The AI power shortage arrives as a household bill before it arrives as a data center delay

Two things are true at once this August. The largest computing build-out in history cannot get the electricity it ordered, and yet almost no data center has actually gone dark waiting for it. The shortage has found a release valve: the cost of the queue, the wires and the spare capacity is landing on household electric bills years before a single delayed server rack becomes news. In Virginia this month, Governor Abigail Spanberger won a State Corporation Commission order requiring data centers to pay the full cost of transmission lines built only for them, a ruling her office says will save ratepayers hundreds of millions of dollars (WSET, Aug 6). The order exists because until now, they did not.

Start with the money already spent. PJM Interconnection runs the wholesale power market across all or parts of thirteen states, including Northern Virginia, the densest data center cluster on earth. Its independent market monitor, Monitoring Analytics, calculated that including existing and forecast data center demand pushed up system costs by 23.1 billion dollars across three capacity auctions (Monitoring Analytics via remio summary, Jul 16). Those auction obligations run through May 2028 and regulators cannot unwind them. Amazon, Google, Meta, Microsoft, OpenAI, Oracle and xAI have each promised their buildings will not raise household bills. A chunk of the bill has already been committed.

The household side is measurable too. The National Energy Assistance Directors Association expects the average American home to spend 792 dollars on electricity from June through September 2026, up 10.5 percent from 717 dollars last summer, and nearly 40 percent above the 2020 level (NEADA via The Queen Zone, Jul 21). The Energy Information Administration projects residential prices rising 5.7 percent during 2026, to about 18.51 cents per kilowatt-hour in the third quarter (EIA via The Queen Zone, Jul 21). Heat and older rates explain most of that. Data centers add pressure exactly where the grid is tightest, which is why the same summer bill looks different in Ashburn, Virginia than in Vermont.

Separate the trigger from the pressure. The trigger is this year's auction cycle and the filings that followed it: PJM's capacity auctions cleared at the regulatory price cap of 329.17 dollars per megawatt-day, roughly ten times what capacity cost two years earlier at 28.92 dollars (PJM auction results, 2025-26 delivery year). The slow pressure underneath is simpler. Data centers account for roughly half of all new American electricity demand (Fortune via SmarterArticles), and Berkeley Lab projects they could take 11.8 percent of all US electricity by 2030 under its reference case (Berkeley Lab via The Queen Zone, Jul 21). Generators, wire owners and fuel suppliers collect the difference between old demand and new. Households currently lend them the capital.

Name the actors. The hyperscalers want speed; every quarter of delay hands a rival share, so they sign whatever rate class lets the substation move. The utilities, Dominion Energy foremost, want to grow the rate base, and a proposed 67 billion dollar merger with NextEra Energy would create the world's largest regulated utility (WSET, Aug 6). State regulators want re-election without headlines about shutoff notices. Grid operators like PJM want reliability and price signals that actually bite. And Washington wants both the AI build-out and the votes of people who pay its externalities, which is why President Trump pledged in July that big tech, not families, will carry data center energy costs (Reuters via US News, Jul 22).

The compute shortage is being paid off in monthly installments by people who will never train a model.

The politics have already moved from complaint to consequence. Voters in Independence, Missouri, a city of 121,000 near Kansas City, vote in September on removing their city manager over a data center land deal, and Yukon, Oklahoma holds a mayoral recall in November for the same reason (New York Times, Aug 19). Opposition to new facilities has jumped sharply among both Democrats and Republicans in recent polling (University of Pennsylvania survey via WHYY). In Georgia last November, Democrats ousted two Republican utility commissioners over electric bills. When the bill arrives before the benefit, the politician who signed off pays next.

History offers one bounded comparison: the Pacific Northwest aluminum smelters around 2000. Smelters built against cheap federal hydropower shut down when California's crisis pulled spot prices up, and the region learned that industrial loads hooked onto subsidized public power leave ratepayers holding the fixed costs when the arithmetic turns. The mechanism rhymes. A large customer negotiates a preferential rate, the utility socializes the network upgrades, and when prices spike the small customers discover they funded the optionality. What differs this time is direction: the AI load is not leaving, it is compounding, which means the subsidy question gets decided while the load keeps growing rather than after it vanishes.

The counter-case argues the other way, and it is not weak. Hyperscalers can and increasingly do buy their own electrons: Microsoft contracted to restart Three Mile Island's reactor for its own load, and Amazon, Google and Microsoft are funding new nuclear and renewables directly rather than through the tariff. In ERCOT, Texas's isolated grid, large flexible loads were designed in from the start to carry their own weight, which is why Governor Greg Abbott could pause new data center approvals there on August 3 without a consumer revolt (TechGolly, Aug 3). Where the market forces the buyer to bring its own generation, the household bill stays clean. The problem is regional, concentrated where legacy rules let big loads ride common tariffs.

Walk the consequences forward. First order: states copy Virginia. The SCC's tariff logic is portable, and Dominion's own proposal for a dedicated data center rate class shows utilities pre-empting regulators rather than waiting for orders. Second order: hyperscalers accelerate self-supply, buying reactors, gas turbines and long-dated renewable contracts, shifting billions from regulated utilities to unregulated power producers and equipment makers like GE Vernova. Third order: if data centers exit the shared tariff, the fixed grid costs they helped justify spread across fewer remaining customers, raising ordinary rates further and feeding the exact backlash everyone is trying to avoid.

Who pays is now legible: residential and small-business ratepayers inside PJM and other tight markets, through capacity charges locked through May 2028. Who profits is legible too: incumbent generators who cleared the auctions, transmission owners whose rate base grew, and gas turbine makers selling into a shortage. The open fight is over the margin between them, and Virginia just moved that line several hundred million dollars toward households.