This coverage is made possible through a partnership between WABE and Grist, a nonprofit environmental media organization.
An expansive former row-crop farm in Lee County, in southwest Georgia, is now home to a flock of sheep with a job to do: eat grass.
“My rule of thumb is if I can walk out here among the panels and still see my boots, then that’s effective grazing management,” said site manager Tyler Huber on a recent visit to the farm.
Shiny solar panels cover this land, providing shade and shelter for the sheep. The flock, in turn, keeps the grass chewed or trampled so it doesn’t interfere with the production of energy.
This modern pastoral scene serves a high-tech purpose: powering Meta’s data center complex in Social Circle, east of Atlanta.
When the company, then known as Facebook, announced its Georgia data center in 2018, it promised the facility would be totally powered by solar energy, part of the tech giant’s overall goal of net-zero carbon emissions by 2030. The first of the vast solar fields followed, and have only expanded since.
“We now have 11 projects with them in the state of Georgia,” said Matt Beasley, chief commercial officer of Silicon Ranch, the solar company that partnered with Meta to build this site. “We’ve built nine of them for them already.”
But as a smattering of data centers like Meta’s has grown into a bona fide boom industry in Georgia over the last seven years, solar isn’t the only new energy getting built.
State regulators last year greenlit new gas- and diesel-fired turbines to meet spiking energy demand that Georgia Power said comes mostly from data centers. Now, the utility is asking to keep coal plants open longer than planned for the same reason. Electric co-ops, too, including Meta’s power supplier, are planning to add non-renewable resources to power data centers and other large power users.
These utilities are still adding solar and other renewables. But the influx of fossil fuels, mostly to serve data centers, is a major shift that could have serious consequences for climate change.
“These numbers became so large, the tools we had or the parts of the solution space we were exploring wouldn’t work anymore,”
University of Pennsylvania professor Benjamin Lee said computer scientists were close to a “carbon-free computing” reality before generative AI moved the goalpost.
Scientists agree on the need to cut greenhouse gas emissions drastically and immediately to combat climate change. That’s the reason tech companies and utilities alike, including Georgia Power’s parent company, made their net-zero promises in the first place. So climate advocates argue that rushing to power data centers by any means necessary – including lots of fossil fuels – is a step in the wrong direction.
“[Tech companies have] gone from promising to be carbon neutral to really just blowing their goals out of the water. And this is uniquely due to AI and generative AI in particular,” said Alex Hanna, director of research at the Distributed AI Research Institute and author of the book “The AI Con.”
All computing takes energy, whether it’s sending an email or storing a document in the cloud. But generative AI like ChatGPT, the kind that can produce lengthy text or even images and videos, requires much more computation. The AI tool converts every prompt into billions of multiplication problems, then solves them to produce its output.
“That takes a lot of energy just because it’s a pretty compute-intensive operation,” Hanna said.
The computing is so intensive, in fact, that it requires special computer chips called Graphics Processing Units. They use a lot more energy than typical computers. And packing lots of them into new, bigger data centers as tech companies compete in an AI arms race is drastically changing the landscape of energy and computing.
University of Pennsylvania professor Benjamin Lee studies computer systems and did research for Meta on powering its work. Before AI, according to Lee, the goal of “carbon-free compute” was well within reach.
“We, as in computer scientists, were optimistic about the prospects of going to net zero, primarily through massive investments in wind and solar,” he said.
Lee and his colleagues felt net zero was attainable. It would require battery storage and moving around the timing of computing work to account for the variability in those renewable energy sources – the fact that there’s no sun overnight, for instance. In 2020, Google even announced a new computing platform that attuned its usage to wind and solar production.
But at the time, the data centers in question needed 25-50 megawatts of power, maybe 100 MW at the most intensive, Lee said.
Now, with the advent of generative AI, new data centers need upwards of 500 or 1,000 megawatts, according to Lee – five or ten times as much energy, equivalent to half a million homes.
“These numbers became so large, the tools we had or the parts of the solution space we were exploring wouldn’t work anymore,” Lee said.
This new demand has sent utilities scrambling, and not only in Georgia. Utilities across the country are rushing to meet the largest increase in demand, known in the industry as load growth, that they’ve seen in decades.
“I think that we’re suffering from different timescales. I think there’s a really urgent need to deploy infrastructure for generative AI,” said Lee. “Investments in energy infrastructure and upgrades to the grid take much longer.”
“[Tech companies have] gone from promising to be carbon neutral to really just blowing their goals out of the water. And this is uniquely due to AI and generative AI in particular,”
Alex Hanna, director of research at the Distributed AI Research Institute and author of the book “The AI Con.”
That’s why power companies are now saying they need to use fossil fuels to fill the need.
In Georgia and many other states, it’s often faster to expand existing gas plants, as Georgia Power is doing, or extend the operation of existing coal plants, as many utilities, including Georgia Power, are proposing, than to go through the lengthy bidding and site selection process of building new solar or wind.
While these utilities are still building renewables as well, they contend that those will take longer, and fossil fuel expansions and extensions are essential to meeting the immediate demand spikes expected over the next few years. (In the mostly deregulated Texas energy market, solar and wind power are expanding rapidly to meet the rising demand because it’s often faster and cheaper when other factors are removed from the equation.)
These shifts have prompted some major tech companies to extend the target dates for their carbon-free promises, though Meta, at least, said in an email to WABE that its goal remains unchanged.
Some companies are exploring new ways to make their own energy on-site, with hyper-efficient fuel cells that run on gas now but could one day use hydrogen, or even with small nuclear reactors, but much of that technology is still years or even decades from full implementation.
That means that in the short term, the new large-scale data centers cropping up across the country are largely reliant on whatever energy utilities can provide, however they can make it.
But what happens if the data center rush slows down? What if the power demand never gets as high as utility forecasts, or it comes on more slowly?
There are already some signs that could happen. Some consumer and environmental advocates have questioned Georgia Power’s forecasts since the company filed its emergency request for more power in 2023, arguing they predicted too much demand too quickly.
AI could get more efficient, Lee said, if programmers develop more specialized models instead of the all-purpose ones currently out there, though that shift hasn’t happened yet.
Or AI could get less popular. Some AI critics, including Hanna, argue there are already signs that generative AI is a bubble that could burst any day.
“The productivity gains that have been promised by AI haven’t panned out,” she said. “And so economic forces may help slow this kind of data center rush.”
Lee and other energy experts argue that it doesn’t matter if the AI data center boom cools off, because even if data centers don’t use it, all that energy will be needed eventually. Electric vehicles, switching appliances from gas to electric and the return of US manufacturing all require electricity, too, meaning demand is going to increase over time.
But that demand is coming more slowly than the AI rush, meaning there’s more time to build out renewables, invest in efficiency measures and explore other approaches. In the meantime, critics worry that utilities are making long-term fossil fuel investments to meet a short-term demand spike from data centers.
“It’s a commitment of resources that’ll last at least 46 years into the future, is what Georgia Power’s projecting, and possibly longer than that,” said Bob Sherrier, senior attorney at the Southern Environmental Law Center, about Georgia Power’s new gas turbines last year. “And we know that we need to transition [away from fossil fuels] much faster than 40 years.”
This story is part of the WABE Newsroom series Server South: What’s powering Atlanta’s data center growth — and what it means for you.
