A 1 GW campus with no large power transformer order placed is a 2031 campus, not a 2027 campus. The lead time on a 345/500 kV class unit is now 150 weeks.
The industry has compressed a six-variable problem into a one-variable conversation. Every press release, every analyst note, every hyperscaler capex slide is denominated in megawatts. MW signed. MW under construction. MW in the queue. The number is treated as the binding constraint on AI buildout, as if generation availability is the wall the entire sector is pushing against.
It is not. Generation is the one variable the market has roughly priced. The other five are what gate energization, and they are the ones that will decide which campuses come online in 2027, 2028, and 2029. An announced gigawatt with no transformer order, no executed water rights, no large-power-transformer slot, and no trades labor plan is a number on a slide. It is not a campus.
What Everyone Is Counting
The MW conversation has the virtue of being legible. A utility interconnection study produces a number. A PPA produces a number. A CAISO or ERCOT queue position produces a number. Analysts can add them up, hyperscalers can announce them, and the financial press can rank operators by total MW under contract.
This is the input that gets priced into valuations and the input that gets repeated on earnings calls. It is also the input that most often makes deals fail at closing, because the MW figure is the assumption that everything downstream is solvable in parallel. It rarely is. The capacity announcements stacking up across the United States, Ireland, Nordics, and Southeast Asia imply a build-out cadence that the equipment supply chain, water rights system, and trades labor market cannot physically support on the timelines being marketed.
A signed interconnection agreement is not energization. It is an option on energization, contingent on five other variables clearing on the same schedule. The market is buying the option and pricing the underlying.
What Actually Decides Energization
Six binding variables, each independently capable of pushing a campus from 2027 to 2030 or killing it outright:
1. Large power transformer availability. The 345 kV and 500 kV class units that step utility transmission down to campus distribution. 2. Switchgear, particularly gas-insulated switchgear (GIS). The medium-voltage distribution gear inside the campus substation. 3. Industrial water. Both the physical resource and the legal right to consume it on a continuous, non-curtailable basis. 4. Fiber redundancy. Specifically, route-diverse dark fiber adjacency to two Tier-1 backbones. 5. Workforce density. Sufficient trades labor, especially journeyman electricians and pipefitters, available within commuting range during peak construction. 6. Jurisdictional speed. The county or municipality's actual cycle time from application to building permit.
The constraint stack is not additive. It is multiplicative. A site that scores well on five of six fails the sixth, and the sixth dictates the energization date. The campuses that close in 2027 through 2029 are not the campuses with the largest MW number. They are the campuses where someone placed transformer orders and locked water in 2024 and 2025, while everyone else was negotiating PPAs.
The Transformer Bottleneck Almost Nobody Is Pricing
Large power transformers are the single most binding physical constraint on AI infrastructure in North America right now. The numbers are not subtle.
Lead times on 345/500 kV class units have moved from roughly 50 weeks in 2020 to 150 weeks and longer in 2025. Some buyers are quoting 180 to 200 weeks for fully specified units in the upper voltage classes. Spot pricing has risen by a factor of 4x to 7x over the same period. A unit that cleared at $4 million in 2020 is now quoted at $18 million to $28 million, when a slot can be found at all.
The concentration of supply is the reason. GE Vernova, Hitachi Energy, Siemens Energy, Hyundai Electric, and Mitsubishi Electric account for nearly all global capacity in the upper voltage classes. Roughly 80 percent of 345/500 kV units installed in the United States are imported. Domestic capacity exists but is small relative to demand, and the long-cycle nature of transformer manufacturing (grain-oriented electrical steel availability, core stacking labor, factory test bay throughput) means it cannot be flexed up in under three to five years.
Hyperscalers, utility transmission operators, federal grid resilience programs, and data center developers are all competing for the same factory slots. The utilities are winning, because their orders are tied to FERC-approved rate base recovery and have political backing. The data center buyers are taking what is left. A campus that places an LPT order in 2026 is taking delivery in 2029. A campus that has not placed the order at all is, functionally, not a 2027 campus regardless of what the press release said.
This is what the MW conversation is not pricing. Generation can be procured. Transformers must be manufactured, and the factory cannot be persuaded.
Water, Fiber, Workforce
Water. A 1 GW liquid-cooled AI campus running predominantly evaporative cooling can require 5 to 15 million gallons per day of make-up water under design conditions. That is the consumption of a small city. Securing new industrial water rights in the Western United States takes 3 to 7 years on a favorable trajectory, longer in adjudicated basins, and is physically impossible in many regions where surface water is fully allocated and groundwater is in critical overdraft. The will-serve letter from a municipal supplier is not water. It is an option on water, subject to curtailment under drought conditions and to political reversal at the next election. Dry cooling is the alternative everyone points at and few choose, because the efficiency penalty at hot ambient temperatures lands precisely where water scarcity is worst.
Fiber. The dark fiber adjacency conversation used to be a coastal data center problem. It is now a binding constraint in the interior United States. Route diversity to two Tier-1 fiber backbones has moved from a nice-to-have to a hard underwriting requirement for hyperscale tenants. Sites in West Texas, the Intermountain West, and the upper Midwest that look perfect on power, land, and water are getting disqualified because the nearest Tier-1 route is 40 miles away and the cost to build redundant lateral fiber pushes project economics underwater. Fiber capacity is not in shortage in the abstract. Adjacency to the right routes, in the right places, increasingly is.
Workforce. Each hyperscale campus requires roughly 1,500 to 3,000 trades workers during peak construction, weighted heavily toward electricians, pipefitters, sheet metal workers, and instrumentation technicians. The United States is short an estimated 500,000 electricians and pipefitters across all sectors. Trades labor cannot be imported on short notice and cannot be trained on the project timeline; apprenticeship programs run 4 to 5 years to journeyman. Campuses are now actively competing for trades inside a 200-mile commuting radius, and the second campus to break ground in a given metro pays a labor premium that the first campus does not. Northern Virginia, Phoenix, Columbus, and the Dallas-Fort Worth metro have already crossed the threshold where labor cost and availability are dictating site selection, not following it.
Jurisdiction As A Variable
Two sites with identical power, water, fiber, and labor profiles can experience permit timelines that differ by an order of magnitude. The county is the variable, and the variance is enormous.
In jurisdictions with experienced staff, pre-application meetings, parallel permit tracks, and a political consensus around industrial development, a hyperscale campus can move from CUP application to building permit in 6 to 12 months. In jurisdictions where staff is undersized, the planning commission is hostile, environmental review is litigated by organized opposition, and every conditional use permit becomes a referendum on broader land use policy, the same project takes 5 to 7 years. The same buildings, the same loads, the same impacts.
Jurisdictional speed is the variable that most consistently surprises capital. Developers underwrite based on the median permit timeline in a region and discover that the actual timeline at their specific site is set by three planning commissioners and one organized opposition group. There is no analytical substitute for jurisdictional intelligence collected on the ground, county by county, year by year.
The Forward Picture
Three claims about where this goes.
First, by 2028 the binding constraint on US AI infrastructure capacity will not be generation. It will be the transformer/switchgear/water/labor stack, with transformer availability as the dominant single factor. Generation queues will continue to grow, and large blocks of nameplate MW will sit in the queue waiting for the physical equipment to step the power down. The narrative will shift from "we need more power" to "we cannot energize the power we have." This shift is already visible in CAISO and PJM interconnection data; it will become consensus by 2027.
Second, equipment-led developers will pull dramatically ahead of land-led and power-led developers. The operators who placed transformer and GIS orders in 2024 and 2025, against project risk and before permit certainty, will deliver campuses in 2027 and 2028 that competitors cannot match before 2030. The capital cost of carrying that equipment exposure will look like a strategic moat in hindsight. The developers who waited for permit certainty before ordering long-lead equipment will discover that the moat is not crossable.
Third, jurisdictional intelligence will become a priced input. The data centers that close in 2027 through 2029 will not be sited primarily on power maps. They will be sited on a composite map of transformer slot availability, water rights status, fiber route diversity, trades labor depth, and county permit cycle time. The firms that build that composite view, before the rest of the market catches up to the six-variable problem, will harvest the spread between MW-denominated underwriting and energization-denominated reality.
The MW conversation will not end. It will simply stop being the conversation that matters. The campuses that energize on schedule in the back half of this decade are the ones where someone, in 2024 and 2025, stopped counting megawatts and started counting transformers, gallons, fiber routes, electricians, and permit days.
That count is the deal.
Autonomous Industries. AI Industrial Infrastructure.