Moderate evidence
More
AI data centers shift 21.4 million tons of CO₂ abroad, study of 247 projects finds
Study title: Extraterritorial Environmental Accountability of AI Data Centers via Transboundary Harm and Due-Diligence Norms
- Evidence type
- Observational study
- Sample
- n = 247
- Sources
- 1 sourcePeer-reviewed
- Published
- 2 min read
- Truvace record
- TRV-2026-2582cite or check it
Study title: Extraterritorial Environmental Accountability of AI Data Centers via Transboundary Harm and Due-Diligence Norms
A study of 247 hyperscale AI data-center projects announced between 2018 and March 2025 found that siting favours high-carbon grids. Each 100 g CO₂ per kWh increase in a host grid raised the chance of attracting new capacity by 8.1 percentage points. The authors estimate first-year offshore leakage of 21.4 million tons of CO₂ and 1.78 billion litres of water for a typical 150-megawatt site.
Read the study abstract
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…

The problem
A study of 247 AI data-center projects found siting choices produce 21.4 Mt of first-year CO₂ leakage and 1.78 billion litres of water withdrawal, while only 18% of covered firms disclose project-level emissions.
In brief
The projects represent 37 gigawatts of planned load across 34 jurisdictions. Moving from low to high water stress cut siting probability by 5.7 percentage points, but the effect vanished where grids were low-carbon, which the authors read as firms putting carbon optics ahead of water risk. Three areas, central Arizona, northern Virginia and Madrid, carried 58% of the cross-border withdrawal burden.
The authors say 71% of the carbon leakage and 64% of the water leakage trace to firms covered by the EU Corporate Sustainability Due Diligence Directive. Yet 18% of those firms disclose project-level emissions and 7% report any water figure. Their estimates rest on announced design loads and average cooling coefficients, so realised impacts may differ.
Main points
- The dataset covers 247 announcements from January 2018 to March 2025, 37 gigawatts of planned load and 34 host jurisdictions.
- Each 100 g CO₂/kWh rise in host-grid intensity raised siting probability by 8.1 percentage points; weighted by capital spending the figure was 11.4.
- Median offshore surplus was 348 kg CO₂ per kWh of delivered compute; the upper quartile reached 612, driven by Alberta, Poland and the US mid-Atlantic.
- The median water differential was 0.49 L per kWh; cutting it with hybrid cooling reduced aggregate leakage by 37%, but residual withdrawals still exceeded replenishment credits in six of nine high-stress basins.
- 24/7 carbon-free procurement cut the carbon surplus by 46%, yet more than one-third of projects lacked the local renewable supply to meet it.
- Of projected AI-linked load, 42% falls in US counties above the national median for both poverty and non-white population share.
- A compute-adjusted border mechanism would have internalised 12.6 million tons of emissions (59%), and a water footprint certificate would have saved 620 million litres (34%).
The full story
The authors combined event-study and spatial-panel models with life-cycle assessment. They matched announced projects to hourly grid carbon factors and basin-level water-stress scores, then compared each project's footprint with running the same workload on the importing client's domestic grid. The carbon total is roughly the 2022 national inventory of Croatia.
The water finding is lopsided: Arizona, northern Virginia and Madrid carried 58% of the burden. The authors' model has 24/7 carbon-free power reducing carbon leakage from 21.4 to 11.6 Mt and hybrid cooling reducing water leakage from 1.78 to 1.12 billion litres. They argue the size and foreseeability of the harm meet the 'appreciable harm' standard in international law, a legal interpretation rather than a court finding.
The authors recommend hourly water reporting, a compute-adjusted border adjustment and a water footprint certificate, and say distributional effects fall on poorer communities.
Sources
1 on file- Environmental Managementdoi:10.1007/s00267-026-02610-1
Talk it over
Who carries this cost, and what would have to change to prevent it?
AI data centers shift 21.4 million tons of CO₂ abroad, study of 247 projects findshttps://truvace.com/item/extraterritorial-environmental-accountability-of-ai-data-centers-via-transbounda
Who carries this cost, and what would have to change to prevent it?
The problem A study of 247 AI data-center projects found siting choices produce 21.4 Mt of first-year CO₂ leakage and 1.78 billion litres of water withdrawal, while only 18% of covered firms disclose project-level emissions.
Answer in the threadHow do we know? What can this kind of evidence show, and what can’t it?
The evidence This rests on an observational study with a sample of 247, from Environmental Management.
Answer in the threadWhat does it leave open? What else would you want to know before acting on it?
The record’s own caveat The leakage estimates rest on announced design loads and assumed high utilisation, use average regional water-efficiency coefficients, and may undercount edge projects below 20 megawatts.
Answer in the threadHave you seen AI used this way where you study, work or live? Did it go the same way?
Answer in the thread
The questions are the same on every record. The details under them are taken from this record as written.
ace
What readers say