AI game changer hits the grid By Peter Behr | 08/13/2024 06:44 AM EDT

Specially trained technicians have been unhooking live, uninsulated high-voltage lines that are crackling with 230,000 deadly volts of electricity. They’re replacing them with advanced cables that boost power delivery by 50 percent.

You wouldn’t do it this way unless you had to, says Matthew Gardner, vice president for transmission for Dominion Electric, the state’s largest utility, which owns and operates the lines.

Dominion says it has to.

Here, on the outskirts of Dulles International Airport, the largest collection of data centers in the United States is drawing power from Dominion’s lines. The eruption of generative artificial intelligence traffic and expectations for more are sending electricity use soaring.

Dominion chose the rare “hot line” replacement. It simply couldn’t shut down existing power lines that feed 24/7 demand for power from the growing hub of data centers. “It’s absolutely essential that these projects are completed to serve the future growth that’s coming our way,” Gardner said in an interview.

It took Dominion 115 years from its founding to reach its current level of power delivery. “The pace of growth is so rapid, driven by data centers, that we’re on pace to double our system load in the next 15 years,” Gardner said.

Here and at other data center clusters around the U.S., a grid industry struggling with a historic shift from fossil fuel generation to renewable energy is suddenly facing predictions of unprecedented demand because of this next phase of America’s digital economy, with no overall game plan to go on.

At the moment, the U.S. grid industry has slipped well off pace to meet President Joe Biden’s zero-carbon generation goals, required to lessen the risk of catastrophic climate events. The 2024 presidential race stands as a referendum on climate policy, given the scale of federal spending on clean energy under Biden. The AI-driven need for energy poses a new barrier to cutting carbon emissions if utilities rely primarily on gas- and coal-fired generation to meet the demand, according to recent studies.

Arshad Mansoor, chief executive of the Electric Power Research Institute (EPRI), said utilities, grid planners, data center operators and their customers should be working together to develop clean energy. 

In a report to Secretary of Energy Jennifer Granholm’s energy advisory board, EPRI urged the nation to “transition to carbon-free electricity sources for data center operations and low carbon technologies for backup power.” 

Data center vs. climate change?

Meeting power needs without losing the fight against global warming has suddenly changed the planning and politics for the U.S grid.

To illustrate digital technology’s ravenous appetite for power, EPRI noted that simple searches today on a laptop browser consume about one-third of a watt of electricity. But advanced AI technology and high-powered computing are training software algorithms to create answers from vast data files, with skyrocketing increases in power needs.

A complex generative AI question over ChatGPT today would require 3 watts, 10 times the simple search amount, EPRI’s report concluded. A Google search with similar generative AI capability could require 9 watts, it added.

Depending on how fast AI expands, data center electricity demands could mushroom from about 4 percent of U.S. grid output to over 9 percent in just six years, EPRI said.

“Nobody has a good understanding of how generative AI will impact every aspect of the society,” Mansoor said in an interview. “And so anything we say now, most likely will be wrong in six months,” he added.

Energy & Environmental Economics (E3), a San Francisco energy research firm, gathered 13 analyses of how much electricity output may have to grow to keep up with data center expansion, under different assumptions.

The projections range from 20,000 megawatts to 100,000 MW of new capacity by 2030. There are plenty of unknowns, including just how “smart” generative AI becomes, how much efficiency can be built into AI chips and whether a strained electric grid can actually deliver the power data centers require.

“There’s still a ton of uncertainty,” particularly concerning the highest forecasts, said E3 senior partner Kushal Patel, a report co-author. “It’s probably not going to be more certain in the future,” he added.

If some utilities wind up overbuilding infrastructure to serve data centers, consumers could be stuck with higher costs, warned Moody’s Investors Service in a recent research report.

The 100,000 MW figure is roughly comparable to duplicating the generation capacity of the entire 94 U.S. nuclear reactors in six years.

While the impact is unmeasured, growth is coming because the potential profits from AI applications has created a sprint for the lead among hyperscale internet cloud infrastructure companies, led by Amazon Web Services, Microsoft, Google, Meta, Apple and TikTok, said Phill Lawson-Shanks, chief innovation officer at Aligned Data Centers, a Texas-based data center builder. 

“I hesitate to use that term, but there’s almost an arms race,” he said.

“They are saying, ‘OK, I have no idea how much, but I know I need power,’” EPRI’s Mansoor said. “’I need to build as much as I can.’”

The cost of new power

E3 pointed to the health care industry to describe the potential scale of artificial intelligence power demand.

There are roughly 32,000 radiologists in the U.S. with a midrange annual wage of $354,000, the firm estimated. If AI could learn to accurately analyze X-ray, ultrasound and MRI images, and replace just 10 percent of those radiologists, the revenue from that substitution could be more than $1 billion a year, according to E3 researchers.

That would pay for a lot of energy-hungry supercomputers.

With long lead times to connect wind and solar to the power grid and a split between the two political parties on federal transmission policy, the quickest source of new power for data centers is from gas-fired generation, E3 noted.

Officials in North Dakota see prospects for a booming data center business — based on discussions with major technology companies — as an opportunity to use some of the natural gas coming off the region’s prolific oil fields. North Dakota Department of Commerce Commissioner Josh Teigen says state officials are looking for ways to develop more gas-fired generation to power data centers and sell surplus electricity to the grid.

Grid operators and utilities are postponing some fossil fuel plant retirements citing threats to electricity reliability. “These rollbacks clash with the customers’ environmental goals and investments, and they threaten state and utility emission reduction targets,” E3 analysts said.

According to E3, delaying the retirement of a typical 1,000-MW coal power plant for a single year, for example, puts 3.8 million metric tons of CO2 into the atmosphere. That would offset the carbon-saving effects of 3,000 MW of utility-scale solar installations, E3 calculated.

What can be done?

As an immediate step, utilities need to make greater use of advanced transmission cables and “grid enhancing technologies” like dynamic line rating sensors that tell grid operators when more power can be moved without overheating cables, Gardner said.

EPRI analysts suggested data center operators could sync AI operations to times of the day or night when renewable power is most plentiful. Or large AI customers could help finance pilot installations of advanced 24/7 carbon-free power sources.

At its July 30 meeting, the advisory board to the secretary of Energy urged utilities, regulators, and data center users and builders to work together to speed up development of wind, solar and batteries. Longer-term technologies include small modular reactors (SMRs), long-term grid batteries, hydrogen hubs, fuel cells and underground storage of fossil plant carbon emissions.

Andy Bochman, senior grid strategist for the Idaho National Laboratory, emphasized small nuclear reactors.

“Coal and gas plants can then be replaced by similar-sized SMR installations on a one-to-one basis,” Bochman said. “This must be complemented by expanding wind and solar and storage on microgrids as quickly as possible.”

“As an industry, we will be forced to look at self-generation until the grid can catch up,” said Lawton-Shanks of Aligned Data Centers.

The hyperscale operators have huge financial resources, EPRI’s Mansoor said. They can partner with utilities to build pilot generation plants at data centers that could remove some of the financial risk of backing emerging technology like hydrogen or carbon capture.

“We can have 10 SMRs and 10 carbon capture and storage projects working by 2030,” Mansoor said.

There could be 10 carbon capture and storage installations at natural gas and coal plants, he continued.

Will the big data center users invest? “They’re talking,” Mansoor said.