Google claims Gemini AI now 33x more energy-efficient; each prompt uses less power than 9 seconds of TV

Google has released a detailed methodology to measure the energy, water, and carbon footprint of its AI models, addressing a critical gap in understanding AI’s environmental impact. It claims to have achieved major efficiency gains: over 12 months, the median energy use per Gemini Apps text prompt fell 33x, and carbon emissions dropped 44x, while still improving output quality, says Google.

To put it in perspective, each prompt now uses less energy than nine seconds of TV. With a 12% reduction in data centre emissions despite rising electricity demand, Google claims it is a "milestone" in sustainable AI development and energy transparency.

"Gemini’s efficiency gains are the result of years of work, but this is just the beginning. Recognising that AI demand is growing, we're heavily investing in reducing the power provisioning costs and water required per prompt," says Google.

The tech major, via its detailed methodology to measure energy, water and carbon emission of its AI models, has found that its models have achieved significant efficiency gains over the past 12 months, reducing the median energy consumption and carbon footprint per Gemini Apps text prompt by factors of 33x and 44x, respectively.

"To put this into perspective, based on Google's analysis, the energy consumed per median prompt by the Google Gemini app text is equivalent to watching TV for less than nine seconds," says Google.

Google has been focused on improving data centre efficiency for a long time now. In 2024, for example, it reduced data centre energy emissions by 12% even as electricity consumption grew by 27% year-over-year on the expansion of its business and services. The search engine giant says the aim of the study findings and methodology is to drive industry-wide progress toward more efficient AI.

Here are the key findings of Google's study on how much energy it takes for an AI search.

1.) Via its detailed methodology for measuring the energy, emissions, and water impact of Gemini prompts, Google estimates the median Gemini Apps text prompt uses 0.24 watt-hours (Wh) of energy, equivalent to watching TV for less than 9 seconds. The mean text emits 0.03 grams of carbon dioxide equivalent (gCO2e), and consumes 0.26 millilitres (or about 5 drops) of water. These figures, claims Google, are substantially lower than many public estimates.

2). Google says its AI systems are getting better with both software and hardware efficiency, with its Gemini text app prompts reducing the energy and carbon footprint by 33x and 44x, respectively, while at the same time, delivering higher quality responses. To continue the pace of improvement, Google says it has developed an approach that considers the realities of serving AI at its scale, including full system dynamic power, factoring this into total energy footprint, optimising CPU and RAM usage, power usage effectiveness (PUE) of the data centre, and optimising water consumption in the data centre.

3). In its full-stack approach to AI, Google claims it has been able to achieve dramatic efficiency gains from custom hardware and highly efficient models, to robust serving systems. This has been brought into every layer of AI, including more efficient model architectures, efficient algorithms and quantisation, inference and serving, custom-built hardware and ultra-efficient data centres.