Data center construction footprints are the visible half of a power demand number that utilities confirm a year later

Two numbers cannot stay true at once. Data center developers have formally asked American utilities for 1,066 gigawatts of grid capacity, a figure larger than the generating fleet of any country on earth except China and the United States; Wood Mackenzie analysts expect utilities to actually commit to serving about 298 GW of it (Bloomberg, Aug 12). That leaves roughly 768 GW filed into interconnection queues by companies with no financing, no permits and often no intention beyond holding a place in line. Yet the queue is what grid planners forecast from, and what they forecast becomes what ratepayers are ordered to fund. The concrete footprints rising across Virginia and Texas are real. The spreadsheet underneath them is mostly fiction, and nobody learns which is which until the meters run a year later.

The actors want opposite things from the same number. Hyperscalers such as Microsoft, Amazon, Google and Meta need power secured years before their GPU campuses open, so each files applications at several candidate sites simultaneously and abandons all but the winner; filing costs a deposit, not a decision. Utilities want to show regulators booming load forecasts, because bigger expected demand justifies bigger rate bases and bigger capital programs. Grid operators such as PJM Interconnection, which coordinates wholesale power across thirteen states including Northern Virginia, must order generators to be ready several years out, so they take the forecasts as gospel because the penalty for underbuilding is blackouts. Governor Abigail Spanberger wants Virginia households shielded from the bill. Each incentive pushes in the same direction: treat every application as real, then argue later about who pays.

The trigger arrived this month. Exelon stripped speculative projects out of its pipeline and cut its data center load forecast by 40 percent, to 11 GW of signed, contract-backed demand (Utility Dive, Aug 17). That is not a demand collapse. It is one of the first large utilities publicly separating interconnection agreements that carry deposits and timelines from wishful filings, and what remains after the cut is the closest thing the market has to a true number. The slow pressure beneath the trigger is older: metered data center draw on the US grid has roughly doubled since 2023, genuine growth that makes every bloated forecast look plausible (Wood Mackenzie data via Bloomberg, Aug 12).

The history that fits is fiber, not rail. Between 1997 and 2001, telecommunications carriers laid tens of millions of miles of long-haul optical cable against traffic projections that never arrived; when the projections failed, Global Crossing went bankrupt and dark fiber sat under highways for a decade while the carriers who had overbuilt destroyed their own balance sheets. The mechanism rhymes: cheap options on future demand, filed by everyone against everyone, until the aggregate option book swamps the real economy. What differs is who holds the obligation. Carriers spent their own capital on fiber. Here, the capacity to serve phantom load gets procured through regulated auctions, which means the cost lands on customers before anyone knows if the load exists.

The counter-case argues the other way, and it deserves its say. Unlike dot-com traffic projections, some of this demand has already signed contracts and paid deposits; PJM's own meters show real load nearly doubling in three years, and the 11 GW Exelon kept after its haircut carries signed interconnection agreements with real dates (Utility Dive, Aug 17). If even a third of the committed 298 GW materializes on schedule, the country is short turbines, transformers and transmission lines, not long on them. General Electric Vernova's gas turbine backlog stands at 116 GW with three-year lead times, pushing new firm capacity to 2028 at the earliest (The GPU Daily, Aug 18), so a campus breaking ground today races a constraint money cannot shortcut. Both overbuild and shortage can be true in different places at different times. That is exactly why the forecasting failure matters.

Walk the consequences forward. First order: utilities build generation and wires for load that never comes, and regulators let them recover those costs from everyone connected to the wire. Second order: the people paying notice. PJM's independent market monitor, Monitoring Analytics, calculated that including existing and forecast data center load added $23.1 billion in system costs across three consecutive capacity auctions, obligations locked in through May 2028 (Monitoring Analytics report via press coverage, Jul 16). The National Energy Assistance Directors Association expects average household electricity spending of $792 this summer, up 10.5 percent from last year (NEADA survey, Jul 2026), and the Energy Information Administration projects residential prices rising 5.7 percent through 2026 (EIA Short-Term Energy Outlook, Jan 13). Third order: politics. Virginia's State Corporation Commission ruled this month that Dominion Energy must write tariffs making data centers pay the full cost of transmission built solely for them, a decision the governor says saves ratepayers hundreds of millions of dollars (WSET, Aug 6). When one state wins that fight visibly, others copy it, and the socialized-cost model that financed the boom starts closing.

Who profits meanwhile sits upstream of the argument. Turbine and transformer makers hold multi-year backlogs priced against the full request book, and they collect whether the campuses open or not. Landowners near substations sell to whichever developer files next. The losers, if the phantom share evaporates, are utility shareholders who capitalized forecasts into stock prices and the households whose rates embedded them. The asymmetry repeats the oldest pattern in infrastructure booms: whoever books the revenue on paper keeps it, whoever funds it physically eats it.

What confirms this read arrives quarterly. Watch more utilities follow Exelon in publishing signed-versus-speculative load splits; each disclosure shrinks the official forecast toward the 298 GW reality. Watch PJM's independent monitor keep attaching dollar figures to data-center-driven auction costs, because named numbers become litigation. Watch further states adopt Virginia-style dedicated-cost tariffs after the Spanberger precedent. What breaks it is simpler: if metered national data center draw keeps accelerating while queues clear faster than five years, the filings were not phantom, they were early, and the shortage camp was right all along.

The tradable expression runs through the divergence itself. Regulated utilities with the heaviest announced data center pipelines, Dominion Energy foremost among them, carry the widest gap between forecast glory and regulatory repricing, while equipment makers with contracted backlogs sit insulated from the correction either way. The pressure resolves where the SCC just pointed: onto the developers' own balance sheets, through tariffs that make them post collateral for the load they claim. The read fails if the next round of utility earnings revisions goes up instead of down.

A year from now, the number utilities confirm will be smaller than the number they were handed. The footprints will still be there, poured and photographed, monuments to a forecast nobody checked until the invoice came due. In this boom, the fog is the product, and the fog is billed monthly.