Digging through articles written during the early years of the internet, you find concerns that can seem oddly familiar. “Dig more coal — the PCs are coming,” screamed a 1999 Forbes editorial. “Somewhere in America, a lump of coal is burnt every time a book is ordered online.”

The warnings scarcely need updating. Today fears of an AI energy apocalypse abound: that video of you eating McDonald’s with Donald Trump is depleting reservoirs; using ChatGPT as your counsellor is overwhelming the grid. Proposals for data centres in southern England have even been rejected over water and energy use fears. How much water and energy does artificial intelligence actually use?

Let us start with water, which has become a point of obsession for AI sceptics. Writing an email with ChatGPT apparently uses a whole bottle, according to a 2024 Washington Post article. Last November Karen Hao was forced to issue a correction after accidentally overstating the water usage of a Chile data centre in her bestselling book Empire of AI. By a factor of 1,000.

Both of the above cite a University of California paper which suggests that by 2027 “water withdrawal of global AI is projected to reach 4.2 to 6.6 billion cubic metres” — which is half the water withdrawal of the UK. It sounds alarming. Except water withdrawal includes water that is, well, recycled. The same paper says that AI’s total water “consumption” — water that is evaporated — will be between 0.38 billion and 0.6 billion cubic metres, about a tenth of the original figure.

To add to the confusion, the figure for consumption includes both water that is used to cool data centres (which is drinkable) and water used in the original electricity generation (which is not). So while each basic ChatGPT request to a US data centre consumes about 17ml, just 2ml of it could be used for drinking.

Put all of this together, as the US scientist and blogger Andy Masley has done, and it seems that AI will consume between 0.12 billion and 0.22 billion cubic metres of drinkable water in 2027. In other words, you get the entire world’s AI water use for a mere 1.5 per cent of Britain’s drinking water consumption.

These estimates, too, may now be out of date. Sam Altman insisted last year that ChatGPT used about 0.3ml per query, and a Google Gemini paper comes to a similar estimate of 0.26ml: that’s 20,000 requests for a toilet flush.

What about energy? It depends on what we ask of AI. For both Gemini and ChatGPT, to go by company reports, it’s one kilojoule of energy per text prompt — the equivalent of running a microwave for a second, or about 2 per cent of an iPhone battery charge.

An MIT paper found that basic image generation required 2-4kJ. But videos can be an order of magnitude more energy-intensive: some reports suggest Sora, OpenAI’s text-to-video tool, used 3,600kJ — an hour of microwaving — to make a single video. But the vast majority of us are not making videos every day. Sora was closed down after a few months partly because it used too much energy relative to how few people were using it.

Zoom out and the total data centre electricity bill is modest. Data centres, including those keeping the internet running, use a mere 2 per cent of global electricity, according to the IEA (of which about a quarter is purely for AI). This will rise to 3 per cent by 2030.

To be clear, we are talking about hundreds of additional terawatts of demand. But the extra power needed for AI this decade is dwarfed by the demand from air-conditioning, electric vehicles and industry — 3 per cent of global energy use is incredibly good value for something we spend many hours a day using. What, then, is all the fuss about?

AI’s problem is not that it is hungry per se: it is that it is a rapidly growing industry with highly concentrated energy use. The demand for air-con is massive but dispersed in people’s homes and workplaces. Data centres can hog local demand if grids are unprepared.

National power grids are not designed for such localised demand. Ours was designed a century ago to deliver coal power to people’s homes. In the US some households near data centres are starting to see higher bills.

The world needs energy: not just for AI but for air-conditioning, transport and food production. Industry tends to find a way of supplying it. Better technology will also help us use what we have more efficiently. In 1999, according to Forbes, it took 1lb of coal — 1kWh — to transmit 2MB of data to buy a book on Amazon. Today 1kWh can generate a studio-quality video. Supply will need to grow to match AI’s ambitions. But the worry should be directed not at how much AI uses but at where we choose to build the hubs that power it.