The Real Bottleneck Was Never Chips. It Was Megawatts.
Utilities are quietly re-rating every gas turbine they can find. The GPUs are on the loading dock; the interconnection queue is measured in years.
Vincent Jiangthree-minute read
The chips arrived. The power did not. Across three continents, AI data centers are sitting on racks of accelerators they cannot switch on, because the one input nobody hedged, grid interconnection, now runs a queue measured in years.
A queue longer than the hardware's useful life
The number utilities do not put in press releases: the typical interconnection wait in the largest US grid now runs past five years.1 A GPU generation is useful for about two.
Data
| Value | |
|---|---|
| Grid interconnection wait | 5+ yrs |
| GPU generation, useful life | ~2 yrs |
So the buyers went around the grid. Behind-the-meter gas4, restarted nuclear like Microsoft's−0.70% — Microsoft, down 0.70 percent today deal to revive Three Mile Island6, and every idle turbine a utility could re-rate are suddenly assets with a bidding war attached.
Turbines are booked like jet engines
Turbine makers are quoting lead times normally reserved for aircraft engines. The order books are full through the end of the decade2, and the buyers are not the utilities. They are the model labs and their landlords.
The bill lands on someone else's meter
The second-order effect lands on your electricity bill. When a data center outbids a city for the same megawatt3, the clearing price moves, and regulators are only now noticing who is paying it5.
Whoever locks firm power on a ten-year contract locks the training roadmap. Everyone else is renting time on someone else's substation.
Get Power by email, free
How AI's hunger for electricity moves utilities, power plants and the grid. In our daily email, every weekday.






Reader comments
0 comments