How much energy does a ChatGPT prompt use?
About 0.3 Wh to answer a typical text prompt. Counting its share of the energy spent training the model, it comes to about 0.5 Wh, which puts about 0.39 g of CO₂e in the air. An hour of streamed video uses about as much as 154 prompts.
One prompt
- Energy to answer it
- 0.3 Wh
- Share of training
- 0.2 Wh
- Carbon
- 0.39 g CO₂e
- Water consumed
- 2.20 mL
- Energy in all
- 0.5 Wh
A month at 10 prompts a day
- Prompts, with AI in search and apps
- 426
- Energy, with training
- 213 Wh
- Carbon
- 167 g CO₂e
- Healthy life lost
- 0.3 sec
- Water consumed
- 937.4 mL
How I got this
The 0.3 Wh is Epoch AI's estimate for a typical ChatGPT reply of a few hundred words. Long chats cost more, because each turn re-reads the whole conversation, and reasoning models write hidden thinking before they answer. When you link your account I measure your own chats instead.
About 40% of machine learning energy goes to training models rather than answering prompts (Patterson et al. 2022, from Google), so each prompt carries its share. AI is new demand on the grid, so I count its carbon at the power plants that ramp up to meet it, mostly gas, at about 1,100 lb of CO₂ a MWh (EPA AVERT). I add 12% for getting the fuel to the plant, 5% for power lost on the wires, and a quarter for making the chips and building the data centers.
Every factor and source is on How it's calculated.
What it costs to cover
A month at 10 prompts a day comes to under 1¢ in donations, including my margin for an estimate. That's below the $1.00 smallest payment, so most people pay once instead. For someone born in 1995, $2.57 with fees covers everything since late 2022 and every year to age 85, with room for use to grow.
The money goes to charities that work on climate policy, clean air, water in dry places and protecting forest.
Measure your own use and see what it would cost to cover.