Post

How much energy does one AI prompt use? We dug through studies: a normal prompt is about 0.3 watt hours, a long document about 2.5. A median DentroChat prompt: 0.2 to 1.3 watt hours. https://dentro.chat/blog/how-much-energy-does-an-ai-prompt-use

A long document (about 10k tokens, so 7 to 8 pages) costs about 2.5 Wh, and a long reasoning answer about 4 Wh. More tokens read and written means more electricity. Caching helps: if the same chat was read minutes before, the second run is about a tenth of the cost.

CO2 doesn't depend on the datacenter but mostly on the grid. On the EU average (210 g/kWh) a normal prompt is 0.06 g. In France or Sweden it is 0.01 g, in Germany 0.1 g, in the US about 0.12 g. Same model, four times the emissions.

Our models are small mixture-of-experts: 6 to 18 billion parameters work per token, against roughly 100 billion for GPT 4o. So a median real DentroChat prompt lands at 0.2 to 1.3 Wh. Reading 41k tokens sounds like a lot, but 30k of them are usually cached.