Just how much power will AI need?

This article is an on-site version of our Moral Money newsletter. Premium subscribers can sign up to get the newsletter delivered twice a week. Standard subscribers can upgrade to Premium here, or explore all FT newsletters.

Visit our Moral Money hub for all the latest ESG news, opinion and analysis from around the FT

Welcome back.

Staff at leading AI companies and institutions are showing worsening signs of mental strain due to worries about the dangers presented by the latest models, the FT reported yesterday.

What about the climate implications of this sector’s breakneck growth? Angst-ridden AI engineers may want to look away now . . .

AI’s growing power problem

How big a deal are the carbon emissions from the AI boom? It depends which high-profile climate advocate you ask.

“Energy-guzzling artificial intelligence is driving up planet-heating pollution from coal, oil and gas,” UN climate change head Simon Stiell told a New York audience on Monday, urging AI companies to power their proliferating data centres with renewables instead of fossil fuels.

A few days earlier, former US vice-president Al Gore told me and my colleague Attracta Mooney that AI emissions “should be put in perspective”, noting that their greenhouse gas impact was still well below that of the world’s uncovered landfills.

Gore has a point: power usage by data centres currently accounts for well below 1 per cent of global emissions, according to the International Energy Agency.

Unfortunately, so does Stiell.

Enter the agents

Tech evangelists like to point out that posing a question to a chatbot like ChatGPT uses far less energy than charging a phone or boiling a kettle. This is true; in fact, the amount of power needed to address a basic AI query has been declining impressively through efficiency gains in both hardware and software.

Yet the aggregate power consumption of data centres has kept rising fast — by 17 per cent in 2025 alone, says the IEA, which projects that the figure will double from that level by 2030.

The rise of agentic AI — through which autonomous bots can perform highly complex tasks without direct supervision or guidance — is now threatening to make that rise even steeper (on top of the fears it’s creating around cyber security and potential human extinction).

In a July paper, researchers at the Korea Advanced Institute of Technology found that an agentic AI system comparable to current commercial services used up to 136.5 times more energy per query than a basic chatbot session.

Last month Zeke Hausfather, climate research lead at payments company Stripe, calculated that his own use of Anthropic’s agentic AI system (he averaged about 20 prompts per day) consumed about a tenth of the power demand of an average US household.

Here’s a very useful chart from Hausfather that puts this in a broader context, and highlights the massive difference in energy demands between this kind of work and a modest ChatGPT habit:

As Hausfather pointed out, he uses “the latest AI tools more than most people”. But his usage pales by comparison with the growing number of businesses that are beginning to deploy agentic AI at an industrial scale. And tech companies are now on a drive for mass consumer adoption.

See, for example, Meta’s launch this month of its “Muse” agent, which will be made available through WhatsApp, Instagram and Facebook. Meta suggests users try asking it to sell their personal belongings or process their medical records, adding: “As you get comfortable, the tasks you hand over can get larger.”

Really large-scale adoption of agentic AI would have big implications for the global energy system. The KAIST team modelled a scenario where the number of daily AI agent requests rises to 13.7bn, roughly the current daily number of Google searches. That would create new power demand of just under 200 gigawatts, the model showed — about half the average load for the entire US power grid today.

That is an illustrative scenario, not a forecast. The crushing failure of Mark Zuckerberg’s previous “metaverse” strategy shows his company can’t be relied upon to identify the next big thing.

But data centres’ rising energy demands have already been pushing electricity prices higher in parts of the US, creating a major political backlash. Rather than be slowed down by the struggle to secure grid electricity, US data centre developers have resorted to commissioning their own power plants at a huge scale. And despite a growing body of research showing that much of this need could be met by systems combining renewable generation and battery storage, they’ve chosen to embark on a gas power procurement frenzy.

Gas is back

Growth in the overall US power sector has shifted overwhelmingly towards low-carbon sources: last year the country’s capacity in renewables and battery storage grew by 51.7GW, compared with a net gain of just 2.7GW for fossil-fired power. Big Tech is now riding to the rescue of the fossil fuel power industry, with announced plans for 189GW of gas generation tied to data centres, according to a report last month by think-tank Global Energy Monitor. Meanwhile, utilities are cancelling planned retirements of coal and gas plants in response to the surging AI demand for grid power.

Analysts at Morgan Stanley this week estimated that annual emissions from data centres — including those linked with their buildings and equipment as well as power usage — would quadruple between 2025 and 2030 to reach 1.2 gigatonnes. That’s more than the annual emissions of Japan, the world’s fourth-largest economy.

While he acknowledged these emissions are “a cause for concern”, Gore — whose asset management firm Generation counts hyperscalers Amazon, Microsoft and Alphabet among its largest shareholdings — pointed to “very persuasive” research on the future emissions reductions that might be driven by advanced AI.

A paper published in Nature last year, by authors including prominent climate economist Nicholas Stern, suggested that these potential savings — through, for example, more efficient transport and advances in alternative proteins — could be as high as 5.4 gigatonnes, well over 10 times this year’s forecast global data centre emissions.

Perhaps. For now, it’s impossible to know whether the climate benefits of AI will materialise at anything approaching that scale. The rapid emissions growth from this sector, in contrast, is already very real.

Connect with 300+ senior delegates and hear from 60+ speakers — including 20+ CEOs — at the Energy Transition Summit in London, 19–20 October. As a Premium subscriber, you will receive a 20% discount on an in-person pass with the code FTpremiumsub.

Smart reads

• A Kremlin-backed fintech moved $6.9bn through global banks including Standard Chartered and Citigroup through a massive forgery operation, the FT has revealed.
• A UK appeal tribunal has struck down a previous ruling against retailer Next for paying workers in its shops less than those in its warehouses. Rightly so, writes John Gapper, arguing that equal pay law cannot ignore the labour market.

A Kremlin-backed fintech moved $6.9bn through global banks including Standard Chartered and Citigroup through a massive forgery operation, the FT has revealed.

A UK appeal tribunal has struck down a previous ruling against retailer Next for paying workers in its shops less than those in its warehouses. Rightly so, writes John Gapper, arguing that equal pay law cannot ignore the labour market.

Recommended newsletters for you

The Climate Graphic: Explained — Understanding the most important climate data of the week. Sign up

The AI Shift — John Burn-Murdoch and Sarah O’Connor dive into how AI is transforming the world of work. Sign up