Electricity Demand will Rise, But How Much?

“It’s now reasonable to project that half of the electric grid will be powering the digital-Internet economy within the next decade.” – Peter Huber and Mark Mills, Forbes, “Dig More Coal, the PCs are Coming”, Forbes, May 31, 1999

There is general agreement that electricity demand is going to grow as we electrify home heating, transportation and industrial sectors.
In somewhat of a curve ball, there’s been a kerfuffle lately about unexpectedly rapid demand growth from Data centers, especially due to the breakout of AI applications.
A number of utilities have revealed plans to meet that demand by building a number of gas turbine generators, and that’s sparked some needed discussion. The controversy is especially intense in the Southeastern US, where Georgia Power, among others, has released a fossil-heavy IRP (Integrated Resource Plan).

Robinson Meyer in HeatMap:

Citing the return of load growth, utilities across the South are trying to go around normal regulatory channels and build a slew of new natural-gas-burning power plants. Across at least six states, utilities have already won — or are trying to win — permission from local governments to fast-track more than 10,000 megawatts of new gas-fired power plants so that they can meet the surge in demand.

These requests have popped up across the region, pushed by vertically integrated monopoly power companies. Georgia Power won a tentative agreement to build 1,400 new megawatts of gas capacity, Canary reported. In the Carolinas, Duke Energy has asked to build 9,000 megawatts of new gas capacity, triple what it previously requested. The Tennessee Valley Authority has plans to add 6,600 megawatts of new capacity to its grid.
–
Utilities compete against each other to secure big-name customers — much like local governments compete with sweetheart tax deals — so when a utility asks regulators to build more capacity, it doesn’t reveal where potential power demand is coming from. (In other words, it doesn’t reveal who it believes will eventually buy that power.) A company might float plans to build the same data center or factory in multiple states to shop around for the best rates, which means the same underlying gigawatts of demand may be appearing in several different utilities’ resource plans at the same time. In other words, utilities are unlikely to actually see all of the demand they’re now projecting.

Even if we did know exactly how many gigawatts of new demand each utility would see, it’s almost impossible to say how much of it is coming from AI. Utilities don’t say how much of their future projected power demand will come from planned factories versus data centers. Nor do they say what each data center does and whether it trains AI (or mines Bitcoin, which remains a far bigger energy suck).

At the same time, we might have been led astray by overly dire predictions of AI’s energy use. Jonathan Koomey, a researcher who studies how the internet and data centers use energy (and the namesake of Koomey’s Law) told me that many estimates of Nvidia’s most important AI chips assume that their energy use is the same as their advertised “rated” power. In reality, Nvidia chips probably use half of that amount, he said, because chipmakers engineer their chips to withstand more electricity than is necessary for safety reasons. 

And this is just the current generation of chips: Nvidia’s next generation of AI-training chips, called “Blackwell,” use 25 times less energy to do the same amount of computation as the previous generation of chips. 

Koomey helped defuse the last panic over energy use by showing that the estimates Huber and Mills relied on were wildly incorrect. Estimates now suggest that the internet used less than 1% of total U.S. electricity by the late 1990s, not 13% as they claimed. Those percentages stayed roughly the samethrough 2008, he later found, even as data centers grew and computers proliferated across the economy. That’s the same year, remember, that Huber and Mills predicted that the internet would consume half of American energy.

These bad predictions were extremely convenient. Mills was a scientific advisor to the Greening Earth Society, a fossil-fuel-industry-funded group that alleged carbon dioxide pollution would actually improve the global environment. He aimed to show that climate and environmental policy would conflict with the continued growth of the internet.

“Many electricity policy proposals are on a collision course with demand forces,” Mills said in a Greening Earth press release at the time. “While many environmentalists want to substantially reduce coal use in making electricity, there is no chance of meeting future economically-driven and Internet-accelerated electric demand without retaining and expanding the coal component.” Hence the headline of the Forbes piece: “The PCs are coming — Dig more coal.”

What makes today’s AI-induced fear frenzy different from 1999 is that the alarmed projections are not just coming from businesses and banks like Morgan Stanley, but from environmentalists like Friends of the Earth. Yet neither their estimates of near-term, AI-driven power shortages — nor the analysis from Morgan Stanley that U.S. data-center use could soon triple within a year — make sense given what we know about data centers, Koomey said. It is not logistically possible to triple data centers’ electricity use in one year. “There just aren’t enough people to build data centers, and it takes longer than a year to build a new data center anyway,” he said. “There aren’t enough generators, there aren’t enough transformers — the backlog for some equipment is 24 months. It’s a supply chain constraint.”

Look around and you might notice that we have many more servers and computers today than we did in 1999 — not to mention smartphones and tablets, which didn’t even exist then — and yet computing doesn’t devour half of American energy. It doesn’t even get close. Today, computers use 1% to 4% of total U.S. power demand, depending on which estimate you trust. That’s about the same share of total U.S. electricity demand that they used in the late 1990s and mid-2000s.

It may well be that AI devours more energy in years to come, but utilities probably do not need to deal with it by building more gas. They could install more batteries, build new power lines, or even pay some customers to reduce their electricity usage during certain peak events, such as cold winter storms. 

There are some places where AI-driven energy demand could be a problem — Koomey cited Ireland and Loudon County, Virginia, as two epicenters. But even there, building more natural gas is not the sole way to cope with load growth.

“The problem with this debate is everybody is kind of right,” Daniel Tait, who researches Southern utilities for the Energy and Policy Institute, a consumer watchdog, told me. “Yes, AI will increase load a little bit, but probably not as much as you think. Yes, load is growing, but maybe not as much as you say. Yes, we do need to build stuff, but maybe not the stuff that you want.”

Latitude Media:

When Georgia Power Company filed an updated integrated resource plan last fall, the utility drastically increased its near-term load growth predictions, based in large part on expected growth from data centers. To maintain the state’s ability to “attract and retain businesses,” the IRP said, Georgia power must take “immediate action” to meet capacity needs by the end of 2025.

In late March the utility reached a preliminary agreement with the state’s Public Service Commission to take that action, meeting the immense load gap with three new natural gas turbines, expected to provide up to 1,400 megawatts of power, a 150MW residential demand response program, and a handful of fossil fuel heavy power purchase agreements.

The ensuing debate over that plan of action is currently playing out in the Public Service Commission’s docket, where customers ranging from Big Tech to federal agencies are seeking increased transparency on Georgia Power’s heavily redacted resource planning filings, and criticizing the utility’s approach to load forecasting.

• The top line: In comments filed with the Georgia Public Service Commission early last week, Microsoft challenged GPC’s modeling methods, outlining concerns that they both undervalue renewable energy and overestimate data center load.
• The market grounding: Microsoft owns and operates three data center campuses in Georgia Power’s service territory, and that footprint is growing — the tech giant has additional projects expected to come online near Atlanta in late 2025, and has also identified two new potential campuses in the state. ‍
• The current take: In its Commission filing, Microsoft made it clear that the company doesn’t underestimate the looming growth of data centers — “Microsoft agrees that most of the load growth associated with data centers in the 2023 IRP Update is real and that GPC does require additional capacity,” the company said. “However, Microsoft has concerns with GPC’s approach potentially leading to over-forecasting near-term load (through 2030) and procuring excessive, carbon-intensive generation.”

One of Microsoft’s chief concerns is that Georgia Power’s approach to forecasting near-term data center load growth differs from that of its neighbors Dominion, Duke, and APS in a key way: the utility includes projects that are considering Georgia as a location, but haven’t yet committed any funds or made a final selection. 

As data center customers look to secure power from a utility, they go through different contracting stages, with progressive increases in commitment, Microsoft explained. An initial request for electric service, for example, may have no financial commitment, while requests from system impact studies through to final agreements for energization have significant commitments.

By not constraining resource requirement predictions to projects that have already selected GPC, the update may overestimate demand, Microsoft said.

Other utilities in the region, the company added, base their new load forecasts primarily on “known projects that have made varying levels of financial commitment.”

Georgia Power’s projection of resource requirements between now and 2030 should be based only on “mature projects with firm commitments,” the company added. And, the utility should provide “greater transparency to stakeholders regarding its large load forecasting methodology and the underlying data used to support it.”