Fresh concerns over the environmental impact of AI development are being raised with the revelation that water use by Microsoft and Google rose dramatically as both developed their AI models. As reported by the Associated Press, Microsoft reported a 34% increase in water use from 2021 to 2022 that amounts to 1.7 billion gallons. Google has reported a 20% increase in the same time frame.
Outside researchers tied the water consumption rise to generative AI development. Shaolen Ren, a researcher at the University of California, Riverside, said ChatGPT slurps up a 16-ounce bottle of water every time it is asked a question involving between five and 50 prompts, with the range varying depending on where the servers are located and the time of year. Water is used to cool power plants that supply servers with electricity. AI computations themselves can be very energy intensive.
Few people are aware of the water consumption requirements. ChatGPT was developed in a rural area west of Des Moines, IA, according to a Microsoft executive, with water pulled from the Raccoon and Des Moines river watersheds to cool a powerful supercomputer. The requirements for the construction of ChatGPT in Iowa were enormous: 285,000 cores of semiconductors and 10,000 graphic processors.
Google also was thirsty in Iowa and Oregon, but perhaps potentially more concerning was the doubling of its use of water at its facilities outside Las Vegas, a region already concerned about water use due to hot weather and the sharp lowering of water levels on the critical Colorado River. Both companies said they were mindful of AI’s energy use and are working on ways to improve water consumption. Location is a big consideration. In Iowa, the weather is cool enough for outside air to regulate temperatures, but if the temperature rises above 85 degrees F, then water is used. In July, 2022, Microsoft’s local usage in Iowa amounted to six percent of water usage and local authorities are pushing back by saying any future expansion must address water usage, especially during peak periods. AI developers have been criticized for their lack of transparency on water usage. One simple, albeit partial, solution might be to run AI training regimens during cooler parts of the day. AI water management now has a higher profile.
This is not the first time AI water usage concerns have been raised. Last spring, university researchers from Colorado Riverside and Texas Arlington estimated ChatGPT3’s training required an estimated 185,000 gallons of water. The study estimated the average data center uses a gallon of water for every kilowatt hour. Fresh water also is used to control humidity. Sea water can’t be used, due to its corrosive and bacterial dangers.