Last October, Dario Amodei, the CEO of the AI company Anthropic, published an online pamphlet entitled Machines of Loving Grace. It’s pretty dizzying stuff. He predicts that “powerful AI” will speed up progress in biology by “10x” or even “100x”, meaning we’ll see centuries of advances in mere decades.
He foresees humans being able to live until 150 by the century’s end, and also dropped in the prospect of existential cyberwars, which would see liberal democracies fight with authoritarian regimes over control of this God-like technology.
Still, when he was interviewed by the BBC’s Amol Rajan from the AI Action summit in Paris last week, Amodei struck a more humdrum note. When Rajan pressed him on what role an unfashionable, liberal democracy like Britain might play in this future, Amodei’s answer was bracingly prosaic. “Energy and data centres,” he said. “It requires a large amount of energy to train and serve the models.”
It’s a little deflating, isn’t it? We wanted robots to do our laundry. What we will get is vast warehouses of polluting computer servers.
For here is the material reality of the current boom. Large-language models such as ChatGPT and Google Gemini need physical infrastructure: microchips, servers, fibre-optic cables and data centres, which are the ultimate destination of all those discarded selfies, unnecessarily CC’d emails and Dall-E artworks.
Data centres already consume 460 terawatt hours of electricity per year, or 1.3 per cent of the world’s entire supply – about as much as France – and the International Energy Agency predicts this will more than double by 2026. This is how the tech sector came to be a greater carbon-emitter than the aviation industry.
“Data is the dirty secret of the tech industry,” says Matt Watts of the data management company NetApp. “You can see an aircraft or a manufacturing plant. But people don’t realise that every time they send an email or forward a TikTok, they’re pushing something across the networks and into the data centres to be stored. And all of that has to have power going to it.”
Landfill of cat memes
One of Watts’s concerns is that about 68 per cent of this data is basically trash. There are different names for this: dark data, dirty data, single-use data. “It’s data that was created for a purpose that is no longer needed,” he says. “The quantities are now getting bigger thanks to AI.”
Last year, as Google unveiled its AI search engine Gemini, it quietly disclosed that its carbon emissions were up by 48 per cent since 2019. It has since discarded its claims to be carbon-neutral. But Google is still seen as one of the cleanest big tech companies. Open AI’s ChatGPT3 is estimated to produce 4.32g of CO2-equivalent emissions with each prompt – ten times as much as a Google query. One reason that Chinese rival, DeepSeek, freaked out the US AI firms is that it uses far less energy.
All of this poses a conundrum for the UK. Sir Keir Starmer has long identified AI as central to Britain’s future prosperity. In January, he promised to increase Britain’s AI computing power by 20 times by the year 2030. Over in South Mimms by the M25, 85 acres of green belt have been allotted to a data centre known as DC01UK, thought to be Europe’s largest.
The dash towards AI is like the 1980s rush to replace glass with plastic or close railway lines in the 1960s
Only this clashes with another ambitious promise: to create a 95 per cent low-carbon national electricity grid, also by 2030. The idea is only to use natural gas when renewable sources struggle to cope in times of high demand. But AI’s power demands are so high that no one thinks we are likely to meet that target any time soon. “We’re going to have to be pragmatic about the sorts of electricity we use,” says Professor Aoife Foley, an expert in net-zero infrastructure at the University of Manchester.
“Scaling up wind is taking far longer than anticipated. We need more nuclear energy if we want that level of AI – and there’s a 10-year queue to buy the components you need to build a nuclear power station. I’d estimate we’re going to have gas in the system for 70 or 80 years.”
We’ve already had a glimpse of the demand that data can place on the National Grid. South-West London experienced a power cut during the heatwave of July 2022 when one of Google’s London data centres overheated.
And, as Matt Watts warns, this is likely to become more common. He cites the US state of Ohio, which is experiencing a data-centre building boom. The only hitch is that the firm in charge of Ohio’s power has calculated that those under construction will consume the electricity of 21 million households, i.e. nearly double the population of Ohio. Another example closer to home is Ireland, where data centres now use up 21 per cent of the country’s electricity supply.
“That’s an indication of where we could go,” says Watts. “It’s at about two per cent in the UK. The estimate is that it will be about six per cent by 2030, but I think it will go far higher due to AI. The data centres we’ve been building in the UK are designed to support traditional business. The new data centres we’re putting in to support AI require 10, 15, even 20 times the power of the ones we were building before.” The potential environmental costs of all this were off the agenda at the Paris AI summit – where the US delegation lobbied to remove all mentions of climate change from joint statements. Britain later refused to sign an international agreement.
It’s enough to make you question whether the apparent advantages of AI are worth it. Watts runs a quick calculation. If 10 per cent of UK power is going to data centres by 2030, and 20 per cent of data-centre energy consumption is used for storage, and 68 per cent of what’s being stored is useless, that means around 1.5 per cent of our entire energy supply will be spent on cat memes, holiday snaps and the like.
Professor Foley compares the dash towards AI to the rush to replace glass bottles with plastic ones in the 1980s, or the closure of railway lines to make way for cars in the 1960s – decisions which proved destructive. “We go headlong into new technologies without thinking anything through.” The ultimate blame, however, lies with our increased demands. “Data is cheap. When something is cheap, people waste it,” says Professor Foley. “It’s a beast. We’re all going to have to work together to tame it.”
