Enthusiastic predictions of an artificial intelligence (A.I.) data center construction boom continue to push calls for large amounts of new generation capacity. President Donald Trump has even called the potential for a power shortfall to limit the industry’s growth an “energy emergency.” However, some researchers studying this expansion say there is a risk of over-building new generation. Even the most A.I.-aggressive reports on the topic describe scenarios in which demand could grow much more slowly than headline figures would indicate.

Many utilities now are focused on understanding what their real exposure could be through efforts to determine which proposals for new facilities in their service territories are most likely to move forward. And they’re also considering new rate structures to help ensure A.I. developers bear the brunt of new generation costs, rather than spreading this expense across their broader rate bases.

How much will demand grow?

When organizations such as the Electric Power Research Institute (EPRI) and companies like Schneider Electric and Deloitte make power-demand predictions, those forecasts rarely result in a single figure. Instead, researchers typically provide three figures based on varying assumptions about where the economy will be, how quickly technology will advance, how quickly products will be adopted by businesses and consumers and the level of confidence that any of these scenarios will actually play out. But news reports often only focus on the number most likely to get attention—and, in the case of A.I. power demand, that’s usually the one at the highest end of the scale.

For example, an EPRI report last year, “Powering Intelligence: Analyzing Artificial Intelligence and Data Center Energy Consumption,” hit network news for predicting these facilities could be consuming more than 9% of total U.S. electricity production by 2030. However, the actual study suggested data centers’ proportion of 2030 U.S. demand could range from 4.6% to 9.1%—compared to 4.4% today. (In fairness, the EPRI news release promoting their findings led with the 9.1% figure, which helped land the research in the headlines.) So, while researchers were suggesting A.I. demand could grow by as little as 0.04% to as much as 97% in five years, the idea that it could almost double is what resonated. 

Skeptics see current market softness

Recently, some have begun suggesting a second look at these power-demand projections. No one is saying the A.I. boom isn’t real, but there is some evidence that the industry’s growth might not be on the straight upward trajectory news coverage has suggested. 

For example, earlier this year, a RAND Corp. report estimated global A.I. data center power demand could hit 327 gigawatts (GW) by 2030, based on anticipated growth in A.I. chip production. However, a recent white paper from Schneider Electric sees growth ranging between 16.5 GW and 65.3 GW, with 33.8 GW being a likely target. 

Some players are showing signs of backing off their most aggressive development plans. Since January, Microsoft has cut leases for up to 2 GW of data center capacity in the United States and Europe. Developer Tract recently dropped plans for a 30-building facility near Phoenix, though opposition was cited as the reason. Joe Tsai, chairman of Chinese tech leader Alibaba Group, has warned that he sees the beginning of a bubble in the data center construction industry.

A challenge for forecasters in these situations is that companies such as Microsoft, Alphabet and Amazon often propose more projects than they know will actually be developed, because they know some will be eliminated due to poor power availability or local permitting disputes. This means that building out national estimates based on corporate announcements could seriously overstate actual future power demand.

Catching a breath

This isn’t to say A.I. isn’t going to have an effect on U.S. power supplies—the lowest EPRI scenario, for example, was based, in part, on a scenario with low public uptake of A.I. technology, which seems unlikely. However, a slowdown or pause in the rate of growth could give utilities and their regulators some breathing room to determine how best to meet future demand without forcing other ratepayers to foot the bill for new generation and transmission infrastructure.

Two recent cases illustrate approaches that could become more common. In Ohio, regulators have accepted a proposal by utility AEP for a special data center tariff applicable to new developments adding more than 25 megawatts of demand. Developers now must sign power-supply contracts for up to 12 years, with penalties for pulling out early or using less power than agreed. 

In Virginia, Dominion and smaller Rappahannock Electric Coop. have designed plans that involve direct power procurement and supply contracts for facilities with financial buffers for their customer base.

Wongstorn / stock.adobe.com