Advertisement
Science
ChinaScience

China’s thirsty data centres, AI industy could use more water than size of South Korea’s population by 2030: report warns

  • Hong Kong think tank says annual water use by Chinese data facilities – almost half of which are in dry regions – could more than double by 2030
  • It takes 20 times more water to have a ChatGPT conversation than to run a Google search, according to the non-profit group’s report

Reading Time:3 minutes
Why you can trust SCMP
2
Data centres consume water directly through cooling processes and indirectly from electricity generation. Photo: Xinhua
Holly Chik
China’s thirsty data centres and the rapid growth of artificial intelligence (AI) could dramatically increase demand on the country’s water resources, according to a new report by think tank China Water Risk.
Advertisement
The Hong Kong-based non-profit estimated the annual water consumption of data centres in China to be around 1.3 billion cubic metres (343 billion gallons) – enough for residential use for 26 million people. By 2030, the figure could reach over 3 billion cubic metres as more data facilities are expected to open, equivalent to the demand of a population greater than that of South Korea.

Data centres consume water directly to prevent information technology equipment from overheating. They also consume water indirectly from coal-powered electricity generation.

The team projected that by the end of the decade, China would have more than 11 million data centre racks, which house servers, cables and other equipment. That is nearly triple the number it had in 2020 of around 4 million.

The boom in generative AI technology is also expected to add to water demands from the information and communication technology (ICT) industry.
“Along with its huge computational power, AI chatbots drink staggering amounts of water to cool themselves down,” said the report released on Friday.
Advertisement

It pointed to a preprint study – meaning it has not yet been peer-reviewed – from last year by researchers in the United States that showed the large language model GPT-3 consumed 500 millilitres (16.9 fluid ounces) of water for every 10 to 50 responses it generates. That is 20 times more than it takes to produce 50 Google searches.

Advertisement