Your AI runs on Denmark's electricity and a river in Spain

Back to Insights

Most of the conversation about AI's environmental cost focuses on carbon. That is the wrong number to watch. The more immediate constraints are electricity, water, land, and the minerals underneath all of it. Here is what the data actually shows.

Electricity first. Microsoft's power consumption more than doubled between 2020 and 2023, reaching approximately 30 TWh by fiscal year 2024, equivalent to Denmark's entire national consumption. Google's most recent environmental report shows an approximately 27% year-on-year rise, putting its consumption at roughly 32 TWh. Amazon's total electricity draw is estimated at around 39 TWh, making it the largest of the four. Meta sits considerably lower, at around 10 to 15 TWh.

Ireland is the clearest indicator of where this is heading. Datacentres consumed 22% of Ireland's total metered electricity consumption in 2024, more than every home in the country combined. Grid constraints have forced tighter controls on new connections in Dublin. That is not a projection for 2030. It is happening now. The rest of Europe is a few years behind. Source: CSO Ireland, June 2025.

Water is the harder constraint. In 2023, Google used enough water for datacenter cooling to supply London for nine days. Microsoft used enough for nearly three days. Google, Microsoft and Meta combined used enough to supply London for thirteen days, and these are direct cooling figures only. The indirect water footprint from generating the electricity that powers the datacenters is estimated at roughly ten times larger. London uses more than 2.6 billion litres every day, per City Hall and Thames Water. Google's 2023 cooling water totalled approximately 24 billion litres. Source: Company sustainability reports 2023; City Hall/Thames Water.

To put this in personal terms: writing and researching this article used an estimated 500ml to 1 litre of water, the direct cooling cost of the AI queries that helped draft, fact-check, and refine it. That is one article. Multiply that by the billions of daily AI interactions happening right now and the aggregate becomes significant even before you count a single server rack.

Land is next. Pre-AI datacenters spanned 100,000 to 300,000 square feet. Today's AI facilities span millions of square feet across hundreds of acres. Microsoft owns approximately 1,575 acres in Mount Pleasant, Wisconsin for a single campus, with plans to acquire a further 1,030 acres at the same site. A Louisiana campus under development covers 2,250 acres and is expected to consume 5 gigawatts of power, larger than many European towns.

Minerals are where it gets geopolitically serious. Datacenters require copper, aluminium, silicon, and rare earth elements for magnets, cooling fans, and storage drives. Over 90% of aluminium, silicon, magnet rare earths, and gallium are sourced primarily from China. The IEA estimates gallium demand alone could exceed 10% of current global supply by 2030, with China controlling 99% of refined production. One lifecycle study found a single datacenter facility requires enough copper, aluminium and rare earth minerals to construct a 12-kilometre railway. That is per facility, before you count the thousands being built simultaneously. Source: IEA 2025.

Electricity gets the headlines because the numbers are dramatic and easy to measure. Water may be the harder practical constraint, particularly as sites cluster in water-stressed regions where land is cheaper. The minerals question is a geopolitical dependency the industry has not yet seriously addressed. The Brookings Institution notes that while datacenters may account for only 3% of global electricity demand by 2030, their share of local demand can reach 42% in Frankfurt or nearly 80% in Dublin. The aggregate number is not the story. The concentration is.