The study’s empirical foundation is a novel dataset of 247 AI-optimized data-center projects announced between 2018 and 2025. Rather than treating these projects as a uniform wave of construction, the authors distinguish between two fundamentally different kinds of computing workloads. AI training, which involves teaching large models on massive datasets, can tolerate delays and can therefore be located almost anywhere with cheap electricity. AI inference, the process of running trained models in response to user queries, is latency-sensitive and must sit close to fiber-optic backbone networks and end users. This bifurcation, the authors argue, is propelling a global wave of hyperscale construction, and it creates a structural incentive: the training workloads, which consume the most energy, are precisely the ones most free to migrate toward dirty grids.

The econometric results quantify that incentive with striking precision. Using event-study methods, the researchers found that every 100 grams of CO2 per kilowatt-hour increase in a host country’s grid carbon intensity raises the probability of that country attracting new AI data-center capacity by 8.1 percentage points. In other words, the dirtier the local electricity mix, the more attractive the jurisdiction becomes for new construction. Water tells a subtler story: moderate water stress deters investment, but only when carbon advantages are absent. Where a location offers sufficiently cheap and abundant power, developers appear willing to accept significant water risks, a trade-off with serious implications for regions already facing scarcity.

To translate these siting decisions into physical impacts, the team coupled their econometric analysis with life-cycle modeling of what they call offshore emissions surpluses, the additional carbon and water burdens generated by locating facilities abroad rather than in cleaner home jurisdictions. The median offshore surplus reached 348 kilograms of CO2 and 0.49 liters of freshwater per delivered kilowatt-hour. Aggregated across the sample volumes, the authors calculate a first-year quantifiable carbon leakage of 21.4 megatonnes of CO2 and 1.78 billion liters of water. These are not abstract projections; they represent the measurable gap between what the AI industry’s footprint would be under cleaner siting and what it actually is.

The legal significance of those numbers hinges on a threshold concept from customary international law: the duty to prevent significant transboundary environmental damage. This norm, anchored in landmark cases such as the Trail Smelter Arbitration between the United States and Canada and the Pulp Mills dispute between Argentina and Uruguay before the International Court of Justice, obliges states to take all necessary measures to prevent activities under their jurisdiction from causing appreciable harm to the environment of other states. The study argues that the 21.4 megatonnes of first-year carbon leakage and the associated water withdrawals satisfy that appreciable-harm threshold, bringing AI data-center siting within the scope of a legal doctrine traditionally applied to smokestacks, rivers, and acid rain rather than cloud computing.

What makes the finding politically explosive is the corporate accountability angle. Seventy-one percent of the quantifiable impacts traced in the study are attributable to companies now covered by the European Union’s Corporate Sustainability Due Diligence Directive, a sweeping regulation that requires large firms operating in the EU market to identify, prevent, and remedy adverse environmental impacts across their value chains. Yet the disclosure record is dismal: fewer than one in five of the companies studied disclose project-level emissions, and only seven percent report their water use. The gap between legal obligation and transparency practice suggests that the directive’s due-diligence machinery, designed with factories and supply chains in mind, has not yet caught up with the geography of cloud infrastructure, where the most consequential environmental decisions are made at the moment a site is selected.

The study frames this as a latent governance gap rather than an unfixable one. Because the harm is quantifiable, threshold-based, and traceable to identifiable corporate actors, the authors argue that existing norms, the customary duty to prevent transboundary harm, the EU due-diligence directive, and emerging border-adjustment instruments, can be extended to cover cloud infrastructure without waiting for entirely new treaties. The parallel to carbon leakage debates in manufacturing is explicit: just as heavy industry has migrated toward jurisdictions with weaker carbon pricing, AI compute is following the same gradient, and the policy toolkit developed for the former may be adaptable to the latter.

Crucially, the researchers did not stop at diagnosis. Their scenario analysis modeled three interventions and found that a combination of 24/7 carbon-free energy procurement, in which companies match clean power to consumption on an hourly basis rather than through annual averages, hybrid cooling systems that reduce freshwater dependence, and compute-adjusted border carbon adjustments that price the embedded emissions of imported computing services, could halve both the quantifiable carbon leakage and the water-withdrawal differentials identified in the dataset. The message is that the environmental cost of AI’s geography is a policy choice, not a technical inevitability, and that the tools to change it already exist in regulatory form.

The broader context is a rapidly escalating global conversation about the material footprint of the digital economy. Recent assessments, including the United Nations Conference on Trade and Development’s Digital Economy Report and a growing body of life-cycle research on AI infrastructure, have documented rising energy and water demands from data centers, but the international-law dimension has remained largely unexamined. By connecting econometric evidence of carbon leakage to the doctrinal architecture of transboundary harm and corporate due diligence, this study effectively hands regulators a quantitative foundation for treating data-center siting as an environmental governance issue rather than a purely commercial one.

For the AI industry, the implications are uncomfortable. The very flexibility that makes training workloads cheap to relocate, their indifference to latency, is what makes them legally visible as a source of exported harm. As due-diligence regimes mature and as climate litigation increasingly tests the boundaries of extraterritorial responsibility, companies that site their most energy-intensive compute in carbon-intensive, water-stressed jurisdictions may find that the savings on electricity come bundled with rising legal exposure. The study’s authors suggest that the path forward is neither to halt AI growth nor to accept its hidden geography, but to make the environmental costs of siting decisions visible, priced, and accountable, before the next 247 projects are announced.

Subject of Research: Extraterritorial environmental accountability of AI data centers under transboundary harm and due-diligence norms

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

Article References: Extraterritorial Environmental Accountability of AI Data Centers via Transboundary Harm and Due-Diligence Norms. (n.d.). https://doi.org/10.1007/s00267-026-02610-1

Image Credits: AI Generated

DOI: 10.1007/s00267-026-02610-1

Keywords: AI data centers, carbon leakage, transboundary harm, international environmental law, due diligence, water stress, carbon intensity, EU Corporate Sustainability Due Diligence Directive, hyperscale computing, life-cycle assessment, environmental governance, cloud infrastructure