Thirsty AI: To make ChatGPT and Bard, Microsoft severely disrupted the water supply of a US city, causing a disaster. Generative AI like ChatGPT, Bard uses a lot of water.
Microsoft and OpenAI almost caused a water shortage in the US state of Iowa because of how much water they used. For example, ChatGPT uses about 0.5 litres of water for every five prompts it gets.
When people try to make AI breakthroughs like ChatGPT and Bard, the cost to people and the world can be shockingly high.
AI uses a lot of water and energy, which is a well-known fact. Even more surprising, though, is how much water it takes to make AI tools.
A new story says that OpenAI and Microsoft went to Iowa’s Raccoon and Des Moines river watersheds and used a lot of water to cool their supercomputer, which was needed to train their GPT AI models.
Tech giants like Microsoft, OpenAI, and Google are in a race to join the creative AI craze. Yet, they talk little about how their AI projects affect the environment, even though they use expensive chips and a lot more water and electricity than they did before.
Des Moines and Raccoon, Iowa, were squeezed dry.
Few knew that central Iowa was a crucial part of making GPT-4, OpenAI’s most advanced large language model until a high-ranking Microsoft official said it was “literally made next to cornfields west of Des Moines.”
Creating a big language model requires a lot of computer work, which uses a lot of electricity and makes a lot of heat. Data centres often use water, which is sent to cooling towers outside their warehouse-sized buildings to keep the temperatures where they need to be.
According to Microsoft’s most recent environmental report, the company’s global water use rose by 34% from 2021 to 2022, hitting nearly 1.7 billion gallons, the same as more than 2,500 Olympic-sized swimming pools.
This increase is primarily due to the company’s study into AI, especially its significant investments in generative AI and its work with OpenAI.
Researchers at the University of California, Riverside, will say in a study published soon that ChatGPT uses about 500 millilitres of water every time a user gives it a series of 5 to 50 prompts.
“Most people don’t know how many resources are used to run ChatGPT,” says Shaolei Ren, a researcher trying to figure out how generative AI products affect the environment.
Google also said that it used 20% more water during the same period, primarily because of its AI projects; this increase in water use needed to be spread out. There were big jumps in Oregon, Las Vegas, and Iowa, where Google’s data centres in Council Bluffs used more potable water than anywhere else.
Not more processing power, but better use of it
In answer to questions from The Associated Press, Microsoft said that it is investing in research to determine how much energy and carbon AI uses and how big its carbon footprint is. It is also working to improve the efficiency of both training and using large systems. The company talked about how committed it was to sustainability, saying that by 2030, it wanted to be carbon-negative, water-positive, and waste-free.
OpenAI agreed with these ideas and said it was committed to responsible computing and working to reduce the damage that big AI models do to the environment.
Microsoft’s original $1 billion investment in OpenAI in 2019 set the stage for their partnership by giving AI models the computing power they needed to learn.
To make this partnership work, the two companies went to West Des Moines, Iowa, where Microsoft has built data centres for its cloud computer services for over a decade. At the time, the city’s mayor said that Microsoft was interested in the town because it invested in public facilities and paid many taxes.
West Des Moines is an excellent place to train AI models, especially compared to Microsoft’s data centres in Arizona, which use a lot of water. Because of the weather in Iowa, the machine can be cooled with air from outside for most of the year. When the temperature goes above 29.3 degrees Celsius, about 85 degrees Fahrenheit, the building uses water to cool it down.
Choosing to stand up to Big Tech
Still, much water is needed, especially in the summer. Just before OpenAI finished GPT-4 training in July 2022, Microsoft pumped about 11.5 million gallons of water into its data centres in Iowa; this was about 6% of the district’s water, which also serves the city’s people.
In 2022, a paper from the West Des Moines Water Works said that future Microsoft data centre projects would only be considered if they could show they had the technology to cut peak water use by a lot.