Big tech companies have significantly increased their use of water to cool their data centers, sparking concerns about the environmental impact of artificialof intelligence.
Microsoft, Google and Meta have increased their water consumption in recent years with the millions of users on their websites services.
Academics say AI demand will increase water consumption (from ground or surface sources) by between 4,2 and 6,6 billion cubic meters by 2027, or about half the amount the UK consumes each year.
Researchers from the University of California, Riverside, they wrote in a publication in Nature this week that it is a "critical time to uncover and address the secret water footprint of AI models amid an increasingly severe water scarcity crisis, worsening prolonged droughts and aging shared water infrastructure."
That concern has grown over the past year as leading technology companies race to launch products that use artificial intelligence, running on large language models capable of processing and generating vast amounts of text, numbers and other data.
Such models require massive amounts of computing power to run, requiring the use of massive infrastructures with servers that use cold water to cool the equipment. Some of the water evaporates during the cooling process.
Water is used in most forms production fuel and energy, for example, to extract oil and natural gas or to generate steam in thermal power plants. It also evaporates from the surface of reservoirs used for hydroelectric power.
In 2022, the most recent period for which data is available, Microsoft increased water consumption by 34%, Google by 22%, and Meta by 3%, as a result of increased use of their data centers.
A month before OpenAI finished training its most advanced model, GPT-4, a data center cluster in West Des Moines, Iowa, was consuming 6 percent of the area's water, according to a lawsuit filed by area residents.
Shaolei Ren, an associate professor at UC Riverside, reported that asking the popular chatbot ChatGPT for between 10 and 50 responses on the older GPT-3 model would be the equivalent of a 500ml bottle of water disappearing.
GPT-4, which has more parameters and requires more power, would likely need more water, Ren said. Detailed information for the model's energy use have not yet been made available.
Researchers have called for more comprehensive data and transparency from AI companies, including an analysis of how much different services consume, for example, search engines versus AI services.
