AI Data Centres and the Coming Energy Crisis: The Environmental Cost of the Intelligence Revolution

Artificial intelligence is often presented as a digital technology with an almost weightless existence: algorithms operate in the cloud, responses appear instantly and innovation seems disconnected from physical infrastructure. In reality, every AI model depends on vast data centers filled with energy-intensive processors, cooling equipment, power systems, networks and industrial buildings.

The rapid expansion of generative AI is therefore creating a new infrastructure challenge. Electricity demand from global data centers is projected to rise from 485 terawatt-hours (TWh) in 2025 to approximately 950 TWh by 2030, equivalent to around 3% of global electricity consumption. Electricity use by AI-focused data centers is expected to triple over the same period.

This growth does not automatically constitute a global energy crisis. Data centers remain a relatively small share of total electricity demand. However, their concentrated location, continuous operation, rapid expansion and high-power density are creating local crises involving grid capacity, water availability, electricity prices, land use and emissions.

The central question is not whether artificial intelligence should be developed. It is whether the intelligence revolution can expand without transferring its environmental costs to electricity consumers, water-stressed communities and future generations.

The scale of AI’s energy demand

Data centers are becoming major electricity users

Global data-center electricity consumption increased by 17% in 2025, while electricity uses by AI-focused facilities rose by approximately 50%. This growth substantially exceeded the increase in global electricity demand, which was around 3% during the same period.

The International Energy Agency expects total data-center electricity demand to nearly double by 2030:

Year	Global data-center electricity demand	Approximate significance
2024	About 415 TWh	Around 1.5% of global electricity consumption
2025	About 485 TWh	Rapid growth driven by AI infrastructure
2030	About 950 TWh	Around 3% of global electricity demand

The absolute percentage may appear modest, but national averages conceal the real pressure. Data centers are not distributed evenly across electricity systems. They are often concentrated in regions with fiber networks, tax incentives, available land and existing technology clusters. A single AI campus can require hundreds of megawatts of dependable electricity, comparable to the demand of a small city.

The resulting problem is geographical and temporal: electricity may be available nationally but unavailable near the proposed data center, or available during some hours but not during periods of peak demand.

AI workloads are unusually demanding

Traditional data centers primarily support websites, cloud storage, enterprise software and video services. AI facilities add highly specialized workloads:

• Training large models with thousands of processors operating simultaneously.
• Running inference systems that respond to user requests.
• Generating video, images, audio and complex reasoning outputs.
• Supporting autonomous AI agents that perform multiple tasks without continuous human input.
• Maintaining large memory and storage systems for model parameters and datasets.

The energy used for a single simple text query may be relatively low, but the total system demand depends on scale and application type. The IEA reports that simple AI text tasks are becoming more energy-efficient, with energy consumption per task falling rapidly. However, video generation, reasoning and agentic applications can consume hundreds or thousands of times more energy per query than simple text generation.

This creates a rebound effect. More efficient AI services reduce the cost of individual tasks, but lower costs encourage more users, more frequent use and more complex applications. Efficiency gains may therefore be overwhelmed by demand growth.

Why efficiency alone will not solve the problem

Technological efficiency is essential, but it does not guarantee lower overall resource consumption. Three trends are operating simultaneously:

1. Efficiency: Chips, software and cooling systems are becoming more efficient.
2. Adoption: More companies, governments and consumers are using AI.
3. Intensity: New applications require much more computation than basic text generation.

The IEA describes energy efficiency per AI task as improving at an exceptional rate. Yet the same report warns that changing model capabilities and the expansion of energy-intensive applications make future demand highly uncertain.iea

This is a classic rebound problem. If the energy required for a task falls by 90% but the number of tasks increases twentyfold, total energy consumption still rises substantially.

A similar pattern has occurred across digital technologies. More efficient processors did not reduce total computing demand; they enabled more services, larger datasets, higher-resolution content and continuous connectivity. AI is extending this trend into a more computationally intensive phase.

Carbon emissions: the electricity mix matters

The climate impact of AI data centers depends less on their physical existence than on the electricity used to power them. A facility supplied by new renewable generation has a different emissions profile from one drawing marginal electricity from coal or gas.

The IEA projects that emissions associated with data centers could reach approximately 350 million tons in 2035, roughly double current levels, although this would still represent about 2% of global electricity-sector emissions.

This global figure should not be interpreted as evidence that AI’s climate impact is insignificant. National and local impacts can be much higher where data centers increase fossil-fuel generation or delay the retirement of coal plants.

Fossil fuels may fill short-term power gaps

Grid infrastructure generally expands more slowly than AI investment. Data-center developers may therefore seek interim power supplies, including onsite natural-gas generation. The IEA estimates that 15–27 GW of onsite natural-gas capacity could supply data centers by 2030, primarily in the United States.

This creates several risks:

• New gas infrastructure may lock in fossil-fuel emissions for decades.
• Gas plants can increase nitrogen oxide and particulate pollution.
• Backup diesel generators add local air pollution during testing and outages.
• Methane leakage across the gas supply chain can increase the climate impact.
• Electricity demand may be met through fossil fuels when renewable power is curtailed or unavailable.

The result is a potential contradiction: companies may purchase renewable-energy certificates while their additional physical demand is met at the margin by gas or coal. Accounting instruments can support renewable investment, but they do not always demonstrate that a data center is operating on clean electricity every hour.

Renewable-energy claims require better accounting

Corporate renewable-energy commitments commonly rely on annual matching, power-purchase agreements or certificates. These approaches can help finance clean generation, but they may not correspond to the hour-by-hour electricity consumed by a data center.

A more rigorous approach would measure:

• Hourly electricity consumption.
• Hourly clean-energy supply.
• Regional grid emissions.
• New renewable capacity attributable to the facility.
• Transmission losses.
• Backup generation.
• Embodied emissions from construction and equipment.

Without such disclosure, the environmental performance of an AI data center cannot be assessed accurately.

The water footprint of artificial intelligence

Electricity is only one part of the environmental cost. AI data centers also require water, particularly where cooling systems use evaporation.

Servers operate continuously and convert much of their electricity into heat. Cooling systems must remove that heat to prevent equipment failure. Water-based cooling is effective because evaporation transfers substantial heat, but it can consume significant volumes of freshwater.

A study published in Patterns estimated that AI systems could have a global footprint of approximately 312.5 billion to 764.6 billion liters in 2025. The estimate covers operational water use and excludes important supply-chain and end-of-life impacts.

These figures should be interpreted cautiously because water accounting varies. Some studies measure water withdrawn, some measure water consumed, and others include water used to generate electricity or manufacture semiconductors. The boundaries are not interchangeable.

Direct and indirect water use

The water footprint of AI can be divided into three categories:

• Direct water use: Water consumed at data centers for cooling.
• Indirect electricity-related water use: Water consumed by power plants that generate electricity for data centers.
• Supply-chain water use: Water used in semiconductor manufacturing, equipment production, construction and fuel extraction.

Indirect water use can be particularly important. Thermoelectric power plants often require water for cooling, meaning that a data center supplied by a fossil-fuel or nuclear plant may have a substantial water footprint even if the facility itself uses air cooling. Research cited in an Oxford Academic review indicates that roughly three-quarters of data-center water use may arise indirectly through electricity generation.

Data-center water consumption is locally concentrated

Global water totals do not reveal the distributional problem. A data center located in a wet region may have a manageable water impact, while the same facility in a drought-prone basin can compete with households, farms and ecosystems.

A large 100-megawatt data center in India may consume approximately 2 million liters of water per day, equivalent to the daily use of around 6,500 households. India’s data-center water consumption has been estimated at 150 billion liters in 2024–25 and is projected to reach approximately 358 billion liters annually by 2030.

The figures illustrate why water use should be treated as a siting and governance issue rather than merely a corporate-efficiency metric.

India’s AI infrastructure dilemma

India is emerging as an important market for cloud computing, digital services and AI infrastructure. Data-center capacity has expanded from approximately 520 megawatts in 2020 to nearly 1.5 gigawatts, with projections of about 6.5 gigawatts by 2030. Electricity use by Indian data centers is expected to increase from roughly 13 TWh in 2024 to 57 TWh by 2030.

The Union Ministry of Power estimates that AI could add 26.3 GW of electricity demand by 2031–32.

This expansion has economic benefits. Data centers can attract investment, create specialized employment, strengthen digital infrastructure and support domestic AI services. They may also reduce latency for Indian users and improve the resilience of digital systems.

But India faces three structural constraints.

Heat

India is already experiencing rising temperatures, urban heat islands and more frequent episodes of extreme heat. Data centers reject large quantities of heat into their surroundings. This can increase local cooling demand and intensify the pressure on electricity systems during periods when households also require more air conditioning.

The issue is circular:

1. Data centers consume electricity.
2. Electricity is converted into server heat.
3. Cooling systems consume additional electricity and water.
4. Fossil-fuel electricity increases greenhouse-gas emissions.
5. Climate change raises ambient temperatures.
6. Higher temperatures increase cooling requirements.

Poorly planned facilities can therefore contribute to the conditions that make their own operation more resource intensive.

Water stress

Several Indian data-center hubs are located in or near areas facing water stress. Mumbai, Hyderabad, Bengaluru, the National Capital Region and parts of Maharashtra, Telangana, Gujarat and Rajasthan face different combinations of groundwater depletion, variable rainfall and competing urban demand.

The lack of mandatory public disclosure is a major governance weakness. Approval authorities and local communities need information on:

• Daily water withdrawals.
• Seasonal water demand.
• Water source and quality.
• Consumption rather than only withdrawal.
• Discharge and treatment.
• Use of recycled or treated wastewater.
• Dependence on municipal supplies.
• Water availability under drought conditions.

Water consumption cannot be evaluated only through annual company-wide sustainability reports. Site-level data are necessary because environmental impacts are local.

Grid congestion

India’s renewable-energy capacity is growing, but transmission infrastructure is not always expanding at the same pace. The Hindu reported that India curtailed around 300 gigawatt-hours of renewable electricity in the first quarter of 2026 because the grid could not carry it. It also reported delays affecting renewable connectivity and transmission projects.

This matters regarding AI facilities because they require uninterrupted electricity. If renewable power cannot be transmitted to demand centers, data centers may rely more heavily on coal, gas or local backup generation.

The challenge is not simply a shortage of installed generation. It is a coordination problem involving:

• Transmission capacity.
• Storage.
• Grid connection queues.
• Flexible demand.
• Renewable curtailment.
• Tariff design.
• Data-center sitting.
• State-level incentives.

Land, minerals and electronic waste

The environmental footprint of AI begins before a server is switched on.

Semiconductor manufacturing

AI accelerators require advanced semiconductor manufacturing, a process involving high-purity water, chemicals, energy-intensive fabrication and complex global supply chains. Semiconductor plants can have significant water and energy requirements, while mining for copper, cobalt, nickel, lithium and other materials affects land, biodiversity and local communities.

AI hardware also has a relatively short economic replacement cycle. New chips may make previous generations less competitive even when the older hardware remains functional. Rapid turnover increases demand for mining, manufacturing and transport.

Construction and land conversion

Data centers require large buildings, substations, cooling infrastructure, access roads, backup generators and transmission connections. Their development can convert agricultural or natural land into industrial use.

Environmental impacts may include:

• Habitat fragmentation.
• Loss of agricultural land.
• Stormwater runoff.
• Construction dust and noise.
• Increased traffic.
• Pressure on local housing.
• Heat-island effects.
• Visual and landscape changes.

Because data centers are frequently treated as strategic or essential infrastructure, planning processes may priorities speed over cumulative environmental assessment.

Electronic waste

Discarded servers, batteries, cooling equipment and networking hardware contribute to electronic waste. Recycling advanced components is technically difficult, and informal recycling can expose workers and ecosystems to hazardous substances.

A credible sustainability assessment should therefore include the full life cycle:

Operational electricity is important, but it is not the entire footprint.

The economics of the AI energy race

The AI boom is generating a new form of infrastructure competition. Technology companies are investing heavily in computing capacity, while utilities, equipment manufacturers, energy developers and governments are responding with new generation and transmission projects.

The IEA reported that capital expenditure by five major technology companies exceeded US$400 billion in 2025 and was expected to increase by a further 75% in 2026. It also noted that technology companies accounted for about 40% of corporate renewable power-purchase agreements signed in 2025.iea

This creates both opportunities and risks.

Potential economic benefits

AI data centers can:

• Increase investment in electricity generation and transmission.
• Accelerate development of batteries and advanced cooling systems.
• Support renewable energy projects.
• Create demand for nuclear and geothermal technologies.
• Improve industrial productivity.
• Enable energy-system optimization.
• Attract high-value digital industries.

The IEA estimates that proven AI applications could reduce energy costs in energy-intensive industries by 3–10 percentage points and could save more than 13 exajoules of energy by 2035 if adoption barriers are overcome.

Risks of socialized costs

The private benefits of AI infrastructure may be concentrated among technology companies and investors, while the costs are distributed across society through:

• Higher electricity tariffs.
• Public expenditure on transmission upgrades.
• Water-system expansion.
• Air pollution.
• Local environmental degradation.
• Reliability risks during peak demand.
• Subsidies and tax concessions.
• Stranded fossil-fuel assets.

If data-center projects receive discounted electricity or tax exemptions without paying their full infrastructure costs, households and small businesses may effectively subsidies the intelligence revolution.

Can nuclear power solve the AI energy problem?

Nuclear power is increasingly being discussed as a source of dependable, low-carbon electricity for AI facilities. Technology companies have announced or explored agreements involving nuclear power and small modular reactors.

Nuclear energy can provide high-capacity-factor electricity with low operational carbon emissions. However, it does not eliminate all environmental concerns. Nuclear projects involve:

• Long development timelines.
• High capital costs.
• Radioactive-waste management.
• Cooling water requirements.
• Construction impacts.
• Regulatory complexity.
• Public acceptance challenges.

Small modular reactors may eventually provide additional flexibility, but they cannot be treated as an immediate solution to data-center growth. In the near term, efficiency, demand flexibility, renewable energy, storage, transmission and better sitting remain essential.

What sustainable AI data centers should look like

A low-impact AI facility requires more than a renewable-energy claim. Sustainability should be designed into the project from the beginning.

Energy measures

Data-center operators should:

• Match electricity consumption with clean generation on an hourly basis.
• Locate facilities near abundant low-carbon electricity.
• Sign power-purchase agreements that add new renewable capacity.
• Install batteries and long-duration storage.
• Provide demand response to the grid.
• Schedule flexible training workloads when renewable electricity is available.
• Publish power-usage effectiveness and carbon-intensity data.
• Avoid relying on diesel and unabated gas for routine operation.

Cooling measures

Operators should:

• Use closed loop cooling where technically suitable.
• Deploy direct-to-chip or immersion cooling for high-density AI servers.
• Prefer recycled or treated wastewater over potable water.
• Avoid evaporative cooling in severely water-stressed basins.
• Publish water usage effectiveness.
• Disclose seasonal and site-level water consumption.
• Reuse waste heat in district heating, agriculture or industrial processes where practical.

The IEA notes that advanced AI server racks are reaching unprecedented power densities. By 2027, an individual rack could have peak demand equivalent to approximately 65 households, while AI-server power density increased elevenfold between 2020 and 2025. Cooling strategies must therefore evolve alongside chip performance.

Circular-economy measures

Responsible operators should:

• Extend server lifetimes where performance permits.
• Reuse equipment for less demanding workloads.
• Establish take-back and certified recycling systems.
• Disclose hardware replacement rates.
• Design facilities for modular upgrades.
• Reduce embodied carbon in construction materials.
• Require suppliers to disclose water, energy and emissions data.

Policy recommendations

1. Require site-level environmental disclosure

Governments should require operators to report:

• Electricity consumption.
• Hourly or monthly carbon intensity.
• Water withdrawal and consumption.
• Cooling technology.
• Wastewater discharge.
• Backup-generator operation.
• Hardware procurement and disposal.
• Scope 1, Scope 2 and relevant Scope 3 emissions.

2. Link incentives to performance

Tax concessions, land benefits and electricity subsidies should depend on enforceable benchmarks rather than broad claims of economic benefit.

Eligibility could be linked to:

• Renewable-energy matching.
• Water-use limits.
• Zero-potable-water cooling.
• Grid flexibility.
• Local employment.
• Waste-heat recovery.
• Transparent reporting.
• Payment for transmission and generation upgrades.

3. Introduce water-sensitive sitting rules

Data centers should not be approved solely on the basis of land and connectivity. Planning must consider basin-level water availability, drought projections and competing municipal and agricultural demand.

High-water-use facilities should be restricted in critically stressed basins unless they use closed-loop systems and non-potable water.

4. Make data centers grid-interactive

Large AI facilities should provide services in exchange for grid connection or market benefits. These may include:

• Load shifting.
• Temporary curtailment.
• Battery storage.
• Frequency regulation.
• Flexible model-training schedules.
• Participation in demand-response markets.

The IEA estimates that global data centers could host around 20–25 GW of battery storage by 2030, potentially turning them from passive loads into grid assets.iea

5. Prevent cost shifting

Regulators should ensure that data centers pay their fair share of grid expansion, backup capacity and system services. Costs should not automatically be recovered through higher household tariffs.

6. Establish an AI energy-efficiency label

An “AI Energy Star” type rating could allow users, businesses and public agencies to compare models based on:

• Energy per task.
• Water per task.
• Carbon per task.
• Model size.
• Hardware efficiency.
• Computational intensity.

Such a system would make environmental performance visible in procurement and product design.

Conclusion

AI data centers are transforming electricity systems from behind the screen. The intelligence revolution requires physical infrastructure: power plants, transmission lines, semiconductor factories, cooling systems, water supplies, land and minerals.

Global data-center electricity demand is projected to rise from 485 TWh in 2025 to 950 TWh by 2030, while AI-focused demand is expected to triple. At the same time, AI systems may consume hundreds of billions of liters of water annually, generate rising emissions and intensify pressure on already constrained grids and water basins.

The problem is not that AI consumes energy. Every modern technology consumes resources. The problem is that AI infrastructure is expanding faster than environmental accounting, grid planning and water governance.

A sustainable intelligence revolution will require more than efficient chips and corporate net-zero pledges. It will require transparent data, hourly clean-energy matching, water-sensitive siting, circular hardware systems, flexible electricity demand and regulations that prevent private gains from creating public environmental liabilities.

AI can help optimize grids, reduce industrial energy use and accelerate climate innovation. But those benefits will not occur automatically. Without strong governance, the same technology promoted as a tool for solving environmental problems may deepen the energy, water and climate pressures it claims to address.

The future of AI will therefore be determined not only by model capability, computing power or investment capital, but by whether societies are willing to place ecological limits on the infrastructure of intelligence.

Frequently Asked Questions

How much electricity will AI data centers use by 2030?

Global data-center electricity consumption is projected to reach approximately 950 TWh by 2030, up from 485 TWh in 2025. AI-focused data-center electricity consumption is expected to triple during that period.

Are AI data centers causing a global energy crisis?

They are not the sole cause of a global energy crisis, but they are creating serious regional and local pressures. Their concentrated demand can strain grids, delay coal-plant retirements, increase gas generation and raise infrastructure costs.

How much water do AI data centers consume?

A 2025 study estimated the global water footprint of AI systems at approximately 312.5–764.6 billion liters. The exact figure varies depending on whether direct cooling, electricity generation, semiconductor manufacturing and other supply-chain activities are included.

Is renewable energy enough to make AI sustainable?

Renewable energy is necessary but not sufficient. Sustainability also requires new clean generation, transmission, storage, hourly matching, efficient cooling, water safeguards, responsible hardware production and transparent reporting.

What is the main environmental cost of AI?

The main costs are electricity consumption, carbon emissions, water use, semiconductor manufacturing, mining, land conversion, waste heat and electronic waste. The most severe impacts occur where AI facilities are built in water-stressed regions or connected to carbon-intensive grids.

What should India do about AI data centers?

India should adopt a national framework requiring site-level electricity and water disclosure, grid-impact assessments, renewable-energy standards, water-sensitive siting, wastewater reuse, demand response and full payment for infrastructure costs. This is especially important because Indian data-center electricity demand is projected to grow from about 13 TWh in 2024 to 57 TWh by 2030.

Sources

• International Energy Agency — Energy and AI
• IEA — Energy demand from AI
• IEA — Energy supply for AI
• U.S. Department of Energy — Data-centre electricity demand
• Lawrence Berkeley National Laboratory — 2024 United States Data Center Energy Usage Report
• Jegham et al. — How Hungry is AI?
• Google — 2025 Environmental Report