Ghost Data Centers Occupy Grid Capacity as U.S. Utilities Move to Screen Projects
nashnova research
Only 23% of data-center grid applications are rated high-confidence; speculative "ghost" projects are hogging planned capacity. Utilities now demand binding financial commitments to screen them out — a shift that could slow AI's real buildout pace.
What are "ghost data centers," and why do they matter?
Data-center developers routinely file power applications with multiple utilities at the same time. A single project's demand gets counted two to four times across different grids.
This means → the demand utilities see on paper far exceeds real demand. Projects that claimed capacity but never broke ground are called "ghost data centers."
Wood Mackenzie, an energy research firm, found that only 23% of current data-center power applications qualify as high-confidence. The rest are speculative.
What did Exelon do, and did it work?
Exelon, a major US utility covering the Mid-Atlantic and Midwest, saw individual project applications leap from a typical 50 MW to over 1,000 MW — roughly the output of a large nuclear plant.
The company first assessed projects one by one, then switched to cluster reviews, yet still couldn't tell real from speculative. EVP of Finance and Strategy Jeanne Jones said: "We were asked internally and externally: 'How do you know which ones are real?'"
Exelon then required developers to post binding financial commitments and collateral before entering detailed engineering. In plain terms = no real money on the table, no spot reserved.
The move cut Exelon's high-probability pipeline from 18,000 MW to 11,000 MW. Projects in the remaining pipeline have posted roughly $1 billion in collateral.
Why should ordinary ratepayers care?
US grid costs are shared across all users. If a data-center project cancels, the stranded infrastructure cost shifts to remaining ratepayers, pushing up electricity bills.
Jones put it bluntly: "If we build out the grid and the data center walks away, someone has to pay — and that's not fair to other customers."
This reflects a broader political flashpoint: AI infrastructure is already driving up power costs, and state regulators plus consumer groups are demanding that data-center developers bear real financial accountability for the capacity they claim.
What does this mean for AI compute expansion?
As utilities tighten grid-access requirements, the timeline from "application filed" to "power delivered" stretches longer.
This means → AI compute's actual expansion may run slower than the bullish forecasts built on raw application volumes.
In plain terms = the headline numbers made it look like compute capacity would explode. Strip out the ghost projects, and the real buildout pace is what emerges — making this the key variable for testing this infrastructure investment cycle.
市场有风险，内容仅供研究参考，不构成投资建议。