The Environmental Impact of Generative AI

Foundations
Written By Terri Hart-Lancaster

15 min read

Every AI query comes with a hidden environmental cost. Generative AI systems consume enormous amounts of electricity and water, produce significant electronic waste, and rely on supply chains with substantial environmental impacts. As deployment accelerates, these costs are intensifying at a scale that threatens to undermine climate commitments. Yet they remain largely invisible to users and unregulated by governments.

 
 

How much energy does AI actually use?

The numbers tell a stark story. Global data centres used approximately 415 terawatt-hours of electricity in 2024. The International Energy Agency (IEA) projects this will more than double to around 945 terawatt-hours by 2030, roughly equivalent to the entire electricity consumption of Japan.

The energy demands of AI come from two distinct activities. Training large models requires massive one-time energy consumption (the upfront cost). Running those models to answer queries requires continuous power (the operating cost that never stops).

Training large language models consumes substantial electricity, but in large-scale deployments, inference often dominates total energy use. Inference is when a trained model actually generates responses to user queries. Every time you ask ChatGPT a question or generate an image, that's inference. Once a model is widely deployed and handling millions of queries daily, running it can consume more energy than the initial training did.

The shift toward reasoning models intensifies this pattern. These models think through problems step-by-step before answering, rather than responding immediately. This internal "chain of thought" process uses significantly more energy per query than simpler models, the computational equivalent of showing your work.

Data centres worldwide consumed roughly 1.5 percent of global electricity in 2024, a figure the IEA projects will nearly double to 3 percent by 2030. Within those data centres, AI-specific workloads currently account for approximately 5 to 15 percent of energy consumption. This share is projected to surge to 35 to 50 percent by the end of the decade as AI deployment accelerates.

These centres don't spread evenly across countries. Instead, they cluster in specific regions, creating concentrated demands that strain local electricity grids and water supplies far beyond what national statistics suggest.

 
 

Where does this electricity come from?

The IEA projects that renewables and natural gas will provide large portions of incremental electricity supply for data centres. Renewables are expected to meet about half of global growth in data centre demand through 2035, an optimistic projection that still leaves half dependent on other sources. Fossil fuels will remain a substantial component of supply.

Major cloud providers claim to match 100 percent of their electricity with renewable energy, but these claims warrant scrutiny. They often rely on market-based accounting using Renewable Energy Certificates. These are financial instruments that don't guarantee the actual electrons powering a data centre came from renewable sources. The actual carbon intensity of the grids they draw from can tell a different story, particularly during peak demand hours when fossil fuel plants typically kick in.

Data centres currently account for around 180 million tonnes of indirect carbon dioxide emissions annually from electricity consumption globally. This represents approximately 0.5 percent of global fuel combustion carbon dioxide emissions per year, a figure that will climb as AI workloads expand.

 
 

What about water consumption?

Water represents one of the most significant yet underreported environmental costs of AI. Data centres consume water through two channels. Direct on-site cooling uses evaporative towers, chilled water loops, and liquid cooling systems (the visible consumption). Indirect consumption comes from electricity generation at thermal power plants (the hidden cost).

The indirect component is frequently the larger of the two. The IEA estimates global data centres consumed approximately 560 billion litres of water in 2023, enough to fill more than 224,000 Olympic swimming pools. This breaks down as 140 billion litres for direct cooling, 373 billion litres for electricity generation, and 47 billion litres for hardware manufacturing.

These figures are projected to rise sharply. The IEA estimates global data centre water consumption could reach around 1.2 trillion litres by 2030, more than doubling in just seven years. To understand the scale at the model level, a University of California study estimated that GPT-3's training in Microsoft data centres consumed approximately 5.4 million litres of water in total, including around 700,000 litres of on-site water.

At the per-query level, the same researchers estimated roughly 500 millilitres of water per 10 to 50 medium-length responses, about the size of a standard water bottle for a brief conversation. The estimates vary significantly depending on where and when inference occurs, with location and grid composition playing crucial roles.

 
 

Where are the most acute water pressures?

The geographic distribution of data centres creates concentrated pressures in specific regions already struggling with water scarcity. Analysis shows that a substantial proportion of new data centres built or in development since 2022 are located in areas already facing high water stress, a concerning pattern that prioritises access to power and connectivity over protecting local water supplies.

Water stress varies significantly by region and local conditions, but the scale of demand is sobering. The IEA notes that a single 100 megawatt hyperscale facility can consume around 2 million litres per day in total water. More than 60 percent of this can be attributable to indirect water use from electricity generation, making grid composition a critical determinant of water impact.

Localised impacts depend heavily on cooling system design, local water availability, and whether facilities use potable or non-potable water sources. In water-stressed regions, the competition between data centres and human consumption creates mounting tensions.

 
 

What happens before and after a data centre operates?

The environmental cost of AI begins long before a data centre powers on. The supply chain involves mining for critical materials, semiconductor fabrication, and hardware manufacturing. Each stage carries its own environmental burden.

AI chips require silicon, gallium, germanium, cobalt, lithium, tantalum, and neodymium. Extracting these materials exacts a heavy toll. The World Resources Institute reports that between 2001 and 2020 the world lost nearly 1.4 million hectares of forest to mining and related activities. Mining these materials causes deforestation, water contamination, soil degradation, and habitat destruction. This is environmental damage paid upfront for every chip produced.

Manufacturing semiconductors (the computer chips that power AI systems) is extraordinarily resource-intensive. Facilities consume large volumes of ultrapure water daily. This water must be extensively treated to remove virtually all impurities, a process that itself requires significant energy and resources. The manufacturing process also uses PFAS. These synthetic chemicals are known as "forever chemicals" because they don't break down in the environment and accumulate over time. Manufacturing creates substantial chemical waste that needs careful disposal. Building the hardware creates a large carbon footprint before the AI system even runs.

 
 

What happens to AI hardware at the end of its life?

AI hardware is generating a mounting wave of electronic waste. A 2024 study published in Nature Computational Science projected that generative AI could produce a cumulative 1.2 to 5 million metric tonnes of e-waste between 2020 and 2030.

AI-specific chips become outdated more quickly than general-purpose hardware. As AI models improve, they need more computing power. Some companies replace their AI chips frequently to stay competitive. They treat expensive chips as disposable, despite their environmental cost.

Globally, 62 million tonnes of e-waste were generated in 2022. Only 22.3 percent was documented as formally collected and recycled. The rest ended up in landfills, incinerators, or informal recycling operations with minimal environmental protections. The Global E-waste Monitor projects this could rise to 82 million tonnes by 2030 under business-as-usual assumptions.

Recycling recovers less than 1 percent of rare earth elements used in chips. Recycling programs show promise but operate at small scale. They cannot keep up with how quickly companies replace AI hardware.

 
 

What does this mean for the UK?

Industry estimates suggest the UK has around 500 to 525 data centres. This makes the UK the third-largest data centre market globally. Data centres currently use about 2 percent of UK electricity. That amounts to around 5 terawatt-hours per year, according to 2023 analysis from the National Energy System Operator (NESO).

NESO projects this could grow to 22 terawatt-hours by 2030. This is more than four times current use. It would represent about 7 to 8 percent of total UK electricity demand. Meeting this surge while cutting carbon emissions creates a major challenge.

The UK's relatively clean electricity grid helps. The Department for Energy Security and Net Zero reports that UK electricity generation averaged 154 grams of carbon dioxide per kilowatt-hour in 2024. This figure excludes imports and supply-chain emissions. The UK performs better than many countries that rely heavily on coal.

However, gas still provides a significant portion of UK electricity. This is especially true during peak demand periods. The UK Clean Power 2030 Action Plan aims to reduce carbon intensity to well below 50 grams of carbon dioxide equivalent per kilowatt-hour by 2030.

Success depends on several factors. The UK must build new clean power generation and manage electricity imports effectively. Most importantly, data centres need to draw power when clean sources like wind and solar are actually generating electricity. Simply buying enough renewable energy certificates to match annual consumption does not reduce emissions in real time.

 
 

What are the water challenges in the UK?

Water use by UK data centres is poorly tracked. The House of Commons Library notes that no official figures exist. Companies do not have to report this information.

A survey by techUK and the Environment Agency looked at 73 commercial data centre sites in England. The survey found that 51 percent used waterless cooling systems. Among all the sites surveyed, only 4 percent reported using over 100 million litres per year.

However, these results show only part of the picture. The survey excludes many enterprise and on-premise data centres. techUK and the Environment Agency emphasise the results are not representative of the entire sector.

Different cooling methods use vastly different amounts of water. The House of Commons Library provides examples from Microsoft data centres. Water-cooled chillers can use 2.3 to 2.8 litres per kilowatt-hour. Indirect evaporative cooling can use 0.8 to 2.1 litres per kilowatt-hour. Direct evaporative cooling can use 0.02 to 0.7 litres per kilowatt-hour. Design choices make an enormous difference.

England already has water supply problems. The Environment Agency warns that England could face a shortage of 5 billion litres per day by 2055 if water use continues at current rates. At the same time, other sectors will need an additional 1 billion litres per day.

Data centres add to these pressures in two ways. They need water for cooling. Power plants need water to generate the electricity that data centres consume. As AI expands, both demands grow together.

 
 

Are there planning disputes emerging?

These water concerns are causing planning conflicts across the UK. Local authorities have rejected data centre proposals over impacts on water supplies. This creates tension between local environmental protection and national infrastructure goals.

In January 2026, two campaign groups challenged the government in the High Court. Foxglove and Global Action Plan contested the approval of a large data centre at Woodlands Park in Iver, Buckinghamshire. Buckinghamshire Council had rejected the project over environmental concerns, including water use and green belt impacts. The then-Deputy Prime Minister and Secretary of State for Local Government, Angela Rayner, overruled that decision and approved the data centre on appeal.

The campaign groups argued the approval was unlawful. They said the Ministry for Housing, Communities and Local Government (MHCLG) had not required an Environmental Impact Assessment for the project. They also argued that environmental protections had not been properly enforced. On 19 January 2026, the Government Legal Department wrote to Foxglove and Global Action Plan to concede that the approval contained a "serious logical error" and should be quashed. The High Court granted permission for the claim to proceed.

Planning authorities have also warned that data centre electricity demand could constrain housing development in some areas, creating a direct trade-off between AI infrastructure and residential growth.

 
 

What policy gaps exist?

Data centres were classified as Critical National Infrastructure in September 2024, acknowledging their strategic importance. Yet there is currently no dedicated mandatory reporting scheme specific to data centres on energy, water, or carbon metrics, a striking gap that makes accountability nearly impossible.

The contrast with Europe is sharp. The EU's Energy Efficiency Directive requires data centres above 500 kilowatts to report energy performance indicators, creating transparency and enabling comparison. The UK has no equivalent requirement, leaving operators free to disclose—or not—at their discretion.

 
 

Can AI help solve environmental problems?

AI could potentially deliver environmental benefits. Possible applications include optimising energy grids, improving weather forecasts, and discovering new materials for clean energy. Companies highlight these benefits in their marketing.

But the International Energy Agency offers an important warning. Despite industry claims that AI will reduce emissions, no evidence shows this happening at meaningful scale. The benefits remain theoretical. They are projections without proof.

Another problem is the rebound effect. When AI becomes more efficient, companies use it more widely rather than less. Lower costs per query drive higher usage. This growth often cancels out efficiency improvements.

This pattern appears throughout energy economics. Efficiency gains often lead to increased total consumption rather than reduction. Making something cheaper and better typically means people use more of it.

 
 

How can data centres reduce water use?

New cooling technologies could reduce water use. Closed-loop liquid cooling recirculates water instead of evaporating it. This cuts water consumption dramatically. Air-cooled facilities use minimal water but often need more electricity. Immersion cooling submerges servers in special fluid that transfers heat more efficiently. Some companies use geothermal systems or deep lake water for cooling. These offer low-carbon alternatives.

However, these technologies take years to deploy at scale. Meanwhile, companies continue building data centres that use evaporative cooling. These facilities will consume large amounts of water for years to come.

Data centre operators face a difficult choice. Optimising for water efficiency often conflicts with reducing carbon emissions. Running AI systems when temperatures are cooler saves water. Running them when solar power is abundant saves carbon. Managing these competing demands requires sophisticated planning and regulatory guidance.

 
 

What can reduce AI's environmental impact?

Reducing AI's environmental impact requires four strategies. First, improving efficiency at every level reduces electricity use per task. This makes each query less costly. Second, using genuinely clean electricity cuts operational emissions. This means using low-carbon power when AI systems actually run, not just buying credits that offset annual consumption. Third, choosing water-smart cooling systems and locations reduces strain on freshwater supplies. Fourth, designing hardware for longevity and recovery reduces mining and electronic waste.

The UK needs three specific actions. Data centres should report their environmental impact publicly. This would enable comparison and accountability. Planning decisions should assess local water stress before approving new data centres. This would prevent new facilities in areas that already face water shortages. Where water cooling is necessary, data centres should use non-potable or reclaimed water. This protects drinking water supplies for people.

AI's environmental footprint is not a future problem. It exists now and grows with every new model and data centre. The conflict between AI expansion and climate goals does not have to be permanent. But avoiding lasting damage requires transparency, regulation, and rapid technological innovation. Current policy falls short.

 
 

Further Reading / Sources

Energy Demand & Grid Impacts

How much electricity AI and data centres use — and what that means for power systems.

Water Use & Cooling

The emerging debate around AI’s water footprint and cooling infrastructure.

Critical Minerals, Hardware & E-Waste

The material footprint of AI infrastructure.

UK Policy, Planning, Infrastructure & Law

Relevant for understanding how AI infrastructure intersects with UK energy and planning systems.

Broader Context & Climate Debate

Exploring whether AI worsens or potentially mitigates climate impacts.

 
 
Terri Hart-Lancaster
Previous

AI Tokens and the Role of Randomness
Next

Using generative AI at work