Think Piece: AI and the environment – a cup of coffee or a catastrophe?

Could one less cup of coffee a week cover your AI use?

There seems to be a lot of confusion about the environmental impact of AI and its associated energy use. Two people can read the same statistics and reach opposite conclusions.  Some voices are reassuring, others predict doom. Now, the interesting fact. Both sides are using the same numbers. Can both be right? 

This is especially important for arts, cultural and heritage organisations, many of which hold public trust, receive public funding, work with communities and make public commitments around sustainability and social justice. 

One widely quoted source is Google’s 2025 paper that matches the energy consumption of one average Gemini text prompt to watching TV for less than nine seconds. Hannah Ritchie, a respected environmental analyst at the University of Oxford talking about ordinary individual use said, ‘asking AI questions every day is still a rounding error on your total electricity footprint.’  

My calculation taking the data average 0.3Wh per prompt (and trying to include hidden costs of training) and assuming 100 prompts per week – comes out just less than the energy to boil water for one cup of coffee per week. Could one less cup of coffee a week cover your AI use? 

So, nothing to worry about? Perhaps not. 

Or does AI herald potentially catastrophic energy costs?

On the other side of the debate leading researchers such as Dr Sasha Luccioni advocates a change in how we use AI. Luccioni gave the Turing Lecture on Making AI (truly) sustainable in May 2026, her view is that AI energy use is far higher than reported and there are real environmental costs.  

A recent report by the International Energy Agency estimates global electricity consumption for data centres will double by 2030 accounting for just under 3% of global electricity consumption. AI is a major driver in data centre growth. Data centres in Ireland use more than 20% of the country’s metered electricity (CSO) and in Virginia the estimate for 2030 is 57% (EPRI). 

Infographic from Ireland's Central Statistics Office titled "Data Centres Metered Electricity Consumption 2024." Three panels compare the percentage of metered electricity consumed by data centres, urban dwellings, and rural dwellings. Data centres grew from 5% in 2015 to 21% in 2023 and 22% in 2024. Urban dwellings fell from 22% in 2015 to 18% in 2024. Rural dwellings fell from 12% in 2015 to 10% in 2024.

Source: https://www.cso.ie/en/releasesandpublications/ep/p-dcmec/datacentresmeteredelectricityconsumption2024/ 

Noman Bashir at MIT argues that we are on an unsustainable path and the demand for new data centres means the bulk of electricity to power them must come from fossil fuels.  

Same data, different narratives

Many authors are choosing the research that frames their narrative.  

Advocacy organisations tend to select the evidence that makes risk more visible. Industry voices tend to put forward the evidence that makes individual use look small. Both can be using real evidence while still directing attention to different scales. For example, Greenpeace’s Environment and AI report foregrounds campaigns such as South Dublin County Council’s call for restrictions on new data centres and the QuitGPT boycott. Both sides argue in good faith from real evidence. So, if they’re not disagreeing about the numbers they must be disagreeing about something else? 

The contradictions change once you see each side is approaching from a different scale. 

That is something I keep running into in my own work. I’ve spent four years researching generative AI and I originally trained as a geographer and environmental scientist. Geography gives us a useful way into this debate, because it teaches us to ask not only what the data says, but where we are standing when we interpret it.

Three different scales, three different pictures

Let’s look at three scales: space, time and social; each changing the picture. 

Diagram illustrating three scales for understanding AI energy use. Spatial scale: a horizontal flow from Individual to Organisation to Region to Global, shown with icons of a person, a building, an electricity pylon, and a globe. Temporal scale: a timeline progressing from Now to Build-out to Lock-in, with a rising graph indicating increasing energy demand over time. Social scale: a horizontal flow from Individual to Community to Society to Future generations, shown with progressively larger groups of people, ending with a green leaf icon.

Spatial scale - changing the boundary changes the problem

Spatial scale, i.e. local, regional, national and global. This is the easiest one to see. Picture one – per person. We’ve got a 0.3Wh per query figure. Picture two – global aggregate. IEA says 1.5% of global electricity is used by data centres. Picture three – regional. Ireland has projected that data centre electricity use will be 32% of the entire country’s use in 2026 (Central Statistics Office).   

A diagram showing that the same data + different areal units = different results

There’s a theory called the Modifiable Areal Unit Problem (MAUP) coined by geographers in the 1970s. This is one of the stickiest problems in spatial analysis. You can’t just take data at one scale and then apply it at another.  The conclusions you draw depend entirely on the boundaries of your unit of analysis.  This means the same data gives different stories at different scales.  You can’t infer individual-scale properties from aggregate data or vice versa. AI using 2% of global electricity tells you nothing about individual use. A query using 0.3Wh cannot be extrapolated into a global picture. 

The mistake is not using the wrong numbers, but asking a number measured at one scale to answer a question at another.  One of the reasons why this debate is messy is that companies are not disclosing the data. Everyone is reaching for proxies. 

I’m a big fan of the late Doreen Massey. She wrote about how privileging one scale becomes a political choice. Framing AI use per capita makes it about personal choice, framing it about South Dublin makes it an environmental justice one. The choice of scale shapes the policy response.  

The scale you foreground shapes the story and the response

Diagram titled "Privileging one scale is a political choice" with the subtitle "The scale you foreground shapes the story and the response." The central topic, AI and energy use, branches into three columns. Individual scale: framing is personal use and responsibility; political effect is guilt, restraint, and efficiency tips. Regional scale: framing is infrastructure, planning, and public consent; political effect is regulation, permits, and local challenge. Global scale: framing is competitiveness, growth, and net carbon claims; political effect is expansion, innovation, and global targets. A footer box states: "Choosing the scale is not neutral. It shapes what counts as the problem, who seems responsible and what action feels justified."

So, no one is telling fibs, they are just standing in different places.  

Temporal scale

Next, the temporal scale. Space isn’t the only scale that flips the picture. Time does too. Averaging energy use across a day or year hides the narrative that matters. Marginal data is far more useful. At 7pm, the marginal electricity needed to serve demand may come from gas rather than solar, depending on the grid. So the same query can have a different carbon impact depending on when and where it is served. The averages hide this data. There’s a lot of confusion, most companies say ‘powered by renewable energy’ but contractually this means ‘matched with renewable energy’ not physically supplied by renewable electricity. 

So a company can say ‘we are powered by 100% renewable energy’, that often means: we bought enough renewable electricity contracts to match our non-renewable consumption. ‘Match’ is the key word here – it does not mean every data centre is physically supplied by carbon free electricity every hour. 

In some regions, data centre growth risks locking in new fossil-fuel infrastructure for decades. Demand is literally outpacing decarbonisation. And this is only the beginning. All the stats so far are based on text prompts. And as AI gets cheaper and more capable, it gets used more, not less, with use shifting to far heavier work: agentic systems that plan and run tasks continuously, reasoning models and video generation. 

Line chart titled "Projected AI growth and the decarbonisation timeline" with the subtitle "When demand grows faster than clean energy transition, fossil-fuel lock-in becomes more likely." The x-axis shows time from Now through Build-out, 2030, 2040, and 2050. The y-axis shows energy demand and supply capacity. A steeply rising blue line represents AI electricity demand. A slowly rising green line represents clean energy supply. From the Build-out point onward, the gap between the two lines widens and is shaded orange, labelled "Gap met by fossil fuel infrastructure." An annotation marks where "Demand outpaces decarbonisation" and another notes "Transition catches up slowly" near 2050. A callout box highlights "Fossil-fuel infrastructure lock-in." A footer states: "If AI demand accelerates faster than clean energy deployment, new fossil infrastructure can be built now and remain locked in for decades."

So, the catastrophe camp is seemingly right about the trajectory in the future and the one cup of coffee camp are right today. Different points in time. 

Social scale

The final scale is the social one: individual > community > society. The costs of AI are affecting communities who didn’t choose to host this massive infrastructure. People are fighting back. The UK Government overruled Buckinghamshire Council’s decision to reject planning permission for a new data centre. However, an action group challenged this - with the argument that the decision ignored the electricity demands and the climate impact. The Government withdrew, accepting that it had made an error (more info in this article). 

xAI’s Colossus 2 data centre needed its own fossil fuel power plant which went up, unpermitted in a predominantly Black community already failing in air quality standards. NAACP has sued xAI, this article on the xAI gas power plant is worth a read.  

If we then apply the temporal scale to the social one - we then have intergenerational lock-in. The people it will affect are not in the room when today’s decisions are made.  

For leaders of arts and cultural organisations, this is where the question changes. The issue is not whether one staff member should feel guilty about using an AI tool to summarise notes, draft ideas or test a line of copy.

The more important question is perhaps whether AI is becoming embedded into your organisational practice without enough thought about purpose and public trust. 

The real issue

Many cultural organisations already make commitments around environmental responsibility and care for communities. AI use needs to sit inside that same frame. It is not only a productivity or creative opportunity, it is also a governance issue. 

Maybe the real issue is that we have been arguing at the wrong scale. We are at what Fengqi You calls the ‘build-out moment’ - the choices made now lock in what happens for decades and decarbonisation is not keeping pace with AI demand. It is not happening in the places where data centres are expanding fastest; these regions are adding fossil fuel generation, not shedding it. Gates and Hassabis counter that AI will save more carbon than it costs.  Even if that turns out to be true, it is a global-net promise, it says nothing about who pays, where and when.  

 So where could the answer lie? Perhaps not at the individual scale: counting prompts, switching models and private guilt. Not at the global scale: broad averages and net promises. The real decisions sit somewhere in the middle, where spatial, temporal and social scales meet - the infrastructure scale: planning permissions, grid connections, water use, fossil-fuel backup generation, procurement and public consent. 

Diagram titled "Where the real decisions sit" with the subtitle "Not at the individual scale. Not at the global scale. At the infrastructure scale, where spatial, temporal and social scales meet." Three circles at the top labelled Spatial, Temporal, and Social feed down with arrows into a central highlighted box labelled Infrastructure scale, which contains six elements: planning permissions, grid connections, water use, fossil-fuel backup, procurement, and public consent. To the left, a box labelled Individual scale lists three items: counting prompts, switching models, and private guilt -  connected to the centre by a dashed arrow pointing right. To the right, a box labelled Global scale lists two items: broad averages and net promises -  connected to the centre by a dashed arrow pointing left. The layout emphasises that infrastructure-level decisions are where meaningful action on AI energy use occurs.

The focus, then, is at the regional and social scales, where councils, action groups, sector support organisations, trade associations and courts are beginning to challenge how AI infrastructure is being planned and approved. 

So what can we do?  

The cultural sector carries public credibility. It brings together communities, focuses attention and takes on the difficult conversations around ethics and limited resources.  A co-ordinated sector position on AI and environmental responsibility could be an approach.  

So perhaps the better questions are at the organisational and sector level. We ask where AI is genuinely useful and we act together. Where does it support our missions and where could it undermine our values. How do we build environmental impact into AI governance alongside bias, copyright, accessibility, jobs and trust.  

Let’s shift the scale of thinking.  

Paul Blundell, Head of Digital Research and Development, Arts Marketing Association

Join Paul on our popular Essentials of AI for Marketing course.