How much does a DeepSeek prompt consume?
A short prompt sent to DeepSeek-V3 uses 2.78 Wh when the model runs on DeepSeek's own servers, and 0.74 Wh when the same model runs on Microsoft Azure: 3.7 times less. For DeepSeek-R1, the reasoning model, the gap goes from 19.25 Wh to 2.35 Wh, or 8.2 times. These are modelled estimates by Jegham and co-authors (the November 24, 2025 version of "How Hungry is AI?"), not measurements published by DeepSeek: this site found no per-prompt figure issued by the company.
Figures published on November 24, 2025, modelled estimate, short prompt (100 tokens in, 300 out), DeepSeek-V3 hosted by DeepSeek · Sources: « How Hungry is AI? » (arXiv 2505.09598)
How much does a DeepSeek prompt use, depending on its length?
Between 0.74 Wh and 29.08 Wh, depending on the model, the host and the prompt size. The study tests three sizes: short (100 tokens in, 300 out), medium (1,000 in and 1,000 out) and long (10,000 in, 1,500 out). A token is a chunk of a word. For DeepSeek-V3 on DeepSeek's servers: 2.78 Wh, 8.86 Wh, then 13.16 Wh. For DeepSeek-R1 in the same place: 19.25 Wh, 24.60 Wh, then 29.08 Wh. On Azure, V3 drops to 0.74 Wh, 2.17 Wh and 3.70 Wh, and R1 to 2.35 Wh, 4.33 Wh and 7.41 Wh. On a short prompt, R1 uses nearly 7 times more than V3 at DeepSeek, and about 3 times more on Azure.
Why does the same model use 3 to 8 times less on Azure?
Because the electricity behind a prompt depends on the machine and the data center that process it, not only on the model. The study derives energy from the generation time observed on public APIs, the power draw of the graphics cards and the data center's efficiency (its PUE). On a long prompt, R1 and V3 use about 74% and 72% less on Azure. The authors conclude that "hardware and data center efficiency, not model design alone" drive real-world energy use. One caution: for DeepSeek, the study assumes H800 cards and takes the average efficiency of the thirty most efficient data centers in China. That is an assumption, not a reading taken on the company's servers.
What about water and CO₂?
For DeepSeek, the study details these two values only for a long prompt (10,000 tokens in, 1,500 out). The text gives, for DeepSeek-R1 at DeepSeek, more than 200 mL of water and about 17 g of CO₂e. For the same model on Azure: 34 mL and 2.5 g of CO₂e, a drop close to 85%. The authors put the gap down to less efficient data centers, less frugal cooling and a more carbon-heavy electricity supply. For a short prompt, the study only shows these two values as a chart, so this site does not reproduce them.
Are these figures reliable, and do they hold over time?
models.pages.deepseek.a4
A DeepSeek prompt in everyday actions
The figures published for DeepSeek, converted with the site's reference actions (every action value is sourced, see the Sources page):
- 1 hour of LED light = 3.6 DeepSeek prompts
- 1 hour of TV = 25 DeepSeek prompts
A model's number doesn't say which host to choose for a given job: consumption is only one criterion among others. That's a different craft: the GHIS studio helps companies choose and scope their AI tools.