Via bloomenergy.com
Most of the power capacity requested for AI data centers will never materialize, and utilities are starting to push back.
The US artificial intelligence industry is supposed to be hungry for power. Developers have submitted requests for 1,066 gigawatts of new electricity capacity to support data center growth. That number is staggering, roughly equivalent to the entire current generating capacity of the US several times over. The catch: most of it is fiction.
Wood Mackenzie, the energy research firm, estimates that only about 28% of that requested capacity is likely to result in anything actually built. The rest falls into a category that grid operators and utility executives have started calling “phantom” projects, speculative or duplicative applications that clog interconnection queues without any realistic path to construction.
How phantom projects work #
The mechanics are straightforward, if a little cynical. A developer with a data center concept submits applications to multiple utilities simultaneously, shopping for the best grid connection terms. Each utility logs the full requested load as a real demand signal. None of them know the others are holding identical applications from the same developer.
The result is a queue stuffed with mirages. Industry estimates put the share of legitimate applications at somewhere between 20% and 30% of all submissions nationwide. On the PJM grid, which serves roughly 65 million people across the mid-Atlantic and Midwest, about half of data center requests may be credible. In Texas, on the ERCOT grid, the legitimacy rate drops to around 14%.
Texas Governor Greg Abbott has directed an audit of ERCOT’s interconnection queue, specifically targeting the nearly 474 gigawatts of proposed data center capacity sitting in the system.
Data centers currently account for roughly 4% to 6% of total US electricity consumption.
Why utilities are raising the drawbridge #
When every project looks real on paper, grid planners have no reliable way to distinguish genuine demand from noise. That creates two distinct risks that compound each other.
The first is overbuilding. If utilities and grid operators take the full queue at face value and invest in transmission infrastructure to serve projected load, they risk constructing billions of dollars worth of power lines and substations for customers who never show up. That cost eventually lands on ratepayers.
The second is underbuilding for the projects that are real. When phantom applications consume interconnection slots and delay the queue, legitimate data center developers get stuck waiting behind speculative filings. A data center that was supposed to come online in 2026 might not connect to the grid until 2029.
Utilities have started responding with structural changes to their application processes. Several are shifting to “first-ready, first-served” frameworks that require meaningful upfront financial commitments before an application is treated as a real demand signal. Higher fees, credit requirements, and collateral deposits are becoming standard filters designed to separate serious developers from queue-stuffers.
What this means for AI infrastructure investment #
Firms financing data center construction need to understand which interconnection queue positions are real and which are placeholders. A queue position held by a developer with no equity commitment and no signed offtake agreement is not an asset.
The broader energy sector faces a forecasting credibility problem. Models built on interconnection queue data will systematically overstate future load growth until utilities clean up their queues. That means demand projections for natural gas peakers, battery storage, and transmission investment may all need downward revision.
Disclosure: This article was edited by Editorial Team. For more information on how we create and review content, see our