Almost every time I’m on stage talking about AI, the same hand goes up: “What about the environmental impact?” And it’s a fair question. It deserves a proper answer, not a shrug, and not a lecture either.
So let me be clear up front: AI’s environmental footprint is real, it’s serious, and it’s going to grow. Data centres are being built at extraordinary pace, and the energy and water they consume is a genuine, legitimate concern. Nothing in this article dismisses that.
But here’s the thing I think gets lost: when we worry about the impact of asking an AI a question, we rarely put it next to the impact of the everyday choices we make without a second thought. The shower we had this morning. The kettle we’ve boiled four times today. The burger we had for lunch. When you do put them side by side, the picture changes quite dramatically, and it turns out the biggest levers aren’t in your chat window. They’re on your plate.
Before the numbers: these are illustrative equivalents based on specific published benchmarks. Google has published a figure for a median Gemini text prompt; Mistral has published a fuller lifecycle figure for a 400-token response. They use different accounting boundaries, so I’ve shown both throughout, treat them as the optimistic and conservative ends of the range, not as directly comparable. Food water footprints also include rainfall as well as extracted water. Approximate equivalents, not precise accounting.
Water: your shower vs your prompts
Two published benchmarks. Google reports a median Gemini text prompt uses 0.26 mL of water, about five drops. Mistral’s fuller lifecycle figure is 45 mL per 400-token response, roughly three tablespoons. Here’s how everyday activities stack up against both:
| Everyday activity | Water footprint | Gemini prompts | Mistral responses |
|---|---|---|---|
| One minute in a standard shower | 10–15 litres | 38,000–58,000 | 220–330 |
| One eight-minute shower | 80–120 litres | 308,000–462,000 | 1,800–2,700 |
| One cup of coffee | 140 litres | 538,000 | 3,100 |
| One 150 g beef burger | 2,310 litres | 8.9 million | 51,000 |
| One person’s food for one day | 2,000–5,000 litres | 7.7–19.2 million | 44,000–111,000 |
To put it another way: 1,000 Gemini prompts use about one glass of water. A thousand Mistral responses use around 45 litres, roughly three to five minutes in a standard shower. And a single minute in that shower is the water equivalent of somewhere between 38,000 and 58,000 Gemini prompts.
Electricity: your telly vs your prompts
Three useful benchmarks here: a median Gemini text prompt at 0.24 Wh, an average AI-generated image at 2.9 Wh, and a complex long-reasoning prompt at roughly 33 Wh or more.
| Everyday activity | Electricity | Text prompts | AI images | Reasoning prompts |
|---|---|---|---|---|
| One hour of a 10 W LED light | 10 Wh | 42 | 3–4 | Less than one |
| One full smartphone charge | 22 Wh | 92 | 7–8 | About two-thirds of one |
| One hour of a 100 W television | 100 Wh | 417 | 34 | 3 |
| Boiling a 3 kW kettle for three minutes | 150 Wh | 625 | 52 | 4–5 |
| Average UK home, one day | 6.85 kWh | 28,500 | 2,360 | 208 |
| Average UK home, one year | 2,500 kWh | 10.4 million | 862,000 | 75,800 |
Worth being honest about the direction of travel, though. An AI-generated image uses around 12 times the electricity of a median text prompt, and one complex reasoning request can use as much as roughly 138 ordinary prompts, three of those and you’ve matched an hour of television. As we ask AI to do heavier lifting, the per-request footprint climbs. That trend matters far more than any single prompt. (The kettle comparison assumes a 3 kW kettle running continuously for three minutes; real kettles and boiling times vary.)
Carbon: your commute vs your prompts
Same two benchmarks: Gemini at 0.03 g CO₂e per median prompt, Mistral at 1.14 g CO₂e per 400-token response.
| Everyday activity | Carbon footprint | Gemini prompts | Mistral responses |
|---|---|---|---|
| Driving a conventional car 1 km | 150–200 g CO₂e | 5,000–6,700 | 130–175 |
| Driving 10 km | 1.5–2 kg CO₂e | 50,000–67,000 | 1,300–1,750 |
| One 150 g beef burger | ~9 kg CO₂e | 300,000 | 7,900 |
| Driving 100 km | 15–20 kg CO₂e | 500,000–667,000 | 13,000–17,500 |
One note on mixing these up: a burger equals roughly 51,000 Mistral responses by water and roughly 7,900 by carbon. Both are correct, but they answer different questions, so resist the temptation to collapse it all into a single “a burger equals X prompts” headline. I’ve deliberately kept the dimensions separate throughout.
What about training the models?
Training is a different beast from asking a question, and it’s where some of the scarier numbers come from. The original GPT-3 training estimate was around 1,287 MWh of electricity and roughly 5.4 million litres of operational water, including the water used in generating that electricity.
- That’s about 515 years of electricity for an average UK home.
- Around 2.2 Olympic swimming pools of water.
- Or the water footprint of roughly 2,350 beef burgers.
Big numbers. But the important distinction is that training is a large, occasional event, while using AI is a comparatively tiny impact multiplied across potentially billions of requests. One is a fixed cost; the other is where the growth curve lives. Which brings me to the actual point.
The real issue isn’t your prompt
If you take one thing from this article, make it this: the environmental question that matters is not “should I feel guilty about asking ChatGPT something?” A single ordinary prompt is, on every published benchmark, a rounding error next to your morning routine.
The real issues are structural: the sheer scale of adoption as billions of requests stack up while AI becomes embedded in how businesses market and operate, the shift towards heavier reasoning and image workloads, and the pace of data-centre expansion, along with where that power and water comes from. Those are the conversations worth having, with energy companies, with policymakers, and with the AI providers we choose to build on. Transparency like Google’s and Mistral’s published figures is exactly what we should be demanding more of.
And if you’re personally worried about your own footprint, the honest answer is that the biggest levers are the unfashionable ones. Skipping one beef burger saves more water than a lifetime of chatting to Gemini. Eating less meat is a choice available to almost everyone, every day, and it moves the needle by orders of magnitude more than abstaining from AI ever could.
So by all means ask me about AI’s environmental impact at the next event, I genuinely welcome it. Just don’t ask me over a burger.
Sources and caveats: AI figures are from Google’s published Gemini efficiency data (median text prompt: 0.24 Wh, 0.26 mL, 0.03 g CO₂e) and Mistral’s lifecycle analysis (400-token response: 45 mL, 1.14 g CO₂e); image and reasoning estimates from independent benchmarking. Food and household figures are widely used averages (Water Footprint Network, UK energy statistics). AI systems, tasks and accounting methodologies vary enormously, so treat every comparison here as an approximate equivalent based on a specific benchmark, not a universal constant.