The recent explosion of generative AI has sparked criticism as the data centers used to train and run the models consume large amounts of energy.
People wonder – if I do a Gemini search, is it the same as driving a car a thousand miles, says Google's chief researcher Jeff Dean in a video clip, and continues:
The answer is: Absolutely not. The data we have collected shows that the carbon emissions are actually quite small.
Like watching TV
Tech companies are engaged in an arms race to meet the growing demand for AI. The largest data centers being built will use as much energy as 2 million households. Last year, they consumed 415 terawatt-hours, more than the entire UK's consumption, and are expected to more than double by 2030, according to the International Energy Agency (IEA).
However, for the individual user of the AI tool Gemini, the consumption is not particularly large, claims Google.
We saw that a median Gemini text prompt uses 0.24 watt-hours and emits 0.03 grams of carbon dioxide. To put it into perspective, it's like watching TV for less than nine seconds, says Savannah Goodman, head of Google's Advanced Energy Labs.
This is significantly less than calculations made by other researchers. Google explains the differences by saying that many calculations do not take into account various efficiency improvements that have taken place.
According to Goodman, a Gemini prompt today only uses 3 percent of the energy and leads to 2.3 percent of the carbon emissions compared to just a year ago.
Difficult to compare
Google is releasing its research methodology and hopes it will become an industry standard. At the same time, there are many shortcomings.
Among other things, the report says nothing about the environmental impact of image and video generation, or the energy consumption of more complex queries. It is also not possible to compare with a regular Google search.
It's a bit like comparing apples and pears. People interact with Gemini in a completely different way than with a Google search, so it's not entirely easy to compare, says Parthasarathy Ranganathan, technical expert.
Google has investigated how much electricity and water a regular text prompt in its own AI tool Gemini requires, and how large the carbon footprint is.
The calculations are based on the consumption of the chips (TPUs and GPUs) in the computer halls used for the calculations themselves, but also the machines that run idle to handle peaks, other chips and memories, overall cooling systems and the like, and the data centers' water consumption.
The figures the company arrived at are 0.24 Wh, 0.03 grams of carbon dioxide equivalents, and 0.26 milliliters of water per prompt.
