The Ratepayer's Subsidy

The Paranoidist | Issue #21 By Paul Morin | June 30, 2026

There is a number coming in two weeks that almost everyone in the power industry will read as the market working, and the reading will be so natural that the second meaning passes without comment. The offer window for the 2028 to 2029 capacity auction run by PJM Interconnection, the grid operator for much of the eastern United States, opens today and closes July 7, with the result posted after the close on July 14. If the last three auctions are any guide, it will clear at or near its price cap, and the commentary will say what it has said each time: that the high price is the system doing its job, a scarcity signal calling new generation onto a grid that needs it, the visible proof that the artificial intelligence build-out is arriving as real, healthy, bankable demand. PJM keeps the lights on for more than 67 million people across 13 states and the District of Columbia, and it has spent two years watching its capacity prices climb and its reserve margin thin; the story it tells, the story many of the utilities and developers tell with it, is that this is simply what growth looks like before supply catches up.

The demand is, in part, genuinely real, and I want to concede that before I question it, because data centers are not even the largest force pushing on the grid. The International Energy Agency estimates they account for only about a tenth of global electricity demand growth this decade, less than the growth from electric vehicles, from electrified heating, and from the driver a warming climate makes relentless: air conditioning. Space cooling is now the fastest growing electricity use in the world's buildings, climbing nearly 4 percent a year, and on current trends the power it draws is set to roughly triple by mid-century, eventually outpacing data centers outright. That load is the opposite of speculative. No one cancels a heat wave. So the question this issue raises is not whether the grid must grow, because it plainly must. It is narrower and sharper: of the growth that PJM is now pricing and building against, which part is the firm kind that arrives whether anyone wants it to, and which part is a forecast that can walk away. The thing that keeps me up is not that the demand is real. It is that a large and growing share of the demand setting that price, and the build it pays for, has not been built, may never be built, and is ordered by a handful of tenants who can delay it, shrink it, or cancel it long after the rest of us have paid to wire it.

Issue #20, "The Look-Through," ended on a falling inflation number the Fed is built to read the flattering way, and on the labor displacement hiding behind it: demand that quietly leaves the economy and never shows up in the index. This issue is the same shape of problem turned inside out. There the worry was demand that vanishes; here it is demand that may never arrive. In both, a forecast that will not resolve cleanly is being treated as settled, and in both the cost of the error lands on a party that did not place the bet. Call the two readings what they are. Firm demand is load the grid is obligated to serve, the rising summer peak of a hotter decade that no household cancels, the factory and the heat pump that show up whether or not anyone wants them to; the kind that justifies building generation and transmission and pricing capacity high enough to procure it; if it is real, the build is necessary and the cost is fair. Phantom demand is forecast load from facilities not yet built, ordered by a small group of tenants who can walk, while the reliability requirement it sets, the capacity price it lifts, and the wires it triggers are locked in and spread across every ratepayer; if it evaporates, the public has already paid for a position someone else opened and can close at will. The same forecast megawatt is a growth dividend or a contingent liability, and the entire question is which one you are looking at.

Start with the official version, because on its own terms it is coherent and largely correct. PJM runs a capacity market designed to send a price signal about three years before the power is needed, so that generators see scarcity coming and build into it. For most of the last decade that signal sat near the floor. Then it broke loose. The auction for the 2025 to 2026 delivery year cleared at roughly 270 dollars per megawatt-day for most of the footprint, up from under 30 dollars the year before, with two zones touching the high 400s. The next two auctions cleared at their caps, about 329 and then about 333 dollars per megawatt-day, the third record in a row. PJM's own forecast attributes almost all of the recent jump in peak load to data centers; in the December auction, of roughly 5,250 megawatts of forecast load growth, nearly 5,100 megawatts traced to them.

Read generously, this is exactly the machine working. High prices are not a malfunction; they are the invitation. They tell anyone who can pour a foundation that there is money in building a power plant in PJM, and the grid operator can point to early signs of supply responding, withdrawn retirements and new project interest, as evidence the signal is doing its work. The load behind the price is not imaginary in the aggregate: the construction is visible, the chips are real, the capital expenditure is enormous and rising. And the political system has supplied a backstop for the household. In March, seven of the largest hyperscalers and AI companies signed a Ratepayer Protection Pledge at the White House, agreeing to build, bring, or buy the generation their data centers need and to pay for the grid upgrades to connect them. The pledge even commits the signatories to pay for the power they reserve "whether they use the electricity or not," which, if it holds, is precisely the firewall the phantom problem would require. On the comfort story's assumptions, the system is pricing real scarcity, the price is summoning real supply, and the people who created the demand will pay for it. There is nothing to worry about that the machine is not already handling.

Now read the same auction from the other side. PJM's independent market monitor, the entity whose job is to tell the truth about the market's plumbing, has said in plain language that data center load growth is "the primary reason" for the recent capacity conditions: the load forecast, the tight balance, the shortfall, and the high prices. Across the last three capacity auctions, by the monitor's tally, forecast data center load drove about 45 percent of more than 47 billion dollars in capacity costs. And here is the detail the comfort story leaves out. In the December auction, of roughly 6.5 billion dollars in data center capacity costs, about 6.2 billion was for data centers that have not been built, that merely could come online by the delivery year. The price everyone pays is being set, in large part, by load that exists today only as an interconnection request and a line on a forecast.

This is the asymmetry that makes the megawatt a trap. A forecast data center counts as firm demand for the purpose of setting the reliability requirement, so the system is obligated to build for it; it counts as firm demand for the purpose of setting the capacity price, so every ratepayer in the zone is obligated to pay for it; but it remains, for the tenant who ordered it, an option. The tenant can phase it, downsize it, delay it, or cancel it as model economics and chip supply shift under it. The reliability requirement does not un-tighten when the project slips. The transmission line, once approved and built, does not un-build. The capacity already procured does not refund. PJM itself has acknowledged the softness in the very number that drives all of this: after the December auction it flagged that its next load forecast could come in "significantly lower" once large loads were vetted more strictly. The ruler that sets the price and the build is a forecast that the forecaster admits may be too high.

And consider what the build actually is, because the mix decides both how fast it can arrive and how long it lasts. The capacity that clears these auctions is overwhelmingly firm and long-lived. In the most recent one, natural gas supplied about 43 percent of it, nuclear about 21 percent, and coal about 20 percent, with wind, solar, and hydro together in the single digits, and essentially all of the new build is gas. Gas is now the choke point. The few manufacturers who make heavy-duty turbines are booked years out; the largest turbine maker alone reported an order backlog near 100 gigawatts early this year, with lead times of roughly three years and many delivery slots already sold into the end of the decade, and the transformers and switchgear needed to connect any of it face multi-year waits of their own. So the asymmetry has a clock on it. A data center can be ordered, and cancelled, in a quarter. The gas plant ordered to serve it takes three to five years to arrive and then runs, and earns its return from ratepayers, for forty. The thing that can vanish moves fast; the thing built to chase it moves slow and stays long, which is exactly why a high price cannot summon supply quickly enough to relieve the squeeze, and exactly why a build triggered by demand that then evaporates is so expensive to have been wrong about.

And the resemblance has the same cruel feature I flagged in the disinflation case. There, the worse the labor displacement, the better the inflation number looked, so the gauge improved fastest in the scenario the Fed should fear most. Here, the more speculative the data center demand, the stronger the scarcity signal it throws off, because a larger forecast tightens the supply-demand balance and lifts the clearing price regardless of whether the load ever energizes. The capacity market's most alarming readings, the cap-clearing prices and the reserve shortfall, are loudest precisely when a large share of the demand behind them is least certain to show up. A market built to price real scarcity cannot, from the inside, tell scarcity driven by load that will arrive from scarcity driven by load that was merely requested.

This is where the cooling story comes back, and it squares the circle rather than breaking it. Nationally and globally, cooling and electrification are the larger tide; even in the United States, where data centers are expected to drive roughly half of all demand growth this decade, heat pumps and electric vehicles and a warming-driven cooling load account for much of the rest, and they are firm. The difference between them and the data centers is tempo. Cooling lifts the baseline peak gradually, year after warming year, and the forecast absorbs it smoothly. Data centers arrived as a step change, a sudden jump of several thousand megawatts in a single forecast cycle that outran the supply queue and tightened the margin the auction prices. So both things are true at once. The firm tide is cooling and electrification; the surge that spiked this price and triggered this build is data centers. And the firm tide is exactly what gives the surge its cover. When rising cooling demand makes it obviously, physically true that the grid must grow, the speculative data center share gets built and socialized alongside the genuine need, because no utility wants to be the one that under-built into a heat wave. The phantom does not have to win the argument. It only has to ride in on the coattails of the load that did.

The reason this is a fragility and not just an accounting quirk is the way the cost is socialized. Capacity charges, the transmission build, and, increasingly, rate-based generation flow through to the general body of ratepayers in a zone, not only to the customer who triggered them. The market monitor has made the point about as bluntly as a regulator-adjacent body can: allowing utilities to build generation for data centers under traditional cost-of-service regulation "shifts costs and risks to other customers." That is the whole thesis in one clause. The tenant holds the option; the ratepayer holds the obligation. The firm whose forecast lifted the price can revise that forecast next quarter; the household whose bill rose to pay the price cannot revise its bill.

Set against this is the pledge, and the pledge is the part of the comfort story that most deserves to be taken seriously, which is why I want to be precise about what it is and is not. It is voluntary. It is non-binding. The White House that convened it has no authority over electricity markets or state rate cases, a limit the administration's own people have acknowledged even while celebrating the commitment. Its operative promises, separate rate structures and full cost recovery from the data center, have to be negotiated and approved one utility and one state commission at a time, in exactly the venues where the cost-shift is hardest to police and easiest to litigate. A promise to pay "whether they use the electricity or not" is only as good as the contract that encodes it and the regulator that enforces it, and most of those contracts do not yet exist. The pledge is a statement of intent layered on top of a market structure that, absent the pledge, defaults to spreading the cost. Intentions are not firewalls.

Then there is the matter of who the tenants are. The largest of them are, by their own internal numbers, deeply unprofitable. OpenAI's own projections, as reported by The Information, point to a loss on the order of 14 billion dollars in 2026 and cumulative losses in the tens of billions before any profit arrives near the end of the decade. I am not predicting any particular firm fails. I am observing that demand anchored to tenants burning capital at that rate is demand with a wider error bar than a steel mill's or a city's, both in whether the forecast materializes and in whether, if a tenant retrenches, the long-term obligation behind its reserved capacity is actually collectible. The grid is being built against the demand curve of a cohort whose own business models are the least settled thing in the economy.

What lifts this above a regional rate dispute is that three things are arriving in the same window and feeding one another. The first is the question I have circled since the spring: how much of AI's apparent demand is genuine end demand and how much is the manufactured demand of a small group of firms financing one another's build-outs. If even part of the compute demand is circular, then the load forecast resting on top of it inherits that uncertainty, and the grid is the one piece of the AI capital stack that cannot be unwound. A model contract can be renegotiated and a chip order can be cancelled, but a transmission corridor and a gas plant, once built into the rate base, are there for forty years.

The second is the politics of the electric bill. By the Energy Information Administration's data, residential electricity prices are up more than a third since 2020, and they have kept rising into an administration that promised to bring them down. The political response to that pressure has been the pledge, a voluntary instrument, and a more aggressive move that tells you how the government itself reads the situation: in January the administration intervened directly in what its own fact sheet called the "failed" PJM market to force new power-plant development. When the body promising to protect ratepayers also describes the market protecting them as failed, and reaches for emergency build-out as the fix, the optimistic read on the auction price, that it is the market healthily summoning supply, sits awkwardly next to the government's own diagnosis.

The third is the thread from Issue #20. The same technology shows up in two different demand-forecasting problems at once. On the labor side it removes income and the demand that income would have created, a demand shadow the inflation gauge cannot see. On the power side it adds load and the build that load demands, a demand surge the grid must price and may overbuild for. One technology, two forecasts that will not resolve to a probability, and in both cases the institution is structurally inclined to read the number the flattering way and to place the cost of being wrong on the party with the least say: the worker who never gets hired, the ratepayer who never ordered a data center. The declaration, the look-through, and now the subsidy are not three economies. They are one build-out seen from three windows, and the windows look onto the same room.

The honest forecast comes in three parts, and only the first is close to mechanical. Near term, the comfort story holds and looks vindicated. The 2028 to 2029 auction clears under its price cap, the headline number lands roughly where the last two did, and because capacity is a relatively small slice of a retail bill and a price collar is in place, household bills do not spike from this auction alone. PJM can correctly say the shortfall does not mean the lights go out, and the build proceeds. Nothing in this stage looks like a problem, which is the point: the most likely near-term outcome is that phantom demand keeps being priced and built as firm demand, because the two are genuinely hard to separate in advance and every incentive favors treating the forecast as real.

The medium term is where the uncertainty lives, and it turns on two questions that have no date on them. The first is the conversion rate: how much of the forecast large load actually energizes versus quietly disappears from the queue as tenants revise. The second is whether the cost-shift gets policed in time, whether enough states stand up enforceable separate rate classes that survive the rate case and the appeal, before the socialized build is already in the rate base. Neither resolves on a schedule. Both grow more consequential the longer the auctions keep pricing speculative load as firm and the longer the build keeps pouring concrete against it.

The counter-move, when it comes, will not be a clean reversal. Expect instead a slow reallocation dressed as something else: capacity and transmission built for tenants who slipped, quietly relabeled as "system reliability" and "grid modernization" that benefits everyone, the stranded portion folded into the general rate base where no single decision is ever blamed for it. So the base case is not a blackout and not a crash. It is a legitimized overbuild whose cost migrates onto the household by degrees, holding as long as a short list of gauges stays quiet. Watch the gap between forecast large load and large load actually interconnected and energized, because that gap is the phantom resolving one way or the other. Watch the share of capacity and transmission cost allocated specifically to data center load versus socialized to everyone. Watch whether states actually approve separate data center rate classes that hold up, rather than pledges that do not. Watch the tenants' own capital spending and losses, because if the AI labs retrench, the load forecast is the first thing that should fall. Separate the weather-normalized peak from the data center forecast, because the first is the firm tide that comes regardless and the second is the part that can evaporate. And watch whether PJM's looming shortfalls trigger the backstop procurement it has floated for the autumn, because that is the moment the system stops pretending the market is clearing on its own. The scoreboard will keep reading as healthy growth right up until one of those gauges says the megawatts everyone paid for never arrived.

Four areas of board exposure, mapped against the phantom-demand reading rather than the comfort story.

Regulated utilities and investors holding utility equity and debt. This is where the reading cuts first and in two directions. A utility that rate-bases generation and transmission against forecast data center load books growth today and carries stranded-asset and disallowance risk tomorrow if the load slips and a commission later refuses to let it recover the cost from the general body of ratepayers. The same buildout that looks like the best rate-base growth story in a generation is also the largest concentration of demand-forecast risk most of these companies have ever taken, tied to a single, financially unsettled customer class. The question for the board is whether your capital plan treats forecast large load as firm for planning and as contingent for risk, or whether it has quietly booked the same megawatt as certain in both places.

Energy-intensive industrials and large commercial power buyers. If you run manufacturing, cold storage, chemicals, or any load that competes for the same interconnection queue and sits in the same socialized cost pool, the data center surge is your cost even though it is not your project. You may be paying a capacity and transmission bill inflated by forecast load you had no part in creating, and waiting behind it for an interconnection you need. The question is whether your energy budget and siting decisions account for the possibility that you are subsidizing a neighbor's speculative build, and whether a long-term supply contract now is cheaper than the socialized path later.

Data center developers, infrastructure funds, and their lenders. The exposure here is the mirror image and just as real. If tenant demand is partly speculative and partly tied to unprofitable counterparties, then the take-or-pay assumptions, the long-dated power contracts, and the debt underwritten against them carry counterparty and stranded-capacity risk that a "demand is infinite" thesis hides. The question for the investment committee is whether your underwriting stress-tests a world in which a marquee tenant phases or cancels, and who, in that world, is left holding the power contract and the substation.

Anyone underwriting or insuring the AI build-out. Lenders, insurers, and pension allocators sitting under this capital stack are all, in the end, pricing the same assumption: that the forecast demand is firm. That assumption is the single load-bearing beam under utility valuations, data center project finance, and a meaningful slice of the equity market's AI premium. The question is whether your models distinguish demand the grid must serve from demand that was merely requested, because the difference is the difference between an infrastructure asset and a contingent claim, and the top line looks the same either way until it doesn't.

The demand may simply be firm. AI compute demand has done nothing but grow, inference use is rising, and the serious risk may be that the forecasts are too low rather than too high; on that reading, the phantom share converts to real megawatts, the build is necessary, the high prices were an accurate signal, and a grid that hedged against demand that then arrived would have failed worse than one that overbuilt. If that is the world, the cost was not a subsidy at all, it was prudent investment, and I am pattern-matching prudence to recklessness.

The pledge and the rate cases may work. The signatories have committed to pay for power "whether they use it or not," tenants are increasingly signing long-dated bilateral contracts, several states are advancing legislation to wall off data center costs, and the market monitor's proposed reforms are designed precisely to stop the cost-shift. If those instruments harden into enforceable, tenant-specific rate structures faster than the build socializes, the firewall holds and the household is protected, which is exactly what the architecture is supposed to deliver.

The bill impact may be smaller than the rhetoric. PJM itself stresses that capacity is a modest part of a retail bill, that a price collar limits the damage, and that the reserve shortfall does not mean outages. If most of the cost is absorbed by collars, self-supply, and the data centers' own contracts, then the "subsidy" is real at the margin but too small to be the systemic fragility I am describing, and I am inflating a rate-design wrinkle into a thesis.

And the overbuild, if it happens, may not be waste, for a reason this issue has already supplied. The firm tide is still rising. Cooling demand climbs with every hotter summer, electrified heating and transport add load that does not cancel, and a grid that built generation and wires for data centers that then slipped may find the cooling-driven peak and the heat pumps have quietly absorbed the headroom. On a tightening, warming grid, spare capacity is not obviously stranded; it is insurance the next heat wave will cash. If the leftover infrastructure is genuinely additive, then "stranded" is the wrong word, and what I am calling a subsidy is closer to a public investment that a private actor happened to trigger.

The DeepStrategy.ai signature method requires sorting risk, which is quantifiable, from uncertainty, which is not, at every major analysis. The numbers here are risk. The capacity clearing prices and their caps, the reserve-margin shortfall of roughly 6.6 gigawatts and the 14.8 percent margin against a 20 percent target, the data center share of capacity costs, the gap between forecast large load and energized large load, the tenants' reported capital spending and losses, and the residential rate trajectory can all be tracked, charted, and put on a board dashboard tomorrow. They will tell you the size and direction of the move.

What will not resolve to a figure is everything that decides what the move means. Whether the forecast load is firm or phantom, and in what mix, is an uncertainty, because the same megawatt sets the same price either way. Whether the tenants honor commitments made through unprofitable years, and whether a voluntary pledge becomes an enforceable obligation before the build is socialized, is an uncertainty about contracts and regulators that no capacity number reports. Whether AI's end demand is genuine or partly circular is an uncertainty sitting underneath all the others, and it is the one that determines whether the grid is building infrastructure or a monument. The clearing price is a risk and tells you what the auction paid; whether that payment bought necessary supply or financed a bet a tenant can abandon is an uncertainty, and the two answers point in opposite directions from the same number. As always, the precise, trackable figure is the announcement, and the part that actually moves the institution is the part that will not reduce to a probability.

The institution that consumes the analytical process as preparation for multiple futures has what the forecast cannot provide: adaptability. When the auction clears in two weeks, the headlines will say the market is working and the AI age is paying its way, and the pledge will be cited as proof that the data centers are subsidizing a stronger grid for everyone. The architecture says something quieter and harder to celebrate. The pledge points one way; the meter points the other. The same forecast megawatt is being counted as firm when it sets the price and the build, and treated as optional by the only party that can cancel it, and the difference between those two readings is being financed, by default, by the household that never ordered a server.

The boards best positioned for this are not the ones reading the cap-clearing price as healthy growth. They are the ones asking, of every megawatt of forecast demand under their plans and their portfolios, whether it is firm or phantom, and who holds it if it turns out to be phantom, and pricing that answer before the rate case tells them which one it was.

A megawatt that may never arrive. A price that everyone pays anyway. And a subsidy running quietly in the direction nobody named, from the ratepayer who cannot cancel to the tenant who can.

The Paranoidist publishes weekly, with flash issues when events warrant.

If this changed how you think about one thing, consider subscribing. If it didn't, tell me what I'm missing.

Paul Morin is the founder of DeepStrategy.ai, author of Uncertainty: When Risk Is Not Enough (a guide to decision-making when probabilities fail), and publisher of The Paranoidist, BoardroomRadar, and ScenarioWatch. He has spent more than three decades in entrepreneurship, finance, risk management, and insurance, which is why he worries about the things that keep other people awake at night.

Researched, written, and edited in collaboration with Claude by Anthropic.