Table of Contents
Last year, social media was awash in Ghibli images. Users animated their photos in a manner that was debatably similar to Miyazaki’s profound and whimsical works.
And then this year, along came the denim jackets. Every user typing the same prompt had the same jacket, bushy hair and flared pants, and reimagined themselves through the decade’s latest party trick: The 1980s AI photo prompt. Giggles, giggles, and more giggles.
The other side of the internet, harboured several concerns. “Well, there goes another glacier,” read one sardonic reply, part of a wave of posts mourning the climate crisis and the ‘inevitable end of the world.’
Others pushed back, unconvinced. “How exactly will my ChatGPT prompt cause an environmental crisis? Is an angry GPU going to come to my home,” one user shot back. And then, cutting through the noise, the question everyone actually wanted answered: “Will someone explain to me like I am 5, how does using ChatGPT cause a climate crisis?”
It’s a fair ask. So, how does a simple AI prompt become a climate concern?
It starts with electricity
According to Mike Weinstein, director of sustainability at Southern New Hampshire University, who conducted doctoral research on the intersection between technology and environmentalism, electricity demand is at the heart of AI’s environmental footprint, and the energy used by data centres can have implications for both carbon emissions and water consumption.
So, where does the power come from? Citing MIT Technology Review, Weinstein noted on the university’s website, that AI data centers are currently located in areas with grids predominantly powered by fossil fuels, meaning coal- and gas-fired power plants, fuel sources that contribute to global climate change by releasing greenhouse gases into the atmosphere when they are burned. Those emissions warm the planet and settle over the communities living beside the data centres, showing up as asthma and heart disease.
In other words, the climate impact depends not simply on how much electricity AI uses, but also on how that electricity is generated.
“If all of our electricity supply was clean and renewable, like that provided by solar or wind power, we could look at the other impacts of AI in a different context,” Weinstein said. “But unfortunately, we as a planet are tremendously behind in transitioning to a clean, non-carbon energy grid.”
And then, there’s water. Data centres also need cooling systems to keep servers from overheating, and some of those systems can require substantial amounts of water. In regions including the Great Lakes, data centres draw huge volumes of it to absorb the heat their servers generate. “What this often means is that a large volume of freshwater from the surrounding environment is diverted to the data center and away from other uses, like drinking water,” Weinstein explains.
It’s never just one prompt
Here’s the trap in the online debate: A single ChatGPT prompt genuinely is close to nothing. The electricity needed to produce one response isn’t, on its own, a climate event.
But AI doesn’t run one prompt at a time, it runs behind the scenes, constantly, through data centres, chips, cooling systems and storage, all working to answer queries within seconds. This means the environmental footprint is not just about the energy used for one interaction, but the infrastructure required to deliver billions of interactions at scale. As usage stretches from text into images, video, reasoning and autonomous agents, that infrastructure has to grow to match.
That’s where the real footprint starts to show, says Dr Balamurugan Balusamy, Dean of the School of Engineering and IT at Manipal Academy of Higher Education (MAHE) Dubai. “AI already has a measurable environmental footprint, particularly because of the computing power required to train and run large language models. The main contributors are electricity consumption, data-centre infrastructure, and the energy and water needed to cool those systems,” he said.
The scale becomes clearer in the numbers. He points to the International Energy Agency’s estimate that data centres consumed around 415 terawatt-hours of electricity in 2024, or about 1.5 per cent of global electricity consumption, and notes that a single typical text-based query using GPT-4o has been estimated at around 0.3 watt-hours.
Individually, that is a small amount of electricity. Multiplied across billions of interactions, the cumulative demand becomes more significant. “When billions of queries are combined with model training, AI agents, image and video generation, and increasingly sophisticated reasoning tasks, the aggregate demand becomes substantial,” he says.
Part of the problem, Balusamy argues, is that the cost is invisible by design. “A user enters a prompt on a laptop or phone and sees an answer within seconds, but does not see the data centres, computing infrastructure and cooling systems operating behind that interaction.” The question is what happens when billions of people make increasingly demanding requests, day after day, and the infrastructure expands to keep pace.
So no, that 1980s photo prompt isn’t personally sending a glacier into an early grave. The real question is what happens when billions of people make increasingly demanding requests, day after day, and infrastructure keeps expanding to keep pace.
Not every request weighs the same, either: “A short text exchange is very different from generating a high-resolution video, training a large model or asking an AI agent to carry out several tasks,” Balusamy notes. “The environmental impact comes from the wider infrastructure required to provide these services at scale. Greater transparency around resource consumption would help users and organisations make more informed decisions about how AI is used.”
A problem that could get bigger or smaller
However, whether this becomes a defining environmental issue or a manageable one, Balusamy says, depends on a race that hasn’t been decided yet. “It can become a much bigger issue if growth in AI continues to outpace improvements in efficiency,” he warns.
Research into greener computing is underway, but AI development itself is advancing at a much faster rate, as he notes, and sustainable infrastructure costs money that companies may be reluctant to absorb, or eager to pass on to customers. “The key question is therefore whether efficiency improvements can become part of mainstream AI development quickly enough, rather than remaining a separate sustainability exercise.”
The irony is that AI could just as easily be part of the fix. Balusamy points to Google’s Flood Hub, which currently provides river-flood forecasts covering more than 2 billion people in over 150 countries, with warnings available up to seven days in advance, and newer urban flash-flood models capable of up to 24 hours of advance warning — potentially the difference between evacuating a community in time and not.
On the air quality side, UNEP’s GEMS Air Pollution Monitoring platform brings together data from more than 25,000 air-quality monitoring stations across more than 140 countries, using AI to support real-time health protection. Zoom out further, and the IEA estimates that widespread adoption of existing AI applications could potentially reduce around 1.4 billion tonnes of CO₂ emissions per year by 2035.
Even so, Balusamy is careful not to oversell it. AI is not a crystal ball, he explains. The environmental systems are complex, and AI predictions are only as good as the data, models and physical understanding behind them. The greatest value is therefore not replacing scientists or decision-makers, but giving them faster, richer and more actionable information.
The infrastructure fix
If the problem lives in the data centre, so does a lot of the solution, and that’s the argument coming from the infrastructure side of the industry.
Rhys Oxenham, Vice President and General Manager, AIat SUSE, leader of enterprise open source software, breaks down why AI workloads are so hungry in the first place: They ‘solve problems’ by crunching billions of interconnected calculations simultaneously to recognise patterns and generate the most statistically logical responses.
Delivering these capabilities at the unprecedented levels of demand that we’ve seen across the industry requires thousands of high-performance chips spanning global data centres running 24/7. “It creates a massive energy appetite, both from the electricity required to fuel them, and the intense cooling needed to keep them from overheating.”
This puts the focus not only on how much computing AI requires, but on how efficiently the underlying hardware and software can deliver it.
However, he remains optimistic that engineering can bend the curve. Across the industry, he says, they are tackling the challenge by pairing renewable clean energy with smarter software and hardware. “By using intelligent scheduling to match the right AI model to the right task, maximising shared compute resources, and advancing silicon design, we can eliminate immense energy waste.”
The condition, he adds, is that efficiency has to be taken as seriously as raw performance: Breakthroughs in hardware design help, but smart software management matters just as much; building efficiency directly into how systems operate and leveraging hybrid environments to run AI where it makes the most energy sense, rather than just where it’s convenient. “When we manage computing power intelligently, energy consumption doesn’t have to grow in lockstep with AI adoption,” he says.
The data no one is cleaning out
There’s a less obvious contributor to AI’s footprint, and it’s been sitting on company servers for years: Data nobody’s using.
Matthias Nijs, Vice President, Sales EMEA at Datadobi, a dynamic unstructured data management company, points out that every terabyte stored has an energy footprint, and most organisations are storing far more than they can. This is a different part of the environmental equation: not the electricity used to answer a prompt, but the ongoing energy and infrastructure required to store data.
The problem compounds as AI enters the picture, because so much of what gets fed into models is unstructured, video files, images, PDFs, emails, presentations, scanned documents, and the duplicate copies of all of these that accumulate over time, Nijs explains.
It’s rarely catalogued, so nobody has a clear picture of how much exists or whether it’s still needed, and every additional copy multiplies the energy cost as companies expand the infrastructure needed to store, protect and move it.
Most organisations, he says, don’t realise the scale of what they’re sitting on. “Storage is often treated as a fixed cost of doing business rather than a growing energy liability,” he says, a blind spot that matters because the country’s data centre IT load is expanding rapidly from a current base of around 500 megawatts, largely to meet AI demand.
Without visibility into what’s actually being stored, companies end up scaling storage, and its energy demands, well beyond what their data actually requires.
Perspective, not panic
Alfred Manasseh, Co-Founder and COO of Shaffra, doesn’t dismiss the concern nor lets it spiral into doom.
It is a real issue and one that the industry needs to take seriously, as he says. “AI requires significant computing power, and computing power requires electricity. As AI adoption grows, that demand will naturally grow with it.”
However, the size of that demand needs context. “Data centres today represent a relatively small percentage of global electricity consumption compared to industries such as transport, manufacturing or buildings. What is important is the speed at which the demand is growing.”
For Manasseh, the question is therefore not whether AI comes with an environmental cost. Of course it does, he asserts. The more important questions are how efficiently we can build and deploy AI, where the energy comes from, and whether the value created by that AI justifies the resources being consumed.
At scale, he says, the numbers become harder to ignore. Small amounts of computing become very large amounts of computing, while every interaction also generates heat that has to be cooled somewhere, bringing implications for carbon and water use.
Yet, echoing Balusamy’s earlier point, Manasseh stresses that AI workloads can vary dramatically. “Asking a model a simple question is very different from generating a video, training a large model or having an AI agent perform multiple complex tasks.”
That is why, he argues, efficiency needs to be addressed “at the system level, not only at the individual prompt level.”
He also separates two stages of AI’s energy consumption that are often discussed together: training and inference. Training, in which models learn from enormous amounts of data, attracts much of the attention because it can be extremely energy intensive. But inference, is what happens when millions of people and businesses actually use those models, could be just as significant because it happens continuously and at enormous scale.
The distinction becomes even more important as AI moves beyond simple question-and-answer interactions. Agents can perform several actions, interact with different systems, retrieve information and reason through multiple steps to complete one task, creating considerably more computing activity than a straightforward exchange.
Manasseh sees evidence for the concern, but argues that the conversation often loses important distinctions. “There is definitely real data behind the concern. Data centre electricity consumption is growing quickly and AI is becoming an important driver of that growth.”
At the same time, he cautions against treating every AI interaction as though it has the same environmental cost. The idea that one AI question consumes X amount of energy cannot be applied uniformly, he says, as AI workloads are simply not that uniform. Different models, hardware, data centres and tasks can have completely different requirements.
What the industry needs, he argues, is greater transparency. “We need better reporting from the industry so that businesses and governments can make decisions based on actual consumption rather than assumptions.”
There is also a larger point to consider. AI has not created the data-centre economy from scratch. “The internet, cloud computing, streaming, social media, cryptocurrency and virtually every digital service we use already depends on data centres and electricity. AI is not creating this infrastructure model from zero. It is significantly increasing the demand placed on it,” explains Manasseh.
So which way does it break, worse, or better? “Probably both,” he says. Usage will keep climbing regardless: “There is no realistic scenario where billions of people and businesses use more AI and computing demand simply disappears.”
But the technology underneath it is also improving, “Chips are improving, models are becoming smaller and more specialised, and companies are getting better at using the right model for the right task instead of always using the largest possible model.”
The goal, in his view, isn’t using less AI for its own sake. “It should be to create much more value from every unit of computing and energy we consume.”
He leaves the conversation on a line that could stand as the whole debate’s epitaph: “AI has an environmental footprint, like every major technology before it. Our responsibility is to make sure the intelligence and value we create grows faster than that footprint.”
