The main concerns regarding AI and the environment are the extensive usage of electricity and water.
High electricity use makes sense. It is almost too redundant to mention that AI uses copious amounts of electricity. However, the extent of how much electricity is used may be hard to grasp.
In the Harvard Business Review article “The Uneven Distribution of AI’s Environmental Impacts,” Shaolei Ren and Adam Wierman discuss how AI consumes a disproportionate amount of resources given how little it contributes to the greater good of humanity.
In the article, the two authors state that “the training process for a single AI model… can consume thousands of megawatt hours of electricity and emit hundreds of tons of carbon. This is roughly equivalent to the annual carbon emissions of hundreds of households in America.”
There is little doubt that it takes a substantial amount of energy to train an AI model, but causing the same amount of environmental damage as a small town highlights the scale of that energy use.
Energy consumption, however, is not the only environmental impact of AI. In fact, it is the lesser of the two main problems. The issue most are concerned about is how much water AI is reported to consume.
In his article “Data Centers and Water Consumption,” written for the Environmental and Energy Study Institute, Miguel Yañez-Barnuevo states that “large data centers can consume up to five million gallons [of water] per day, equivalent to the water use of a town populated by 10,000 to 50,000 people.”
Orange City, Iowa, has a population of 6,267, according to the 2020 U.S. Census. When training their models, AI data centers can consume more freshwater than a town at least twice the size of Orange City.
Yañez-Barnuevo also sites a study from the University of California, Riverside which says that each 100-word prompt given to an AI model consumes an estimates 519 milliliters of water. An amount of water that is roughly the same as what’s in the average plastic bottle.
As significant as that may sound, Yañez-Barnuevo explains in his article that this level of water consumption does not have to be the status quo. He notes that there are alternative ways to cool AI data centers. He goes further to say that there are more efficient ways, though those methods are more expensive for the companies that operate them.
With that in mind, the issue appears to be less about AI itself and more about the infrastructure and cost-saving decisions made by the companies that run large data centers.
As Yañez-Barnuevo says, “If the United States moves toward 100% renweable energy generatino and the retirement of fossil fuel plants, the water savings would be enormous, with billions of gallons of water saved, and more freshwater would be available for both human consumption and natural ecosystems.”
