The AI technology behind ChatGPT was born in Iowa and uses a lot of water

The cost of building an AI product like ChatGPT is hard to measure. Microsoft-backed OpenAI's technology requires vast amounts of water from the Raccoon and Des Moines river basins in central Iowa to cool a powerful supercomputer that helps its AI system learn how to mimic human writing.

As leading technology developers like Microsoft, OpenAI, and Google race to capitalize on the generative AI boom, they also acknowledge that the growing demand for their AI tools comes with high costs, from expensive semiconductors to increased water consumption.

But they tend to keep specifics secret. Few people in Iowa knew that it was the birthplace of OpenAI's most advanced large-scale language model, GPT-4, until a Microsoft executive said in a speech that the model was "actually developed next to a cornfield west of Des Moines."

Building large language models requires analyzing patterns in large amounts of human-written text. All of these calculations consume a lot of electricity and generate a lot of heat. To stay cool on hot days, data centers need to pump water – usually to cooling towers outside warehouse-sized buildings.

In its latest environmental report, Microsoft disclosed that from 2021 to 2022, its global water consumption surged by 34% (to nearly 1.7 billion gallons, or more than 2,500 Olympic-sized swimming pools), a sharp increase in water consumption compared to previous years, which outside researchers believe is related to Microsoft's artificial intelligence research.

Shaolei Ren, a researcher at the University of California, Riverside, said: "It can be said that most of the increase in water consumption is caused by artificial intelligence, including its significant investment in generative artificial intelligence and its collaboration with OpenAI. "

In a paper to be published later this year, Ren's team estimates that every time you ask ChatGPT 5 to 50 prompts or questions, ChatGPT drinks 500 milliliters of water (close to the capacity of a 16-ounce water bottle). This range varies depending on the location and season of the server. This estimate also includes indirect water consumption that the company does not measure, such as cooling the power plant that powers the data center.

"Most people don't understand ChatGPT's resource usage," Ren said. "If you don't know how resources are being used, then we can't help you conserve resources."

Google's water consumption increased by 20% over the same period, which Ren attributed mainly to its AI work. Google's water use growth has been uneven — in Oregon, Google's water use has remained stable, with public concern in the state, while outside of Las Vegas, Google's water use has doubled. In Iowa, Google's devices are just as water-hungry, and its Board Bluffs data center uses more drinking water than anywhere else.

In response to questions from The Associated Press, Microsoft said in a statement this week that it is investing in research to measure the energy and carbon footprint of artificial intelligence, "while looking at how to make large systems more efficient in training and applications."

"We will continue to monitor our emissions and accelerate progress while increasing the use of clean energy to power data centers, purchasing renewable energy, and making other efforts to achieve our sustainability goals of achieving negative carbon emissions, positive water emissions, and zero waste by 2030." The company's statement said.

OpenAI echoed the same sentiment in its own statement on Friday, saying it was "seriously thinking" about how to make better use of computing power. "We recognize that training on large models can be energy and water intensive," it said, and are working to improve efficiency.

Microsoft first invested $1 billion in San Francisco-based OpenAI in 2019, more than two years after the startup launched ChatGPT and drew global attention to AI advancements. As part of the deal, the software giant will provide the computing power needed to train AI models.

To get at least some of that done, the two companies set their sights on West Des Moines, Iowa, a city of 68,000 that has been home to Microsoft's cloud computing services for more than a decade. Microsoft's fourth and fifth data centers will open here later this year.

Steve Gaer said, "They're building these data centers as fast as they can. What attracted the company was the city's commitment to building public infrastructure and contributing an "amazing" amount of money to support the investment through taxes. He added: "But, you know, they were pretty secretive about what they were doing there. "

For the first time, Microsoft said it was working on one of the world's most powerful supercomputers for 2020's OpenAI, when the company declined to disclose its location but described it as a "single system" with more than 285,000 traditional semiconductor cores and 10,000 graphics processors — chips that have become key to AI workloads.

Experts say it makes sense to "pre-train" AI models in a single location because large amounts of data need to be transferred between computing cores.

It wasn't until late May that Microsoft President Brad Smith revealed that Microsoft had established an "advanced AI supercomputing data center" in Iowa specifically for OpenAI to train its fourth-generation model, GPT-4. The model now powers advanced versions of ChatGPT and some of Microsoft's own products, and accelerates discussions about curbing the societal risks of AI. It was made by extraordinary engineers in California, but it was actually made in Iowa.

In a way, West Des Moines is a relatively efficient place to train powerful AI systems, especially compared to Microsoft's data center in Arizona, which consumes far more water for the same computing needs.

"So if you're going to develop AI models in-house at Microsoft, you should schedule the training in Iowa, not Arizona," Ren said. There is no difference in terms of training. But in terms of water consumption or energy consumption, the difference is significant. "

With the weather cool in Iowa for most of the year, Microsoft can use outside air to keep the supercomputer up and running and push heat out of the building. Microsoft takes water only when the temperature exceeds 29.3 degrees Celsius (about 85 degrees Fahrenheit), the company said in a publicly disclosed filing.

This still consumes a lot of water, especially in summer. In July 2022, a month before OpenAI said it completed GPT-4 training, Microsoft pumped about 11.5 million gallons of water into its Iowa data center cluster, according to West Des Moines Water Works. This accounts for about 6% of the total water used in the area, which also provides drinking water to the city's residents.

In 2022, a West Des Moines Water plant document said that if Microsoft's data center projects can "demonstrate and implement technologies that significantly reduce peak water use at current levels" to protect residential and other commercial water supplies, then West Des Moines Water and the city "will consider these projects."

Microsoft said Thursday that it is working directly with the water authority to address issues with its feedback. The plant said in a written statement that Microsoft has been a good partner and has been working with local officials to reduce its water footprint while meeting water needs.