You may have heard about AI data centres guzzling insane amounts of energy, and that using ChatGPT to write an email is equivalent to emptying out a bottle of water.
But is that true? And how worried should we be about GenAI’s environmental impact?
It’s difficult to measure the exact scale of the problem, as AI companies have been shy to share information on their resource use.
But let’s look at what we do know.
GenAI models are too large, complex, and hungry for energy to run on ordinary computers, so they need specialised data centres. This year, data centres are expected to be the fifth-largest electricity consumer in the world if ranked as a country, right between Japan and Russia.
Discussion of GenAI energy consumption long focused on the electricity it took to “train” and develop models. But today, according to research by MIT, our prompts are the problem: 80–90% of computing power for AI is used for users’ day-to-day interaction with the models.
How much energy one prompt uses varies widely depending on its length, the kind of output, and the size of the model. A recent report by the United Nations University Institute for Water, Environment and Health (UNU-INWEH) found that generating an image, for example, can use 1,500 times more energy than a simple text prompt. Generating a short video can use 200,000 times more.
Where you live also affects the carbon cost: if the data centre processing your request is hooked up to a fossil-heavy grid, that means higher emissions. Because renewable energy availability peaks during the day, a query processed at noon could have a lower carbon cost than it would if processed at midnight.
The (very important) part we don’t know
Last year, Google disclosed that a median text-only prompt processed by Gemini (right in between the lowest and highest-usage prompts) used as much electricity as it takes to run a microwave for one second, and about five drops of water, while emitting 0.03g of CO₂. Google did not respond to our question of whether these figures are still current.
What does this tell us?
Unfortunately, very little. Without knowing the full scale of demand for a model, it’s impossible to assess its real impact. While the cost of a single query could be low, and even falling, increasing demand means AI models' absolute emissions are likely surging. But that part is still Big Tech's well-kept secret.
The UNU-INWEH, however, estimated that ChatGPT processes around 2.5 billion prompts per day. In a year, this adds up to roughly 400GWh in energy use (roughly the same as a small European city), water equivalent to the domestic needs of 500,000 people in Sub-Saharan-Africa, and 800 football fields worth of land.
According to research by the Vrije Universiteit van Amsterdam (VU), AI systems emit 30-80 million tonnes of CO₂ per year, roughly as much as a global city like New York.
“The precious few numbers that we have may shed a tiny sliver of light on where we stand right now, but all bets are off in the coming years,” said AI researcher Sasha Luccioni.
New tech, good old fossil fuels
Let’s look deeper into that sliver.
The problem with GenAI's increasing electricity demand is that renewable energy sources simply can't keep up.
Instead, countries like Ireland are building new coal or gas power plants and re-activating or extending those set to retire. Tech companies are abandoning their climate goals and constructing fossil plants in the US built exclusively to power their data centres. Resources previously earmarked for the energy transition are being diverted towards the data centre boom.
Over the next decade, Bloomberg New Energy Finance estimates that data centre demand will increase total global power sector emissions by 10%.
But even if the AI “bubble” bursts tomorrow, we'll be locked into new fossil fuel infrastructure until 2050 and beyond. “Even if you were to pause everything – no more new models, no more deployment, not one single new person using generative AI – it just gets frozen in time,” energy analyst Ketan Joshi told us. “The energy demand still goes up, because you have to continue retraining the models based on new information being published on the internet.”
The footprint beneath the footprint
Yet carbon emissions are far from being the whole story.
Data centres use water to cool down computing equipment – 4.5 trillion litres in 2025, enough to fill 1.8 million Olympic-sized pools, according to the UNU-INWEH. It's projected to more than double by 2030. The VU estimates that AI systems alone could consume as much water as all bottled water drunk worldwide in a year.
How dangerous these big numbers actually are is highly dependent on how much water is available in a given area. “The real concern is where and when water is used, and how local water systems are affected,” Kaveh Madani, UNU-INWEH's director who led the investigation team of their recent report, told us.
“Data centres can place pressure on watersheds that are already drought-prone, stressed, or facing competing needs” he added.
Their research found that some “green” solutions can also make resource pressures worse: while generating electricity from bioenergy instead of coal reduces carbon emissions by 70%, it uses 30 times more water and 100 times more land.
“The burden is most likely to fall on communities near data centres, energy infrastructure, water-stressed regions, and mining or manufacturing sites linked to AI hardware, while the benefits of AI are often captured by companies, consumers, and institutions elsewhere,” said Miriam Aczel, lead author of the UNU-INWEH report. “This is why AI’s environmental footprint is also an environmental justice issue.”
Coming to a grid near you
Europeans are not spared from these effects.
The EU intends to triple its data centre capacity by 2030. Today, many AI-focused data centres are clustered around major urban hubs like Frankfurt, London, Amsterdam, Paris, and Dublin, but demand is set to accelerate in Northern and Southern Europe.
In Ireland, data centres used 22% of the country’s total energy in 2024 – more than all urban households combined in that same year. An incredible 80% of Dublin's electricity use can be attributed to data centres, according to a Greenpeace estimate based on data from McKinsey, the International Energy Agency, and other sources.
As infrastructure costs get passed on to consumers, data centre expansion has raised Irish households’ annual energy bills by €100. It has also led to higher oil and gas consumption, delayed electrification plans, and reallocated grid capacity away from housing projects.
Emerging research suggests data centres raise temperatures in their immediate surroundings by between 2-9°C.
“Similar pressures could appear elsewhere if data centre growth is concentrated faster than grids, planning systems, and clean electricity supply can keep up,” Madani and Aczel warn. New demand should be matched with genuinely additional clean power, they added.
While the EU's AI Act requires companies to report on model energy use and efficiency, only 36% of relevant data centres have done so, according to a European Commission document seen by POLITICO.
In April, Investigate Europe found that Big Tech successfully lobbied the EU to classify individual data centre information as confidential and commercially sensitive. In the end, the suggestion made by Microsoft and DigitalEurope (a group including Amazon, Google, and Meta) was added to the AI Act almost word for word, and the Commission encouraged EU member states to refuse public requests for this information.
Can't Claude save the climate?
Listening to most tech companies and some researchers, you could be convinced that GenAI's environmental impacts are no big deal – down the line, AI will help us solve the climate crisis, and its net impact will have been positive! Right?
A 2025 LSE study found AI could reduce global emissions annually by 3-5 billion tonnes of CO₂-equivalent by 2035 – roughly all of the EU’s emissions in 2023. Google has claimed AI could mitigate 5-10% of global emissions by 2030.
Sounds impressive, until you realise what they're actually talking about. Ketan Joshi investigated claims of climate benefits through AI, and found that 97% are based on potential applications of “traditional” AI, not GenAI tools like chatbots.
The difference is crucial: “Both of these things get called AI, but they operate in very different ways,” Joshi told us. Traditional AI uses machine learning technology for niche tasks such as data analysis for weather forecasting, without generating anything new. GenAI completes much broader tasks and has between 6-14 times higher energy use than traditional AI models.
Yet the potential of the old is being used to justify the new – which is the very one pushing up our energy demand and driving the data centre boom in the first place.
As AI journalist Karen Hao put it: “The term ‘AI’ is so vague that it’s like the term ‘transportation’. You could be talking about a bicycle or a rocket. But that doesn’t mean that the benefit from using the bicycle also justifies us building the rocket.”
What's more, Joshi found that 74% of climate benefit claims cite no published academic research, while 36% cite no evidence at all. Google’s 5-10% figure traces back to a blog post by consulting firm BCG providing no evidence. Google did not respond to our request to any further research behind that figure.
“Claims that deployment of AI, regardless of type, can bring about gigatonne-scale reductions in global emissions are, at best, deserving of more scrutiny and at worst catastrophically overstated,” according to Joshi.
So are we cooked?
The UNU-INWEH researchers told us that “public debate can overstate the issue when it treats every AI query as catastrophic, but it also understates the issue when it focuses only on model training and ignores everyday deployment, data centres, cooling, grids, and hardware supply chains.”
Huge numbers without context can scare people, they argue. But comparing AI's environmental footprint against other benchmarks to imply that its impact is negligible is equally misleading.
Instead, the UNU-INWEH team argues that AI within planetary limits is achievable: “The risk is not that AI alone overwhelms planetary systems today, but that rapid deployment adds new electricity, water, land, hardware, and e-waste pressures at the same time that societies are trying to decarbonise,” Madani and Aczel said.
To move the needle, we'll have to collectively get our act together. The researchers suggest concrete measures including “mandatory environmental disclosure for major AI systems and data centres, energy and water reporting by location, and stronger permitting rules for data centres in water- or grid-stressed areas,” alongside “requirements for additional clean electricity rather than paper-only claims.”
Thankfully we asked them – not Chat – how cooked we are.