Environment · AI Infrastructure · Data Investigation
The Invisible
Furnaces
How the AI infrastructure boom is supercharging a global energy and water crisis — and why the industry's net-zero pledges are already collapsing under the weight of their own data.
PublishedMarch 2025
Primary SourcesIEA · Lawrence Berkeley Lab · EESI · Corporate Filings
The servers never sleep. In industrial estates on the edges of cities, in converted warehouses across the American midwest, in purpose-built campuses outside Dublin and Singapore and São Paulo, vast halls of humming machines process and transmit the world's data around the clock — and they are growing faster than at any point in history.
The catalyst is artificial intelligence. The International Energy Agency, in its landmark April 2025 report on Energy and AI, estimated that global data centre electricity consumption reached 415 terawatt-hours in 2024 — roughly equivalent to the entire annual electricity consumption of the United Kingdom. That figure has grown at 12% per year since 2017, but the AI surge is dramatically steepening the curve. Electricity consumption in accelerated servers — the GPU clusters powering large language models — is now growing at 30% annually, against 9% for conventional compute.
"A single ChatGPT query draws nearly ten times the electricity of a standard Google search. Multiply that across hundreds of millions of daily users and thousands of competing models."
In the IEA's central scenario, data centre demand will more than double to around 945 TWh by 2030 — equivalent to Japan's entire current electricity consumption. The United States alone accounts for 44% of current global consumption, and will drive nearly half of all growth through the decade. By 2030, the IEA calculates, the US will consume more electricity processing data than it uses to manufacture aluminium, steel, cement, chemicals, and all other energy-intensive goods combined.
Water is the other cost that the industry has been slower to discuss. A typical data centre consumes around 300,000 gallons of water per day for cooling; large hyperscale facilities can use up to 5 million gallons — equivalent to a town of 50,000 people. Collectively, US data centres consumed an estimated 449 million gallons per day in 2021. Northern Virginia, home to more than 300 operational data centres, consumed close to 2 billion gallons in 2023 alone — a 63% increase from 2019.
Key Finding — Carbon Accounting Gap
Guardian analysis of corporate sustainability disclosures found actual data centre emissions from Google, Microsoft, Meta and Apple were likely around 7.62 times higher than officially reported between 2020–22. The gap is explained by industry use of renewable energy certificates (RECs) rather than location-based accounting — a method that allows firms to claim carbon neutrality by purchasing offsets from anywhere in the world, regardless of where servers consume electricity.
The carbon picture is deeply contested. The IEA estimates that electricity use by data centres generated around 180 million metric tonnes of CO₂ in 2024 — but this relies on companies' self-reported market-based figures. When assessed using location-based accounting — which measures the actual carbon intensity of the grids being drawn upon — the numbers look very different. Microsoft's location-based scope 2 emissions more than doubled between 2020 and 2024, rising from 4.3 million to nearly 10 million metric tonnes. Google's location-based data centre emissions rose 22% between 2023 and 2024 alone, even as the company reported a 12% "reduction" in data centre energy emissions through certificate accounting.
Sources: IEA Energy & AI (Apr 2025); Carbon Brief (Sept 2025); Policy Review Institute (2025); Google Environmental Report 2024; Microsoft Sustainability Report 2024.
Nearly half of US data centre capacity is concentrated in five regional clusters. In Dublin, data centres account for 79% of the city's electricity demand. Globally, the IEA warns that clustering makes grid integration disproportionately challenging relative to diffuse demand sources like electric vehicles.
Global Data Centre Clusters — Energy Intensity, 2024
High (>20 TWh)
Medium (5–20 TWh)
Emerging (<5 TWh)
Source: IEA Energy & AI (2025); Carbon Brief analysis; Oeko-Institute (Dublin 79% figure); incorrys.com compilation. Cluster positions are approximate centroids. Northern Virginia alone accounts for ~26% of Virginia state electricity demand.
The geography of this expansion is not accidental. Developers follow cheap land, political permissiveness, and — above all — power. Northern Virginia, already the world's densest concentration of data centres, continues to expand even as its grid strains. In 2024, a single disturbance in Fairfax County caused 60 data centres to switch simultaneously to diesel backup generators, abruptly withdrawing 1,500 megawatts from the grid — roughly the entire power demand of Boston.
Ireland tells a starker story. Data centre electricity use has more than tripled since 2015, now accounting for 21% of national demand. The IEA estimates that share could rise to 32% by 2026. In Dublin itself, data centres already account for 79% of the city's electricity consumption. The country's grid operator EirGrid has warned that planned expansion is structurally incompatible with Ireland's renewable energy transition.
"By 2030, the US is set to consume more electricity for data processing than for the production of aluminium, steel, cement, chemicals and all other energy-intensive goods combined." — IEA, April 2025
Water stress compounds the energy problem. Amazon's facilities in Aragon, Spain consume 500 million litres of drinking water annually. In December 2024, Amazon applied to increase its water consumption permit in the region by 48%, explicitly citing climate change as the reason cooling demands would intensify. Three months later, Aragon applied for EU drought emergency aid. Community group Tu Nube Seca Mi Río — "Your Cloud is Drying My River" — is calling for a moratorium on new data centre construction in the region.
Regulators are beginning to respond. The EU's recast Energy Efficiency Directive now requires data centres above 500kW to report energy use, water consumption and waste heat. The EU AI Act will mandate resource reporting for high-risk AI systems. In the US, Senator Ed Markey's Artificial Intelligence Environmental Impacts Act — which would require the EPA to investigate AI's contribution to climate change — has remained stuck in committee since February 2024. Meanwhile, 17 regulatory bills introduced in Virginia's 2024 legislative session were all defeated or indefinitely postponed.
Sources: WEF (Dec 2025); EthicalGEO (July 2025); Sierra Club investigation (2024); IEA Energy Efficiency Directive briefing.