The utilities, not the chipmakers, now ration America's AI build-out
The chips arrive in months and the megawatts arrive in years, so the queue outside the substation is where the frontier is being decided.
In Central Ohio, a utility told its regulators this year that the new power demand it once projected at 30,000 megawatts had fallen to about 5,700, after a first-in-the-nation tariff made data centers put money down before they connect (Ohio Capital Journal, Feb 20). The chips were never the bottleneck. The waitlist at the substation is.
Half the AI data centers hoping to go live in 2026 now face delay or cancellation, and the cause is power, not capital (Gartner forecasts, via industry reporting, Aug 2026). For readers with a brokerage account, the shift matters because the scarce asset in this boom has quietly changed: it used to be Nvidia's output, and it is now the utility's permission slip. Whoever holds the queue holds the ration.
PJM, the grid operator stretching from the Mid-Atlantic through thirteen states, held its latest capacity auction and cleared at $329.17 per megawatt-day, the federal price cap, for the third straight year (Environ Energy, Aug 19). The independent market monitor attributed $6.3 billion of that auction's $16.4 billion total cost to data centers, 38 percent of the bill (Monitoring Analytics, Jul 23).
Down the road, in Texas, the state grid's large-load queue reached 438 gigawatts in June, roughly five times the entire state's peak demand, and about nine-tenths of it is data centers (ERCOT large-load queue, Jun 2026).
Virginia's regulators smelled the problem and moved. The State Corporation Commission ordered Dominion Energy, which serves the world's densest data-center cluster in Loudoun County, to come back this fall with a plan shifting more transmission cost onto the data centers themselves (Potomac Local, Aug 6).
And in Washington, the Federal Energy Regulatory Commission voted unanimously on June 18 to order six of the country's largest grid operators to explain how they would connect large loads faster, then gave them sixty days to rewrite their rules, a window that closed around August 17 (FERC order, via industry reporting, Jun 18 and Aug 5).
The hidden tail on the bill
AEP Ohio's own filing shows the failure the standard forecast does not see: utilities were carrying speculative data-center projects, backed by an LOI rather than a shovel, and pricing the grid for all of them. The projected 30,000 megawatts that collapsed to about 5,700 was never real load (Ohio Capital Journal, Feb 20). It was an option written by ratepayers, who pay for capacity built against phantom demand through the same bills that already carry $9.3 billion of this auction's cost across PJM (gpuleaseindex analysis of PJM data, 2026). Nobody sized that exposure on a Bloomberg terminal, because it does not live on one.
The economics that built this arrangement go back years. Utilities earn a regulated return on the plants and wires they build, so more load meant more capital, and utility executives courted hyperscalers the way ports court shippers. Dominion contracted 51 gigawatts of data-center capacity by early this year and accelerated a grid-modernization plan past $65 billion (Dominion filings, via industry reporting, May 5). The hyperscalers, in turn, priced compute against a grid that had never said no. Both sides of that bargain are now being repriced at once, by regulators in Ohio and Virginia and by a price cap in PJM that has become less a ceiling than a routine.
The 30,000 megawatts that collapsed to 5,700 was never real load; it was an option written by ratepayers (Ohio Capital Journal, Feb 20).
The analogue is the railroad land-grant era, when the carriers had the routes and the shippers had the demand, and the money was made by whoever controlled the junction. The counteranalogue: rails could be built anywhere, while a substation has to sit next to a transformer that takes three years to arrive. Dominion's own queue for new large commercial service stretches beyond 36 months for substation connections (construction-industry survey, 2026), which means the constraint is physical inventory, not paperwork alone.
This is why the OpenAI answer tells you the shape of the rest. Stargate plans roughly 7 gigawatts across seven American sites, and at least three of them bypass the grid entirely with on-site natural-gas plants (Epoch AI site tracker, 2026). At its Abilene, Texas campus the company has installed gas turbines by the dozen (SemiAnalysis reporting, 2025). Sam Altman told a Cleveland crowd the blunt version: "We're burning gas to run this data center" (The Land Cle, 2026). A builder who can sign a turbine order in a quarter has stopped asking the utility for permission; a utility with a five-year transformer backlog cannot match him.
The speed line is the whole story here. A hyperscaler with a turbine order is measured in quarters; a household whose bill carries the capacity charge is on a schedule it did not choose, and a regulated utility moving through a rate case is measured in years.
Who pays for the quiet middle
The consequence run lands in three places. The hyperscalers pay first, in delays and in capital shifted from chips to turbines and batteries, which is a margin story for GE Vernova-class turbine makers and for gas producers supplying behind-the-meter plants. The utilities pay second, in regulatory risk, because the same data-center revenue that lifted Dominion's growth story now drags its political exposure. The ratepayer pays third and last, through PJM capacity charges and transmission riders, and the groups collecting signatures in Ohio to ban most data centers (Cleveland.com, Aug 18) are that cost turning into votes.
Watch one number from here: the next PJM capacity auction clearing price. A third year pinned at the $329.17 per megawatt-day federal cap confirms that demand keeps outrunning supply and that regulators will keep moving cost onto the hyperscalers (Environ Energy, Aug 19). A clear below the cap, with the market monitor's data-center share shrinking alongside it, breaks the read, because it would mean the speculative load was real all along and the grid caught up faster than Ohio's queue suggested it could.
The utilities learned to say no this year. The lesson of the queue is that the rationing was never going to happen in Washington or on Nvidia's pricing page; it happens one interconnection agreement at a time, and the meter is running on everyone else's bill.