Traffic flowed past a Microsoft data center on Interstate 35 in West Des Moines, Iowa, on Tuesday, September 5, 2023. Microsoft has spent over a decade building a cluster of data centers to support its cloud computing services. The city is set to open its fourth and fifth data centers later this year. (Credit: AP Photo/Charlie Neibergall)
Measuring the cost of developing an artificial intelligence product like ChatGPT can be a complex task.
However, one crucial requirement for Microsoft-backed OpenAI in its AI technology development was a significant supply of water. This water was drawn from the Raccoon and Des Moines rivers’ watershed in central Iowa to cool a powerful supercomputer used in teaching AI systems to emulate human writing.
Leading technology companies, including Microsoft, OpenAI, and Google, have recognized the substantial expenses associated with meeting the growing demand for their AI tools. These expenses range from costly semiconductors to increased water consumption. Nevertheless, they often remain discreet about the specifics of these costs. Few individuals in Iowa were aware of the city’s role as the birthplace of OpenAI’s most advanced large language model, GPT-4, until a top Microsoft executive revealed in a speech that it had been created “literally next to cornfields west of Des Moines.”
Developing a large language model involves the analysis of patterns in extensive amounts of human-written text. The computational power required for this task is immense and generates significant heat. To keep these systems cool, data centers must pump in water, often directed to cooling towers outside their warehouse-sized buildings.
According to Microsoft’s latest environmental report, its global water consumption surged by 34% from 2021 to 2022, reaching nearly 1.7 billion gallons (equivalent to more than 2,500 Olympic-sized swimming pools). This sharp increase is largely attributed to the company’s AI research, especially its investment in generative AI and partnership with OpenAI, as stated by Shaolei Ren, a researcher at the University of California, Riverside.
Ren’s research estimates that ChatGPT consumes approximately 500 milliliters of water (similar to the content of a 16-ounce water bottle) each time a user asks a series of 5 to 50 prompts or questions. This estimate takes into account indirect water usage, such as the water required to cool the power plants supplying electricity to data centers, which the companies do not typically measure.
“Most people are not aware of the resource usage underlying ChatGPT,” Ren noted. “If you’re not aware of the resource usage, then there’s no way that we can help conserve the resources.”
Google also reported a 20% increase in water use during the same period, which Ren largely attributes to the company’s AI work. This water use increase varied across locations, with noticeable spikes in Oregon, Las Vegas, and Iowa, where Google’s data centers are located.
In response to inquiries from The Associated Press, Microsoft stated its commitment to researching and measuring AI’s energy and carbon footprint while working on enhancing the efficiency of large systems in both training and application. The company emphasized its aim to meet sustainability goals, including becoming carbon negative, water positive, and zero waste by 2030.
OpenAI echoed these sentiments in its statement, emphasizing its focus on making efficient use of computing power.
Microsoft’s initial $1 billion investment in OpenAI in 2019 laid the foundation for their partnership. As part of the deal, Microsoft would provide the necessary computing power to train AI models. Their collaboration led them to West Des Moines, Iowa, where Microsoft has established a cluster of data centers to power its cloud computing services for over a decade. The city is poised to open its fourth and fifth data centers later this year.
West Des Moines proved to be a relatively efficient location for training powerful AI systems, particularly when compared to Microsoft’s data centers in Arizona, which consume significantly more water for the same computational demands.
Microsoft’s president, Brad Smith, disclosed in late May that the company had constructed an “advanced AI supercomputing data center” in Iowa exclusively for OpenAI’s use in training its fourth-generation model, GPT-4. This model now powers premium versions of ChatGPT and some Microsoft products, sparking discussions about managing AI’s societal implications.
In terms of training, Iowa’s climate is generally conducive to efficient operations. Most of the year, the cool weather allows Microsoft to use outside air to maintain the supercomputer’s performance and dissipate heat. Only when the temperature exceeds 29.3 degrees Celsius (about 85 degrees Fahrenheit) does the company resort to using water for cooling.
Nonetheless, even in this efficient environment, water consumption can be significant, especially during the summer months. In July 2022, the month preceding the completion of GPT-4’s training, Microsoft utilized around 11.5 million gallons of water for its Iowa data centers, accounting for approximately 6% of the district’s total water consumption, which also supplies drinking water to the city’s residents.
In 2022, a document from the West Des Moines Water Works stated that future data center projects from Microsoft would only be considered if they could demonstrate and implement technology to substantially reduce peak water usage, preserving the water supply for residential and commercial needs.
Microsoft has expressed its commitment to collaborating with the water works to address their feedback and reduce their water footprint while still meeting operational needs.