The demand side is real and quantifiable

Data centres used around 485 terawatt hours of electricity in 2025, a little over 1.5% of global supply. Demand grew 17%, against 3% for electricity as a whole. The IEA expects it to roughly double again to 950 TWh by 2030 (close to 3% of global demand) with AI-focused facilities tripling. Its modelling puts 2035 somewhere between 700 and 1,700 TWh. 

The more useful number is what generates it. Around 60% of the electricity serving data centres came from fossil fuels in 2024. 1 The IEA expects that ratio roughly to invert by 2035, but gas generation serving data centres still more than doubles along the way, from 120 TWh to 293 TWh, and emissions roughly double to around 350 Mt. Cleaner in proportion but considerably larger in absolute terms. 

Water is less visible and harder to pin down. US data centres consumed an estimated 66 billion litres directly in 2023, with a further 800 billion litres consumed generating the electricity they bought 2 . The more telling point is that disclosure barely exists. Among the largest operators, I believe only Google publishes facility-level figures. An unmeasured cost that material is a governance gap. 

In the UK, the story is one of queueing. Nearly 59 GW of data centre capacity is waiting for a transmission connection – that is roughly equal to the entire peak demand of the UK's national grid. However, the National Energy Systems Operator (NESO), expects around 5.2 GW of it to actually be connected by 2030, on a range of 3.7 to 6.3 GW, and notes that a number of the remaining projects are likely to be non-viable. 

None of this is particularly speculative. It is sited, financed and sitting in somebody’s capital plan. 

The optimistic case is stronger than the sceptics allow

On the other side of the argument, the IEA estimates that well-documented AI use cases could save over 13 exajoules of energy by 2035, equivalent to 3% of global final energy consumption. On its own modelling, savings on that scale would more than cancel out the emissions of the data centres producing them. Grid optimisation and load balancing, demand forecasting that lets a system operator hold less spinning reserve and materials discovery that shortens the path to better batteries. 

Some of this is already demonstrated, not just promised. The European Centre for Medium-Range Weather Forecasts (ECMWF) put its own AI model into operational service in February 2025, reporting gains of up to 20% on measures including tropical cyclone tracks. That matters directly for integrating intermittent generation into a grid, one of the hardest operational challenges in a high-renewables system. 

There is a larger version of the case. We manage planetary systems we can barely sense: emissions counted annually, aquifers understood only after the damage is done, decisions taken on electoral cycles about systems that move over centuries. A forest under stress can move resources toward the part that is struggling; a civilisation cannot, because it cannot feel itself. The interesting case for AI (made by Laurie Menoud, At One Ventures) is that it could become a kind of nervous system, showing which water table is falling and which grid node is close to failing, so that millions of separate actors can act before a problem becomes a crisis. That would be an entirely new capability, not a faster version of an old one. However, it likely meets the same wall: a sensing layer changes nothing unless we collectively act on what it shows. 

But note which word is doing the work

Existing. 

The big prize comes from applying solutions we already have, at scale, in sectors that have not adopted them. And the same analysis carries a caveat: the energy sector is not yet making the most of AI, and there is no adoption momentum that would unlock those savings. 

The bottleneck is not invention, it is deployment. 

Which is precisely the bottleneck the transition already had

Climate is a wicked problem partly because deploying it allocates identifiable costs to identifiable people on a known timetable, while the benefits are diffuse, shared and later. It is a political problem masquerading as a technical one. 

AI does not really change that arithmetic. No model allocates the cost of decarbonisation, and the reductions above are unlocked by exactly the adoption decisions that governments and firms have already spent a decade not making. 

There is one honest exception. If AI drives the cost of the low-carbon option far enough below the alternative, that option needs less political capital to survive contact with an electorate. Cost advantage is one of the only arguments that has reliably beaten political resistance. But that is an economic route, not a technological one, and it moves at the speed of industrial cost curves. 

What this means now

For banks, data centre and power generation lending is becoming a concentration risk that presents as growth. It arrives with financed emissions attached, in a sector whose collateral value depends on grid connections that may never be granted. Fifty-nine gigawatts of requests against five of expectation is a credit question before it is an energy one. 

For mid-market corporates, Scope 2 is about to be repriced by someone else’s compute. You do not set the power price, and your reported emissions will move with a generation mix you did not choose. 

For private equity, grid connection and water permits have become diligence items that bite inside a hold period, and the energy cost trajectory of a portfolio company belongs in the model, not the ESG annex. 

Underneath all of this sits a measurement problem. An organisation spending millions on AI cannot currently compare the carbon, energy or water footprint of the models and services it is buying. As Richard Tarboton of the World Resources Institute puts it, disclosure sits at company or facility level, which tells a buyer very little about the product they are choosing. Procurement is running ahead of measurement, the same governance gap we touched on earlier. 

That gap is also where a buyer has some leverage. At Climate Week in New York this week, WRI, Climate Group and The B Team convened some of the world’s largest corporate buyers of AI to ask whether collective purchasing power could shift AI onto a more sustainable path. Product-level disclosure is the precondition for that, and for any market mechanism that would reward lower-carbon AI. It is a rare intervention that works through procurement instead of politics. 

The point

Both sides of the AI story are real, but they are not the same kind of thing. 

One is a cost that is already sited, financed and scheduled (but not visible to consumers). The other is an option with an uncertain strike price and no delivery date. 

Boards should price the certain half properly, and treat the other half as upside, not mitigation. 

And pricing it properly starts with being able to see it at all. 

 

Sources

Figures in this piece are drawn from the following. Where two editions of the same report are cited, the year of each figure is given in the text. 

  • Carbon Brief, AI: https://www.carbonbrief.org/ai-five-charts-that-put-data-centre-energy-use-and-emissions-into-context 
  • A. Shehabi et al., 2024 United States Data Center Energy Usage Report, Lawrence Berkeley National Laboratory, December 2024. https://eta.lbl.gov/publications/2024-lbnl-data-center-energy-usage-report 
  • P. McCauley and M. Scanlan, Center for Water Policy, University of Wisconsin-Milwaukee, August 2025. Water disclosure practice among major data centre operators. https://theconversation.com/data-centers-consume-massive-amounts-of-water-companies-rarely-tell-the-public-exactly-how-much-262901  
  • NESO, written evidence to the House of Commons Environmental Audit Committee, DCU0081. https://committees.parliament.uk/writtenevidence/166090 
  • ECMWF, ECMWF's AI forecasts become operational, 25 February 2025. https://www.ecmwf.int/en/about/media-centre/news/2025/ecmwfs-ai-forecasts-become-operational 
  • Richard Tarboton, Director, Corporate Transition, World Resources Institute.  
  • Laurie Menoud, Founding Partner, At One Ventures. 

 

A note on two figures

  • The indirect water figure is the standard citation but it is contested. It counts evaporation from hydroelectric reservoirs and does not net off the renewable power purchase agreements the largest operators hold, and some analysts put the realistic figure materially lower.  
  •  The IEA published a new edition of its energy and AI work in April 2026. The demand and emissions figures here follow that edition. The 2035 range and the fuel mix follow the April 2025 edition, which the 2026 report does not restate.