Half of America’s planned 2026 data centers are stalled, and not for lack of chips or capital. The companies best positioned for the next phase of AI are the ones that treated grid hardware as a strategic asset two years ago.
Ask a boardroom what’s gating its AI ambitions and you’ll hear “GPUs” or “talent.” Ask the person who actually has to energize the building, and you’ll get a different answer: there are no transformers, and the ones you order today won’t land until 2030.
That gap, between what executives think constrains AI and what actually does, now shows up in the construction data. Bloomberg reported this spring, citing Sightline data, that roughly half of the US data centers planned for 2026 have been delayed or canceled, and that only about a third of the 12 gigawatts slated for this year is under active construction. The reason isn’t a chip shortage but a shortage of transformers, switchgear, and the grid capacity to feed them.
The bottleneck is moving from chips you can buy to infrastructure you have to wait for. Compute is no longer the only scarce input in the AI buildout. For many projects, the gating input has moved downstream: grid hardware and interconnection queue position now decide whether the compute ever turns on. You could push back, and the bull case is simple: capital this large usually clears bottlenecks eventually. But “eventually” is doing a lot of work in that sentence. In power infrastructure, it often means years, not quarters.
The Data Center Race Is Hitting The Grid
For two years the industry measured progress in chips: order books, allocation fights, export controls. That was the dominant metric for 2023. When half of a year’s planned capacity slips and two-thirds of the rest isn’t in the ground, the binding constraint has moved.
Silicon supply is hard, but it’s a problem the technology industry knows how to attack: more fabs, more packaging, more suppliers. Delivered electricity doesn’t respond to that playbook. It depends on transmission, substations, transformers, permitting, and a utility’s ability to energize a specific site on a specific date. The companies optimizing only for chip allocation are solving yesterday’s bottleneck.
A Transformer Is Not A Software Problem
A high-voltage transformer is not a commodity you reorder. A handful of manufacturers worldwide build the largest units, most are imported, and Wood Mackenzie puts lead times for the biggest at 80 to 210 weeks, with power-transformer prices up about 77% since 2019. At the high end, that’s four years of waiting. Software companies are used to compressing timelines. Transformers don’t compress.
Demand isn’t pausing for supply to catch up. The International Energy Agency estimates data centers used about 415 terawatt-hours of electricity in 2024 and projects that more than doubles to roughly 945 terawatt-hours by 2030, slightly more than Japan uses today. Even discounted heavily, that outruns what the grid can quickly add.
I’ve watched this play out over and over. Companies pick a site the way they always have: optimize for tax abatement, land, fiber, and labor, then treat power as a formality to sort out later. That order is now backwards. A utility study, a transformer slot, and a written energization date now matter more than the ribbon-cutting incentive package. You can order more GPUs. You can’t order a substation into existence in a quarter. A generous tax package on a site that can’t get power for years isn’t a deal. It’s a stranded asset with good optics.
The Next Fight Is Who Pays For The Upgrade
The hardware shortage and the cost fight are the same story, and regulators are scrambling. On June 9, FERC approved a PJM “expedited interconnection” track to fast-track shovel-ready power projects, effective July 31. Still pending is the bigger question: by the end of June, FERC has said it will act on how large loads, generally those above 20 megawatts, connect to the grid, and on who pays for the upgrades they require.
That pressure is already showing up in prices. In PJM, the largest US grid region, wholesale power costs rose almost 76% year over year in the first quarter, according to Monitoring Analytics, the grid’s independent market monitor, which named data-center load growth as the primary driver. Once the cost of serving hyperscale load starts appearing in residential bills, the issue stops being a technical debate and becomes a political one. Regulators are under pressure to make large loads absorb more of the upgrade costs they create. If they do, it reprices every megawatt in the pipeline.
Power Geography Is Rewriting The AI Map
Watch where the disciplined money goes. Microsoft signed a 20-year deal with Constellation tied to restarting Three Mile Island Unit 1, now the Crane Clean Energy Center. Amazon contracted for up to 1.9 gigawatts from Talen’s Susquehanna nuclear plant. These are supply deals, not green gestures, from companies that concluded electrons are the real constraint and chose to buy them directly. Look at the structure. A 20-year contract tied to a specific reactor isn’t how you buy a commodity; it’s how you buy certainty when the commodity might not be there. These buyers are taking on plant-level risk to escape queue-level risk, a trade that only makes sense if you believe the queue is the bigger threat.
Location is reorganizing around the same logic. The old map followed fiber and customers; the new map follows deliverable megawatts. Northern Virginia won the last era on fiber and power; Texas now competes on delivering power on the operator’s timeline; around Dublin, grid limits capped new connections and left projects waiting. Same technology, opposite outcomes, decided by the grid, not the model.
The Question Every Board Should Ask
The buildout has hit the part of the stack you can’t iterate: physical infrastructure with multi-year lead times. None of this means compute stops mattering. It means power decides whether the compute you bought ever runs.
So re-baseline your build plan against your actual energization date, not your capex calendar, and put one question to your infrastructure team: what’s our transformer delivery date and interconnection queue position, and what does the coming FERC rule do to our cost per megawatt? If they can’t answer, the AI strategy isn’t yet connected to the infrastructure strategy. You may be ordering GPUs you can’t plug in.
