Drive west out of Washington, D.C. on the Dulles Greenway and within twenty minutes you cross into a stretch of Loudoun County that looks, at first, like any other Northern Virginia exurb. The forest opens. The road widens. Then the buildings begin — long, low, windowless boxes the colour of cold pewter, set back from the road behind quiet berms and chain-link fences. There are no signs you would recognise. No corporate logos, no glass lobbies. Just the buildings, the chillers humming on the rooftops, and the unmistakable architecture of an outdoor 230 kV switchyard breaking the tree line every few miles.
This is "Data Center Alley," and depending on how you count, somewhere between 60 and 70 percent of the world's internet traffic moves through it. The campus density is so extreme that Dominion Energy, Virginia's largest utility, now treats Loudoun and neighbouring Prince William County as a single planning region with its own electrical climate — one where a few square miles of farmland can pull more power than the entire downtown of a mid-sized American city.
For two decades the bottleneck for hyperscalers in this corridor was land, then fiber, then water for cooling. Today it is something much less photogenic and far harder to solve: the large power transformer. The buildings can be poured and stood up in nine months. The 230/34.5 kV step-down transformer that feeds them now has a quoted lead time of 130 to 160 weeks. That gap — between a building's construction schedule and the iron that energises it — has quietly become the single most important number in American hyperscaler planning.
How Loudoun Became the World's Densest Load
The story usually starts with AOL in the late 1990s and the cluster of peering exchanges that grew up around its Dulles campus. But the real reason Loudoun pulled away from every other US data centre market is electrical, not historical. Dominion's 230 kV network into Northern Virginia was overbuilt in the 1970s for industrial loads that never fully materialised. When the first wave of colocation operators arrived, they found a grid that could already absorb hundreds of megawatts without needing new transmission. That latent capacity, combined with a friendly tax environment and dark-fiber abundance, did the rest.
By 2018 the Alley was consuming around 1,700 MW. By 2023 the number had passed 3,500 MW. Dominion's most recent integrated resource plan assumes Loudoun-and-Prince-William load will reach roughly 9,500 MW by 2030 and continue climbing past 13,000 MW by 2035 — and almost every utility planner who has looked at the underlying queue believes that number is conservative. For context, 13 GW is roughly the entire summer peak of New England.
A single new hyperscaler campus in the corridor now routinely requests 300 to 750 MW at a single point of interconnection. Five years ago a 100 MW request was considered aggressive.
The Substation Inside the Fence
What changed inside the fence is just as dramatic. The traditional model — utility owns the transmission substation outside the customer's property, customer steps down from 34.5 kV inside — has been replaced almost everywhere in the Alley by customer-owned 230 kV substations sitting directly on the campus. The hyperscaler builds and owns the GIS or AIS yard, the 230/34.5 kV main transformers, the network protectors, and frequently a portion of the upstream 230 kV line itself.
This shift happened because the utility could no longer keep up. When a single campus needs four 138 MVA main transformers plus spares, plus a dozen unit substations behind them, the customer simply has more incentive — and more cash — to procure and energise that equipment than Dominion does. The result is that the largest data centre operators in Loudoun now run, in effect, private transmission utilities. They schedule their own outages, hold their own spares, and increasingly negotiate directly with manufacturers for slot allocations.
The 130-Week Number
That last point is the heart of the bottleneck. A 230/34.5 kV main transformer of the size hyperscalers buy — typically 80 to 150 MVA, ONAN/ONAF/ONAF rated, with on-load tap changer and a copper-wound core suitable for the harmonic profile of a UPS-fed IT load — was, in 2019, a 50-week order. By 2022 the quote had drifted to 78 weeks. Today, depending on the manufacturer and the slot, it is 130 to 160 weeks, with deposits required well before drawings are finalised.
Three things are happening at once. First, demand has doubled globally while the number of factories capable of building units above 80 MVA has barely moved. Second, grain-oriented electrical steel — the magnetic core material — comes from a small number of mills in Japan, Korea, and increasingly India, and the data centre boom is competing for the same tonnage as the wind and HVDC backbones being built across the United States and Europe. Third, the skilled labour required to wind a large core-form transformer takes years to train, and several manufacturers lost a generation of winders during the 2010s consolidation.
The hyperscalers responded the only way they could: they started ordering before they had buildings. The largest campuses in Loudoun now place transformer orders in parallel with site selection, sometimes for sites that have not yet been entitled. That changes the relationship with the manufacturer fundamentally. It is no longer a transactional purchase; it is a multi-year capacity reservation, often tied to a portfolio of identical units the operator commits to absorbing whether or not every site comes online on schedule.
What the Transformers Themselves Look Like
The unit being built for an AI training campus today is not the same unit Dominion was buying for substation reinforcement ten years ago. A few characteristics stand out.
The load is non-linear. UPS rectifiers, even modern transformer-less topologies, push significant harmonic current back into the primary. K-factor ratings of 13 and above are now standard, and operators routinely specify a third winding — a stabilising tertiary — to absorb zero-sequence and triplen harmonic content rather than letting it propagate onto the 230 kV side.
The load is also extraordinarily steady. Unlike a utility distribution transformer that sees a 3-to-1 swing between night and day, a hyperscaler main transformer rarely drops below 70 percent of its nameplate rating. That changes the thermal design. Insulation is specified for continuous high-temperature operation, oil preservation systems use sealed nitrogen blankets rather than conservator tanks, and the cooling stages are sized for the ONAF2 rating to be the everyday operating point, not the emergency rating.
Finally, the redundancy philosophy is N+1 or 2N at every level, and the customer pays for it. A 400 MW campus will typically install five 100 MVA main transformers when four would technically suffice — not because the engineers cannot do the math, but because losing a unit and waiting 130 weeks for a replacement is operationally unacceptable. Spare transformers, fully oil-filled and energised on standby, are now treated as a normal capital line item.
The Queue Beyond the Queue
Dominion's interconnection queue for Loudoun is, depending on which month you check, somewhere between five and seven years deep for new large loads. That queue is itself constrained by transmission — there is only so much 500 kV import capacity into Northern Virginia, and PJM's regional expansion projects move on their own timescale, currently measured in years even with fast-track designations.
The hyperscalers' answer has been to push further west, into Prince William, then Fauquier, then across the Blue Ridge into the Shenandoah Valley, chasing whichever substation has spare 230 kV transformer capacity that can be unlocked with a comparatively modest upgrade. Each move adds fiber latency, but in many AI training topologies the round-trip to Northern Virginia is still acceptable, and the alternative — waiting in the Loudoun queue — is not.
A second response is on-site generation. Several recent campus designs in the corridor include behind-the-meter gas turbines sized to carry a meaningful fraction of the IT load, not as backup but as a permanent peak-shaver during the years the customer is waiting for transmission upgrades to land. Whether that ends up as a footnote or a structural feature of American hyperscaler design will depend on how quickly PJM clears the queue.
What This Means for Everyone Else
For an industrial customer in the same Dominion territory — a chip fab, a battery plant, a hydrogen electrolyser project — the consequences are immediate. Transformer slots, transmission capacity, and substation real estate are all competing inputs, and the hyperscalers have proven willing to pay premiums that distort the curve for everyone else. A 138 MVA unit that a utility would have ordered for grid reinforcement now sits in a procurement queue behind a hyperscaler order for an identical specification.
The longer-term implication is more interesting. The transformer industry is responding — capacity expansions are underway at several major manufacturers, including new core-cutting lines in North America — but the response is measured in years. In the meantime, the hyperscalers are vertically integrating into power. They are buying equipment, holding spares, owning substations, and increasingly co-locating with generation. Data Center Alley is, in 2026, less a real-estate market than an industrial-scale electrical system that happens to also run the internet.
The buildings will still go up in nine months. The transformers will still take three years. Until that gap closes, the bottleneck in American hyperscaler growth is not silicon, and it is not concrete. It is a five-hundred-tonne stack of laminated steel sitting on rails in a railcar somewhere west of Chicago, waiting for its turn.
