Artificial intelligence is having its “gold rush” moment. Everyone’s using it — from students cranking out essays, to businesses rewriting their marketing in seconds, to creatives generating jaw-dropping art at the click of a button.
You ask ChatGPT a question, and it fires back with a perfectly worded response in seconds.
Need help with writing? Business strategy? Dinner ideas? Done.
It feels almost free.
But it isn’t. There’s a side we’re not talking about enough.
The cost of running this tech: not in dollars but in energy, resources and environmental impact.
I really got to thinking about this recently especially now – at a time when climate concerns are driving glossy sustainability pledges by businesses, the rapid adoption of AI in those same businesses creates an awkward contradiction. When you embrace both sustainability and energy-hungry tech, are you working at cross purposes?
Behind every chat, every image render, every “write me a witty social post” request is a bank of power-hungry servers doing some serious work. And those servers don’t run on magic — they run on electricity. Lots of it. So I did some research…
How much electricity? Exactly?
For a text-based AI like ChatGPT, each query uses roughly 2.9 watt-hours of electricity (The Growing Energy Footprint of Artificial Intelligence, Alex de Vries, 2023, Joule). That’s about the same as running an LED light bulb for nine hours, or boiling a third of a kettle, or watching Netflix in HD for 20 to 30 minutes. (By comparison, a standard Google search reportedly uses 0.3 watt-hours of electricity.)
Sounds small, but multiply that by billions of queries every day, and you’re looking at around 87GWh/day – the daily power consumption of entire small countries like Ireland or New Zealand.
Graphics-based AI like Midjourney or DALL·E are heavier. Estimates suggest a single image request could be 10 to 30 times more energy-intensive than a text query (Quantifying the Carbon Emissions of Machine Learning, Luccioni et al., 2023). That’s because image generation models have more parameters, require more complex calculations, and often involve multiple inference steps to create those “wow” visuals. That’s up to 300 watt-hours per image depending on size, server load, and quality settings. In real terms: boiling a full kettle; running a small fan for several hours; or driving a petrol-powered car 2-3km.
Scale that to millions of daily prompts – images being generated every second for ads, content, art, and if we’re honest, random aesthetic curiosity – and you’ve got an environmental footprint bigger than most people realise.
In May 2025, The Guardian reported that AI may account for up to 49% of global data centre power consumption by the end of 2025.
And then there’s the carbon cost…
And if that 2.9 watt-hour query was powered by coal or gas, all that power means emissions. For ChatGPT, each prompt might emit around 1.1–1.5 grams of CO₂ (depending on the grid mix). Your 2.9 watt-hour query if powered by coal equates to around 78 grams of CO₂ – that’s like charging your phone 10 times. Think about our billions of queries a day, and you’re looking at around 78,000 tonnes of CO2 a day. For image AI, it can be up to 15–45 grams — still tiny per request, but enormous in aggregate.
If powered by renewables, the emissions drop close to zero depending on the set up but manufacturing, cooling and infrastructure still carry embodied environmental costs, even if the electricity is ‘green’.
… And the indirect impacts …
All of the above is before we consider the indirect environmental impact – it’s not just about electricity!
Water: Data centres use millions of litres of water for cooling because servers pump out heat. Some systems run on closed loops, recycling the same water, but many use evaporative cooling, where the water is lost into the atmosphere.
Even with efficiency improvements, absolute water consumption has been climbing as AI workloads grow. One 2023 study estimated that training GPT-3 used about 700,000 litres of clean freshwater (Luccioni et al., 2023), much of it evaporated and never returned to the system. In 2022, Google’s AI servers alone used nearly 5 billion litres – that’s 2,000 Olympic sized pools. In 2023, total water consumption (data centres and offices) reached 24 billion litres (22% of data centre water was from reclaimed wastewater). Of this, 3 billion litres was replenished. (Google Environmental Reports 2022, 2023).
Some tech giants now use reclaimed wastewater or rainwater where possible, but in many regions, AI is literally competing with communities and ecosystems for fresh water. Even when water is recaptured, it can pick up contaminants or chemicals from cooling towers, meaning it needs treatment before reuse — which adds cost and complexity.
Resources: AI hardware relies on rare earth metals and semiconductors, which require energy-intensive and often environmentally destructive mining.
E-waste: The pace of AI development means chips and servers are replaced quickly, contributing to electronic waste — which is often exported to developing countries with poor recycling infrastructure.
This is before we even get into the question of all that generated content clogging up the web…
So whilst AI is talked about as ‘in the cloud’ we should remember, that cloud is tethered to the ground, and it leaves a footprint.
Can AI be part of a sustainability solution?
AI has the potential to do a lot of good — if we use it wisely. AI can optimise energy use in buildings and factories, manage sustainable farming practices, design more efficient products, and reduce waste and emissions in supply chains (Google Sustainability Report, 2023). But there’s a balance to strike between using AI to solve problems and using AI because it’s convenient.
If we’re going to live in an AI-powered world, we can’t ignore the electricity bill — or the carbon bill and other environmental impacts. And we certainly can’t claim we are a business that cares about the planet if we’re also using AI daily for work “efficiencies”.
AI isn’t inherently good or bad. It’s a tool – and tools have impacts. Just because those impacts are hidden in server farms and cooling towers doesn’t mean they’re not real.
What you can do
Next time you spin up 50 AI-generated images “just to see which one you like best” or fire off a dozen reworded versions of the same email, maybe pause. Not because we shouldn’t use AI, but because we should use it intentionally. Take a moment and ask yourself: Is this worth a third of a kettle boil? If not, maybe step outside, breathe in some actual air, and let the planet have one less watt-hour to worry about.
Just like we became more conscious of how often we fly, or how we use plastic, we can be more thoughtful about how we use AI.
✅ Ask: Is this query useful or just curiosity-fuelled noise?
✅ Support AI companies that invest in renewable energy and transparent sustainability reporting
✅ Push for greener digital infrastructure, not just faster and bigger tools
✅ Choose quality over quantity in how we interact with tech
And when you’re using AI to improve real outcomes — for your business, your systems, or your sanity — that’s a smart use of energy.
AI isn’t free — we’re just not paying the full price yet. The planet is.
