AI's environmental costs threaten water, land and climate

United Nations · article · 2026-06-04 Reader 🖥️

AI Environmental Footprint Compute & Energy Infrastructure Jevons Paradox

AI's environmental footprint extends well beyond carbon emissions to substantial and unevenly distributed water, land, and e-waste costs — driven overwhelmingly by everyday inference use rather than training and resistant to efficiency gains via a rebound effect — such that averting resource strain now depends on deliberate governance choices by governments, companies, and users, not on technological improvement alone.

practical · expository news report summarizing a single research study, with an advocacy-adjacent policy conclusion — UN News transmitting findings of a UN University (UNU) report. Structurally it is reporting, not argument in its own voice, but the piece adopts the study's framing wholesale and ends on an explicit call to action ('governance choices made today') rather than staying neutral-descriptive.

Structure

Part	Role
Opening hook	frames AI's footprint as broader than climate, and accelerating
Rethinking how sustainability is measured	introduces the UNU study and its core measurement critique: carbon-only accounting overlooks water/land, and green trade-offs exist between them
Daily use is the main culprit	relocates the energy problem from training to everyday inference and introduces the rebound effect as a reason efficiency alone won't help
Local burdens, global benefits	establishes the distributional/environmental-justice dimension: costs concentrated, benefits diffuse
A widening digital and environmental divide	extends the distributional argument to global compute concentration (US/China) and infrastructure-poor nations
Towards responsible AI	resolves the piece into a policy call: a 'responsible AI ecosystem' framework and shared governance across governments, companies and users

Key terms

water footprint / land footprint
Resource costs of AI infrastructure analogous to carbon footprint but for water (cooling, energy production) and land (power generation, supply chains) — proposed as necessary complements to carbon accounting
rebound effect
The mechanism by which cheaper, more efficient AI drives higher usage that outpaces per-unit efficiency gains, so total resource consumption keeps rising
responsible AI ecosystem
UNU's proposed governance framework built on transparency, efficiency by design, equity, lifecycle responsibility, global cooperation, and sustainable use
environmental justice (implicit framing)
Costs (water draw, e-waste, land) concentrated in specific — often lower-income — regions while economic and technological benefits of AI accrue globally

Propositions

explicitcorroboratedAI's environmental footprint extends well beyond carbon emissions to substantial water and land footprints, which current sustainability measurement (prioritizing GHG accounting) largely overlooks.
supports ←I'm done using AI · p7 — Qualitative, unsourced corroboration (data centers, water usage, hidden costs) of the report's core claim that AI's environmental footprint goes well beyond carbon and is largely invisible to users.
explicitGlobal AI data-centre electricity consumption could reach 945 terawatt-hours annually by 2030 — nearly triple the combined annual electricity use of Pakistan, Bangladesh, and Nigeria (650+ million people).
extendsGuest post: AI Inference Is Breaking Unit Economics · p10 — Adds an environmental-footprint dimension (electricity, and by extension water and land) to the same 2030 AI-infrastructure investment horizon ($6.7T total, $5.2T AI-tied) cited in the inference unit-economics piece.
explicitAI-related water consumption could by 2030 equal the basic annual domestic water needs of 1.3 billion people, and AI's land footprint may exceed 14,500 square kilometres (roughly twice the Jakarta metropolitan area).
explicitSwitching to certain renewable energy sources to cut AI's carbon emissions can significantly increase water consumption and land use — 'green' solutions on one metric can worsen pressure on another.conf: high
explicitDay-to-day AI usage (inference), not model training, accounts for roughly 80-90 per cent of total AI energy demand.
explicitEnergy cost varies enormously by AI task type: generating a single AI image can require more than a thousand times the energy of simple text classification, with video generation costing even more.
explicitdevelopedEfficiency improvements in AI models and infrastructure are unlikely to offset rising resource demand, because a rebound effect — lower costs and better performance driving higher usage — increases total consumption faster than efficiency gains reduce it.
supportsYour Company Needs an Energy Strategy for AI's Next Phase · p4 — Both independently argue that AI efficiency gains will not reduce total energy/resource demand because cheaper, better AI drives higher usage — the rebound effect / Jevons paradox applied to AI, from two unrelated sources.
extends ←This Week in AI: When agents outnumber people · p4 — Both treat AI infrastructure demand as a hard physical constraint rather than a solvable efficiency problem; this piece adds the concrete, site-specific angle (seasonal cooling/water tradeoffs, EU regulation) to the UN report's macro resource figures.
explicitAI's environmental costs are unevenly distributed — concentrated in specific regions via electricity share of national grids and water draws sometimes amid drought conditions — while its economic and technological benefits accrue globally.
extendsThe growing AI backlash · p2 — Supplies concrete distributional/environmental-justice detail (local burden vs. global benefit, e-waste falling on lower-income countries) for the backlash piece's general claim that GenAI is 'environmentally damaging' via data centers.
supportsThe Trump administration's AI doomer moment · p7 — Substantiates the political-backlash claim with concrete data-centre resource burdens (electricity share, water draws amid drought) plausibly driving the local opposition described.
explicitAI infrastructure could generate up to 2.5 million tonnes of e-waste annually by 2030, a burden likely to fall disproportionately on lower-income countries with limited capacity for safe disposal; critical-mineral extraction for AI hardware raises related environmental degradation and social-inequity concerns.
explicitMore than 90 per cent of AI-specialised computing capacity is concentrated in just two countries (the United States and China), while over 150 nations lack significant domestic AI infrastructure — compounding economic and environmental inequity.
implicitThe report's findings are not meant as an argument against AI itself, but as a call for AI development to proceed within planetary limits — implying the resource problem must be solved through governance (by governments, companies, and users) rather than through technology alone.
refinesYour Company Needs an Energy Strategy for AI's Next Phase · p12 — Suggests AI's energy/environmental problem needs a broader governance frame (equity, lifecycle responsibility, global cooperation, planetary-limits framing across governments/companies/users) than the HBR piece's narrower procurement-and-architecture framing handled by existing CIO/CFO/procurement structures.

Arguments

Urgent, coordinated governance action — by governments (energy/water/land-use planning), companies (design for minimal resource consumption), and users (choosing lower-impact applications) — is needed to keep AI's environmental footprint within planetary limits, since technological efficiency improvements alone will not resolve it. strength: medium

Explicit premises:

Data-centre electricity, water, and land footprints are large and growing toward 2030 (p2, p3)
Current measurement practice overlooks non-carbon costs (p1)
Daily-use dominance and the rebound effect show that efficiency-focused strategies are structurally insufficient (p5, p7)
Costs are unevenly distributed, concentrating burdens (e-waste, water stress, grid pressure) on specific — often lower-income — regions while benefits are global (p8, p9, p10)

Implicit premises:

AI's continued growth and use is not itself in question — the argument is for managing it, not curtailing it
Governments, companies, and individual users each have sufficient leverage to change the trajectory described, though this is asserted rather than demonstrated

Inferences:

Because the problem is structural (measurement gaps, usage-driven demand, unequal distribution) rather than a simple technology deficit, the remedy must be structural too — a shared governance framework spanning policy, corporate design, and individual choice.

Narratives

Photographs of an Irish data centre and server racks with captions naming the location and describing server function.
→ illustrative/atmospheric — grounds an abstract statistical argument in a physical image of the infrastructure being discussed, not offered as evidence
Comparing projected 2030 AI data-centre electricity use (945 TWh) to the combined annual electricity consumption of Pakistan, Bangladesh, and Nigeria (650+ million people), and comparing AI's projected land footprint (14,500 km²) to twice the size of Jakarta.
→ rhetorical scaling device — makes an abstract large number viscerally comprehensible; not independent corroborating evidence, since both figures derive from the same study
'One widely used AI service' (unnamed, plausibly a reference to a major chatbot such as ChatGPT) is estimated to process around 2.5 billion prompts per day, consuming hundreds of gigawatt-hours annually.
→ concrete anecdotal grounding for the abstract claim that daily use dominates AI energy demand — weakened as evidence by the fact that the service is not named and no source/methodology is given

Questions addressed

✅ Is AI's environmental cost limited to carbon emissions?
No — the report argues water and land footprints are comparably significant and currently under-measured.
✅ Is model training or everyday AI use the bigger driver of AI energy demand?
Everyday use (inference), accounting for 80-90% of total demand, despite public debate centering on training.
✅ Will efficiency improvements in AI resolve its environmental impact over time?
No — the rebound effect means efficiency gains are outpaced by usage growth.
✅ Are AI's environmental costs and benefits evenly distributed globally?
No — costs (electricity share, water stress, e-waste, mineral extraction) concentrate in specific regions while benefits are enjoyed globally; compute capacity itself is concentrated in two countries.
⭕ Could AI's own capabilities (e.g., grid optimization, climate modeling, materials discovery) offset some of the environmental costs described?
⭕ Given that renewable energy sources trade carbon reduction against water/land increases, how should that tradeoff actually be weighed or resolved?

Gaps

incompleteThe piece is a one-sided, advocacy-adjacent summary of a single UNU study. It never engages the counter-case that AI is also being used to address environmental and climate problems (grid optimization, climate modeling, materials discovery, disaster response) that could offset some of the described costs, nor does it cite any independent critique or methodological scrutiny of the UNU report's projections. For a reader assessing whether this is rigorous accounting or advocacy framing, the absence of any dissenting or complicating voice is the central weakness.
Read next: IEA data-centre energy reports; academic AI-carbon-footprint methodology critiques (e.g., variation in assumptions about PUE and grid carbon intensity); vendor/hyperscaler sustainability disclosures and rebuttals; case studies of AI applied to climate/environmental problems.
incompleteThe central 'rebound effect' claim — that efficiency gains won't offset rising demand — is asserted without any quantitative comparison of the rate of efficiency improvement versus the rate of usage growth. No numbers are given for how much rebound offsets how much efficiency gain, leaving the reader unable to judge whether this is an established empirical pattern or a directional risk being stated as inevitability.
Read next: Empirical Jevons-paradox / rebound-effect literature from historical energy-efficiency episodes (lighting, computing); compare against extract-hbr-energy-strategy-for-ai's parallel claim about DeepSeek-style efficiency gains and induced demand.
uninformedEvery headline statistic in the piece (945 TWh, 1.3 billion people's water needs, 14,500 km² land footprint, 2.5 million tonnes of e-waste, >90% compute concentration) traces to the same single UNU study, with no independent replication, no peer-review status noted, and no discussion of estimate uncertainty or ranges. Readers cannot assess how contested or provisional these projections are relative to other bodies (IEA, academic groups, hyperscalers' own disclosures).
Read next: Cross-check against IEA 'Electricity' data-centre reports, academic AI energy/carbon-footprint studies (e.g., Luccioni et al.), and hyperscaler sustainability reports to triangulate the numbers and see how much they diverge.

Closes gaps elsewhere

Guest post: AI Inference Is Breaking Unit Economics · g3 — Supplies exactly the environmental/energy-implications-of-inference data that gap called for — UNU study data on data-centre electricity (945 TWh by 2030), plus water and land footprints, tied to the same AI-investment growth story.
Baseten CEO Tuhin Srivastava on the AI Inference Crunch, Custom Models · g4 — Partial, cross-framing closure: adds the missing physical-infrastructure/sustainability dimension (energy, water, and land constraints on data centres) that gap flagged as absent, though this document frames it as environmental cost to the planet rather than as a bottleneck constraining AI's ability to scale — the angle the original gap asked about.