Order the transformer before you pick the architect

Here is a sequencing problem most pro formas get backwards. A development team finalizes the site, finalizes the design, gets the entitlement, and only then goes shopping for the electrical equipment that determines when the campus can actually turn on. Wrong order. By the time that order goes in, the answer may already be 2029.

Power transformers were running 128 weeks from order to delivery on average in 2025, generator step-up units 144 weeks, and large units in constrained parts of the market have stretched to four years. Prices moved with the lead times: power transformers up 77% since 2019, generator step-ups up 45%, some distribution transformers up 95%. None of that is a rounding error you absorb after the fact. It is a schedule and a cost that has to be locked in before the building it serves has a final design.

Why did equipment procurement move to the front of the project?

Because manufacturing capacity does not flex the way a construction schedule does. Build.inc's framing is blunt: large power transformers require specialized steel, copper, engineering, testing, and factory capacity, and new manufacturing capacity takes years, not months, to bring online. Utilities are competing for the same units to upgrade their own transmission systems. A developer waiting on a transformer order is really waiting behind the grid operator it is trying to interconnect to, a queue inside a queue.

The demand side explains why this crept up on the industry. Since 2019, demand for generator step-up transformers is up 274%, substation power transformers up 116%, and overall power-transformer demand up 119%, against a supply base that had 30% and 10% shortfalls in power and distribution transformers respectively in 2025. More than 40 million distribution units nationally are already operating beyond their expected service life. Data-center load is the newest and fastest-growing claimant on a supply chain that was already behind before hyperscale demand showed up.

Build.inc's actual argument is the one worth sitting with: "Electrical equipment is no longer a procurement package opened after design development. It is a front-end feasibility input." Read that sentence twice, because it inverts the order almost every legacy development process assumes. A team now has to commit real capital to transformers and switchgear before entitlement is certain and before a tenant lease is signed, because waiting for certainty means waiting behind everyone else who also wanted 2027 delivery.

Not everyone agrees the shortage is structural. Patrick Tarver of Bolt Electrical argues the real bottleneck is procurement practice and vendor-qualification rules, not physical scarcity, and that units can arrive in 12 to 14 months once engineering is approved. Even on his more optimistic read, the lesson holds: engineering approval, and the order that follows it, has to happen early, not after the building is designed.

Does cooling architecture lock in just as early?

Yes, and for a similar structural reason: retrofitting the wrong choice later is expensive enough that it changes which projects are worth building at all. Air-cooled facilities top out around 20 to 30 kW per rack, with the best row-based air designs reaching about 29.5 kW. Direct-to-chip liquid cooling supports 60 to 175 kW per rack today, and Nvidia's Vera Rubin NVL144 racks are expected to need 300-plus kW in 2026, with the Rubin Ultra NVL576 pushing past 600 kW by 2027. An air-cooled shell is not a smaller version of a liquid-cooled one. It is a different building, and the density gap between the two is widening every product generation, not narrowing.

Liquid cooling penetration reached roughly 37% of new deployments in 2026, up from about 3% in 2021, and the direct-to-chip cooling market alone is projected to grow from $3.33 billion in 2026 to $17.31 billion by 2032. That growth curve is a proxy for how many campuses are choosing liquid cooling from day one rather than bolting it on afterward, because bolting it on afterward is genuinely painful. Raised floors built for legacy loads support about 150 pounds per square foot; liquid-cooled racks exceed 3,000 pounds, and floor reinforcement runs about $200 per square foot and requires facility downtime to install. Retrofitting a facility from air to liquid cooling can cost roughly $2 million per megawatt, against upwards of $11 million per megawatt to build liquid-cooled capacity greenfield, which sounds like the retrofit wins on price until you count the downtime, the structural work, and the fact that some operators have concluded it is cheaper to demolish and rebuild than to fight an existing floor plate.

A raised floor engineered for 150 pounds per square foot cannot host a rack that weighs 3,000 pounds. That is not a retrofit. That is a different building wearing the old one's address.

What does it look like when a team gets the sequencing right?

CoreWeave's Kenilworth, New Jersey campus shows the retrofit path done deliberately rather than as a rescue. The $1.8 billion project combines a 284,500-square-foot retrofit of a former Merck research building with 108,100 square feet of new construction, for a combined 250 MW. The reason this retrofit pencils where others don't is sequencing: power, fiber, and cooling infrastructure at the former NEST campus were already characterized before the project began, because the site had been an industrial research campus, not a speculative shell waiting for a tenant. The retrofit decision was made because the early infrastructure inputs already existed, not despite the fact that they didn't.

Applied Digital's Ellendale, North Dakota campus shows the greenfield version of the same discipline. The on-site substation was energized in the fourth quarter of 2025, ahead of the campus's own building schedule, and Building 1 hit its Phase I 50 MW milestone on the contractual timeline CoreWeave had signed up for, then reached full 100 MW energization in November 2025. The campus is designed for 400 MW and phased explicitly: a second 150 MW building targets mid-2026, a third, optional 150 MW building targets 2027. Nothing about that schedule is speculative. It is staged against a signed, 15-year, roughly $7 billion CoreWeave lease, and the substation went in before the second and third buildings had a reason to exist yet.

Two different playbooks (retrofit a characterized industrial site, or energize a rural substation ahead of the building it will serve), and both work for the same reason. Neither team waited for the building design to finish before locking the input that actually gates energization.

What should a development team change about its own sequencing?

Move the electrical equipment order and the cooling-architecture decision into the feasibility phase, not the construction phase, because both now behave like long-lead entitlement items rather than construction line items. Ask, before the design is final, whether the campus is being built for the rack densities Nvidia is shipping today or the densities it will ship in two years, because a shell designed for 30 kW racks is not a cheaper version of a 150 kW shell. It is a stranded asset with an extra step. And treat phasing as a financing tool, not a fallback: Applied Digital's staged 50-then-100-then-150-then-150 megawatt build let the campus match construction spend to a signed lease schedule instead of guessing at demand three buildings ahead.

My bet, and it is a specific one: within 18 months, the standard hyperscale RFP requires bidders to show a signed or letter-of-intent transformer order before the site's zoning application is even filed, inverting the sequence that has governed data-center development for a decade. If in 2028 the industry norm is still "design first, order equipment after entitlement," I misjudged how fast the supply constraint would reshape the process, and I'll own that. Check your own pipeline right now. If your electrical equipment order sits behind your entitlement filing on the critical path, you have the sequence exactly backward.

This analysis is a source-cited research summary drawn from public records, not legal advice. It can contain errors and should be verified independently before any investment decision.