The Environmental Cost of 1 Million AI Prompts
AI’s footprint is tiny per prompt and enormous at scale. Here is what one million typical AI prompts actually cost the planet — by model class, text versus image — modelled from published factors with a fully open methodology. Free to cite (CC BY 4.0).
1,000,000 mid-size prompts cost roughly
A typical ~700-token exchange on a mid-size model. Modelled estimate, ±40%.
of electricity
before offsets
5,569 litres
The model you pick moves the footprint 5× or more
Per 1 million prompts of ~700 tokens each. Only the energy-per-token differs by class; overhead, grid and water are constant.
| Model class | Energy (kWh) | CO₂ (tonnes) | Water (litres) |
|---|---|---|---|
| Small / fast | 874 | 0.35 | 2,970 |
| Mid-size | 1,638 | 0.65 | 5,569 |
| Frontier / large | 4,150 | 1.64 | 14,109 |
| Reasoning-heavy | 4,368 | 1.73 | 14,851 |
A million images cost far more than a million chats
Per 1 million generated images at 1024×1024.
| Workload | Energy (kWh) | CO₂ (tonnes) | Water (litres) |
|---|---|---|---|
| Standard image | 3,120 | 1.23 | 10,608 |
| Higher-quality image | 5,460 | 2.16 | 18,564 |
| 1M mid-size chats (for comparison) | 1,638 | 0.65 | 5,569 |
What Ecoia actually measured and offset
This section reports anonymised, aggregate data from real Ecoia usage — the figures only Ecoia has. Populate each quarter from production analytics:
“One million typical AI prompts on a mid-size model produce roughly 0.65 tonnes of CO₂ and consume about 5,569 litres of water — before any offsetting.”
Cite as: Ecoia AI, “The Environmental Cost of 1 Million AI Prompts”, ecoia.ai/reports/ai-environmental-cost. Released under CC BY 4.0.
Report FAQ
How much CO2 do 1 million AI prompts produce?
On a mid-size model, one million typical (~700-token) prompts produce roughly 0.65 tonnes of CO2, use about 1,638 kWh of electricity and consume around 5,569 litres of water. Frontier and reasoning models are several times higher; small models several times lower. These are modelled estimates from published factors.
Is image generation worse than text?
Per item, yes. A single standard image is billed per image, not per token, and works out to roughly the energy of a multi-thousand-token chat answer. At scale a small share of image traffic can dominate a workload’s footprint.
Where does this data come from?
The per-model figures are derived from published, peer-reviewed and government factors (energy per token, PUE, grid intensity, water intensity), applied to a defined reference workload. The full methodology, sources and confidence range are published openly.
Can I cite or reproduce these numbers?
Yes. The report is released under CC BY 4.0 — cite “Ecoia AI” with a link. Every factor and formula is on the methodology page so the numbers are fully reproducible.
Keep reading
Use AI that gives back more than it takes
Ecoia.ai runs Claude, GPT & Gemini for chat, images and an API, and offsets over 200% of the water usage and carbon emissions your AI creates.