In a first, Google has released data on how much energy an AI prompt uses

So some Gemini claims use energy much more than this: gives a debt an example to feed dozens of books in Gemini and request a detailed summary of its content. “This is the type of thing that may take more energy than the medium claim,” says Dean. The use of the thinking model can also be the order -related demand because these models take more steps before producing an answer.

This report was also strictly limited to text claims, so it is not required to create a picture or video. (Other analyzes, including one in Massachusetts Institute Technology Review TechnologyPower Hungry series earlier this year, showing that these tasks can require more energy.)

The report also finds that the total energy used to create Gemini has decreased significantly over time. The medium Gemini claim was used 33 times in May 2024 of what it was in May 2025, according to Google. The company indicates progress in its models and other program improvements for improvements.

Google also estimates greenhouse gas emissions associated with medium demand, which they placed at 0.03 grams of carbon dioxide. To reach this number, the company doubled the total energy used to respond to demand the average emissions for each unit of electricity.

Instead of using emissions based on the average American network, or the average network in which Google operates, the company instead uses a based market estimate, which takes into account the electricity purchases that the company performs from clean energy projects. The company has signed agreements to purchase more than 22 gigawatts of energy from sources, including solar projects, winds, thermal energy, thermal and advanced energy since 2010. Because of these purchases, Google emissions for each unit of electricity on paper are nearly a third of those on the middle network in which it operates.

Artificial intelligence data centers also consume water for cooling, and Google estimates that each router consumes 0.26 milliliters of water, or about five drops.

The aim of this work was to provide users with a window in the use of energy for their interactions with artificial intelligence, says Dean.

“People use [AI tools] For all kinds of things, they should not have great concerns about the use of energy or the use of Gemini models, because in our actual measurements, what we was able to show was that it is already equal to things that you do not do until you think about it on a daily basis, such as watching a few seconds of TV or its five water consumption. “