Clean Commitments, Clean Data Centres?

The rapid expansion of artificial intelligence is driving a parallel expansion of the infrastructure on which it depends. Data centres consumed around 415 terawatt-hours (TWh) of electricity in 2024, roughly 1.5% of global demand, and the International Energy Agency's base case projects about 945 TWh by 2030. Governments in the Gulf and Europe have responded by treating data centres as strategic assets while remaining committed to decarbonising their energy systems. Yet a commitment to clean energy does not necessarily amount to clean data centres.

This distinction is becoming increasingly important because the timelines differ so sharply. A data centre can be built in one to two years, whereas transmission takes far longer and new clean generation, nuclear in particular, longer still. When demand arrives before supply, the gap tends to be filled by gas. Ireland illustrates the dilemma: data centres accounted for 23% of its metered electricity in 2025, up from 5% in 2015, and the regulator now requires new large facilities to bring their own generation or storage. The central question, therefore, is whether AI infrastructure and the green transition can advance together.

The rapid expansion of artificial intelligence is creating a parallel expansion of the physical infrastructure needed to develop and run it. Data centres are not new, but generative AI has changed the scale and intensity of their energy requirements. Training large models and answering billions of user requests depend on clusters of specialised processors running continuously, which makes AI one of the fastest-growing sources of data-centre demand.

According to the International Energy Agency (IEA), data centres used about 415 terawatt-hours (TWh) of electricity in 2024, roughly 1.5% of global consumption. In its base case, demand more than doubles to about 945 TWh by 2030, and electricity demand from AI-optimised facilities more than quadruples. Beyond 2030 the uncertainty widens: the IEA’s 2035 range runs from roughly 700 to 1,700 TWh, depending on how fast AI adoption and efficiency gains evolve. The investment behind this demand is on a similar scale. McKinsey estimates that data centres will need about $6.7 trillion of capital expenditure worldwide by 2030, of which $5.2 trillion is for AI-capable facilities. This includes chips and servers as well as buildings. The intensity reflects a basic difference between conventional computing and AI workloads. AI systems rely on graphics processing units and other accelerators that perform vast numbers of calculations but also generate considerable heat, so AI facilities need more power per rack and more sophisticated cooling.

The footprint does not end with training, because every later interaction consumes electricity too. Per-query estimates are contested. A widely cited 2023-era figure of about 3 Wh per ChatGPT request implied a tenfold gap with a conventional search, but a more recent Epoch AI analysis puts a typical GPT-4o query at about 0.3 Wh, and reasoning-heavy queries use considerably more. The point holds either way: the footprint depends less on any single query than on the volume of inference multiplied across billions of daily interactions.

Data centres also use water, mainly for cooling, through chilled-water or evaporative systems. One estimate puts on-site cooling at around two litres per kilowatt-hour, but actual use varies widely with facility design, climate, cooling technology and water source. Indirect water use, embedded in the electricity supplying the site, can be larger than on-site use. This gives AI’s footprint a geographical dimension: the same facility imposes a very different burden in a water-abundant region than in one facing drought.

That geography is also what makes the electricity impact concentrated. Locally, a single 100 MW data centre can use as much power as 100,000 households. In Ireland, data centres consumed 7,663 GWh in 2025, 23% of metered electricity, up from 5% in 2015. The regulator expects the share could reach about 31% by 2034, and in 2021 it imposed a de facto halt on standard new connections in the Dublin region. In Frankfurt, data centres account for up to 40% of the city’s power demand, and the local operator told new applicants to expect connection dates beyond 2032. These are the clearest cases of digital demand colliding with physical grids.

Whether this demand harms the climate depends on what generates the electricity. Where grids are largely renewable or low-carbon, operational emissions can be modest. Where growth is met by coal or gas, AI expansion prolongs fossil dependence. The IEA expects renewables to supply an additional 450 TWh to data centres by 2035, but natural gas to grow by 175 TWh, notably in the United States. A company that buys renewable certificates or power purchase agreements may still draw fossil power from the physical grid at the hours its servers run, which is why market-based and location-based emissions accounting can tell different stories about the same facility.

At the global level, the IEA estimates that emissions from data-centre electricity use are about 180 million tonnes of CO₂ today, rising to 300 million tonnes in the Base Case by 2035 and up to 500 million tonnes in the Lift-Off Case, a rise of about two-thirds in the base case. Even the higher figure stays below 1.5% of energy-sector emissions, but data centres are among the few sectors whose emissions are still rising. The IEA also notes that widespread adoption of existing AI applications could cut emissions by more than the data-centre sector adds. That is a potential offset, not a guaranteed one, and it depends on adoption happening in energy-intensive sectors.

Other local and upstream effects add to this picture. Cooling relocates heat rather than eliminating it, and suggest this can contribute to local warming. Backup diesel generators add intermittent but concentrated air pollution where clusters are dense. Facilities occupy land that may previously have been agricultural or natural, and the servers themselves depend on semiconductors and metals whose mining, processing and manufacturing carry their own energy, pollution and ecosystem costs.

Taken together, these impacts are unevenly distributed. The services are consumed globally, while the electricity, water, land and air-quality burdens fall on specific grids, watersheds and communities. Whether AI infrastructure can grow without undermining climate goals therefore depends less on efficiency improvements to individual models than on where facilities are built and what power and water they draw on. The next sections test this in two regions with ambitious green-transition agendas: the Gulf and Europe.

Whether AI infrastructure is compatible with the green transition depends not only on how much electricity it uses, but on when and where that electricity is produced. The key question is whether new AI demand is matched by new low-carbon electricity at the same time and in the same location. This is a major challenge because the infrastructure is built on very different timelines. A data centre can be completed in around 18 months, while new transmission lines can take four to eight years, and new clean generation, particularly nuclear power, can take even longer. If AI demand arrives before new clean electricity and grid infrastructure are ready, the gap will have to be filled by existing generation, which in the Gulf region for example is still largely based on natural gas. The IEA’s base case illustrates this tension as renewables are expected to provide more than 450 TWh of additional electricity for data centres by 2035, but natural gas is also expected to increase by 175 TWh.

This makes four conditions particularly important. The first is firm clean supply, whether AI demand is met by additional low-carbon generation rather than by redirecting electricity that is already being used by other consumers. The second is grid access and timing: whether electricity connections and transmission infrastructure can be built quickly enough to match the construction of new data centres. The third is the water-cooling-energy trade-off: in hot and water-scarce regions, reducing water use can require more electricity, while reducing electricity use for cooling can increase water consumption. The fourth is procurement and accounting rules: whether data-centre operators can help finance new clean electricity and whether their overall energy and water footprint is properly measured. These challenges differ across regions. The Gulf is energy-rich but water-scarce and still heavily dependent on hydrocarbons, while Europe has cleaner electricity systems but faces constraints in grids, permitting and infrastructure development. The Gulf therefore provides an important test of whether rapid AI expansion can be aligned with the green transition.

The UAE, Saudi Arabia and Qatar have all identified AI infrastructure as a national priority. However, all three are starting from electricity systems that remain heavily dependent on hydrocarbons. This creates a challenge because rapid AI growth could increase demand for fossil-fuel-based electricity. At the same time, it creates an opportunity: these governments have significant control over energy planning, investment and infrastructure development, giving them greater ability to coordinate AI expansion with the transition to cleaner energy. The starting position differs across the three countries.

The UAE has the strongest starting position because its AI plans can draw on a combination of nuclear, solar and gas, and Abu Dhabi has approved a transmission corridor connecting the planned campus to the grid and Barakah. However, the key issue is additionality. Barakah already supplies electricity to existing users, so its contribution to the AI expansion should not automatically be treated as new clean supply. The UAE plans to add around 21.7 GW of generation capacity by 2030, including 12 GW of solar, and the extent to which this new capacity keeps pace with AI demand will determine how much of the new load can genuinely be supported by low-carbon electricity.

Saudi Arabia starts from a much more fossil-fuel-dependent position. Renewable capacity nearly doubled to around 12.3 GW in 2025, but renewables accounted for less than 1% of generation in 2024. At the same time, around two-thirds of the generation capacity under construction in early 2025 was gas-fired. This means that the country’s 50% renewable target by 2030 will already require a major acceleration in clean-energy deployment before the additional demand created by AI is taken into account. The scale of the planned AI expansion therefore makes the timing of renewable investment particularly important.

Qatar faces a different challenge. Its AI proposition is partly based on access to relatively cheap and abundant electricity, but that electricity remains overwhelmingly gas-based. Qatar’s national strategy projects electricity demand rising from around 51 TWh in 2021 to 80 TWh in 2040, a projection that predates the current AI expansion. At 1.5–2 GW, data centres alone could therefore account for a substantial share of the projected increase in demand. The issue for Qatar is whether the cost advantage created by its gas resources can continue to support AI growth without increasing dependence on a carbon-intensive electricity system. The gap between announced and forecast AI capacity is particularly important because the two figures can differ significantly. In the UAE, for example, announced capacity is several times larger than some current forecasts. This creates a planning problem. If governments plan infrastructure around the largest announcements, they risk building more generation and transmission capacity than is ultimately required. If they plan around lower forecasts, they could face shortages if data-centre projects are delivered faster than expected. The challenge is therefore to ensure that generation, transmission and data-centre construction move forward on compatible timelines.

Water is an additional constraint across all three countries because they rely heavily on desalination. In Qatar, desalination is integrated with gas-fired cogeneration, meaning that water used for cooling is indirectly linked to gas consumption. The choice of cooling technology therefore involves a trade-off between water and electricity use. This is particularly important in the Gulf, where high temperatures increase cooling requirements and water resources are limited. Current efficiency levels also show that there is room for improvement: data centres in the Middle East have an average power usage effectiveness (PUE) of 1.79, compared with 1.56 globally, while newer UAE facilities are targeting levels below 1.5. However, improving efficiency alone will not remove the underlying pressure on electricity and water systems. One estimate puts Saudi Arabia’s data-centre water use at around 15 billion litres in 2024, although data on water consumption remain limited.

The ability of data-centre operators to directly support new clean-energy investment is another important factor. Wood Mackenzie notes that UAE regulations currently prevent data-centre operators from signing corporate power purchase agreements, meaning that hyperscalers cannot directly use their electricity demand to finance new clean capacity. Europe provides a different model: dedicated renewable PPAs for European data centres increased from around 1.5 TWh in 2021 to almost 15 TWh in 2025. The comparison highlights how procurement rules can influence whether AI demand simply increases pressure on the existing electricity system or helps create additional clean generation.

Reading across,The binding constraint differs across the three countries. In the UAE, the central issue is whether AI demand is matched by additional clean generation and whether procurement rules allow data-centre operators to contribute directly to that expansion. In Saudi Arabia, the key challenge is the speed of renewable deployment relative to rapidly growing electricity demand and continued investment in gas-fired generation. In Qatar, the main question is whether an AI sector built around cheap gas-based electricity can be reconciled with longer-term decarbonisation objectives. The Gulf’s challenge is therefore not simply whether it can build AI infrastructure, but whether clean electricity, grid capacity and water infrastructure can be developed quickly enough to support it. If AI infrastructure grows faster than these systems, fossil fuels are likely to fill the gap in the short term. If the two are developed in parallel, the region has an opportunity to make AI expansion part of, rather than a constraint on, its broader energy transition.

On the other side of the world, Europe starts from the opposite end of the spectrum from the Gulf. Its grids are cleaner, its climate commitments are binding in law (the EU’s 2030 renewables target is 42.5%, to verify), and its regulation of data centres is the most developed anywhere. What it lacks is what the Gulf has in abundance, spare network capacity. In June 2026 the Commission proposed the Cloud and AI Development Act (CADA), which aims to triple EU data-centre capacity in five to seven years with roughly €200 billion of mostly private investment, using simplified permitting and priority grid connections for projects meeting sustainability criteria. At the same time, Europe must electrify transport, heating and industry on the same grids.

Data centres currently account for around 2.5% of EU electricity consumption, and EU capacity is expected to more than double by 2030, from 12 GW to 28 GW. Estimates of future demand diverge, partly because scopes differ. ENTSO-E puts European data-centre demand at 87 TWh in 2024, 134 TWh by 2030 and 199-254 TWh by 2035, while Ember, working from ICIS data, has 96 TWh, 168 TWh and 236 TWh. S&P Global is higher still, with 145 TWh in 2025 rising to 238 TWh by 2030. The competition for clean power is direct: Ember puts data-centre demand growth to 2030 at 72 TWh, more than electric vehicles (67 TWh) and comparable to electrified industry (80 TWh).

The main risk is the marginal source of generation. In countries where the extra electricity still comes from gas or coal, each added terawatt-hour of AI demand locks in higher emissions. Developers are responding by moving toward clean supply. Rabobank notes that renewable availability and the maturity of PPA markets now determine where large facilities can secure decarbonised power, and Ember projects the share of capacity in the five main hubs (Frankfurt, London, Amsterdam, Paris, Dublin) falling from 62% today to 51% by 2035. Grid access and timing is where Europe binds hardest. A data centre takes one to two years to build, but connection queues in the main hubs average seven to ten years, and Dublin and Amsterdam have already paused new projects. Data centres use 33-42% of electricity in Amsterdam, London and Frankfurt and almost 80% in Dublin. Nationally, EirGrid forecasts Irish data centres reaching 13.3 TWh, or 30% of national demand, by 2032. The IEA expects European installed capacity to grow by only about 70% by 2030, partly because of these delays.

Ireland’s regulator has tried a “bring your own generation” approach. New data centres must provide generation or storage matching their requested connection capacity, and that capacity must participate in the electricity market. This eases the grid problem, but the proposal set no minimum requirement for renewable electricity procurement, and several operators have looked to on-site generation or gas network connections to get around constraints. Grid adequacy and decarbonisation are therefore not the same objective.

Cooling, water and heat is Europe’s distinctive lever is waste heat, which can feed district heating. Germany’s Energy Efficiency Act required heat reuse and tight efficiency limits, but on 24 June 2026 the federal cabinet approved an amendment that raises the efficiency thresholds for existing data centres, replaces mandatory heat reuse with a cost-benefit analysis, and pushes the 100% renewable requirement from 2027 to 2030. On water, the Commission’s new rating scheme will measure it for the first time, and around half of reporting data centres exceeded the suggested reference values for energy efficiency, water efficiency or both.

Relating to Procurement and accounting, Germany’s rule requiring 100% renewable electricity is met on a balance-sheet basis, and academic analysis cautions that energy certificates are no guarantee a facility actually consumes renewable energy. The EU’s rating scheme, which applies to facilities above 500 kW, assesses power-use efficiency, water-use efficiency and low-emission energy use, and the Commission says it can recognise nuclear. However, it imposes no limits on energy or water use and does not require disclosure of total power consumption. Critics note that a facility with a high efficiency score could still draw an undisclosed gigawatt, and minimum performance standards are not due until 2027.

The expansion of AI infrastructure and the green transition are not inherently opposed, but they do not align automatically. Data centres can be built in one to two years, while the grids and clean generation that serve them take far longer. Where that gap opens, it tends to be filled by gas. Compatibility therefore depends less on how much electricity data centres use than on whether clean supply arrives at the same pace as the buildings.The two cases show different versions of this problem. The Gulf has abundant energy, capital and planning capacity, but its grids remain largely fossil-fuelled, its water depends on desalination, and announced capacity far exceeds independent forecasts. Europe has cleaner grids and stronger regulation, but connection queues of up to a decade and an aim to triple capacity make grids and permitting the main constraint. Both also share a transparency gap: where facilities do not disclose how much electricity they use, forecasts diverge and planners cannot see the demand they must serve.

Three steps follow. Connections and permits can be linked to additional clean supply. Disclosure of electricity use, water use and carbon intensity can be built in from the start, which is a particular opportunity for markets still building capacity. And capacity can be phased against delivered clean generation instead of announcement dates. These findings are indicative rather than global, but they identify the questions any region expanding AI infrastructure alongside a green transition will need to answer.

Alvarez & Marsal. “Financing Saudi Arabia’s Data Center Build-Out.” August 6, 2026. https://www.alvarezandmarsal.com/node/97616.

Central Statistics Office. “Data Centres Metered Electricity Consumption 2025.” July 7, 2026. https://www.cso.ie/en/releasesandpublications/ep/p-dcmec/datacentresmeteredelectricityconsumption2025/keyfindings/.

Commission for Regulation of Utilities. Large Energy Users Connection Policy Decision Paper. CRU/2025236. December 12, 2025. https://cruie-live-96ca64acab2247eca8a850a7e54b-5b34f62.divio-media.com/documents/CRU2025236_Large_Energy_User_connection_policy_decision_paper.pdf.

DLA Piper. “New Irish Large Energy Users Connection Policy.” January 7, 2026. https://www.dlapiper.com/en/insights/publications/2026/01/new-irish-large-energy-users-connection-policy.

Ember. “Grids for Data Centres: Ambitious Grid Planning Can Win Europe’s AI Race.” June 19, 2025. https://ember-energy.org/latest-insights/grids-for-data-centres-ambitious-grid-planning-can-win-europes-ai-race/.

International Energy Agency. “AI and Climate Change.” In Energy and AI. Paris: IEA, 2025. https://www.iea.org/reports/energy-and-ai/ai-and-climate-change.

———. “CO2 Emissions Associated with Electricity Generation for Data Centres by Case, 2020-2035.” Chart, last updated April 3, 2025. Paris: IEA. https://www.iea.org/data-and-statistics/charts/co2-emissions-associated-with-electricity-generation-for-data-centres-by-case-2020-2035.

Reuters. “EU to Require Data Centres to Disclose Energy and Water Efficiency.” September 21, 2026. https://www.thestar.com.my/tech/tech-news/2026/09/21/eu-to-require-data-centres-to-disclose-energy-and-water-efficiency.

S&P Global Market Intelligence. “European Datacenter Power Demand to Double by 2030, Straining Grids.” July 2025. https://www.spglobal.com/market-intelligence/en/news-insights/articles/2025/7/european-datacenter-power-demand-to-double-by-2030-straining-grids-91488155.

Zewe, Adam. “Explained: Generative AI’s Environmental Impact.” MIT News, January 17, 2025. https://news.mit.edu/2025/explained-generative-ai-environmental-impact-0117.

Zhou, Yanxi, and Tao Zou. “Who Pays for AI? Data Centres, Electricity Costs and Regional Energy Equity.” Grantham Research Institute on Climate Change and the Environment, London School of Economics, August 20, 2026. https://www.lse.ac.uk/granthaminstitute/news/who-pays-for-ai-data-centres-electricity-costs-and-regional-energy-equity/.