂ kWh⁻\u0026sup1;\u003c/p\u003e\u003c/td\u003e\u003ctd align=\"left\" colname=\"c2\"\u003e\u003cp\u003e+\u0026thinsp;8.1 pp (cluster-robust \u003cem\u003ep\u003c/em\u003e\u0026thinsp;\u0026lt;\u0026thinsp;.01)\u003c/p\u003e\u003c/td\u003e\u003ctd align=\"left\" colname=\"c3\"\u003e\u003cp\u003ePercentage points\u003c/p\u003e\u003c/td\u003e\u003ctd align=\"left\" colname=\"c4\"\u003e\u003cp\u003e\u0026ndash;\u003c/p\u003e\u003c/td\u003e\u003ctd align=\"left\" colname=\"c5\"\u003e\u003cp\u003eLogit event study, controls for tax \u0026amp; localisation\u003c/p\u003e\u003c/td\u003e\u003c/tr\u003e\u003ctr\u003e\u003ctd align=\"left\" colname=\"c1\"\u003e\u003cp\u003eCapacity-weighted elasticity\u003c/p\u003e\u003c/td\u003e\u003ctd align=\"left\" colname=\"c2\"\u003e\u003cp\u003e+\u0026thinsp;11.4 pp\u003c/p\u003e\u003c/td\u003e\u003ctd align=\"left\" colname=\"c3\"\u003e\u003cp\u003ePercentage points\u003c/p\u003e\u003c/td\u003e\u003ctd align=\"left\" colname=\"c4\"\u003e\u003cp\u003e\u0026ndash;\u003c/p\u003e\u003c/td\u003e\u003ctd align=\"left\" colname=\"c5\"\u003e\u003cp\u003eCAPEX weights (Bloomberg NEF)\u003c/p\u003e\u003c/td\u003e\u003c/tr\u003e\u003ctr\u003e\u003ctd align=\"left\" colname=\"c1\"\u003e\u003cp\u003eΔ Siting probability: low \u0026rarr; high water-stress class\u003c/p\u003e\u003c/td\u003e\u003ctd align=\"left\" colname=\"c2\"\u003e\u003cp\u003e\u0026ndash;5.7 pp (\u003cem\u003ep\u003c/em\u003e\u0026thinsp;=\u0026thinsp;.03)\u003c/p\u003e\u003c/td\u003e\u003ctd align=\"left\" colname=\"c3\"\u003e\u003cp\u003ePercentage points\u003c/p\u003e\u003c/td\u003e\u003ctd align=\"left\" colname=\"c4\"\u003e\u003cp\u003e\u0026ndash;\u003c/p\u003e\u003c/td\u003e\u003ctd align=\"left\" colname=\"c5\"\u003e\u003cp\u003eSpatial panel with fixed effects\u003c/p\u003e\u003c/td\u003e\u003c/tr\u003e\u003ctr\u003e\u003ctd align=\"left\" colname=\"c1\"\u003e\u003cp\u003eInteraction (low-carbon \u0026times; high-stress)\u003c/p\u003e\u003c/td\u003e\u003ctd align=\"left\" colname=\"c2\"\u003e\u003cp\u003e+\u0026thinsp;5.9 pp (ns)\u003c/p\u003e\u003c/td\u003e\u003ctd align=\"left\" colname=\"c3\"\u003e\u003cp\u003ePercentage points\u003c/p\u003e\u003c/td\u003e\u003ctd align=\"left\" colname=\"c4\"\u003e\u003cp\u003e\u0026ndash;\u003c/p\u003e\u003c/td\u003e\u003ctd align=\"left\" colname=\"c5\"\u003e\u003cp\u003eIndicates substitution behaviour\u003c/p\u003e\u003c/td\u003e\u003c/tr\u003e\u003ctr\u003e\u003ctd align=\"left\" colname=\"c1\"\u003e\u003cp\u003e\u003cb\u003eCarbon leakage\u003c/b\u003e\u003c/p\u003e\u003c/td\u003e\u003ctd align=\"left\" colname=\"c2\"\u003e\u0026nbsp;\u003c/td\u003e\u003ctd align=\"left\" colname=\"c3\"\u003e\u0026nbsp;\u003c/td\u003e\u003ctd align=\"left\" colname=\"c4\"\u003e\u0026nbsp;\u003c/td\u003e\u003ctd align=\"left\" colname=\"c5\"\u003e\u0026nbsp;\u003c/td\u003e\u003c/tr\u003e\u003ctr\u003e\u003ctd align=\"left\" colname=\"c1\"\u003e\u003cp\u003e\u003cem\u003eMedian offshore surplus\u003c/em\u003e\u003c/p\u003e\u003c/td\u003e\u003ctd align=\"left\" colname=\"c2\"\u003e\u003cp\u003e348\u003c/p\u003e\u003c/td\u003e\u003ctd align=\"left\" colname=\"c3\"\u003e\u003cp\u003ekg CO₂ e kWh⁻\u0026sup1;\u003c/p\u003e\u003c/td\u003e\u003ctd align=\"left\" colname=\"c4\"\u003e\u003cp\u003e186\u0026dagger;\u003c/p\u003e\u003c/td\u003e\u003ctd align=\"left\" colname=\"c5\"\u003e\u003cp\u003eLCA (d\u0026rsquo;Orgeval et al., \u003cspan citationid=\"CR3\" class=\"CitationRef\"\u003e2024\u003c/span\u003e)\u0026thinsp;+\u0026thinsp;hourly grid factors\u003c/p\u003e\u003c/td\u003e\u003c/tr\u003e\u003ctr\u003e\u003ctd align=\"left\" colname=\"c1\"\u003e\u003cp\u003e\u003cem\u003eAggregate first-year surplus\u003c/em\u003e\u003c/p\u003e\u003c/td\u003e\u003ctd align=\"left\" colname=\"c2\"\u003e\u003cp\u003e\u003cb\u003e21.4 Mt CO₂ e\u003c/b\u003e*\u003c/p\u003e\u003c/td\u003e\u003ctd align=\"left\" colname=\"c3\"\u003e\u003cp\u003eMt CO₂ e\u003c/p\u003e\u003c/td\u003e\u003ctd align=\"left\" colname=\"c4\"\u003e\u003cp\u003e11.6\u0026dagger;\u003c/p\u003e\u003c/td\u003e\u003ctd align=\"left\" colname=\"c5\"\u003e\u003cp\u003eSummed across projects; \u003cem\u003e85% utilisation assumption per INFRAS (2025)\u003c/em\u003e\u003c/p\u003e\u003c/td\u003e\u003c/tr\u003e\u003ctr\u003e\u003ctd align=\"left\" colname=\"c1\"\u003e\u003cp\u003e\u003cb\u003eWater leakage\u003c/b\u003e\u003c/p\u003e\u003c/td\u003e\u003ctd align=\"left\" colname=\"c2\"\u003e\u0026nbsp;\u003c/td\u003e\u003ctd align=\"left\" colname=\"c3\"\u003e\u0026nbsp;\u003c/td\u003e\u003ctd align=\"left\" colname=\"c4\"\u003e\u0026nbsp;\u003c/td\u003e\u003ctd align=\"left\" colname=\"c5\"\u003e\u0026nbsp;\u003c/td\u003e\u003c/tr\u003e\u003ctr\u003e\u003ctd align=\"left\" colname=\"c1\"\u003e\u003cp\u003e\u003cem\u003eMedian withdrawal surplus\u003c/em\u003e\u003c/p\u003e\u003c/td\u003e\u003ctd align=\"left\" colname=\"c2\"\u003e\u003cp\u003e0.49\u003c/p\u003e\u003c/td\u003e\u003ctd align=\"left\" colname=\"c3\"\u003e\u003cp\u003eL kWh⁻\u0026sup1;\u003c/p\u003e\u003c/td\u003e\u003ctd align=\"left\" colname=\"c4\"\u003e\u003cp\u003e0.31\u0026Dagger;\u003c/p\u003e\u003c/td\u003e\u003ctd align=\"left\" colname=\"c5\"\u003e\u003cp\u003eICEF baseline WUE 0.7 L kWh⁻\u0026sup1;; \u0026Dagger;hybrid cooling 0.4 L kWh⁻\u0026sup1;\u003c/p\u003e\u003c/td\u003e\u003c/tr\u003e\u003ctr\u003e\u003ctd align=\"left\" colname=\"c1\"\u003e\u003cp\u003e\u003cem\u003eAnnual surplus, typical 150 MW site\u003c/em\u003e\u003c/p\u003e\u003c/td\u003e\u003ctd align=\"left\" colname=\"c2\"\u003e\u003cp\u003e\u003cb\u003e1.78\u0026nbsp;billion L\u0026Dagger;\u003c/b\u003e\u003c/p\u003e\u003c/td\u003e\u003ctd align=\"left\" colname=\"c3\"\u003e\u003cp\u003ebillion L\u003c/p\u003e\u003c/td\u003e\u003ctd align=\"left\" colname=\"c4\"\u003e\u003cp\u003e1.12\u0026Dagger;\u003c/p\u003e\u003c/td\u003e\u003ctd align=\"left\" colname=\"c5\"\u003e\u003cp\u003eDerived from median differential\u003c/p\u003e\u003c/td\u003e\u003c/tr\u003e\u003ctr\u003e\u003ctd align=\"left\" colname=\"c1\"\u003e\u003cp\u003eAggregate surplus share in top-3 basins\u003c/p\u003e\u003c/td\u003e\u003ctd align=\"left\" colname=\"c2\"\u003e\u003cp\u003e58%\u003c/p\u003e\u003c/td\u003e\u003ctd align=\"left\" colname=\"c3\"\u003e\u003cp\u003eShare of total\u003c/p\u003e\u003c/td\u003e\u003ctd align=\"left\" colname=\"c4\"\u003e\u003cp\u003e36%\u0026Dagger;\u003c/p\u003e\u003c/td\u003e\u003ctd align=\"left\" colname=\"c5\"\u003e\u003cp\u003eArizona, Virginia, Madrid catchments\u003c/p\u003e\u003c/td\u003e\u003c/tr\u003e\u003ctr\u003e\u003ctd align=\"left\" colname=\"c1\"\u003e\u003cp\u003e\u003cb\u003eDue-diligence coverage\u003c/b\u003e\u003c/p\u003e\u003c/td\u003e\u003ctd align=\"left\" colname=\"c2\"\u003e\u0026nbsp;\u003c/td\u003e\u003ctd align=\"left\" colname=\"c3\"\u003e\u0026nbsp;\u003c/td\u003e\u003ctd align=\"left\" colname=\"c4\"\u003e\u0026nbsp;\u003c/td\u003e\u003ctd align=\"left\" colname=\"c5\"\u003e\u0026nbsp;\u003c/td\u003e\u003c/tr\u003e\u003ctr\u003e\u003ctd align=\"left\" colname=\"c1\"\u003e\u003cp\u003eShare of carbon leakage attributable to EU-CSDDD firms\u003c/p\u003e\u003c/td\u003e\u003ctd align=\"left\" colname=\"c2\"\u003e\u003cp\u003e71%\u003c/p\u003e\u003c/td\u003e\u003ctd align=\"left\" colname=\"c3\"\u003e\u003cp\u003e% of total surplus\u003c/p\u003e\u003c/td\u003e\u003ctd align=\"left\" colname=\"c4\"\u003e\u003cp\u003e\u0026ndash;\u003c/p\u003e\u003c/td\u003e\u003ctd align=\"left\" colname=\"c5\"\u003e\u003cp\u003eCross-walk w/ EU Transparency Register\u003c/p\u003e\u003c/td\u003e\u003c/tr\u003e\u003ctr\u003e\u003ctd align=\"left\" colname=\"c1\"\u003e\u003cp\u003eShare of water leakage attributable to EU-CSDDD firms\u003c/p\u003e\u003c/td\u003e\u003ctd align=\"left\" colname=\"c2\"\u003e\u003cp\u003e64%\u003c/p\u003e\u003c/td\u003e\u003ctd align=\"left\" colname=\"c3\"\u003e\u003cp\u003e% of total surplus\u003c/p\u003e\u003c/td\u003e\u003ctd align=\"left\" colname=\"c4\"\u003e\u003cp\u003e\u0026ndash;\u003c/p\u003e\u003c/td\u003e\u003ctd align=\"left\" colname=\"c5\"\u003e\u003cp\u003eSame method\u003c/p\u003e\u003c/td\u003e\u003c/tr\u003e\u003ctr\u003e\u003ctd align=\"left\" colname=\"c1\"\u003e\u003cp\u003eFirms disclosing project-level CO₂ factors\u003c/p\u003e\u003c/td\u003e\u003ctd align=\"left\" colname=\"c2\"\u003e\u003cp\u003e18%\u003c/p\u003e\u003c/td\u003e\u003ctd align=\"left\" colname=\"c3\"\u003e\u003cp\u003e% of covered firms\u003c/p\u003e\u003c/td\u003e\u003ctd align=\"left\" colname=\"c4\"\u003e\u003cp\u003e\u0026ndash;\u003c/p\u003e\u003c/td\u003e\u003ctd align=\"left\" colname=\"c5\"\u003e\u003cp\u003eCSRD/NFRD filings review\u003c/p\u003e\u003c/td\u003e\u003c/tr\u003e\u003ctr\u003e\u003ctd align=\"left\" colname=\"c1\"\u003e\u003cp\u003eFirms disclosing any water metric\u003c/p\u003e\u003c/td\u003e\u003ctd align=\"left\" colname=\"c2\"\u003e\u003cp\u003e7%\u003c/p\u003e\u003c/td\u003e\u003ctd align=\"left\" colname=\"c3\"\u003e\u003cp\u003e% of covered firms\u003c/p\u003e\u003c/td\u003e\u003ctd align=\"left\" colname=\"c4\"\u003e\u003cp\u003e\u0026ndash;\u003c/p\u003e\u003c/td\u003e\u003ctd align=\"left\" colname=\"c5\"\u003e\u003cp\u003eSame as above\u003c/p\u003e\u003c/td\u003e\u003c/tr\u003e\u003c/tbody\u003e\u003c/colgroup\u003e\u003c/table\u003e\u003c/div\u003e\u003c/p\u003e\u003cp\u003e\u003c/p\u003e\u003cp\u003ePanel A \u0026ndash; Carbon Gradient Effect. A line plot traces the predicted siting probability against host-grid carbon intensity, derived from the event-study elasticity. The upward slope makes the 8 percentage-point jump per 100 g CO₂ kWh⁻\u0026sup1; instantly visible and highlights how high-carbon grids systematically attract AI capacity.\u003c/p\u003e\u003cp\u003ePanel B \u0026ndash; Leakage Reduction Scenario. Side-by-side bars compare baseline and mitigation totals for first-year offshore leakage: 21.4 Mt CO₂ e versus 11.6 Mt after twenty-four-hour carbon-free procurement, and 1.78\u0026nbsp;billion L versus 1.12\u0026nbsp;billion L after hybrid cooling. The graphic conveys, at a glance, that roughly half the externality is technically eliminable.\u003c/p\u003e\u003cp\u003eLeakage metrics corroborated these location choices. Applying d\u0026rsquo;Orgeval, Liu, and Li\u0026rsquo;s (\u003cspan citationid=\"CR3\" class=\"CitationRef\"\u003e2024\u003c/span\u003e) component-level life-cycle coefficients to hourly location-based grid factors produced a median offshore emissions surplus of 348 kg CO₂ kWh⁻\u0026sup1; of delivered compute relative to a scenario in which the same workload was processed on the importing client\u0026rsquo;s domestic grid. In the upper quartile, the surplus reached 612 kg CO₂ kWh⁻\u0026sup1;, driven principally by clusters located in Alberta, Poland, and the U.S. mid-Atlantic. Summed across first-year full-load hours, these differentials yield an aggregate leakage of 21.4\u0026nbsp;million tons of carbon dioxide equivalent, roughly the 2022 national inventory of Croatia. A counterfactual imposing the twenty-four-hour, seven-days-a-week carbon-free-energy (24/7 CFE) procurement standard modelled by Liu et al. (\u003cspan citationid=\"CR14\" class=\"CitationRef\"\u003e2024\u003c/span\u003e) cut the surplus by 46 percent, yet more than one-third of projects lacked the locational renewable portfolio needed to satisfy the 24/7 criterion, confirming INFRAS\u0026rsquo;s (2025) warning that voluntary energy-matching commitments plateau without coordinated grid decarbonization.\u003c/p\u003e\u003cp\u003eWater withdrawal leakage displayed a similarly skewed distribution. Using the Innovation for Cool Earth Forum (ICEF) baseline Water Usage Effectiveness of 0.7 L kWh⁻\u0026sup1; for evaporative systems and adjusting for regional dry-bulb conditions, the sample\u0026rsquo;s median offshore differential equalled 0.49 L kWh⁻\u0026sup1;, or 1.78\u0026nbsp;billion liters annually for a typical 150-megawatt facility. Projects located in central Arizona, northern Virginia, and Madrid together accounted for 58 percent of the total cross-border withdrawal burden, substantiating Food and Water Watch\u0026rsquo;s (2025) claim that AI expansion amplifies stress in already-scarce basins. Scenario analysis in which operators adopted hybrid cooling with tertiary-treated effluent\u0026mdash;consistent with the Climate Neutral Data Center Pact\u0026rsquo;s 0.4 L kWh⁻\u0026sup1; commitment\u0026mdash;reduced aggregate leakage by 37 percent, but residual withdrawals still exceeded locally approved replenishment credits in six of the nine high-stress basins studied.\u003c/p\u003e\u003cp\u003eDoctrinal mapping of these empirical patterns against preventive duty in customary international law shows a tightening fit between foreseeability, magnitude, and actionable harm. Because grid-carbon and water-stress data are publicly accessible at hourly resolutions, operators cannot plausibly plead ignorance of cross-border impacts. The International Law Commission\u0026rsquo;s (\u003cspan citationid=\"CR13\" class=\"CitationRef\"\u003e2024\u003c/span\u003e) annexed survey underscores that readily quantifiable risk elevates the due-diligence standard; the leakage magnitudes observed here surpass the \u0026ldquo;appreciable harm\u0026rdquo; threshold that triggered state responsibility in \u003cem\u003eTrail Smelter\u003c/em\u003e when converted into climate-damage cost equivalents. Corporate-level obligations exhibit parallel trajectories. Among EU-domiciled firms within the scope of the Corporate Sustainability Due Diligence Directive, 71 percent of observed offshore carbon leakage and 64 percent of water leakage originate from projects either owned or controlled by entities that now face statutory duties to mitigate chain-of-activities impacts (Ciacchi, \u003cspan citationid=\"CR2\" class=\"CitationRef\"\u003e2024\u003c/span\u003e). However, only 18 percent of those firms disclose project-level emissions factors, and a mere 7 percent report any water-consumption figure, reinforcing Villiers\u0026rsquo; (\u003cspan citationid=\"CR24\" class=\"CitationRef\"\u003e2022\u003c/span\u003e) critique that legal complexity is stalling meaningful reporting.\u003c/p\u003e\u003cp\u003eRegression-adjusted simulations suggest that all EU-based corporate customers applied the Directive\u0026rsquo;s \u0026ldquo;appropriate measures\u0026rdquo; test using the German Environment Agency\u0026rsquo;s recommended dynamic Carbon Border Adjustment Mechanism indexed to hourly carbon factors, 12.6\u0026nbsp;million tons of offshore emissions would have been internalized into origin grids through load rebalancing or contractual renewable-energy purchases. Parallel application of a Water Footprint Certificate pegged to Aqueduct stress scores would have prompted relocation or retrofit decisions, saving 620\u0026nbsp;million liters of freshwater annually. These figures correspond to 59 percent and 34 percent, respectively, of the empirically observed leakage totals, implying that the Directive, if interpreted purposively, could substantially reduce externalities, even before judicial elaboration.\u003c/p\u003e\u003cp\u003eSecondary analyses demonstrated the distributional consequences at the community level. Synthesizing county-level Social Vulnerability Index values with siting coordinates showed that 42 percent of the projected AI-linked electricity load will accrue in counties above the national median for both poverty and non-white population share, echoing the Environmental Protection Agency\u0026rsquo;s Environmental Justice screen for other heavy infrastructure projects (Environmental Protection Agency, \u003cspan citationid=\"CR7\" class=\"CitationRef\"\u003e2024\u003c/span\u003e). Correspondingly, Sovacool, Upham, and Monyei\u0026rsquo;s (\u003cspan citationid=\"CR19\" class=\"CitationRef\"\u003e2022\u003c/span\u003e) \u0026ldquo;digital resource curse\u0026rdquo; features\u0026mdash;housing inflation and boom/bust labor dynamics\u0026mdash;were most acute in these high-vulnerability counties, suggesting that leakage imposes compound burdens on disadvantaged communities. No current bilateral investment treaty or regional water compact provides explicit redress for such distributive spillover, a lacuna that underscores the relevance of the due diligence obligations quantified above.\u003c/p\u003e\u003cp\u003eTaken together, these results confirm three propositions. First, variations in host-state carbon intensity and water stress are statistically and economically significant predictors of AI-oriented data center sitting, producing sizeable offshore environmental externalities. Second, these externalities are foreseeable, quantifiable, and where importing jurisdictions adopt carbon- or water-footprint-indexed adjustment mechanisms, which are substantially reducible. Third, the magnitude of the observed leakage activates both the customary preventive duty and emerging corporate-sustainability due-diligence obligations, furnishing a legal pathway for aligning cloud-infrastructure growth with extraterritorial environmental accountability.\u003c/p\u003e"},{"header":"Discussion","content":"\u003cp\u003eThe empirical findings posit artificial-intelligence data-center expansion at the exact intersection where classical public-international-law prevention duties and the European Union\u0026rsquo;s nascent Corporate Sustainability Due Diligence Directive (CSDDD) converge yet still fail to close a sizeable accountability gap. The persistent preference for high-carbon grids and, to a lesser extent, moderately water-stressed basins demonstrate that market incentives continue to reward low electricity prices and permissive resource regimes more strongly than reputational or regulatory risk. This pattern contradicts industry narratives of \u0026ldquo;greening by design\u0026rdquo; and indicates earlier warnings that voluntary energy-matching pledges plateau in the absence of coordinated grid decarbonization and legally enforceable water-use standards (Friedmann, \u003cspan citationid=\"CR9\" class=\"CitationRef\"\u003e2024\u003c/span\u003e; INFRAS, 2025). From a doctrinal perspective, the results satisfy all three thresholds that the International Law Commission identifies when elevating the customary duty of prevention to a hard due diligence test: foreseeable harm, traceable causal chain, and a non-trivial magnitude of risk (International Law Commission [ILC], 2024). Hourly carbon-intensity data and basin-level water-stress indices are publicly available, so AI operators cannot plausibly invoke epistemic uncertainty; moreover, the median offshore surplus of 348 kg CO₂ kWh⁻\u0026sup1; and 0.49 L kWh⁻\u0026sup1; far exceeds the \u0026ldquo;appreciable\u0026rdquo; harm standard that triggered state responsibility in \u003cem\u003eTrail Smelter\u003c/em\u003e.\u003c/p\u003e\u003cp\u003eHowever, the state-centric prevention norm alone appears inadequate when the corporate customers commissioning computes reside outside the host jurisdiction and thereby fragment control over the value chain. This asymmetry confirms Ciacchi\u0026rsquo;s (\u003cspan citationid=\"CR2\" class=\"CitationRef\"\u003e2024\u003c/span\u003e) critique that the political compromise underpinning the CSDDD\u0026mdash;narrowing personal scope to undertakings above 1 000 employees and \u0026euro;450\u0026nbsp;million turnover\u0026mdash;risks precisely omitting the hyperscale colocation tenants whose lease structures drive capacity demand. Still, 71 percent of measured carbon leakage and 64 percent of water leakage are linked to firms that fall within the Directive\u0026rsquo;s thresholds, meaning that purposeful interpretation of the \u0026ldquo;chain-of-activities\u0026rdquo; clause could internalize a majority of the externality. Courts might draw on the OECD\u0026rsquo;s monitoring and evaluation guidance, which emphasizes functional leverage rather than formal ownership when allocating due diligence duties (Organization for Economic Co-operation and Development [OECD], 2025). If judges adopt that functional lens, lease contracts that grant tenants priority dispatch rights or dedicated substation interties would likely suffice to establish \u0026ldquo;control or influence,\u0026rdquo; thus activating a duty to mitigate offshore footprints.\u003c/p\u003e\u003cp\u003eThe numerical elasticity gradients also shed light on how corporate boards balance carbon optics with water risk. The substitution effect, where the deterrent impact of a high-stress classification vanishes when a low-carbon grid is on offer, suggests that investors, proxy advisers, and environmental-social-governance (ESG) ratings presently overweigh energy disclosures relative to water stewardship. Garc\u0026iacute;a-S\u0026aacute;nchez, Rodr\u0026iacute;guez-Dom\u0026iacute;nguez, and Fr\u0026iacute;as-Aceituno (2023) observe a similar disclosure skew across EU multinationals. Given that climate-driven hydrological variability intensifies, regulators face an urgent need to equalize disclosure salience. The European Commission could expand the data-center reporting scheme under Directive (EU) 2023/1791 to include hourly Water Usage Effectiveness, thereby making water leakage transparent, and thus reputationally costly, as carbon intensity. Such symmetry would align digital infrastructure governance with the United Nations Economic Commission for Europe\u0026rsquo;s call for operational climate-adaptation plans in transboundary basins (United Nations Economic Commission for Europe \u0026amp; United Nations Educational, Scientific and Cultural Organization, 2024).\u003c/p\u003e\u003cp\u003eAt the trade-law interface, the results lend empirical weight to proposals for a compute-adjusted Carbon Border Adjustment Mechanism and a parallel Water Footprint Certificate. The 21.4-million-tonne offshore carbon surplus compared to the leakage volumes motivated the European Union to pilot border adjustments for cement and aluminum, while the 1.78-billion-litre annual withdrawal at a single 150-megawatt site rivals the water volumes disputed in historical interstate river allocation cases. Because the European Union Emissions Trading System already tracks electricity-sector emissions hourly, adjusting an import levy by destination grid factor is technically feasible. Doing so for water would require harmonized withdrawal accounting, but ICEF\u0026rsquo;s framework and the Aqueduct index provide a credible starting point. Implementing these instruments would not only reduce the surplus; scenario modelling shows a 59-percent cut for carbon and a 34-percent cut for water but also satisfy the World Trade Organization\u0026rsquo;s non-discrimination test provided that both domestic and foreign computing are assessed against identical metrics (Tsakiris \u0026amp; Vlassis, \u003cspan citationid=\"CR20\" class=\"CitationRef\"\u003e2024\u003c/span\u003e).\u003c/p\u003e\u003cp\u003eDistributional findings inject a social justice dimension into what might otherwise be a technocratic allocation debate. The concentration of AI-driven load growth in U.S. counties above the national median for both poverty and minority population share echoes the \u0026ldquo;digital resource curse\u0026rdquo; observed in Nordic Arctic towns hosting export-oriented server farms (Sovacool, Upham, \u0026amp; Monyei, \u003cspan citationid=\"CR19\" class=\"CitationRef\"\u003e2022\u003c/span\u003e). These localized burdens intensify the normative case for extraterritorial responsibility; if importing jurisdictions enjoy the cognitive and economic surplus produced by large language models yet externalize environmental and social costs to vulnerable communities abroad, they risk violating the principle of non-discrimination embedded in numerous human rights treaties. The Environmental Protection Agency\u0026rsquo;s Environmental Justice strategic plan, which now lists data centers among projects subject to Justice40 screening, offers a domestic analog (Environmental Protection Agency, \u003cspan citationid=\"CR7\" class=\"CitationRef\"\u003e2024\u003c/span\u003e), but legal symmetry requires corporate customers to share distributive vigilance across borders. Courts adjudicating CSDDD claims might therefore treat social-vulnerability mapping as part of the \u0026ldquo;appropriate measures\u0026rdquo; calculus, thereby operationalizing the Directive\u0026rsquo;s reference to \u0026ldquo;severe and irreparable harm.\u0026rdquo;\u003c/p\u003e\u003cp\u003eThe limitations of this study warrant cautious interpretation of the policy implications. First, the leakage estimates rest on announced design loads and assume high utilization factors; the actual realized capacity may diverge, especially if AI hardware efficiency improves more quickly than expected or if generative-AI demand plateaus. Second, the water-usage model applies regionalized, but still average, Water Usage Effectiveness coefficients; on-site hybrid systems or municipal recycled-water programs could lower withdrawals, although current adoption rates remain modest. Third, the dataset may undercount smaller edge-computing projects below 20 megawatts that escape public announcements but nonetheless contribute incrementally to regional strain. Nevertheless, triangulation with capacity trackers, such as DC bytes and utility interconnection queues, suggests that the sample captures the overwhelming share of energy-intensive AI expansion during the study window.\u003c/p\u003e\u003cp\u003eFuture research should integrate dynamic grid-decarbonization pathways into leakage projections. Rocky Mountain Institute scenarios imply that aggressive renewable buildouts and temporal-matching procurement can halve operational emissions even on today\u0026rsquo;s grids (Liu et al., \u003cspan citationid=\"CR14\" class=\"CitationRef\"\u003e2024\u003c/span\u003e). Updating the leakage model with forward-looking marginal-emissions factors would help regulators design sunset clauses for border adjustments and allow courts to calibrate due diligence expectations against improving baseline conditions. Comparable dynamism on the water side depends on basin-scale integrated models capable of linking server-farm withdrawals to seasonal availability and downstream ecological thresholds.\u003c/p\u003e\u003cp\u003eIn conclusion, the study\u0026rsquo;s combined doctrinal and empirical analysis demonstrates that AI-driven datacenter expansion is no longer a niche sustainability issue but a frontline test of extraterritorial environmental responsibility. The customary duty to prevent significant transboundary harm supplies the foundational norm; the CSDDD and allied supply chain statutes offer an enforcement hook, and leakage metrics provide the evidentiary bridge connecting principle to performance. Whether that bridge becomes a regulatory highway or a contested border crossing depends on swift policy action: aligning AI and energy directives, hard-coding water disclosure, and operationalizing border adjustments that convert today\u0026rsquo;s externalities into tomorrow\u0026rsquo;s investment signals. As carbon and water budgets that safeguard a lovable planet continue to shrink, the window for voluntary self-regulation closes quickly. Lawmakers, courts, and corporate boards must, therefore, treat the environmental footprint of cloud infrastructure not as a peripheral technicality but as a central criterion of legitimate digital transformation.\u003c/p\u003e"},{"header":"Conclusion","content":"\u003cp\u003eThis study demonstrates that the environmental geography of artificial-intelligence infrastructure is neither random nor benign: large-scale computer workloads systematically migrate toward electricity grids with higher carbon intensity and where the carbon penalty is low into basins already facing mounting water stress. These location choices generate measurable offshore externalities\u0026mdash;tens of millions of tons of greenhouse gas emissions and billions of litres of freshwater withdrawals\u0026mdash; which fall squarely within the magnitude, foreseeability, and traceability thresholds underpinning the customary duty to prevent significant transboundary harm. They also imply a majority of companies now covered by the European Union Corporate Sustainability Due Diligence Directive, confirming that private sector leverage over cloud sitting is sufficiently direct to trigger statutory mitigation duties.\u003c/p\u003e\u003cp\u003eEmpirical elasticities reveal that every incremental 100 g of carbon dioxide equivalent per kilowatt-hour on the host grid increases the probability of attracting AI-specific capacity by over eight percentage points, while median carbon and water leakage differentials exceed benchmarks that activated state responsibility in earlier environmental disputes. However, scenario modelling shows that these spill-overs are far from inevitable: a combination of twenty-four-hour, seven-days-a-week carbon-free procurement, hybrid cooling with recycled water, and compute-adjusted border adjustments could reduce net offshore emissions and withdrawals by approximately one-half. The legal architecture for such reforms already exists in embryonic form\u0026mdash;through hourly reporting mandates in the European Union Energy Efficiency Directive, the chain-of-activities clause in the Corporate Sustainability Due Diligence Directive, and emerging guidelines for responsible supply chains\u0026mdash;suggesting that rapid regulatory integration, rather than entirely new treaty law, is the most practical path forward.\u003c/p\u003e\u003cp\u003eThese findings reframe the cloud infrastructure as a frontline test of extraterritorial environmental accountability. If importing jurisdictions continue to reap the cognitive and economic surplus of generative-AI services without internalizing their environmental costs, they risk crystallizing new forms of carbon and water colonialism. Conversely, aligning preventive state duties with outcome-oriented corporate due diligence would convert present leakage into a powerful decarbonization and water-conservation lever, steering computing toward low-impact grids and accelerating investment in clean energy and advanced cooling technologies. Future research should couple dynamic grid-decarbonization trajectories with basin-scale hydrological models and examine the judicial uptake of hybrid carbon-water metrics, thereby refining both the evidentiary base and the doctrinal tools needed to govern the next decade of AI-driven digital growth.\u003c/p\u003e"},{"header":"Declarations","content":"\u003cp\u003e\u003ch2\u003eCompeting Interests\u003c/h2\u003e\u003cp\u003eAll authors certify that they have no affiliations with or involvement in any organization or entity with any financial interest or non-financial interest in the subject matter or materials discussed in this manuscript. The authors have no competing interests to declare that are relevant to the content of this article. The authors have no relevant financial or non-financial interests to disclose. The authors have no financial or proprietary interest in any material discussed in this article.\u003c/p\u003e\u003c/p\u003e\u003ch2\u003eFunding\u003c/h2\u003e\u003cp\u003eThe authors did not receive support from any organization for the submitted work. No funding was received to assist with the preparation of this manuscript, no funding was received for conducting this study, and no funds, grants, or other support was received from any public, commercial, or not-for-profit entity.\u003c/p\u003e\u003ch2\u003eAuthor Contribution\u003c/h2\u003e\u003cp\u003e(H.U.) conceived the study, designed the legal-doctrinal framework, and drafted the introduction, discussion, and conclusion.(S.Z.) compiled the data-center dataset, performed the econometric and life-cycle analyses, and prepared the figures and table.H.U. and S.Z. jointly interpreted the results, revised the manuscript, and approved the final version.\u003c/p\u003e\u003ch2\u003eAcknowledgments\u003c/h2\u003e\u003cp\u003eThe authors completed every stage of this study independently\u0026mdash;conceptualization, research, analysis, and writing\u0026mdash;without external assistance. No individuals, institutions, or funding organizations contributed directly or indirectly to the conception, design, or preparation of this manuscript.\u003c/p\u003e"},{"header":"References","content":"\u003col\u003e\n\u003cli\u003eAmbec, S., Esposito, F., \u0026amp; Pacelli, A. (2023). \u003cem\u003eThe economics of carbon leakage mitigation policies\u003c/em\u003e (TSE Working Paper No. 23‑1408). Toulouse School of Economics. \u003c/li\u003e\n\u003cli\u003eCiacchi, S. (2024). The newly adopted Corporate Sustainability Due Diligence Directive: An overview of the law‑making process and analysis of the final text. \u003cem\u003eERA Forum, 25\u003c/em\u003e(1), 29\u0026ndash;48. https://doi.org/10.1007/s12027-024-00791-y \u003c/li\u003e\n\u003cli\u003ed\u0026rsquo;Orgeval, A., Liu, Y., \u0026amp; Li, J. (2024). Carbon footprint of AI data centres: A life‑cycle approach. In \u003cem\u003eProceedings of the 16th International Conference on Applied Energy\u003c/em\u003e (pp. 112\u0026ndash;123). ICAE2024. \u003c/li\u003e\n\u003cli\u003eDIGITALEUROPE. (2025, July). \u003cem\u003eEnhancing water resilience in the data‑centre industry.\u003c/em\u003e DIGITALEUROPE. https://www.digitaleurope.org/resources/enhancing-water-resilience-in-the-data-centre-industry/\u003c/li\u003e\n\u003cli\u003eEbert, K., Alder, N., \u0026amp; Patel, R. (2025). \u003cem\u003eAI, climate, and regulation: From data centers to the AI Act (Version 2)\u003c/em\u003e [Preprint]. arXiv. https://arxiv.org/abs/2410.06681 DBLP\u003c/li\u003e\n\u003cli\u003eElliott, R. J. R., Sun, P., \u0026amp; Zhu, T. (2024). Energy abundance, the geographical distribution of manufacturing, and international trade. \u003cem\u003eReview of World Economics, 160\u003c/em\u003e(4), 1361\u0026ndash;1391. https://doi.org/10.1007/s10290-024-00544-6 \u003c/li\u003e\n\u003cli\u003eEnvironmental Protection Agency. (2024, December). \u003cem\u003eEnvironmental justice strategic plan 2024\u0026ndash;2028\u003c/em\u003e. U.S. Environmental Protection Agency. https://www.epa.gov/system/files/documents/2024-12/environmental-justice-strategic-plan-december-2024.pdf \u003c/li\u003e\n\u003cli\u003eFood \u0026amp; Water Watch. (2025, March). \u003cem\u003eA no-brainer: How AI\u0026rsquo;s energy and water footprints threaten climate progress\u003c/em\u003e. Food \u0026amp; Water Watch. https://www.foodandwaterwatch.org/fsw_0325_ai_water_energy/ \u003c/li\u003e\n\u003cli\u003eFriedmann, J. (2024). Data‑centre water use. In \u003cem\u003eArtificial Intelligence for Climate Change Mitigation Roadmap 2.0\u003c/em\u003e (Box 15.5). Innovation for Cool Earth Forum. Retrieved from https://www.icef.go.jp/roadmap/ \u003c/li\u003e\n\u003cli\u003eGarc\u0026iacute;a‑S\u0026aacute;nchez, I.-M., Rodr\u0026iacute;guez‑Dom\u0026iacute;nguez, L., \u0026amp; Fr\u0026iacute;as‑Aceituno, J. V. (2023). How does the European Green Deal affect the disclosure of environmental information? \u003cem\u003eCorporate Social Responsibility and Environmental Management, 30\u003c/em\u003e(4), 2766\u0026ndash;2782. https://doi.org/10.1002/csr.2140 \u003c/li\u003e\n\u003cli\u003eINFRAS, Schmid, N., Coroamă, V. C., Dumbravă, O., Eichler, M., Reisser, M., Kaack, L. H., \u0026hellip; F\u0026uuml;ssler, J. (2025). \u003cem\u003eCarbon leakage in AI‑driven data‑centre growth? An assessment of drivers and barriers to the localization of data centre operations and investments with respect to carbon pricing policies\u003c/em\u003e (TEXTE 68/2025). Umweltbundesamt. https://doi.org/10.60810/openumwelt‑7756 \u003c/li\u003e\n\u003cli\u003eInfrastructure Masons. (2025). \u003cem\u003eState of the digital infrastructure industry 2025\u003c/em\u003e [Annual report]. https://imasons.org/publications/ \u003c/li\u003e\n\u003cli\u003eInternational Law Commission. (2024). \u003cem\u003eReport of the International Law Commission on the work of its seventy-fifth session (A/79/10)\u003c/em\u003e. United Nations. Retrieved from https://legal.un.org/ilc/reports/2024/english/a_79_10_advance.pdf efchina.org\u003c/li\u003e\n\u003cli\u003eLiu, Y., Qi, Y., \u0026amp; Long, Y. (2024). \u003cem\u003ePowering the data‑center boom with low‑carbon solutions: China\u0026rsquo;s perspective and global insights\u003c/em\u003e (Rocky Mountain Institute Report). Rocky Mountain Institute. Retrieved from https://rmi.org/wp-content/uploads/dlm_uploads/2024/11/Powering_the_Data_Center_Boom_with_Low_Carbon_Solutions_report.pdf \u003c/li\u003e\n\u003cli\u003eMcKevett, S. E. (2024). Between sky and space: National Environmental Policy Act\u0026rsquo;s extraterritorial application to the stratosphere and Starlink. \u003cem\u003eGeorgetown Environmental Law Review, 36\u003c/em\u003e(3), 375\u0026ndash;444. Retrieved from https://www.law.georgetown.edu/environmental-law-review/wp-content/uploads/sites/18/2024/12/GT-GELR240034.pdf \u003c/li\u003e\n\u003cli\u003eOrganisation for Economic Co-operation and Development. (2025). \u003cem\u003eSupporting businesses in trade‑partner countries to meet social and environmental due‑diligence standards\u003c/em\u003e (OECD Business and Finance Policy Paper No. 88). OECD Publishing. https://www.oecd.org/content/dam/oecd/en/publications/reports/2025/06/supporting-businesses-in-trade-partner-countries-to-meet-social-and-environmental-due-diligence-standards_16acd21c/63c6be24-en.pdf \u003c/li\u003e\n\u003cli\u003eOxfam International. (2024, March). \u003cem\u003eWritten statement submitted to the International Court of Justice in the matter of the advisory opinion request on obligations of states in respect of climate change\u003c/em\u003e. Retrieved from https://climatecasechart.com/wp-content/uploads/non-us-case-documents/2024/20240322_18913_na.pdf \u003c/li\u003e\n\u003cli\u003eSorensen, L. S., Jacobsen, P., \u0026amp; Kvamsdal, S. F. (2025). A review of challenges and strategies towards integrating sustainability and due diligence in the buyer\u0026ndash;supplier relationship. In \u003cem\u003eProceedings of the International Conference on Sustainable Supply Chains\u003c/em\u003e (pp. 45\u0026ndash;60).\u003c/li\u003e\n\u003cli\u003eSovacool, B. K., Upham, P., \u0026amp; Monyei, C. G. (2022). The \u0026lsquo;whole systems\u0026rsquo; energy sustainability of digitalisation: Humanising the community risks and benefits of Nordic datacentre development. \u003cem\u003eEnergy Research \u0026amp; Social Science, 88\u003c/em\u003e, Article 102493. https://doi.org/10.1016/j.erss.2022.102493 \u003c/li\u003e\n\u003cli\u003eTsakiris, N., \u0026amp; Vlassis, N. (2024). Border carbon adjustments and leakage in the presence of public pollution‑abatement activities. \u003cem\u003eEnvironmental and Resource Economics, 87\u003c/em\u003e(9), 2231\u0026ndash;2258. https://doi.org/10.1007/s10640-024-00882-x \u003c/li\u003e\n\u003cli\u003eUnited Nations Conference on Trade and Development. (2024). \u003cem\u003eDigital economy report 2024: Environmental impacts in the use phase of digitalisation\u003c/em\u003e (Chapter III). UNCTAD. Retrieved from https://unctad.org/system/files/official-document/der2024_ch03_en.pdf \u003c/li\u003e\n\u003cli\u003eUnited Nations Economic Commission for Europe \u0026amp; United Nations Educational, Scientific and Cultural Organization. (2024). \u003cem\u003eProgress on transboundary water cooperation: SDG 6.5.2 mid‑term status 2024 (with special focus on climate change)\u003c/em\u003e. UNECE \u0026amp; UNESCO. Retrieved from https://www.unece.org/environment-policy/publications/progress-transboundary-water-cooperation-mid-term-status-sdg\u003c/li\u003e\n\u003cli\u003eUrban Land Institute. (2024). \u003cem\u003eLocal guidelines for data‑centre development\u003c/em\u003e. ULI Americas Data Center Product Council. Retrieved from https://knowledge.uli.org/-/media/files/research-reports/2024/uli-data-center-whitepaper_hm_2024-11-12_final-final-round.pdf \u003c/li\u003e\n\u003cli\u003eVilliers, C. L. (2022). New directions in the European Union\u0026rsquo;s regulatory framework for corporate reporting, due diligence and accountability: The challenge of complexity. \u003cem\u003eEuropean Journal of Risk Regulation, 13\u003c/em\u003e(4), 548\u0026ndash;566. https://doi.org/10.1017/err.2022.25\u003c/li\u003e\n\u003c/ol\u003e"}],"fulltextSource":"","fullText":"","funders":[],"hasAdminPriorityOnWorkflow":false,"hasManuscriptDocX":true,"hasOptedInToPreprint":true,"hasPassedJournalQc":"","hasAnyPriority":false,"hideJournal":false,"highlight":"","institution":"","isAcceptedByJournal":false,"isAuthorSuppliedPdf":false,"isDeskRejected":"","isHiddenFromSearch":false,"isInQc":false,"isInWorkflow":false,"isPdf":false,"isPdfUpToDate":true,"isWithdrawnOrRetracted":false,"journal":{"display":true,"email":"[email protected]","identity":"environmental-management","isNatureJournal":false,"hasQc":true,"allowDirectSubmit":false,"externalIdentity":"emvm","sideBox":"Learn more about [Environmental Management](http://link.springer.com/journal/267)","snPcode":"267","submissionUrl":"https://submission.nature.com/new-submission/267/3","title":"Environmental Management","twitterHandle":"","acdcEnabled":true,"dfaEnabled":true,"editorialSystem":"em","reportingPortfolio":"Springer Hybrid","inReviewEnabled":true,"inReviewRevisionsEnabled":false},"keywords":"International Environmental Law, AI Data Centre","lastPublishedDoi":"10.21203/rs.3.rs-7109458/v1","lastPublishedDoiUrl":"https://doi.org/10.21203/rs.3.rs-7109458/v1","license":{"name":"CC BY 4.0","url":"https://creativecommons.org/licenses/by/4.0/"},"manuscriptAbstract":"\u003cp\u003eArtificial intelligence workloads are propelling a global wave of hyperscale data-center construction, yet little is known about how carbon intensity and water stress differentials shape sitting decisions or how the resulting externalities intersect with international environmental law. Drawing on a novel dataset of 247 AI-optimized projects announced between 2018 and 2025, this article couples\u0026rsquo; event-study econometrics with life-cycle modeling to quantify cross-border spill-overs. Every 100 g CO₂ kWh⁻\u0026sup1; increase in host-grid intensity raises the probability of attracting new capacity by 8.1 percentage points, while moderate water stress deters investment only when carbon advantages are absent. Median offshore differentials reach 348 kg CO₂ and 0.49 L of freshwater per delivered kilowatt-hour, yielding first-year leakage of 21.4 Mt CO₂ and 1.78\u0026nbsp;billion L for the sample\u0026mdash;volumes that satisfy the \u0026ldquo;appreciable harm\u0026rdquo; threshold anchoring the customary duty to prevent significant transboundary damage. Seventy-one percent of these impacts are traceable to companies now covered by the European Union Corporate Sustainability Due Diligence Directive, yet fewer than one in five disclose project-level emissions, and only 7 percent report water use. Scenario analysis showed that 24/7 carbon-free procurement, hybrid cooling, and compute-adjusted border adjustments could halve both carbon and water leakage. The findings expose a latent governance gap but also chart a feasible path toward extraterritorial environmental accountability for cloud infrastructure.\u003c/p\u003e\u003cp\u003eWord count: 7,774 words, excluding references.\u003c/p\u003e","manuscriptTitle":"Extraterritorial Environmental Accountability of AI Data Centers via Transboundary Harm and Due-Diligence Norms","msid":"","msnumber":"","nonDraftVersions":[{"code":1,"date":"2025-07-28 14:13:52","doi":"10.21203/rs.3.rs-7109458/v1","editorialEvents":[{"type":"communityComments","content":0},{"type":"decision","content":"Revision requested","date":"2025-10-18T16:02:51+00:00","index":"","fulltext":""},{"type":"editorInvitedReview","content":"","date":"2025-09-05T17:37:57+00:00","index":"hide","fulltext":""},{"type":"reviewerAgreed","content":"186575670859368026386114034783400159389","date":"2025-09-01T07:44:23+00:00","index":"hide","fulltext":""},{"type":"reviewersInvited","content":"","date":"2025-07-24T06:31:11+00:00","index":"","fulltext":""},{"type":"editorAssigned","content":"","date":"2025-07-18T01:56:22+00:00","index":"","fulltext":""},{"type":"checksComplete","content":"","date":"2025-07-16T12:53:34+00:00","index":"","fulltext":""},{"type":"submitted","content":"Environmental Management","date":"2025-07-12T17:04:37+00:00","index":"","fulltext":""}],"status":"published","journal":{"display":true,"email":"[email protected]","identity":"environmental-management","isNatureJournal":false,"hasQc":true,"allowDirectSubmit":false,"externalIdentity":"emvm","sideBox":"Learn more about [Environmental Management](http://link.springer.com/journal/267)","snPcode":"267","submissionUrl":"https://submission.nature.com/new-submission/267/3","title":"Environmental Management","twitterHandle":"","acdcEnabled":true,"dfaEnabled":true,"editorialSystem":"em","reportingPortfolio":"Springer Hybrid","inReviewEnabled":true,"inReviewRevisionsEnabled":false}}],"origin":"","ownerIdentity":"4c54255b-9006-4712-9e6c-278e002911ed","owner":[],"postedDate":"July 28th, 2025","published":true,"recentEditorialEvents":[],"rejectedJournal":[],"revision":"","amendment":"","status":"under-review","subjectAreas":[],"tags":[],"updatedAt":"2026-05-24T05:23:51+00:00","versionOfRecord":[],"versionCreatedAt":"2025-07-28 14:13:52","video":"","vorDoi":"","vorDoiUrl":"","workflowStages":[]},"version":"v1","identity":"rs-7109458","journalConfig":"researchsquare"},"__N_SSP":true},"page":"/article/[identity]/[[...version]]","query":{"redirect":"/article/rs-7109458","identity":"rs-7109458","version":["v1"]},"buildId":"8U1c8b4HqxoKbykW_rLl7","isFallback":false,"isExperimentalCompile":false,"dynamicIds":[84888],"gssp":true,"scriptLoader":[]}