How much energy does Ai consume? One Gemini prompt is equivalent to 9 seconds in front of the TV

As Partha Ranganathan, who is primarily concerned with the company's computing infrastructure and data centres, with a focus on efficiency and sustainability, pointed out, the study only focuses on average textual queries. The analysis spans a 12-month period and thus includes a wider diversity of prompts. In particular, the 'median prompt' was measured, which is defined as the prompt that ranks at the 50th percentile in terms of energy consumption. The process works like this: for each AI model, the average energy consumed to process a single prompt (a single user request) is calculated. A ranking is created: all user prompts are put into a list and sorted according to the energy consumption required to process them. At this point, the 50th percentile is identified, i.e. the prompt that is exactly in the middle of this ranking. This is the 'median prompt'. The limitation of this study is that it does not tell us anything about text-to-image and text-to-video systems, those, to put it bluntly, that generate images and videos and for this very reason are considered more energy-intensive. The researchers explained that there is little consensus on how to measure the impact of other types of generation. Moreover, it would be scientifically incorrect to provide estimates to compare in terms of energy impact a query to Google's search engine with a query to Gemini.

'The interaction patterns are different,' said Ranganathan, 'it would be like comparing apples with pears.

Il grafico confronta diverse metodologie di misurazione del consumo energetico dei modelli di intelligenza artificiale, mettendo a confronto gli approcci esistenti con una metodologia proposta (Proposed Approach). La tabella scompone il consumo in diverse categorie: “Chip Power”, “Utilization”, “CPU & RAM”, “Idle Machines” e “Overhead”. Questo approccio, definito “Comprehensive Approach”, mira a una misurazione più completa e realistica del consumo energetico.

An average query i.e. a median Gemini prompt uses 0.26 ml of water (about 5 drops). Google experts pointed out that the use of water can reduce energy consumption by 10 per cent compared to air cooling. At sites in water-stressed areas, such as Mesa in Arizona, air cooling is chosen so as not to burden local resources. More than 25% of data centres use non-potable or recycled water.

0.24 Wh per query might not seem much, but this figure is only for a very simple prompt, such as 'hello'? Energy consumption increases as a conversation develops, queries become more complex and the memory required increases accordingly. Many analysts argue that measuring the environmental impact of generative AI per single query is crucial information, but it does not help to estimate the real impact of AI in terms of the volume of queries users make every day. Reading through Google's quarterly figures for the second quarter of 2025, we know about users but not about traffic volumes. The Gemini app has more than 450 million monthly active users while daily queries increased by more than 50 per cent compared to the first quarter. So we do not know how much energy Gemini consumes. We have to make do with having a unit of measurement, which has the merit, however, if adopted, of helping to compare LLM in the market.

The size of AI models, such as Gemini, has grown exponentially (doubling every 3.5 months or so). Google, they explained, is innovating to improve energy efficiency, implementing intelligent controls and redesigning power distribution systems. They are also shifting AI workloads to use electricity during periods of least stress on the grid.

In 12 months, Google reduced its footprint per query by 33 times in terms of energy and 44 times in terms of CO₂ emissions.