Operators can lease land and pour concrete faster than they ever could. They still cannot get megawatts to the door. AI data centers are being designed on the assumption that power will arrive, and across much of the world it will not.
Interconnection queues tell you how hard this is. In large parts of the United States, a project can wait five years or more just to connect to the grid. Building a new transmission line can take longer than building a new chip factory, which is really saying something.
The demand is enormous, the supply is not
Analysts have spent two years revising forecasts upward. Load growth that utilities thought would take a decade is now arriving in three or four years. Some of it is data centers. Some of it is electrification of transport and industry. A lot of it is AI training and inference, which runs around the clock and does not shed load at peak.
That makes it a bad match for a system built to handle a slow, predictable rise. Utilities plan for a few percent growth a year. They are now being asked to absorb double-digit growth in the span of a few quarters, plus the simultaneous retirement of older gas and coal plants.
The bottleneck is not generation. It is everything between the plant and the data center, and the permits to build it. Transformers are a hidden problem. The large power transformers the industry needs have long lead times measured in years, and demand has outpaced the suppliers that make them. You can buy a generator tomorrow and still wait two years to step its output down to the voltage your racks use. There is also a pricing problem. Most utilities bill large customers for peak demand, not just total energy. A facility that pulls a huge burst during a system-wide peak gets punished for it on the invoice, so peak load becomes something to design around rather than assume on a curve.
Power is the new real estate
For years the constraint on data center growth was floorspace or fiber. Now it is capacity. Operators talk about it openly, and you can see it in their choices. They are buying up old industrial sites, refineries, mills, even abandoned buildings, not because the location is ideal but because the power hookup already exists.
This changes the economics of a build. A site with a large substation nearby is worth more than a site with a perfect network backbone. The question a hyperscaler asks before it signs a lease is no longer just how much internet is here. It is how many megawatts can I actually switch on.
Developers are also sitting on large banks of capacity they cannot energize. Some have spent enormous sums locking in power rights and paying for leases to ensure they can build later. That has led to a standoff. The megawatts are contracted, but the infrastructure to deliver them is not finished. It has also distorted local markets, because a small number of projects can hold a big parcel of potential load for years.
What actually moves us forward
The honest answer is that this does not resolve quickly. A utility cannot speed up transformer deliveries. A government cannot shorten environmental review and keep public confidence at the same time. The grid was built for a different age, and it is hard to extend it.
The parts that do move are the ones that do not need the grid at all. On-site generation, batteries, and a willingness to trade latency for availability all matter. Many operators are planning their own gas turbines to cover the gap, because a few years of on-site generation beats years of waiting for a transmission upgrade.
There is a real chance the legacy generation-heavy regions will not land most of the next AI buildout. The ones that will are the ones with spare capacity, fast permitting, and a grid that can already deliver. This is the single largest force shaping where AI infrastructure gets built over the next several years, and it has nothing to do with the chips. Learn more about Data center demand.