This is the second article in our four-part series on data centers and electricity demand. Read part 1.
There’s no question that data centers are about to cause U.S. electricity demand to spike. What remains unclear is by how much.
Right now, there are few credible answers. Just a lot of uncertainty — and “a lot of hype,” according to Jonathan Koomey, an expert on the relationship between computing and energy use. (Koomey has even had a general rule about the subject named after him.) This lack of clarity around data center power requires that utilities, regulators, and policymakers take care when making choices.
Utilities in major data center markets are under pressure to spend billions of dollars on infrastructure to serve surging electricity demand. The problem, Koomey said, is that many of these utilities don’t really know which data centers will actually get built and where — or how much electricity they’ll end up needing. Rushing into these decisions without this information could be a recipe for disaster, both for utility customers and the climate.
Those worries are outlined in a recent report co-authored by Koomey along with Tanya Das, director of AI and energy technology policy at the Bipartisan Policy Center, and Zachary Schmidt, a senior researcher at Koomey Analytics. The goal, they write, “is not to dismiss concerns” about rising electricity demand. Rather, they urge utilities, regulators, policymakers, and investors to “investigate claims of rapid new electricity demand growth” using “the latest and most accurate data and models.”
Several uncertainties make it hard for utilities to plan new power plants or grid infrastructure to serve these data centers, most of which are meant to power the AI ambitions of major tech firms.
AI could, for example, become vastly more energy-efficient in the coming years. As evidence, the report points to the announcement from Chinese firm DeepSeek that it replicated the performance of leading U.S.-based AI systems at a fraction of the cost and energy consumption. The news sparked a steep sell-off in tech and energy stocks that had been buoyed throughout 2024 on expectations of AI growth.
It’s also hard to figure out whose data is trustworthy.
Companies like Amazon, Google, Meta, Microsoft, OpenAI, Oracle, and xAI each have estimates of how much their demand will balloon as they vie for AI leadership. Analysts also have forecasts, but those vary widely based on their assumptions about factors ranging from future computing efficiency to manufacturing capacity for AI chips and servers. Meanwhile, utility data is muddled by the fact that data center developers often surreptitiously apply for interconnection in several areas at once to find the best deal.
These uncertainties make it nearly impossible for utilities to gauge the reality of the situation, and yet many are rushing to expand their fleets of fossil-fuel power plants anyway. Nationwide, utilities are planning to build or extend the life of nearly 20 gigawatts’ worth of gas plants as well as delaying retirements of aging coal plants.
If utilities build new power plants to serve proposed data centers that never materialize, other utility customers, from small businesses to households, will be left paying for that infrastructure. And utilities will have spent billions in ratepayer funds to construct those unnecessary power plants, which will emit planet-warming greenhouse gases for years to come, undermining climate goals.
“People make consequential mistakes when they don’t understand what’s going on,” Koomey said.
Some utilities and states are moving to improve the predictability of data center demand where they can. The more reliable the demand data, the more likely that utilities will build only the infrastructure that’s needed.
A “wild west” of data center hot spots
In recent years, the country’s data center hot spots have become a “wild west,” said Allison Clements, who served on the Federal Energy Regulatory Commission from 2020 to 2024. “There’s no kind of source of truth in any one of these clusters on how much power is ultimately going to be needed,” she said during a November webinar on U.S. transmission grid challenges, hosted by trade group Americans for a Clean Energy Grid. “The utilities are kind of blown away by the numbers.”
A December report from consultancy Grid Strategies tracked enormous load-forecast growth in data center hot spots, from northern Virginia’s “Data Center Alley,” the world’s densest data center hub, to newer boom markets in Georgia and Texas.
Koomey highlighted one big challenge facing utilities and regulators trying to interpret these forecasts: the significant number of duplicate proposals they contain.
“The data center people are shopping these projects around, and maybe they approach five or more utilities. They’re only going to build one data center,” he explained. “But if all five utilities think that interest is going to lead to a data center, they’re going to build way more capacity than is needed.”
It’s hard to sort out where this “shopping” is happening. Tech companies and data center developers are secretive about these scouting expeditions, and utilities don’t share them with one another or the public at large. National or regional tracking could help, but it doesn’t exist in a publicly available form, Koomey said.
To make things more complicated, local forecasts are also flooded with speculative interconnection requests from developers with land and access to grid power, with or without a solid partnership or agreement in place.
“There isn’t enough power to provide to all of those facilities. But it’s a bit of a gold rush right now,” said Mario Sawaya, a vice president and the global head of data centers and technology at AECOM, a global engineering and construction firm that works with data center developers.
That puts utilities in a tough position. They can overbuild expensive energy infrastructure and risk whiffing on climate goals while burdening customers with unnecessary costs, or underbuild and miss out on a once-in-a-lifetime economic opportunity for them and for their community.
Getting a grip on what’s coming onto the grid
In the face of these risks, some utilities are trying to get better at separating viable projects from speculative ones, a necessity for dealing with the onslaught of new demand.
Utilities and regulators are used to planning for housing developments, factories, and other new electricity customers that take several years to move from concept to reality. A data center using the equivalent of a small city’s power supply can be built in about a year. Meanwhile, major transmission grid projects can take a decade or more to complete, and large power plants take three to five years to move through permitting, approval, procurement, and construction.
Given the mismatch in timescales, “the solution is talking to each other early enough before it becomes a crisis,” said Michelle Blaise, AECOM’s senior vice president of global grid modernization. “Right now we’re managing crises.”
Koomey, Das, and Schmidt highlight work underway on this front in their February report: “Utilities are collecting better data, tightening criteria about how to ‘count’ projects in the pipeline, and assigning probabilities to projects at different stages of development. These changes are welcome and should help reduce uncertainty in forecasts going forward.”
Some utilities are still failing to better screen their load forecasts, however — and tech giants with robust clean energy goals such as Amazon, Google, and Microsoft are speaking up about it. Last year, Microsoft challenged Georgia Power on the grounds that the utility’s approach is “potentially leading to over-forecasting near-term load” and “procuring excessive, carbon-intensive generation” to handle it, partly by including “projects that are still undecided on location.” Microsoft contrasted Georgia Power’s approach to other utilities’ policies of basing forecasts on “known projects that have made various levels of financial commitment.”
Utilities also have incentives to inflate load forecasts. In most parts of the country, they earn guaranteed profits for money spent on power plants, grid expansions, and other capital infrastructure.
As a matter of fact, many utilities have routinely over-forecasted load growth during the past decade, when actual electricity demand has remained flat or increased only modestly. But today’s data center boom represents a different kind of problem, Blaise said — utilities “could build the infrastructure, but not as fast as data centers need it.”
Making data centers pay — and promise to stick around
In the face of this gap between what’s being demanded of them and what can be built in time, some utilities are requiring that data centers and other large new customers prove they’re serious by putting skin in the game.
The most prominent efforts on this front are at utility American Electric Power. Over the past year, AEP utilities serving Ohio as well as Indiana and Michigan have proposed new tariffs — rules and rates for utility customers — to deal with the billions of dollars of data center investments now flooding into those states.
