The Wrong Kind of Scarcity

Since Issue 05 we have followed the AI buildout to its physical limit, the megawatt, and this issue settles whether the data center is the reason the residential bill is rising. Over the past decade it is not.

Read the companion deck (PDF)→

Issue 06 and its companion are our attempt to understand one piece of a large and fast-moving question, how much of the rising electricity bill traces to AI data centers. We are not utility regulators or grid engineers, and a specialist may see nuance we have missed. What we have done is follow the primary evidence where it leads, the state price panels, the capacity-auction results, the interconnection queues, and the tariff filings, and flag every figure that is an estimate or rests on a single source. We hold the conclusions as a working hypothesis open to correction.

The decade's rate increases are delivery, and they predate the data center

The residential bill has been rising for reasons that predate the first AI data center, and the shape of the rise is the first clue. The US average was nearly flat across the 2010s and then rose 32 percent in nominal terms in the five years to 2025, much of it the general inflation of 2021 and 2022, but the increase was not evenly shared. Grouped by state and set against the CPI, the high-regulation states, California, New England, and New York, ran well ahead of inflation, toward 30 cents a kilowatt-hour, while the top data-center states, near 15, roughly kept pace with it (Exhibit 1). A break that sharp, and that concentrated in the policy states, is not a load story. It is the 2022 fuel and inflation shock landing on top of a delivery-cost cycle that was already under way.

EIA's utility accounts show a growing share of spending going into delivery while production's share falls (Exhibit 2). Over the decade to 2024, spending on transmission and distribution rose sharply while the cost of producing power, mostly fuel, was flat to lower. The longer series runs the same way. From 2003 to 2023, real generation spending fell about 24 percent while transmission nearly tripled and distribution rose 160 percent, to a record $50.9 billion in 2023. A commission approves those increases so a utility can recover the capital it puts in the ground and earn its allowed return, which means a rising bill is a cost-recovery decision on an aging grid, taken in a rate case, before any question of who draws the power.

On EIA's panel, ranked by the change in electricity sales from 2015 to 2024, the fastest-growing states rose about as much as the nation, Virginia 27 percent and Texas 29 against a US average near 30. The steepest increases landed on states with shrinking load and heavy climate and wildfire cost recovery, California 88 percent, Maine 56, Massachusetts 48 (Exhibit 3). The correlation between demand growth and price change is weak and slightly negative, near -0.37, so the states adding the most load saw marginally smaller increases than average, the same message the delivery accounts give.

A Rutgers panel of 22,834 areas across 24 states finds no statistically significant effect of a data center on residential bills at the level where rates are set, the utility territory, through 2024, a null its authors read as provisional, a "not yet," because the data ends before the hyperscaler campuses now measured in hundreds of megawatts. An instrumental-variables study by EPRI economists goes further, estimating that data centers modestly lowered average US retail rates from 2015 to 2024, through the same economies of scale in wires and generation our panel shows, though those authors caution that future supply constraints could reverse the dilution they measure. Both are working papers outside our own, and both point the same way. Whatever has lifted the residential bill over the decade, the data center is not it.

Where the data center raises the bill, it is recent and mostly forward

Where the data center does lift the bill, the effect is recent and concentrated. In the twelve months to March 2026 the states of the PJM data-center belt led the country on residential price, the District of Columbia up 23 percent, New Jersey 18, Ohio and Maryland 17, Virginia 15, against an EIA US-total figure near 9. We measure each belt state against the national power average, because general inflation and fuel costs move every state's bill together, so the distance a state opens beyond that average is the part specific to it. With the CPI up 3.3 percent over the same year, the national residential bill rose about 5 percent after inflation and the belt's leaders far more, from Virginia near 11 percent to the District of Columbia near 19. A few non-PJM states rose as fast on their own fuel and policy pressures, New Hampshire 18 and Washington 14, so the belt leads the country as a cluster, and it alone carries a capacity-market story. In every one of those belt states the annual pace had run near zero through the 2010s, so this is a sharp recent break (Exhibit 4).

The channel is the capacity auction, which is why a wholesale number reached retail bills first. PJM does not buy electricity in that auction. It buys a commitment to be available three years out, and it buys it for the whole region at a single clearing price. Every megawatt of capacity that clears is paid that price, and the total is billed back across all load in the zone. So when data-center demand lifts the amount of capacity the region must secure, the clearing price rises for everyone, and the charge lands on the household a full year or more before the plant it pays for is built.

PJM's most recent capacity auction, held in December 2025 for the 2027/28 delivery year, cleared at the FERC-approved cap of $333.44 per megawatt-day, a third straight record. The grid's independent market monitor attributes $6.5 billion of the $16.4 billion cleared cost, about 40 percent, to data-center demand (Exhibit 5). In the earlier 2025/26 auction, which cleared lower at $269.92, the monitor tied roughly $9.3 billion of the cost increase billed back to customers to the same source. That is scarcity pricing a constrained input, applied to firm capacity before the megawatts exist.

Virginia's legislative auditor, JLARC, commissioned an independent grid model and a cost-of-service study, and it found that current rates allocate costs correctly, so the shift onto other customers has been small so far. The larger cost is forward, and it lands even after the data center is billed its share, because meeting the load requires building at 3.6 gigawatts a year for fifteen years against a state single-year high of 2.2, and those net-new generation and transmission costs would not otherwise be incurred, so they raise the system cost for every customer (Exhibit 6). JLARC's consultant puts the effect on a typical Dominion residential customer at $14 to $37 a month in real terms by 2040.

Virginia is only the first state to have modeled that forward cost, and the rest of PJM faces the same turn as its own build accelerates. The delivery share that explains the past is no alibi for the future, because the load forces new transmission of its own, and that figure is generation and transmission together.

The demand in the queue is mostly optionality

The forecasts that justify the building overstate what will connect. Interconnection queues have filled with requests that behave like options. A developer files for one project across several territories, or files early for a site it may never build, and each request is counted. By March 2026 ERCOT's large-load interconnection queue had reached about 239 gigawatts of cumulative requests, roughly 2.8 times the grid's record peak of about 85 gigawatts (Exhibit 7). RMI estimates developers file five to ten times more requests than the projects that get built, and utilities serving eight of the ten largest data-center markets have already rewritten their rules to screen the speculative ones. The number under every capacity forecast is inflated by design, and the operators running the grids know it.

The grid has room if the load flexes

The demand that is real does not require the grid the forecasts assume, because a data center is an unusually flexible large load. Modeling by Duke's Nicholas Institute finds a large cushion. If new loads curtail just 0.5 percent of their annual energy, shed across about 177 hours a year in mostly partial reductions, the existing grid can absorb close to 100 gigawatts of new load with no new generation, 18 gigawatts of it in PJM alone (Exhibit 8). A training run can pause and shift in a way a steel mill or a city cannot, so the binding limit is institutional, a tariff that would pay a data center to flex, which most utilities have not written.

Inference is the exception. It serves live traffic and cannot curtail the way a training run can, so as it grows as a share of the load, the flexible fraction is smaller than the headline.

The fight has moved to the tariff docket and the federal seam

Because the exposure is real and mostly forward, the argument has moved from the auction to the rate case, where the terms of who pays are being set now. In July 2025 the Public Utilities Commission of Ohio approved a data-center tariff for AEP Ohio. It makes a large load pay for 85 percent of the capacity it reserves whether it draws the power or not, ramp over four years, post about three years of minimum charges as an exit fee, and put up cash collateral if it is rated below A-/A3, a bar that catches many investment-grade firms (Exhibit 9). These are the covenants a lender writes against a borrower it does not fully trust, and a commission is writing them for every other ratepayer.

The federal regulator has now moved too. On December 18, 2025 FERC found PJM's rules for large loads that co-locate at a power plant unjust, and ruled that netting a data center's on-site generation against its grid draw shifts cost onto other customers, ordering PJM and the transmission owners to file new terms by February 2026. Ohio sets the template for the retail rate case, and FERC sets it for the wholesale seam where the largest loads are trying to leave the grid altogether. How many states follow Ohio, and how the FERC compliance lands, decides whether the buildout's cost falls on the firms that ordered it or on the households around them.

The net load is a fraction of the queue, and the docket sets the cost

The figure that matters is the net load the grid will connect, well under the gross queue. Requests run five to ten times the projects built and Grid Strategies marks its own data-center number down about 40 percent, so what connects is a fraction of what is announced. The cost of serving it is decided in the tariff docket, where Ohio-style take-or-pay terms and the FERC co-location compliance set whether data centers or households carry the net-new build, and those terms move power costs more than any gigawatt headline. The exposure itself is a duration problem, firm forty-year generation built against speculative, sub-investment-grade load whose commitment is short, so take-or-pay minimums and offtaker credit are what separate a protected cash flow from a stranded one. Flexibility is the cheapest capacity on the system, since a load willing to curtail 0.5 percent of its annual energy unlocks much of the roughly 100 gigawatts of headroom on the grid already built.

Next the series turns from the power asset to the hardware that draws on it. Issue 07 finds that a GPU's economic life is set by the price of power: where watts are cheap the old chip keeps earning, and where they are scarce, the condition this issue has traced, a kilowatt earns several times more on new silicon and the old chip gives way. That economic life is the collateral under the roughly $450 billion of data-center debt now sitting in private credit, so the power constraint priced here is the variable that decides how fast that collateral erodes.

Residential price levels and the cross-state change in electricity sales from EIA state panels, with Exhibit 1 indexed to 2001 = 100 against the CPI; the delivery-versus-production spending split from EIA's utility financial accounts, with the 2003 to 2023 series (real generation down about 24 percent, transmission nearly tripled, distribution up 160 percent to a record $50.9 billion in 2023); the two working papers are a Rutgers panel of 22,834 areas across 24 states and an instrumental-variables study by EPRI economists, both through 2024; PJM capacity-auction results, the December 2025 auction for 2027/28 clearing at $333.44 per megawatt-day and the earlier 2025/26 auction at $269.92, with the data-center attribution from PJM's Independent Market Monitor; the Virginia forward-cost figures from the JLARC study and its consultant's cost-of-service and grid model; interconnection-request multiples from RMI; the ERCOT large-load queue of about 239 gigawatts as of March 2026 from ERCOT; the flexibility headroom from Duke University's Nicholas Institute; the AEP Ohio data-center tariff approved by the Public Utilities Commission of Ohio in July 2025; and the FERC order of December 18, 2025 on co-located large loads. Demand-forecast figures are Grid Strategies'.

The two academic results are working papers whose data ends in 2024 and predates the largest current campuses; their authors flag the limitation and the possibility of reversal, and we carry that caveat. The market-monitor attribution of auction cost to data-center demand is an estimate, as is the $14-to-$37 Virginia figure, a modeled real-terms projection to 2040. Queue totals are gross request counts, not committed load. Second Order is an independent research briefing, for discussion only and not investment advice.