A new data-center campus can request load that used to belong to a mid-sized city. Those requests arrive in clusters from companies used to shipping code weekly. The system that has to answer them was designed for the opposite tempo. Connecting a facility to the grid can take four to ten years in congested markets; the data centers themselves are typically planned and built in two to three. That gap is not an oversight. It is the residue of a system built to keep the lights on for everyone else while a handful of operators race to train the next model.
You can bring your own generation. Plenty of people already do. Portable and aeroderivative gas turbines have become a common bridge precisely because the interconnection queue will not move on a software schedule. Elon Musk’s xAI operation in the Memphis area ran dozens of them for months while the permanent plant was still being permitted; SpaceX later committed billions more to the same approach. The turbines still have to be interconnected, coordinated, and accepted by a network that was never designed to absorb sudden, concentrated demand the way a software team absorbs a new feature branch. The lag remains structural.
The physical choke points are even more stubborn. Large power transformers that once took roughly two years now stretch to four or five; manufacturers report multi-year order backlogs and a shortfall measured in double-digit percentages of annual need. Electrical gear is a modest line item in a data-center budget, yet it can decide when the building actually turns on. At the end of 2025 more than two thousand gigawatts of generation and storage sat waiting in U.S. interconnection queues—roughly twice the entire installed fleet. In ERCOT alone the large-load queue has ballooned past two hundred gigawatts against a grid whose all-time peak is a fraction of that figure. The queue itself has become the binding constraint.
This tension is old. Early electric utilities faced factories and cities that wanted power faster than the networks could safely expand. The same pattern appeared with the railroads, the interstate highways, the early telephone system. Demand spikes. The physical layer lags. People who live in the lag get frustrated. People who maintain the lag get accused of obstruction. Both sides are usually partly right. The last time U.S. electricity use grew at five percent a year or faster was the long mid-century wave of air conditioning, refrigerators, and household appliances. The current surge is sharper, more geographically concentrated, and arriving from operators who treat physical constraints as temporary inconveniences.
What is different is the scale and the concentration. Pattern-prediction models are powerful. They are not magic. At some point they run into the same hard limits everything else does: energy, cooling, copper, transformers, transmission capacity, and the human systems that keep those things from cascading. Grid operators have already begun revising near-term demand forecasts downward because interconnection is slower than the announcements implied; some have added new emergency tools simply to manage the risk of large, sudden load drops or ramps. The current growth curve still assumes the grid will stretch. That assumption is looking thinner by the quarter.
The deliberate pace of the utility is not romantic. It is not exciting. It is also the reason most of us can still open the fridge without thinking about it. That boring reliability is the actual foundation everything else is standing on. The question is whether we treat it as a constraint to be respected or an obstacle to be overridden.
What happens if the grid does not stretch on the timeline the models need? Do the labs keep building their own generation and accept higher costs, local friction, and the occasional lawsuit over unpermitted turbines? Do we start seeing more creative—and riskier—workarounds? And what does “enough power” even look like once the next generation of models arrives? The last one was already surprising. The next one will be hungrier.
I’m curious what the lag feels like from where you sit—utilities, data centers, infrastructure planning, or just watching the demand curves. How do you see the gap closing, if it closes at all?