For two years the AI infrastructure conversation was about chips — who could get them, how many, at what price. That conversation has quietly changed. The hyperscalers and their financing partners now describe power as the scarce input, and the paperwork backs them up: grid interconnection requests from data center developers have multiplied at every major US transmission operator.
The numbers involved are large enough to show up in state-level energy planning. Single campuses now request service in the gigawatt range — the output of a nuclear reactor — and utilities that planned for flat demand are being asked to serve load growth not seen since the postwar electrification era.
Why it matters
Who pays for the grid buildout is becoming a contested political question. Utilities traditionally spread infrastructure costs across all ratepayers; regulators in several states are now creating special rate classes that put data center costs on data center customers, with minimum-take contracts and exit fees. The outcomes will shape where AI infrastructure is economical to build.
There is also a reliability dimension. Concentrating gigawatts of new load in specific transmission zones stresses systems designed for gradual growth, and grid operators have begun warning publicly about resource adequacy in the regions where data center development is densest.
How the bottleneck actually works
Three separate constraints stack. First, generation: new power plants — gas, nuclear, renewables with storage — take years to permit and build. Second, transmission: moving power to the site requires lines and substations with their own queues. Third, equipment: large power transformers and switchgear have lead times that stretched from months to years as global demand spiked.
Developers are responding with workarounds that would have been unusual five years ago: behind-the-meter gas generation, fuel cells, co-location at nuclear plants, and site selection driven by available substations rather than by fiber routes or tax incentives.
Evidence
Grid operators' public queue data shows the scale: PJM, which covers the densest data center region in the country, has reported interconnection requests totaling multiples of its current peak load. Utility earnings calls that once mentioned data centers in passing now lead with them, and several utilities have raised capital-expenditure plans by billions specifically for data center service.
Equipment manufacturers confirm the squeeze from the other side: transformer makers have announced factory expansions in the US and abroad, citing order books that extend years out.
The competing read
Skeptics argue the announced demand is inflated — that developers file speculative interconnection requests to hold queue positions, and that real load will come in well below the headline figures. Grid planners counter that even heavily discounted, the queue implies a buildout the current system cannot deliver on the industry's preferred timeline.
Both sides agree the uncertainty itself is a problem: utilities must commit capital against load forecasts that could be off by half in either direction.
What happens next
Watch state rate cases, where the cost-allocation fight will be decided; watch FERC's interconnection reform, which aims to clear speculative requests from queues; and watch whether behind-the-meter generation at data center campuses becomes standard practice or remains a bridge. The next two years of utility capital plans will reveal how much of the announced demand the industry actually believes.