- Utah residents express concerns over data centers' environmental impact and resource use.
- A Deseret News-Hinckley Institute poll shows over half oppose Box Elder project.
- Experts challenge exaggerated water consumption claims, citing more efficient AI models today.
SALT LAKE CITY — Last week in Springville, Utah, a group of young moms sat in Memorial Park, watching their toddlers play. In the high-70-degree weather, atop the park's green grass, the conversation turned to data centers.
One mom had mused how she was going to keep her new infant cool this summer, and Matty Shmitz responded with worry that the approved data center in Box Elder County was going to make the summer's heat worse.
In a later conversation with the Deseret News, Shmitz described the data center as a "constant numb pain in the back of my brain."
Through TikTok, Instagram and news reports, Shmitz said she has developed serious concerns about the data center.
Will it use an inordinate amount of water? Where will that water come from? Will it drive up the cost of electricity and increase the temperature? And at the root of it, why are we risking Utah's beautiful land to fuel artificial intelligence when it seems like artificial intelligence will hurt American society?
The Stratos Project is one of more than 1,800 U.S.-based data centers in various stages of development. Once built, it will join more than 3,100 data centers already in operation across the United States, the earliest of which were built in the 1990s.
Backed by "Shark Tank" investor Kevin O'Leary, the project as planned is massive compared to other data centers in the U.S. At full buildout, it will have a power capacity of 9 gigawatts — roughly 90 to 225 times larger than the average data center, which uses 40 to 100 megawatts.
These existing data centers power the digital infrastructure behind the internet.
If you store images in the cloud, buy anything with a debit card, ask Siri a question, stream a show on Netflix, or load directions on Google Maps, you're using a data center.
But data centers' everyday utility has been lost in a haze of anxiety about new proposals. Shmitz is not alone in her concerns. More than half of Utah's residents say they oppose the data center in Box Elder County, according to a Deseret News-Hinckley Institute poll conducted in mid-May.
So it makes sense to ask: Are concerns about data centers well-founded?
The water consumption confusion
In 2024, The Washington Post released a report claiming that a 100-word email written by ChatGPT consumes an entire bottle of water, or 519 milliliters.
The article proceeded to scale up that number: One ChatGPT-generated email a week for a year consumes 27 liters of water; one ChatGPT-generated email a week for a year from 10% of the U.S. population (16 million people) requires more than 435 million liters — "equal to the water consumed by all Rhode Island households for 1.5 days."
Readers were then left to fill in the blanks about usage. In 2025, ChatGPT was queried about 2.5 billion times a day.
Based on the Post's measurements, 2.5 billion queries (if they were all the size of an email) would require nearly 343 million gallons of water — more than 520 Olympic-sized swimming pools — a day.
When Andy Masley, a former physics teacher turned writer, saw this report, it didn't sit well with him. So he started looking into the article's methodology, then reached out to the researcher tapped for the calculation, Shaolei Ren, an associate professor of electrical and computer engineering at the University of California, Riverside.
"The majority of that bottle of water, even in (Ren's) own estimates, isn't actually used in the data center itself. It's used in off-site power plants to generate the electricity," Masley told the Deseret News.
About half of the water-bottle estimate comes from evaporation off lakes dammed by hydro plants to generate hydropower.
"I've been emailing back and forth with (Ren), who made this estimate, and he agrees with me that the actual estimate is way, way smaller," Masley said.
Masley said more current estimates, from experts like EcoLogits, find that individual prompts cost between 1 and 10 milliliters of water, about 99% smaller than the estimate published by the Washington Post. The amount of water actually used in the data center is smaller still — about 0.2-2 milliliters.
In a conversation with the Deseret News, Ren said the Washington Post's report should not be considered an accurate measure of today's artificial intelligence water demands.
"We cannot just use a number from two years ago to describe today's system. And we cannot use specific models' results to generalize other models. I could give you a super-efficient AI model that uses almost zero resources, or I could give you a very large model that can — it's just never correct to say, 'AI uses this much water,'" Ren said.
Even describing ChatGPT as a single model with universal resource needs is incorrect, he continued. "ChatGPT is not a single model. Nobody knows what exactly they're doing under the hood. So I don't know the resource efficiency," Ren said.
A chatbot's efficiency depends on a spread of factors, he explained. To make AI more efficient, engineers optimize its hardware, algorithms and how workloads are scheduled. Its efficiency can change dramatically if it's in thinking mode, reasoning mode or plain text output, as well as a myriad of other ways.
