AI has a hidden cost for everyone

Mark Zuckerberg said last week that Meta would spend hundreds of billions of dollars to build data centres for artificial intelligence as it races other tech giants for dominance.

So what? He said nothing of the energy it would require. AI may be getting smarter, but it is also

• power hungry;
• thirsty for water resources; and
• a boon for those who would benefit from environmental rollbacks.

Powering the future. The global network of roughly 12,000 data centres is described as the central nervous system of AI. Most are located in the US, with more than 500 in the UK. Many are enormous: one planned by Meta will cover an area nearly the size of Manhattan. They can use as much electricity as aluminium smelters due to the energy required for operating servers.

Find the power. The planned closure of some coal-fired plants in the US has already been put back to help meet demand. This is only part of the picture of data centre consumption, which is

• here now: the indirect carbon emissions of Amazon, Microsoft, Alphabet and Meta rose on average by 150 per cent from 2020 to 2023;
• going global: electricity demand from centres is forecast to double by 2030 to slightly more than the entire consumption of Japan; and
• likely to encourage more damaging choices: a US Department of Energy report published this month said the risk of power outages could increase 100 times by 2030 if the US “continues to shutter” oil and gas.

Slow down. Ireland is a European hub for AI infrastructure, with its network of about 80 data centres consuming more than a fifth of the country’s electricity supply. This has led to a moratorium on new data centres in Dublin until 2028.

Going green. Tech giants don’t want to see this happen elsewhere, so are now in a race to source cleaner energy for AI.

• Google last week agreed a $3 billion deal to modernise two hydropower plants in Pennsylvania.
• Meta said in June that it had struck a 20-year deal with a nuclear plant in Illinois to power its data centres.
• Microsoft is preparing to reopen a nuclear reactor at Three Mile Island in Pennsylvania, the site of the most serious nuclear meltdown in US history.

Water, water everywhere. It is not clear that these agreements will substantially stem emissions, nor do they deal with the issue of water, which is required in huge amounts for cooling. One study estimates that data centres could consume 1.7 trillion gallons of water globally by 2027, with some using millions of gallons a day. This is already raising costs and threatening shortages for those living nearby.

Informed choice. It is hard to tell how much water and energy is required for a typical AI inquiry on the consumer end. According to the OpenAI chief Sam Altman, the average ChatGPT query uses the equivalent of one fifteenth of a teaspoon of water, or the same amount of energy as an oven turned on for about a second. He did not provide a source.

Conversely, MIT Technology Review calculated that a charity campaign planned out with AI could use the same amount of energy needed to run a microwave for more than three hours.

Part of the solution. There are ways in which AI may mitigate its own needs, by helping to optimise water resources and providing blueprints for more climate-efficient buildings.

But for now… More energy will be required. Altman warned in Davos last year that AI would consume vastly more power than people expected, and that there would be no way to get there without an energy breakthrough. Zuckerberg had no answer for that.

This article was amended on 24 July 2025. An earlier version said that the indirect carbon emissions of Amazon, Microsoft, Alphabet and Meta rose by 150 per cent from 2020 to 2023. They rose on average by 150 per cent.

Photograph by Jason Alden/Bloomberg via Getty