Google says Gemini AI uses as little energy as watching TV for 9 seconds, cuts carbon footprint 44x in a year

Gemini

How much energy does it really take to power an AI like Gemini, the model behind Google’s apps? Google has finally given us a clear picture, and the numbers are smaller than many people might expect.

According to a new study shared by Google, each Gemini text prompt – that’s every time someone types in a question or request – uses just 0.24 watt-hours of energy. To put that into perspective, it’s about the same as watching television for less than nine seconds. Along with that, a single prompt is responsible for 0.03 grams of carbon emissions and consumes 0.26 milliliters of water – roughly five drops.

The announcement is important because until now, there has been very little public data on the environmental footprint of AI systems. With demand for AI growing rapidly, questions about energy use have become louder, and Google’s disclosure offers one of the most detailed looks yet.

What’s even more striking is how quickly Gemini has become more efficient. Over the past year, Google says the median energy use per text prompt dropped by 33 times, while the carbon footprint fell by 44 times – all while responses got better. This progress builds on Google’s ongoing work to make its data centers more energy efficient. In 2024, for example, the company cut its data center emissions by 12 percent even though electricity use grew by 27 percent.

So how is Google managing this balance? The company says it comes down to designing efficiency into every part of the system – from the models themselves to the hardware they run on. Gemini models are built with structures like Mixture-of-Experts, which only activate the parts of the model needed to answer a specific question. This reduces unnecessary computation. On top of that, Google has been developing custom hardware like its Tensor Processing Units (TPUs), which are far more efficient than traditional chips. Its latest TPU, codenamed Ironwood, is 30 times more energy efficient than the first one the company released publicly.

Google’s data centers also play a big role. They are among the most efficient in the industry, with an average Power Usage Effectiveness (PUE) of 1.09, meaning very little energy goes to waste. The company is also working toward running on carbon-free energy around the clock and aims to replenish more water than it consumes.

Of course, the energy use figures come with some caveats. Many earlier estimates of AI’s footprint only looked at the power used by active chips like GPUs or TPUs, ignoring other factors such as idle machines, CPUs, cooling systems, and water use. Google says its new methodology is more comprehensive because it accounts for all of these elements.

By publishing these findings, Google says it hopes to encourage more transparency across the tech industry. With AI becoming part of everyday life, understanding its hidden costs is crucial. And if Google’s numbers hold up, they suggest that AI models like Gemini might not be as power-hungry per use as some fear – though the real challenge is scaling this efficiency as millions, and eventually billions, of people use AI every day.