The queue
Right now, more than a thousand gigawatts of power projects sit in interconnection queues, waiting years for permission to plug into the grid — more capacity stuck in line than parts of the country currently generate. That backlog, not a shortage of ideas, is the real bottleneck. A company building its own power plant in the backyard isn’t a triumph of innovation. It’s what you do when the shared system is too jammed to wait for.
We run the wires, and here’s the unglamorous truth: it takes longer to permit a transmission line than to build the plant it connects. The fixes that would absorb much of this demand are boring — faster interconnection, demand response that lets data centers throttle when the grid is tight, scheduling AI training for the hours when power is slack, and chips that do more per watt. None of it makes a headline. All of it works.
The nuclear renaissance is right that we need firm power, and the abundance builders are right that supply matters. But if every big load islands onto its own plant, we lose the pooling that made the grid cheap and reliable to begin with. And someone should ask the question under the deniers-versus-believers shouting: is a frontier model worth a Japan’s worth of electricity, when the 1960s taught us what we ignore until the river catches fire?
Where we concede ground: Efficiency gains get eaten by more usage — Jevons’ paradox is real, and we can’t optimize our way out alone.
What would change our mind: Fixing interconnection, permitting, and demand response, and still finding a hard supply gap only new dedicated plants can close.
Read the full synthesis: Should AI data centers get their own power plants?