White Paper on: AI and Environment
An ISOC LIVE Summary
Publication Date: July 2026 Organizations: Nature and People Foundation; China Institute, Fudan University; Pahle India Foundation — in partnership with the Dynamic Coalition on Environment of the Internet Governance Forum Authors: Polina Burkova; Sofia Denisova; Liu Dian; Nitish Dogra; Victoria Elias; Alexey Kokorin; Sergey Rybakov; Jeffrey D. Sachs; Zhang Weiwei
Overview
This 90-page white paper examines the relationship between artificial intelligence and the environment as a single problem rather than two. Its stated contribution is to be the first report to bring AI’s environmental footprint and AI’s environmental benefits into one analytical framework, and to do so from the vantage of the BRICS countries and the Global South rather than from US and European policy perspectives alone.
Its central argument is that AI’s environmental impact is not a fixed cost but a governance choice. Whether AI’s contributions to solving environmental problems outweigh its contributions to creating them will be determined, the authors argue, not by technology but by the design decisions, disclosure rules, and policy frameworks adopted in the next few years — while infrastructure investments with decades-long consequences are still being made.
The paper opens with a foreword by Jeffrey D. Sachs, who describes AI as a general purpose technology in the company of the steam engine and electricity, and identifies the paper’s three virtues as timeliness, comprehensiveness, and the multipolar character of its authorship.
The occasion
The paper is written for presentation at two events in 2026: the inaugural Global Dialogue on AI Governance and the AI for Good Global Summit, convening together in Geneva 7–10 July 2026, and the World Artificial Intelligence Conference (WAIC) in Shanghai. The authors chose the moment deliberately: the institutional architecture for AI governance is being assembled now — the WSIS+20 review made the IGF a permanent UN forum in December 2025, and UN General Assembly Resolution A/RES/79/325 established the Independent International Scientific Panel on AI and the Global Dialogue in August 2025 — while the environmental dimension of that architecture remains undefined.
A multipolar perspective
The paper combines contributions from experts associated with India, China, Russia, and international organizations, alongside analysis of European and US approaches. The authors argue that earlier work on AI and the environment was shaped mainly by US, European, and to some extent Chinese perspectives, and that AI governance cannot be designed by one group of countries and exported to the rest. Each partner institution contributes a nationally grounded case: India’s tension between digital-infrastructure growth, water stress, and an environmental-clearance system never designed for cloud-based AI; China’s attachment of environmental requirements directly to computing-power expansion; and Russia’s framing of low-carbon energy as the enabling foundation for AI compute, together with cold-region monitoring applications underrepresented in global datasets.
Understanding AI’s environmental footprint
The paper’s analytical foundation is lifecycle assessment across energy, carbon, water, critical materials, and electronic waste — not carbon alone. Its key finding is that AI’s footprint is not determined by model size but by choices within human control: model architecture, hardware efficiency, grid carbon intensity, and data-center siting, each capable of changing impact by large factors.
The numbers it assembles include:
Global data-center electricity consumption reached approximately 415 TWh in 2024, around 1.5% of global demand, growing 12% annually since 2017. The IEA projects 945 TWh by 2030, with data-center emissions potentially reaching 300 Mt CO2 by 2035 in the base case and up to 500 Mt in a high-growth scenario.
Training emissions for frontier models rose from 0.01 tons for AlexNet (2012) to 588 tons for GPT-3 (2020), 5,184 tons for GPT-4 (2023), and 8,930 tons for Llama 3.1 405B (2024).
AI systems’ total water footprint could reach 312.5–764.6 billion litres in 2025; the OECD estimates AI-related water consumption could reach 6.6 billion cubic metres by 2027.
UNEP estimates that producing a single 2-kilogram computer requires roughly 800 kilograms of raw materials. Global e-waste reached 62 million tons in 2022, and AI-specific e-waste is tracked separately by no national or international reporting system.
Two lifecycle studies anchor the analysis. The BLOOM assessment (2023) found a total footprint of approximately 50.5 tons CO2eq — roughly double the training-only figure of 24.7 tons — with 49% from active GPU computation, 29% from idle data-center infrastructure, and 22% from hardware manufacturing. Mistral AI’s assessment of Mistral Large 2, published in July 2025 with Carbone 4 and ADEME, reported 20.4 kt CO2eq over 18 months of use, 281,000 m³ of water, and 660 kg Sb eq of resource depletion — and became the first commercial disclosure of marginal inference impacts: a single 400-token response generates 1.14 g CO2e, consumes 45 mL of water, and uses 0.16 mg Sb eq.
The paper argues that attention has fixed on training while inference has become the dominant source of operational emissions — approximately 65% of lifecycle emissions according to production data from Meta, and 80% under Schneider Electric’s 2023 estimate. It also notes that not all AI use carries equal weight: image generation is roughly 1,500 times more energy-intensive than text classification.
Four environmental domains
The report structures its analysis around four domains, each aligned with the corresponding multilateral framework.
Climate change (UNFCCC)
AI is presented as both problem and solution. Against rising infrastructure demand, the paper sets AI’s contributions to weather and climate modelling (DeepMind’s GraphCast, Microsoft’s Aurora, Huawei’s Pangu-Weather, and a Yandex–HSE model forecasting El Niño up to 18 months ahead), independent emissions monitoring through Climate TRACE, which tracks over 350 million assets globally, grid optimisation, wildfire prediction, and carbon-capture materials discovery. The paper is direct about the gap between corporate pledges and outcomes: Microsoft’s emissions rose 29% between 2020 and 2024, and Google’s rose 48% against 2019, in both cases driven by AI infrastructure expansion.
Biological diversity (CBD)
The paper covers species identification, bioacoustics, camera traps, environmental DNA, biodiversity finance, and the Biodiversity Policy Analyzer for mapping national strategies against the Kunming-Montreal Global Biodiversity Framework. It gives particular attention to the Cali Fund for Digital Sequence Information, established at CBD COP16 in Cali in 2024 and launched in Rome in February 2025, which directs at least 50% of its resources to indigenous peoples and local communities. Indigenous data sovereignty is treated as a rights question, not a consultation question.
Desertification and land degradation (UNCCD)
Up to 40% of the world’s land is already degraded, with restoration needs estimated at $2.6 trillion by 2030. The centrepiece initiative here is the International Drought Resilience Observatory (IDRO), the first global AI-driven platform for proactive drought management, whose prototype was unveiled at UNCCD COP16 in Riyadh in December 2024 and whose full version is expected to launch at COP17 in Mongolia in 2026. The paper is candid that this domain faces harder constraints than the others: data scarcity in the most affected regions, limited technical capacity, and funding at a fraction of assessed need.
Water resources (SDG 6)
Water receives sustained attention as what the paper calls a “dual relationship.” Around 4 billion people experience severe water scarcity for part of the year, and in India half of all data centres sit in extremely water-stressed regions. Against this, the paper sets Google’s Flood Hub, which forecasts riverine flooding up to seven days ahead across more than 80 countries for roughly 460 million people and was extended in April 2026 with a flash-flood model forecasting 24 hours out; AI leak detection in networks where 30–40% of treated water is lost before reaching consumers; precision irrigation; water-quality monitoring; and China’s digital-twin basin management.
Measuring sustainable AI
The paper argues that measurement is the precondition for governance, and documents a standards convergence now taking shape:
ITU-T L.1801 (February 2026) — the first international standard specifically for assessing the environmental impact of AI systems.
ISO/IEC TR 20226:2025 — sustainability considerations across the full AI lifecycle.
UNEA Resolution 7/9 (December 2025) on the environmental sustainability of AI systems, the highest-level multilateral mandate for this work.
The Coalition for Environmentally Sustainable AI, launched by France, UNEP, and the ITU in February 2025 with over 100 partners.
UNEP’s Sustainable Procurement Guidelines for Data Centers and Servers (June 2025), the first international procurement framework with environmental performance criteria.
Open-source tooling has matured alongside: CodeCarbon, and from Russia, Sber and AIRI’s Eco2AI for emissions tracking and Eco4cast for carbon-aware scheduling, which the paper cites as achieving up to 90% CO2 reduction in some scenarios. The barrier, it concludes, is adoption rather than capability — most companies still report aggregate data-center metrics without separating AI workloads.
Comparing governance approaches
The paper compares six jurisdictions — the BRICS grouping, China, the European Union, India, Russia, and the United States — and identifies three contrasting models:
China’s infrastructure-planning-led model, embedding environmental conditions into the approval and siting of computing capacity. The “Eastern Data, Western Computing” programme establishes eight national computing-hub nodes supported by ten data-centre clusters; the green-development action plan targets average PUE below 1.5 by 2025 and 10% annual growth in renewable-energy utilisation, with reported average PUE across the hub clusters at approximately 1.3 and the most advanced facilities at 1.04.
The European Union’s law-centered model, distributing obligations across the AI Act, the Energy Efficiency Directive (mandatory reporting for data centres of 500 kW or more since May 2024), and the CSRD. Germany’s Energy Efficiency Act goes furthest, mandating a 50% renewable electricity share for data centres, rising to 100% from January 2027.
The United States’ corporate-led model, relying on the voluntary commitments of a small number of dominant firms in a country hosting over 33% of the world’s data centres.
The paper’s position is that comparing these models is more instructive than ranking them, and that an effective international framework will likely need elements of all three.
Six regulatory gaps common to every jurisdiction
Across all six jurisdictions examined, the paper identifies the same structural gaps:
Disclosure gap — no country requires AI-specific environmental disclosure.
Inference gap — the growing impact of inference is largely unmeasured and unregulated.
Water gap — water governance lags far behind energy governance.
Geographic equity gap — the uneven distribution of AI’s costs and benefits is rarely addressed by any framework.
Application accounting gap — there is no widely adopted method for verifying claimed environmental benefits.
Public accountability gap — AI-assisted environmental decisions often lack explainability, appeal mechanisms, and independent audit.
Six cross-cutting challenges
Synthesizing across domains rather than jurisdictions, the paper sets out six challenges any credible response must confront:
Data on AI’s energy, carbon, and water footprint remain incomplete and non-comparable — and the asymmetry runs deeper than that, since AI’s environmental costs are becoming measurable while its claimed benefits remain unverifiable for want of any recognized validation framework.
Environmental impacts are geographically uneven and create equity problems, including through the underrepresentation of non-English and indigenous knowledge.
AI infrastructure is expanding faster than regulatory and measurement systems.
Efficiency gains can produce rebound effects — the Jevons paradox — so governance must address absolute resource use, not only relative efficiency.
Environmental AI requires high-quality data and long-term institutional maintenance, and must be treated as continuous public infrastructure rather than a one-time deployment.
International cooperation is complicated by digital sovereignty, export controls, and uneven access to computing power.
Roles, recommendations, and seven proposed initiatives
The closing sections assign responsibilities by actor. Governments are urged to adopt mandatory environmental disclosure disaggregated by AI workload, integrate environmental criteria into AI governance, apply environmental performance criteria in public procurement, and account for regional energy mix and water stress in data-centre siting. AI developers are urged to publish full lifecycle assessments covering training and inference, and to treat environmental performance as a primary evaluation criterion. Hardware manufacturers are singled out for the most conspicuous data gap in the field: no major GPU manufacturer publishes environmental product declarations for AI accelerators, leaving embodied carbon unquantifiable by anyone outside the manufacturer. International organizations are asked to accelerate standards adoption and support capacity building; civil society and research institutions to maintain independent measurement tools and hold corporate claims to account.
Building on these, the Coalition proposes seven concrete initiatives for follow-up in Geneva and Shanghai: a common reporting template for data-centre energy, carbon, and water indicators; a comparative policy map; a BRICS+ and Global South case repository; dual disclosure of both AI’s footprint and AI’s claimed benefits; pilot projects on carbon- and water-aware workload scheduling; use of international forums to test the measurement framework; and a common glossary of AI-environment terminology and metrics.
Key takeaways
The paper rejects the idea that AI’s environmental impacts are an unavoidable consequence of technological progress. Its recurring lesson across all four domains is that AI’s environmental benefits are real but conditional — credible only when the additional computing demand required to produce them is measured against the savings they deliver, and only when applications that genuinely reduce environmental pressure are distinguished from those that merely shift costs elsewhere. As governments and companies commit to long-lived investments in data centres and AI infrastructure, the authors treat the present moment as a closing window for embedding environmental governance before current practices become entrenched.
RESOURCES
Nature and People Foundation — Russian co-author; observer at UNCCD COP
China Institute, Fudan University — Chinese co-author, directed by Zhang Weiwei
Pahle India Foundation — Indian co-author
IGF Dynamic Coalition on Environment — the paper’s institutional home
ITU-T L.1801 — first international standard for assessing AI’s environmental impact (in force, February 2026)
ISO/IEC TR 20226:2025 — sustainability across the full AI lifecycle
UNEA Resolution 7/9 — multilateral mandate on the environmental sustainability of AI systems (December 2025)
UNEP Sustainable Procurement Guidelines for Data Centres and Servers — June 2025
IEA, Energy and AI — source of the 415 TWh and 945 TWh projections (April 2025)
Global Dialogue on AI Governance — inaugural session, Geneva, July 2026