(Paris) A request to Gemini, the artificial intelligence (IA) of Google, consumes less energy than watching nine seconds of television, said Thursday the American tech giant which quotes a study conducted by its researchers while the energy voracity of AI remains a major problem in the sector.
On average, a textual “prompt” request (called “gemini applications consumes 0.24 watt hours (WH), less than viewing nine seconds of modern television which corresponds to around 100 Wh, estimate Google researchers.
A request also emits 0.03 grams of carbon dioxide equivalent and consumes 0.26 milliliter of water, the equivalent of about five drops, they add. This study relates to data from May 2025 collected through the various applications where Gemini is present.
Google does not detail what it means by “prompt”, if it is a word or a series of words, and refuses to communicate the number of requests formulated on Gemini every day, which would allow an idea of the overall energy consumption of the model.
Google also sees its greenhouse gas emissions increase.
They reached 15.2 million tonnes of CO2 In 2024, more than 50 % more than in 2019, its reference year, according to its last annual environmental report, largely because of its data centers, essential for the development of AI.
For its study, the group has counted the active calculation phases of AI models to generate responses to user requests, the real energy consumed by fleas and that of processors and data centers as a whole.
Google explains that it has taken into account the energy of inactive machines, which must be supplied constantly to be ready to manage traffic peaks at any time, the cooling systems of very energy -consuming servers and other indirect charges of its data centers.
These figures do not, however, integrate all the training of AI models, and should be taken with precaution, the study having not been verified by an independent third party, as Google itself recalls in a note accompanying the study.
Calculating the environmental footprint of an AI model is an extremely complex task because there is no global standard of measurement.
The tech giants mostly maintain opacity on the functioning of these programs and their manufacturing conditions, which complicates the task of independent researchers and international institutions that lack data.
The boss of Openai, Sam Altman, for example revealed in June that each request sent to Chatgpt consumed on average 0.34 Wh electricity, the energy necessary to operate an oven for a second, and 0.3 ml of water.
However, the manager provided no explanation on how Openai had reached these figures, which makes any comparison impossible.
The question of EA energy consumption is however crucial.
The data centers, a base of this technology, should weigh around 3 % of the world’s electricity needs by 2030, double the current proportion, according to the International Energy Agency (AIE).