DeepSeek energy and water use per prompt
What DeepSeek publishes, what independent research measures, and where the two do not meet.
Water
DeepSeek has not published a per-prompt water figure. This page does not estimate one: a made-up number would look identical to a published one and would be impossible for a reader to tell apart.
Energy
Benchmarked above 33 Wh for a long prompt, among the most energy-intensive models measured.
“A single short GPT-4o query consumes 0.43 Wh ... o3 and DeepSeek-R1 emerge as the most energy-intensive models, consuming over 33 Wh per long prompt, more than 70 times the consumption of GPT-4.1 nano.”
Which boundary this page uses
Every figure above is stated with its scope, because the same model can be described as using a third of a millilitre or ten millilitres of water depending on whether power generation is counted. Where DeepSeek publishes on-site cooling only, that is the narrower number, and it is not comparable with a life-cycle study without saying so.
What is left out
Training the model, building the accelerators, and keeping a facility provisioned for peak load are all excluded from per-prompt figures. They are real costs, and they are not attributable to your message, which is why this site keeps them out of the calculator and mentions them here instead.
Putting it in perspective
An ordinary message costs somewhere between a fifth and two-thirds of a watt-hour. A year of twenty messages a day is a few kilowatt-hours, which is a rounding error on a household bill. The figure that changes policy is the total across a platform, not the figure attached to your question.