As artificial intelligence transforms every sector of society—from healthcare to finance, education to transportation—its unprecedented environmental footprint is emerging as one of the most critical sustainability challenges of the 2020s.
The AI boom is driving soaring energy consumption, with AI-related workloads accounting for roughly 20% of total data centre electricity use in 2025, a share expected to rise to 40% by the end of the decade.
These AI data centres are notoriously resource-intensive, consuming vast quantities of electricity for computing, water for cooling massive server farms, and significant land for infrastructure.
A groundbreaking 2025 Cornell University study projects that by 2030, the current rate of AI growth would annually put 24 to 44 million metric tons of carbon dioxide into the atmosphere—the emissions equivalent of adding 5 to 10 million cars to U.S. roadways—while draining 731 to 1,125 million cubic meters of water per year, equal to the annual household water usage of 6 to 10 million Americans.
The United Nations has issued a stark warning, urging countries to confront these escalating environmental costs, predicting that total electricity demand from data centres could roughly double by 2030, producing nearly 400 million tonnes of planet-heating carbon dioxide and consuming 9.3 trillion litres of water, meeting the drinking water needs of Earth's 8.1 billion people for approximately 1.6 years.
Artificial intelligence systems generate enormous amounts of heat through a fundamental physical process: electrical resistance. When AI processors (particularly GPUs and specialized AI accelerators) perform complex calculations, electrical energy converts into computational work and heat. The more computation, the more heat generated.
| Metric | Value |
|---|---|
| Heat output per AI server | 30-50 kW (high-performance) |
| Heat density per rack | 100-400 kW (AI data centres) |
| Traditional data centre rack | 5-15 kW (normal servers) |
| Heat transfer efficiency | Liquid 1,000 times better than air |
AI data centres operate at significantly higher temperatures than traditional facilities:
Standard data centres: 18-27°C operating temperature
AI data centres: 30-45°C (higher heat density)
Three Critical Factors:
Computational Intensity
AI models perform billions of calculations per second
Large language models (LLMs) require massive parallel processing
Each training run can consume as much electricity as 100+ households use annually
Hardware Density
AI servers pack 8-64 GPUs per machine
GPUs consume 250-500 watts each during operation
Traditional servers use 1-2 CPUs at 100-200 watts each
Continuous Operation
AI systems run 24/7 without breaks
Training and inference workloads constantly active
No downtime for cooling recovery
AI data centres must remove heat to prevent:
Hardware failure (processors overheat above 85°C)
Performance degradation (thermal throttling reduces speed)
Equipment damage (permanent failure from sustained high temperatures)
Cooling energy: Up to 40% of total data centre electricity
Heat disposal: Billions of BTUs per facility per day
Also Read: Global Warming in 2026: Major Causes and Their Impact on Our Planet
New research from the European Commission (March 2026) shows AI data centres could become water-positive and carbon-negative by using waste heat for:
Water purification (distillation processes)
Carbon capture (activating capture technologies)
District heating (warming nearby buildings)
This transforms data centres from purely consumptive to potentially beneficial infrastructure, though implementation remains limited.
Baseline Data:
| Metric | 2025 Value |
|---|---|
| Global data centre electricity | 448 trillion watt-hours (kWh) |
| AI-related electricity share | ~20% of total |
| Carbon emissions from electricity | 208 million tonnes CO₂ |
| Equivalent country emissions | Argentina's annual output |
The 208 million tonnes of CO₂ produced by global data centres in 2024 rivals the emissions of some of the world's largest countries.
According to the UN University Institute for Water, Environment and Health (UNU-INWEH) report (June 2026):
AI electricity share: Will rise from 20% (2025) to 40% by 2030
Total data centre demand: Could roughly double by 2030
AI-linked consumption: Could meet residential electricity needs of all 1.3 billion people in sub-Saharan Africa for more than two years
2025 Electricity Use: ████████████████░░░░░░ 20% AI share 2030 Projected Use: ████████████████████████████████ 40% AI share (2x total demand, 2x AI share)
Carbon footprint: Nearly 400 million tonnes of CO₂ by 2030
Tree offset requirement: 6.7 billion trees grown over a decade
Comparison: About twice the number of trees in the entire UK
AI industry CO₂ emissions: 24-44 million metric tons annually by 2030
Car equivalent: Adding 5-10 million cars to U.S. roadways
Net-zero impact: Current growth puts AI industry's net-zero targets out of reach
Even if a data centre is powered by "low carbon" electricity generation, this does not necessarily make AI generation less impactful because it could still be linked to significant:
Water use (for cooling and generation)
Land use (for infrastructure and renewable farms)
This means carbon neutrality alone is insufficient—comprehensive environmental planning must address water, land, and carbon simultaneously.
2030 carbon dioxide: Would drop only ~15% compared to baseline
Residual emissions: Approximately 11 million tons of CO₂ would remain
Required capacity: 28 gigawatts of wind OR 43 gigawatts of solar to reach net-zero
The Solution:
"The solution is to accelerate the clean-energy transition in the same places where AI computing is expanding" — Fengqi You, Cornell Engineering
AI data centres consume water primarily for:
Cooling systems (evaporative cooling, water loops)
Humidity control (maintaining optimal server environment)
Power generation (if on-site or nearby hydroelectric/nuclear)
By 2030, data centres will use approximately 9.3 trillion litres of water annually:
| Comparison | Value |
|---|---|
| Drinking water equivalent | Earth's 8.1 billion people for 1.6 years |
| Household usage equivalent | 6-10 million Americans' annual water use |
| Volume | 9,300,000,000,000 litres |
731-1,125 million cubic meters per year (731-1,125 billion litres)
Note: Slightly lower than UN estimate, but same magnitude of concern
Rapid clustering strains local infrastructure and water resources
High water-stress region (limited water supply, high demand)
Competition with residential and agricultural users
Slash water demands by ~52% through strategic siting alone
Total water reductions of 86% when combined with grid and operational best practices
| State | Water Stress Level | Data Centre Water Impact |
|---|---|---|
| Virginia | High | Extreme (665+ facilities) |
| Texas | Medium-High | High (413 facilities) |
| California | High | High (321 facilities) |
| Illinois | Medium | Moderate (123 facilities) |
| Georgia | Medium | High (141 planned facilities) |
Texas (wind energy, moderate water)
Montana (low water stress, renewable energy)
Nebraska (low water stress, wind power)
South Dakota (low water stress, renewable capacity)
These states offer:
Lower water-stress (more available water)
Abundant renewable energy (wind)
Cooler climates (reduced cooling needs)
Available land for infrastructure
Active Data Centres by State:
| State | Active Facilities | Planned Facilities |
|---|---|---|
| Virginia | 665+ (highest) | 287 (highest) |
| Texas | 413 (2nd) | 170 (2nd) |
| California | 321 (3rd) | Not specified |
| Illinois | 123 (4th) | 123 |
| Georgia | Not specified | 141 (3rd most planned) |
Total U.S. Data Centres: Over 4,500 active facilities as of 2026
Northern Virginia's Advantages:
Historical Hub: "Data Centre Alley" built since 1990s
Proximity to DC: Government and corporate clients nearby
Fiber Infrastructure: Dense network connections
Power Availability: Large electricity capacity
Tax Policies: Favorable commercial rates
But Critical Problems:
Water stress: Limited water supply for cooling
Grid strain: Rapid clustering overwhelms infrastructure
Competition: Conflict with residential/agricultural water users
Pew Research Finding (April 2026):
Most new data centres in the U.S. are coming to rural areas:
Top States for Planned Data Centres:
| Rank | State | Planned Facilities | Rural Focus |
|---|---|---|---|
| 1 | Virginia | 287 | Mixed urban/rural |
| 2 | Texas | 170 | Rural expansion |
| 3 | Georgia | 141 | Rural counties |
| 4 | Illinois | 123 | Rural areas |
Rural Area Benefits:
Lower land costs
Less water competition
Cooler climates (reduced cooling)
Available power capacity
Tax incentives from local governments
According to the UN report:
Only 32 nations host AI-specialised cloud infrastructure
90% of capacity is in the U.S. and China combined
150+ countries have no sovereign AI computing at all
The Digital Divide:
| Nation Type | Characteristics |
|---|---|
| AI Builders | U.S., China, select 30 nations |
| AI Consumers | 150+ countries with no infrastructure |
| Environmental Burden | Excluded countries face mineral extraction and e-waste while strategic benefits flow elsewhere |
This creates an unequal environmental footprint distribution, where developing nations bear resource extraction and waste disposal costs while wealthy nations control AI benefits.
Chip production (semiconductor manufacturing)
Critical minerals (lithium, cobalt, copper, rare earth elements)
Electronic waste (discarded hardware)
Chip Manufacturing Impact:
Water use: Semiconductor Fab requires 4-10 million gallons daily
Chemicals: Toxic solvents and acids in production
Energy: Extreme ultra-pure manufacturing processes
Waste: Hazardous byproducts requiring special disposal
AI infrastructure is predicted to generate up to 2.5 million tonnes of e-waste each year by 2030.
Context:
Current global e-waste: ~50 million tonnes annually (2022)
AI contribution by 2030: 5% of total global e-waste
Growth rate: Exponential (AI hardware refresh cycles: 3-5 years)
Performance demands: Older hardware can't run new models
Training wear: GPUs degrade under continuous heavy load
Short lifecycle: 3-5 years average (vs. 10+ years for traditional servers)
| Mineral | Use | Environmental Impact |
|---|---|---|
| Lithium | Batteries, cooling systems | Water depletion, soil contamination |
| Cobalt | Electronics | Child labor, toxic waste |
| Copper | Wiring, connectivity | Acid mine drainage, deforestation |
| Rare Earth Elements | Magnets, processors | Radioactive waste, ecosystem destruction |
Extraction Locations:
Cobalt: Democratic Republic of Congo (70% of supply)
Lithium: Chile, Australia, Argentina ("Lithium Triangle")
Rare Earths: China (85% of supply)
Excluded countries (150+ without AI infrastructure) often face mineral extraction and e-waste burdens while strategic benefits flow to AI-building nations.
73% carbon dioxide reduction
86% water reduction
compared with worst-case scenarios.
The solution requires three pillars working together:
Smart Siting (location strategy)
Faster Grid Decarbonization (clean energy)
Operational Efficiency (technology improvements)
Strategic Location Principles:
Low water-stress regions: Access abundant water without competing with residents
Renewable energy zones: Proximity to wind, solar, hydroelectric power
Cooler climates: Reduce cooling energy needs
Available land: Minimize environmental disruption
Grid capacity: Ensure electricity infrastructure can handle demand
Best Locations:
Cornell identified the Midwest and windbelt states as optimal:
Texas: Wind energy + moderate water
Montana: Low water stress + renewable energy
Nebraska: Wind power + available water
South Dakota: Renewable capacity + low stress
Accelerating Renewable Deployment:
To reach net-zero by 2030, the AI industry needs:
28 gigawatts of wind capacity OR
43 gigawatts of solar capacity
State-by-State Opportunities:
Texas: Wind belt (largest U.S. wind capacity)
Montana: Hydroelectric + wind
Nebraska: Wind + solar potential
South Dakota: Wind energy abundant
Critical Insight:
"There isn't a silver bullet. Siting, grid decarbonization and efficient operations work together—that's how you get reductions on the order of roughly 73% for carbon and 86% for water" — Fengqi You, Cornell
Liquid cooling is ideally suited to cool next-generation AI accelerators for greater efficiency, sustainability, and density.
Energy and Cost Savings:
| Metric | Air Cooling | Liquid Cooling | Improvement |
|---|---|---|---|
| CO₂ emissions | 8,700 tons/year | 1,200 tons/year | 87% reduction |
| Annual cost per server | $254.70 | $45.99 | 86% savings |
| 5-year chassis power | Baseline | 14.9% less | 14.9% reduction |
| Performance per kW | Baseline | 20.7% higher | 20.7% improvement |
Instead of pushing cool air across racks, direct-to-chip liquid cooling circulates coolant directly over heat-generating components, efficiently removing heat through a closed-loop system.
Physics Advantage:
Liquids transfer heat 1,000 times more effectively than air
Enables tighter rack configurations
Higher performance density
Dramatically improved cooling efficiency
Energy Reduction Breakdown:
Direct-to-chip systems: Cut cooling energy use by 30-60%
Temperature increase: When cooling supply temperature rises by 20°C, total cooling energy drops by 40% and water use by 60%
Combined with smart siting: 86% total water reduction possible
Other Advanced Cooling Technologies:
Immersion cooling (servers submerged in liquid)
Two-phase cooling (evaporation-based systems)
Membrane-based cooling (water-efficient technology)
Heat recovery systems (waste heat for water purification)
Total Reductions When Combined:
According to Cornell research:
Carbon reduction: 73% (7% from tech + 66% from siting + grid)
Water reduction: 86% (32% from tech + 54% from siting + grid)
Technology-only impact:
Carbon: 7% reduction (from advanced cooling + server utilization)
Water: 29% reduction (from liquid cooling + efficiency)
Combined total: 32% water reduction when tech combined with other strategies
This demonstrates that technology alone cannot solve the problem—strategic location and clean energy deployment are essential.
Integrated Planning Requirements:
To ensure AI grows sustainably, the UN recommends:
For Governments:
Integrate AI infrastructure decisions with:
Energy planning
Water governance
Land-use permitting
Treat environmental footprints as material investment risks:
Electricity consumption
Carbon emissions
Water usage
Land requirements
Include communities and civil society early in location decisions for data centres
Model selection (lighter models need less computing)
Default outputs (optimize for efficiency)
Routing decisions (send queries to efficient servers)
Improve efficiency by design rather than as an add-on
Approach location and energy procurement decisions as environmental choices
For Users:
Choose the lightest model that meets your task requirements
Select lowest-energy format available
Example: Using a smaller AI model instead of the largest available reduces computational requirements by 50-80%.
For Investors:
Treat electricity, carbon, water, and land footprints as material financial risks that must be factored into investment decisions.
The environmental cost of data centres is rising, prompting the QuitGPT movement to gain momentum.
Core Question:
"As the QuitGPT movement gains momentum, should people concerned about the environmental impacts of AI consider opting out?"
Movement Principles:
Reduce AI usage when not essential
Choose manual alternatives to AI-powered tools
Support sustainable AI companies with transparent environmental practices
Advocate for regulation of data centre environmental impacts
Practical Steps:
Limit AI queries: Use AI only when necessary, not for trivial tasks
Choose efficient models: Select smaller, optimized AI models
Reduce image generation: AI image creation is extremely energy-intensive
Avoid unnecessary training: Personal AI model training consumes massive resources
Support green AI: Companies investing in renewable-powered data centres
The environmental cost of artificial intelligence represents one of the most significant sustainability challenges of the 2020s. As AI transforms society, its massive resource demands—consuming 20% of data centre electricity in 2025 (projected to reach 40% by 2030), generating 400 million tonnes of CO₂ emissions, using 9.3 trillion litres of water annually, and producing 2.5 million tonnes of e-waste by 2030—threaten natural resources for billions.
Key Takeaways:
Heat crisis: AI data centres produce 30-50 kW per server (10x traditional); liquid cooling reduces energy 87%
Power explosion: Data centre electricity demand will double by 2030; AI share grows from 20% to 40%
Water dragon: 9.3 trillion litres by 2030 equals 1.6 years of drinking water for 8.1 billion people
Carbon footprint: 24-44 million tonnes CO₂ annually by 2030 (5-10 million cars equivalent)
E-waste crisis: 2.5 million tonnes/year by 2030 from AI hardware refresh cycles
Geographic concentration: 90% of AI capacity in U.S. and China; 150+ countries have no sovereign AI
Solution exists: 73% carbon reduction and 86% water reduction achievable through smart siting, grid decarbonization, and liquid cooling
The Path Forward:
The solution is not to stop AI development but to accelerate the clean-energy transition in the same places where AI computing is expanding. By combining:
Smart siting in low-water-stress, renewable-energy zones (Texas, Montana, Nebraska, South Dakota)
Liquid cooling technology delivering 87% CO₂ reduction and 86% water savings
Grid decarbonization with 28GW wind or 43GW solar capacity
The AI industry can achieve net-zero emissions while continuing to deliver transformative benefits. However, this requires urgent action from governments, industry, and users—all must prioritize environmental sustainability in AI development.
The UN's warning is clear: without immediate intervention, AI's environmental footprint will make net-zero targets impossible to reach. The choice is between unsustainable growth and strategically managed expansion that balances technological advancement with planetary health.