AI's drinking habit has been grossly overstated, according to a newly published report from Google, which claims that software advancements have cut Gemini's water consumption per prompt to roughly five drops of water -- substantially less than prior estimates.
Flaunting a comprehensive new test methodology, Google estimates [PDF] its Gemini apps consume 0.24 watt hours of electricity and 0.26 milliliters (ml) of water to generate the median-length text prompt.
Google points out that's far less than the 45ml to 47.5ml of water that Mistral AI and researchers at UC Riverside have said are required to generate roughly a page's worth of text using a medium-sized model, like Mistral Large 2 or GPT-3.
However, Google's claims are misleading because they draw a false equivalence between onsite and total water consumption, according to Shaolei Ren, associate professor of electrical and computer engineering at UC Riverside and one of the authors of the papers cited in the Google report.
To understand why, it's important to know that datacenters consume water both on and off site.
The power-hungry facilities often employ cooling towers, which evaporate water that's pumped into them. This process chills the air entering the facility, keeping the CPUs and GPUs within from overheating. Using water is more power-efficient than refrigerant. By Google's own estimates, about 80 percent of all water removed from watersheds near its datacenters is consumed by evaporative cooling for those datacenters.
But water is also consumed in the process of generating the energy needed to keep all those servers humming along. Just like datacenters, gas, coal, and nuclear plants also employ cooling towers which - spoiler alert - also use a lot of water. Because of this, datacenters that don't consume water directly can still have a major impact on the local watershed.
...