Somewhere in the outskirts of a mid-sized American city, a windowless warehouse the size of several football fields hums through the night. Inside, thousands of processors are running inference queries, training neural networks, and generating text, images, and video on demand. The servers run hot. Cooling systems pump millions of gallons of water. The electricity draw never stops.
This is what powering artificial intelligence actually looks like – and in 2025, the world got its most detailed reckoning yet of what it costs.
The Numbers Have Changed – and Not for the Better
In April 2025, the International Energy Agency (IEA) published its landmark Energy and AI report, the most comprehensive data-driven analysis of AI’s energy footprint to date. The findings set the terms of the debate for the rest of the year.
Global investment in data centres nearly doubled since 2022, reaching half a trillion dollars in 2024. The IEA projects that electricity demand from data centres worldwide is set to more than double by 2030 to around 945 terawatt-hours (TWh) – slightly more than the entire electricity consumption of Japan today.
More striking still, just weeks ago the IEA released a follow-up report that confirmed the surge is already happening. Electricity demand from data centres soared by 17% in 2025, with AI-focused facilities growing even faster – well outpacing growth in global electricity demand of just 3%.
To put that in perspective: US data centres alone consumed 183 terawatt-hours of electricity in 2024, roughly equivalent to the annual electricity demand of the entire nation of Pakistan – and by 2030, this figure is projected to grow by 133% to 426 TWh.
AI-optimised data centres are not merely large – they are energy-intensive in ways conventional facilities are not. AI-focused data centres can draw as much electricity as power-intensive factories such as aluminium smelters, but they are far more geographically concentrated. Nearly half of data centre capacity in the United States sits in just five regional clusters. This geographic compression places acute strain on local electricity grids.
The Carbon Paradox: Big Tech’s Broken Promises
Perhaps the most damning finding of the past year has come not from academic research, but from the companies themselves.
In June 2025, the United Nations’ International Telecommunication Union (ITU) released a report assessing the greenhouse gas emissions of the world’s 200 largest digital companies from 2020 to 2023. The results were striking. Operational emissions for Amazon grew 182% against 2020 levels, while emissions for Microsoft grew 155%, Meta grew 145%, and Google’s parent Alphabet grew 138% over the same period.
These are companies that have all pledged net-zero or carbon-negative futures. Google, which once styled itself a leader in clean energy, reported a 48% rise in greenhouse gas emissions since 2019, driven mostly by energy consumption at its data centres. Microsoft’s emissions rose 23.4% since 2020, despite its pledge to be carbon negative by 2030, mainly due to skyrocketing energy demands from AI and cloud computing.
Investor pressure is mounting. Shareholders have pushed Amazon, Microsoft, and Google for more data on water and power use at US data centres, with one investor group noting that a major tech company’s emissions rose 51% despite pledges to halve them – leaving investors “in the dark” about how it planned to meet its stated goals.
A peer-reviewed study published in Cell Reports Sustainability in December 2025, by VU Amsterdam researcher Alex de Vries-Gao, offered one of the first rigorous estimates of AI-specific environmental impact. The carbon footprint of AI systems alone could be between 32.6 and 79.7 million tonnes of CO₂ emissions in 2025 – comparable to the annual emissions of a small European country – while the water footprint could reach 312.5 to 764.6 billion litres. That water figure is roughly equivalent to the global annual consumption of bottled water.
The study underscored a deeper problem: no company currently reports any AI-specific environmental metrics, despite several acknowledging AI as a key driver of increased energy consumption. Without disaggregated disclosure, independent researchers and policymakers are left working with approximations.
Water: The Crisis Nobody Is Talking About
Energy is visible in electricity bills and carbon pledges. Water is not. Yet data centres are quietly becoming some of the most significant consumers of freshwater in the regions where they operate.
The IEA estimated that global data centres were responsible for 560 billion litres of water consumption in 2023. As AI workloads intensify, that figure is climbing. Research from the University of California, Riverside, projected that combined scope-1 and scope-2 water withdrawal from global AI could reach 4.2 to 6.6 billion cubic metres by 2027.
Unlike electricity, water stress has immediate local consequences. Data centres tend to cluster in regions already facing scarcity – the American Southwest, parts of Europe, and increasingly, water-stressed parts of Asia. The competition between digital infrastructure and basic human needs for freshwater is not a distant hypothetical. In several US states and Ireland, data centres already account for over 10% – and in Ireland’s case over 20% – of total national electricity consumption. The water picture is similarly concentrated.
The CO₂ Trajectory: Peaking, But When?
The IEA’s modelling offers a conditional note of optimism. CO₂ emissions from electricity generation for data centres are projected to peak at around 320 million tonnes by 2030, before entering a shallow decline toward 300 million tonnes by 2035 – as nuclear and renewables increasingly displace coal in the energy mix.
But that decline is contingent on decisions that have not yet been made. The IEA’s World Energy Outlook 2025 found that demand for coal, driven primarily by China, is growing 50% faster than demand for natural gas, resulting in spikes in carbon emissions – even as renewable deployment accelerates.
The tech sector is aware of this tension. It accounted for around 40% of all corporate power purchase agreements for renewables signed in 2025, and has become a major source of momentum for the nuclear and advanced geothermal industries. The pipeline of conditional agreements between data centre operators and small modular reactor (SMR) nuclear projects has grown from 25 gigawatts at the end of 2024 to 45 gigawatts today.
SMRs – compact, factory-built nuclear reactors – are increasingly favoured by tech firms for their round-the-clock baseload power and relatively small physical footprint. Google has signed an agreement with Kairos Power for several SMRs to power its AI infrastructure, while Amazon has committed to an SMR project in Washington state with Energy Northwest. Critics, however, note that most SMR designs are still unproven at commercial scale, and that nuclear projects have historically run over cost and schedule.
The Efficiency Paradox
One nuance that often gets lost in environmental debates about AI is the efficiency improvement happening within the technology itself. The IEA noted in its April 2026 update that power consumption per AI task is declining rapidly, with efficiency improving at a rate unprecedented in energy history.
This is real progress. Newer AI chips perform far more operations per watt than their predecessors. Smaller, more targeted AI models are emerging as viable alternatives to enormous general-purpose systems.
Yet total consumption keeps rising. The reason is what economists call the rebound effect: as AI gets cheaper and faster, more people use it for more tasks – including computationally intensive applications like AI agents, video generation, and scientific simulation. More people are using AI, and energy-intensive uses such as AI agents are on the rise. Efficiency gains are being outpaced by volume growth.
Regulation Steps In – Starting in Europe
The most concrete policy response to date is emerging from the European Union. Under the EU AI Act’s energy provisions, providers of large general-purpose AI models are now required to document energy consumption data, accessible to EU regulators without prior notice. Germany has already enacted a national Energy Efficiency Act requiring data centres with a capacity of 300 kW or more to source 50% of their electricity from renewables – rising to 100% from January 2027.
The European Commission is expected to release a broader Data Centre Energy Efficiency Package in early 2026, alongside a Cloud and AI Development Act that will tie public support for data centre expansion to binding sustainability requirements. This would include measures to sustainably integrate data centre electricity demand into the broader energy system, with the goal of tripling EU data centre processing capacity over five to seven years – but only where operators comply with requirements on energy efficiency, water use, and circularity.
Outside Europe, mandatory disclosure frameworks remain largely absent. In the United States, no federal law currently requires AI companies to report energy or water consumption at the model or system level.
What Comes Next
AI is not going to stop. The capital expenditure of the world’s five largest tech companies surpassed $400 billion in 2025 and is projected to grow by a further 75% in 2026. The infrastructure race is accelerating, not slowing.
What can change is how that infrastructure is built, powered, and accounted for. The path forward involves parallel action across several fronts: accelerating the transition to clean baseload energy, deploying next-generation cooling technologies that dramatically reduce water use, moving to mandatory and AI-specific environmental disclosure, and investing in smaller and more efficient AI architectures that deliver comparable capability at lower computational cost.
The IEA’s conclusion is precise: AI has the potential to transform the energy sector in the coming decade, driving a surge in electricity demand from data centres around the world while also unlocking significant opportunities to cut costs and reduce emissions. That potential is real. So is the risk of squandering it by allowing AI’s own footprint to grow unchecked.
The data is no longer ambiguous. The question now is whether the industry, regulators, and users will treat it with the urgency it demands.