WATER CONSUMED
—litres
About — cups of water (250 mL each)
A LITTLE PERSPECTIVE
Your AI usage,
in context.
Every prompt has a footprint. Get a rough idea of what yours could look like, beyond the screen.
01 / YOUR USAGE
Let's keep it simple.
Choose your AI usage and the estimate to apply.
Your usage stays in your browser.
02 / YOUR ESTIMATE
The bigger picture.
Based on 100 prompts per day.
ChatGPT average query reference.
ENERGY USED
—kWh
Like a 10 W LED bulb for — hours
CARBON PROXY
—g CO₂
Comparable to — km driven in an average US car
IN TREE TERMS
—tree-days
Approximate time for one urban tree to absorb this CO₂
BREAD
—slices
Approximate carbon equivalent of bread slices (40 g each)
BEER
—beers
Approximate carbon equivalent of beers (330 mL each)
BEEF BURGER
—patties
Approximate carbon equivalent of beef patties (100 g each; patty only)
Estimates, not measurements. Food comparisons use different emissions boundaries; see the note below.
03 / YOUR USAGE OVER TIME
A little perspective over time.
The same usage and estimate, averaged over a day, month and year. Assumes steady usage, 365 days per year and 12 equal months. Food comparisons use different emissions boundaries; see the note below. A dash means the metric is unavailable.
1 day
— prompts on average
- Water
- — litres
- Energy
- — kWh
- Carbon proxy
- — g CO₂
- Tree time
- — tree-days
- Bread (40 g slices)
- — slices
- Beef burger (100 g patty only)
- — patties
- Beer (330 mL)
- — beers
1 month
— prompts on average
- Water
- — litres
- Energy
- — kWh
- Carbon proxy
- — g CO₂
- Tree time
- — tree-days
- Bread (40 g slices)
- — slices
- Beef burger (100 g patty only)
- — patties
- Beer (330 mL)
- — beers
1 year
— prompts on average
- Water
- — litres
- Energy
- — kWh
- Carbon proxy
- — g CO₂
- Tree time
- — tree-days
- Bread (40 g slices)
- — slices
- Beef burger (100 g patty only)
- — patties
- Beer (330 mL)
- — beers
A NOTE ON THE NUMBERSUseful perspective,
not a precise footprint.
A NOTE ON THE NUMBERS
Useful perspective,
not a precise footprint.
We multiply your prompt count by a published reference: ChatGPT's reported average query (0.34 Wh and 0.000085 US gallons of water), Google's median Gemini Apps text prompt (0.24 Wh and 0.26 mL), Mistral Large 2's 400-token response (45 mL of water), BLOOM's average query (4 Wh), or a LLaMA 65B benchmark response (approximately 0.3 Wh). The BLOOM and LLaMA figures are reported in Google's research summary. Estimates with no water or energy figure leave those results unavailable.
These studies use different methods and boundaries and are not directly comparable. BLOOM was measured without batching; LLaMA was benchmarked on specific hardware. Mistral's water figure includes upstream impacts. Long prompts, reasoning, images, model choice, and infrastructure can change the result substantially.
Carbon is an illustrative proxy, not actual emissions: energy is multiplied by the US EPA's 2022 average electricity factor of 394 g CO₂/kWh. Driving uses the EPA's 393 g CO₂e per mile for an average US gasoline vehicle. Tree-days use its approximately 60.5 kg CO₂ per urban tree per year averaged over 10 years; trees vary and do not instantly offset emissions. The 10 W bulb and 250 mL cup are illustrative.
Food comparisons divide the carbon proxy by approximate food greenhouse gas emissions. Bread assumes a 40 g slice of conventional baguette at 0.56 kg CO₂e per kg; beer assumes 330 mL at 0.771 kg CO₂e per litre. Both come from one cradle-to-gate study, not global averages. Beef burger means only a 100 g beef patty, not the bun, toppings or cooking, using the global average of 99.48 kg CO₂e per kg of beef from beef herds. Serving sizes are illustrative. Food CO₂e includes other greenhouse gases and has broader boundaries than the electricity-only CO₂ proxy, so these are rough comparisons, not measurements or offsets.
Results are not a full lifecycle assessment: training and other impacts are not included. Water estimates have different boundaries; Google's figure covers data-centre cooling, Mistral includes upstream impacts, and OpenAI's disclosure does not explain its methodology. Mistral also reports lifecycle emissions, but those are not used in the electricity-based carbon proxy shown here.