DES MOINES, Iowa (AP) – The cost of developing an artificial intelligence product like ChatGPT can be difficult to measure.
But one thing the Microsoft-backed OpenAI needed for its technology was copious amounts of water, drawn from the watershed of the Raccoon and Des Moines rivers in central Iowa, to cool a powerful supercomputer that was teaching its AI systems to imitate human writing.
Leading technology developers like Microsoft, OpenAI, and Google, in the race to embrace generative AI, have recognized that the growing demand for their AI tools comes with high costs, from expensive semiconductors to increased water consumption.
However, they often remain silent about the details. Few people in Iowa knew of its status as the birthplace of OpenAI’s most advanced major language model, GPT-4, before a senior Microsoft executive said in a speech that it “literally sprang up next to cornfields west of Des Moines “.
Building a large language model requires analyzing patterns in a vast body of human-written text. All that computing power consumes a lot of power and generates a lot of heat. To keep it cool on hot days, data centers must pump water in – often into a cooling tower outside of their warehouse-sized buildings.
In its latest environmental report, Microsoft announced that global water use increased 34% from 2021 to 2022 (to nearly 1.7 billion gallons, or more than 2,500 Olympic-size swimming pools), a sharp increase from previous years cited by outside researchers his AI research.
“It’s fair to say that most of the growth has come from AI,” including “their heavy investments in generative AI and their partnership with OpenAI,” said Shaolei Ren, a researcher at the University of California, Riverside who has tried to calculate the environmental impact of generative AI products like ChatGPT.
In an article to be published later this year, Ren’s team estimates that ChatGPT gulps down 500 milliliters of water (roughly the equivalent of a 16-ounce water bottle) each time you give him a series of 5 to 50 prompts or asks questions. The range varies depending on the location of the servers and the season. The estimate includes the indirect water consumption that the companies do not measure – for example to cool the power plants that supply the data centers with electricity.
“Most people are unaware of the resource usage underlying ChatGPT,” Ren said. “If you are not aware of resource use, we cannot contribute to resource conservation.”
Google reported a 20% increase in water use over the same period, which Ren also largely attributes to his AI work. Google’s rise hasn’t been uniform — it’s been steady in Oregon, where water use has garnered public attention, while doubling outside of Las Vegas. It was also thirsty in Iowa, supplying the data centers in Council Bluffs with more drinking water than anywhere else.
In response to questions from The Associated Press this week, Microsoft said in a statement that it is investing in research to measure the energy and carbon footprint of AI “while working on ways to make large systems more efficient, both in the training as well as in the application.”
“We will continue to monitor our emissions, accelerating progress while increasing the use of clean energy to power data centers, renewable energy purchases and other efforts to meet our sustainability goals of being carbon negative, water positive and zero waste by 2030,” said the states in the company’s statement.
OpenAI reiterated those comments in its own statement on Friday, saying it has “considerable concern” about how to make the best use of computing power.
“We recognize that training large models can be energy and water intensive” and are working to improve efficiency, it said.
Microsoft first invested $1 billion in San Francisco-based OpenAI in 2019, more than two years before the startup launched ChatGPT and sparked global fascination with AI advances. Under the agreement, the software giant would provide the computing power needed to train the AI models.
To do at least some of that work, the two companies turned to West Des Moines, Iowa, a city of 68,000 where Microsoft has built data centers for its cloud computing services for more than a decade. The fourth and fifth data centers are scheduled to open there this year.
“They’re building them as quickly as possible,” said Steve Gaer, who was the city’s mayor when Microsoft came to town. Gaer said the company was impressed by the city’s commitment to building public infrastructure and contributed an “amazing” amount through tax payments that support that investment.
“But you know, they’ve been pretty secretive about what they’re doing out there,” he added.
Microsoft first announced in 2020 that it was developing one of the most powerful supercomputers in the world for OpenAI, but declined to reveal its then location to AP, instead describing it as a “single system” with more than 285,000 cores made of conventional semiconductors and 10,000 Graphics processors – a type of chip that is crucial for AI workloads.
Experts say that because of the large amounts of data that need to be transferred between computing cores, it may make sense to pre-train an AI model at a single location.
Just in late May, Microsoft President Brad Smith announced that the company had built its “state-of-the-art AI supercomputing data center” in Iowa solely to enable OpenAI to train its fourth-generation model, GPT-4. The model, now based on premium versions of ChatGPT and some Microsoft-owned products, has accelerated a debate about mitigating the societal risks of AI.
“It was made in California by these extraordinary engineers, but it was really made in Iowa,” Smith said.
In some ways, West Des Moines is a relatively efficient place to train a powerful AI system, especially compared to Microsoft’s data centers in Arizona, which use far more water for the same computing needs.
“So if you’re building AI models at Microsoft, you should plan your training in Iowa instead of Arizona,” Ren said. “In terms of training, there is no difference. There’s a big difference in terms of water usage or energy usage.”
The weather in Iowa is cool enough most of the year that Microsoft can use outside air to keep the supercomputer running properly and remove heat from the building. Water is only withdrawn when the temperature exceeds 29.3 degrees Celsius (approximately 85 degrees Fahrenheit), the company announced in a public announcement.
That can still be a lot of water, especially in summer. According to West Des Moines Water Works, Microsoft pumped about 11.5 million gallons of water into its data centers in Iowa in July 2022, a month before OpenAI said it completed its GPT-4 training. This corresponded to about 6% of the total water consumption in the district, which also supplies the city’s residents with drinking water.
In 2022, a West Des Moines Water Works document states that it and the city government will consider “future data center projects” from Microsoft only if those projects can “demonstrate and implement technology to significantly reduce peak water usage from current levels.” to reduce”. Ensuring water supply for residential and other commercial purposes.
Microsoft said Thursday that it is working directly with the water company to address its feedback. In a written statement, the water company said the company has been a good partner and has worked with local authorities to reduce its water footprint while meeting its needs.
—-
O’Brien reported from Providence, Rhode Island.
——
The Associated Press and OpenAI have a licensing agreement that allows some of AP’s text archives to be used to train the tech company’s large language model. AP receives an undisclosed fee for use of its content.