AI CapEx Bottlenecks at a 160-Week Transformer Queue

Every AI dollar eventually has to become electricity, and electricity has to arrive through specific, physical things. Most investors will never see the most important of them: a generator step-up transformer, the bus-sized tank of copper and steel that ties a power plant into the grid, engineered for one site and qualified for years. In 2020 and 2021 you could order one and wait about a year. By the first quarter of 2026, lead times for the large units had stretched past 160 weeks, and transformer prices were up about 77% since 2019. A data center can raise the money and buy the chips and still miss its opening date, because the equipment that brings it power is not currently for sale.

That is what the "grid" story is really made of. Roughly 2,300 gigawatts of generation and storage — more than the country's entire installed generating capacity — are sitting in interconnection queues, the review process that decides who gets to plug in. The median wait from first request to actually running has more than doubled, from under two years for projects filed in 2000–2007 to more than four years for 2018–2024. And only 13% of the projects that entered the process before 2020 have ever reached commercial operation; three-quarters were withdrawn. The queue is not a delay; it is a permanent feature of the system. Goldman Sachs Research projects data-center power demand up 165% by 2030, and roughly $720 billion of grid investment this decade.

Now add the third force. Communities are saying no. In the first quarter of 2026, about $130 billion of U.S. data-center projects were blocked or delayed by local opposition — 75 projects, in one quarter matching the total for all of 2025. Grassroots opposition groups more than doubled in a few months, from 396 to 833, spread across 49 states. A Gallup survey this spring found 71% of Americans oppose a data center near them. Denver imposed a one-year moratorium on data-center zoning in May; a statewide moratorium bill has cleared both chambers of New York's legislature; federal lawmakers have proposed a national pause. Disapproval is rising faster than electricity demand.

Here is the turn. None of that stops the spending — it redirects it. Rather than wait on a queue they don't control or a zoning board they can't persuade, hyperscalers are building their own gas-fired power on their own sites, behind the meter, bypassing the grid. Proposals for new U.S. gas capacity tripled in 2025, and the five largest hyperscalers are on track to spend $745–775 billion of capex in 2026. Follow each route and the bottleneck re-converges: whether the power comes from the grid or from a developer-owned gas plant, it must pass through the same scarce components — transformers, switchgear — and the gas route adds turbines on top. The narrowest node is the gear, and its queue is not a review process. It is how long the product literally takes to build.

The order books show the wave landing. GE Vernova's power-segment orders were up 134% in the second quarter; its gas-turbine backlog hit 116 gigawatts, with management targeting 125 gigawatts by year-end and already taking reservations for 2031 delivery; total backlog reached $176 billion, up from $129 billion a year earlier; and data-center orders for its electrification business passed $5 billion year-to-date, more than double all of 2025. Eaton's Electrical Americas orders rose 41% on a rolling twelve-month basis, its electrical backlogs up 33% and 103% year over year, and the combined electrical businesses were booking $1.20 of orders for every dollar of sales — with its chief executive calling data centers a "key growth driver." Wood Mackenzie frames the scale: data centers were about 2% of the U.S. electrical-equipment market in 2020, and could be 40% by 2030. None of it is only about AI — roughly four-fifths of GE Vernova's gas-turbine orders still go to traditional utilities — which is exactly why the wave has a floor under it.

Why can't the market simply build its way out? A large power transformer is engineered to order, safety-critical, and qualified one customer at a time. And the constraint sits one layer down as well: roughly 80% of large U.S. transformers are already imported, and the grain-oriented electrical steel laminated into every transformer core has exactly one U.S. producer, Cleveland-Cliffs. That is the part of the map the AI story usually skips — even the fix is bottlenecked.

The discipline is to name what breaks a chokepoint. Three things are already moving. Capacity: close to $2 billion of new North American transformer plants comes online around 2028 — too late for projects being wired today, but it is the countdown clock, and the gear makers themselves are expanding. Substitution: one California municipal utility reported roughly three-quarters of recent transformer bids arriving from China and South Korea at shorter lead times than domestic suppliers. And some industry voices argue a meaningful share of the "shortage" is self-inflicted — over-specified orders placed too late.

Then there is price, where a correct map stops being an automatic trade. GE Vernova is up roughly 45% this year near $950 — a $254 billion company selling at about six times trailing revenue, still a fifth below its 52-week high. Eaton is up about 32% to roughly $420, about 31 times this year's adjusted-earnings guidance and nearer 40 times on a GAAP basis. Both prices assume the multi-year backlog converts into margin without incident. Hubbell, a transformer and switchgear maker in the same wind, is up barely 6% this year. Same exposure, different price. The market has paid up for the two big "AI grid" narratives; it has not paid up for grid exposure in general.

The structure is confirmed down to the order books, and the economics are converting now: GE Vernova raised its free-cash-flow guidance to $11.5–12.5 billion, and Eaton is guiding Electrical Americas margins up 450 to 500 basis points in the second half. The open question is the countdown. New capacity and foreign suppliers will thin the node on a fairly well-known schedule — around 2028. The variable with no clock is the local board: data-center cancellations rose from six in 2024 to twenty-five in 2025 and set a record pace in the first quarter of 2026. Pushback is an accelerant for the equipment makers while it merely redirects demand; it only becomes a brake when it cancels projects. That is the single number worth watching as everything else in the chain gets scheduled, budgeted, and priced in.