Headlines continue to highlight the need for a massive growth in generation capacity to support the booming artificial intelligence (A.I.) market. Coal and nuclear plants previously set to retire are being kept running as utilities and independent power producers rush to get new natural gas generators online. But they face a question amid this scramble: are they overbuilding?

This isn’t a hypothetical concern. The U.S. Energy Information Administration (EIA) recently lowered its 2026 load-growth forecast due to a slowdown in aggressive data center development forecasts. While the drop is largely driven by shifts in one region, it still illustrates the challenge of matching 30- to 40-year investments against an A.I. market that’s changing by the day.

Oversupply—round one

Looking back a quarter century or so, we can see parallels in the downturn the power industry faced following the dot-com bust of 2001. In the late 1990s, utility planners and regulators believed a tremendous boost in electricity demand was just over the horizon, and they had good reason. In 1999, the personal computer market was booming, online retail sales hit an estimated $15 billion and Amazon gained its 10 millionth customer. Investors saw large potential in this new startup economy in which companies valued growth over profit. Telecommunications companies alone spent more than $500 billion building new fiber networks between 1996 and 2001.

Not surprisingly, electricity demand growth projections became bullish—one influential 1999 report from the Greening Earth Society (a group funded by the Western Fuels Association, an organization of coal-burning utilities) predicted that, by 2010, up to half the U.S. grid’s output could be driven by “the direct and indirect needs of the internet.” 

It turned out, though, that earlier warnings of “irrational exuberance” proved prophetic. Within two years, the dot-com bust was well underway. The power industry, especially independent developers, lost billions in the early 2000s as predicted demand failed to materialize. The resulting generation oversupply drove power prices below companies’ ability to make interest payments on their capital investments. The result was a rash of bankruptcies among merchant power operators.

Today’s situation is different though. The electric utility industry was then in the early days of deregulation. In many states, utilities were facing new competition for the first time from independent power producers. New rules requiring them to buy electricity from the least expensive supplier provided strong market incentives for investors to pour money into natural gas generation, which offered cheaper production than coal or nuclear options. So, while the resulting over-development was stoked by highly optimistic demand predictions, this regulatory shift created a boom of its own that contributed to driving down wholesale power costs below what companies needed to maintain high debt loads.

Today’s situation

This doesn’t mean that concerns that we’re in the middle of a similar boom-bust cycle today are irrational, though, as recent reports illustrate. Grid planners now are facing confusing data that can shift at a rapid pace. The EIA’s recent revision, for example, which pushed U.S. demand growth projections for 2026 down to 1.7% from the previous 3%, is largely due to market changes in Texas, where planners now expect demand to grow by 9.6% instead of the previous 15.7%. That’s a 38% drop in anticipated demand in just one month.

And Texas isn’t alone. A September report from the World Resources Institute looked at a range of projections for U.S. demand growth out to 2030 that were developed by national labs, top-tier financial companies and other respected sources. They found a difference of more than 850 terawatt-hours per year between the lowest and highest figures—for comparison, that figure represents about 20% of all U.S. generation in 2023. 

The report’s authors noted several variables contributing to the difficulty of developing accurate power ­demand projections. Topping the list is the challenge posed by A.I. developers submitting speculative requests for utility connections for projects that won’t get built, as well as price-shopping the same project to multiple utilities, resulting in double counting.

Beyond this, planners also face shortages of power and computing equipment that could push out today’s aggressive construction timelines, making new generation less urgent. Additionally, each new generation of A.I. chips is becoming more efficient—with Stanford University recently documenting 40% annual efficiency improvements between 2022 and 2024, which also could limit future power needs. 

Finally, WRI also noted the possibility that the market is simply growing too quickly, citing Operation Stargate, announced January 2025 by President Trump, as an example. The $500 billion initiative backed by OpenAI, SoftBank and Oracle has struggled to meet its aggressive timeline, with other prospective investors questioning its leaders’ financial assumptions. Will the current rate of growth prove prescient? Or are we just creating a glut of generating assets with an accounting lifespan of 20 years or more?

obe.com/Erfindadwi