
Quantifying the costs of constructing an artificial intelligence product like ChatGPT may present a considerable challenge, yet it’s clear that a primary resource that Microsoft-backed OpenAI needed for its technology was water—ample amounts drawn from the watershed of the Raccoon and Des Moines rivers situated in central Iowa. This resource was integral to cooling the powerful supercomputer engaged in equipping OpenAI’s AI systems with the skills to emulate human writing styles.
On the fringes of the generative AI revolution, prime tech developers such as Microsoft, OpenAI, and Google have conceded that the surging demand for their AI instruments incurs considerable costs–from pricey semiconductors to escalating water usage. However, the specifics surrounding these costs often remain enshrouded in secrecy. For instance, the pivotal role of Iowa as a nucleus for the development of OpenAI’s most advanced large language model, GPT-4, was relatively unknown until an eminent Microsoft executive highlighted that it was “literally made next to cornfields west of Des Moines.”
The construction of an extensive language model hinges on analysis of patterns discernible across vast volumes of human-written text, a process that necessitates a copious amount of electricity and yields considerable heat. On sweltering days, data centres must therefore pull in water—often to an exterior cooling tower to maintain an optimum temperature within their warehouse-sized establishments.
According to Microsoft’s latest environmental report, global water consumption soared by 34% from 2021 to 2022, reaching approximately 1.7 billion gallons – over 2,500 Olympic-sized swimming pools. This noticeable surge, compared to preceding years, is typically attributed to AI research by independent researchers. University of California-based researcher, Shaolei Ren, suggested that this is predominantly due to hefty investments in generative AI and collaborations with OpenAI.
Ren and his team predict that every instance of ChatGPT handling a sequence of 5 to 50 queries or prompts consumes around 500 millilitres of water, comparable to a 16-ounce water bottle. Variables such as server locations and seasons cause oscillations in this range, with a majority of this water consumption attributed to cooling power plants supplying electricity to data centres—an aspect often overlooked by companies.
Moreover, Google also reported a 20% increase in water usage during the same period, which Ren hypothesizes mainly stems from its work on AI. However, usage rates vary across different locations, maintaining a steady rate in Oregon while doubling just outside of Las Vegas. Google’s Iowa-based Council Bluffs data centres are the most water-intensive, drawing more potable water than any other facility.
In a recent statement, Microsoft mentioned that they are financing research to assess AI’s energy and carbon footprint, while pioneering ways to enhance the efficiency of large systems, both in instruction and application. The company also affirmed its commitment to remain vigilant of its emissions while expediting the progression toward clean energy for data centres and renewing its dedication toward achieving its 2030 sustainability goals.
This sentiment was echoed in a subsequent statement by OpenAI, who admitted to investing a great deal of thought into optimising computing power usage while recognising the energy and water-intensive nature of these models. Since their initial $1 billion investment into San Francisco-based OpenAI in 2019, Microsoft is committed to supplying the requisite computing resources to train AI models.
In pursuit of this endeavor, the two tech giants focused their attention on West Des Moines, Iowa—a city of 68,000 residents highly regarded by Microsoft for its data centres which undergird their cloud computing services. Steve Gaer, the ex-mayor, noted that Microsoft had been an integral part of its community, owing to the ‘staggering’ local infrastructure investment through tax payments.
In 2020, Microsoft unveiled plans for developing one of the world’s most powerful supercomputers for OpenAI, comprised of more than 285,000 semiconductor cores and 10,000 graphics processors. Later, in May, Microsoft’s president Brad Smith announced the existence of this sophisticated data centre in Iowa, crafted exclusively to aid the training of OpenAI’s fourth-generation model, GPT-4, which now serves premium versions of ChatGPT and Microsoft’s own products.
Given the favourable climatic conditions for much of the year, Iowa proves an efficient choice for training powerful AI systems, especially when contrasted against Microsoft’s data centres in Arizona which consume significantly more water. As a result, experts recommend scheduling AI training in Iowa, as despite no noticeable difference in training, there are stark contrasts in water and energy consumption.
Despite the seemingly efficient usage of resources, data centres can still consume large quantities of water, especially during the summer months. For instance, in July 2022, right before OpenAI completed its training of GPT-4, Microsoft utilised nearly 11.5 million gallons of water for its set of Iowa data centres, representing about 6% of all water used in the district.
In light of these circumstances, the West Des Moines Water Works declared that any future data centre projects from Microsoft must demonstrate and implement technology to significantly reduce peak water usage. In response, Microsoft affirmed that it is actively working with the water works to address their concerns and is committed to reducing its water footprint while meeting its requirements.