The limiting factor on AI capacity expansion is no longer chip supply. It is electricity, and specifically the ability to connect a large new load to a grid that was not planned for it.
The scale of the ask
A large AI training facility draws power comparable to a small city, continuously, with little tolerance for interruption. Grid operators handling connection requests of that size have to answer a question that used to be theoretical: whether the local network can deliver it at all.
In many places the answer is not yet, and the timeline for upgrading transmission is measured in years rather than months.
What operators are doing
Building where the power already is, which increasingly means near generation rather than near users. For training workloads that is acceptable, because latency to the end user does not matter when nobody is waiting on a response.
Signing long-term agreements directly with generators, which provides revenue certainty to the supplier and price certainty to the operator. And in several cases, contracting for nuclear output specifically, because it is the only source that is both carbon-free and continuously available.
The inference shift
The load profile is changing. Training is bursty and location-flexible. Inference – actually serving model responses to users – is continuous, latency-sensitive, and has to happen near the people using it.
As deployment grows relative to training, the power demand moves back towards population centres, which is precisely where connecting new load is hardest.
Why efficiency has not solved it
It has improved substantially. Energy per unit of computation keeps falling. Total consumption keeps rising anyway, because cheaper computation produces more computation rather than less spending. That pattern is old and reliable.
What it means practically
Electricity prices in regions absorbing large new loads face upward pressure, and the political response to that will shape where the next round of building happens. The constraint on AI over the next few years is less likely to be model capability than a substation somewhere with a five-year upgrade queue.