Introduction: The 700-Gigawatt Question
In the first days of September 2026, one number captured the strange new economics of America’s artificial-intelligence infrastructure boom better than any earnings call, any chip announcement, or any model release: 700 gigawatts.
A Reuters review published on September 1, 2026 found that requests from very large electricity users—predominantly datacenters—had exceeded 700 gigawatts across portions of the Midwest, the Mid-Atlantic and the South, which is more than ten times industry estimates of the electricity currently consumed by all United States datacenters combined.[1] To make the number vivid, Reuters observed that datacenters had requested roughly as much electricity across the middle swath of the United States as it takes to power every home in the country. The requests were so large that, taken literally, they implied an extraordinary industrial transformation occurring almost simultaneously across multiple American electricity markets—an electrification event without precedent in the postwar era, compressed into a planning window of less than a decade.
Yet much of the requested load may never appear. Some proposed projects are duplicated across jurisdictions. Some developers have sought electricity from several locations while intending to construct only one facility. Some have land but no customer. Others have an announced customer but no secured generation, no final permit, insufficient financing, or an uncertain construction schedule. Still others may be little more than attempts to secure a valuable position in a suddenly scarce electricity queue—a placeholder in line for a commodity whose scarcity has made the place in line itself an asset. Reuters described the growing problem as “ghost demand,” and reported that consumer advocates and regulators had warned that many requests are likely to be duplicative or filed by companies without the funds or expertise to see the projects through, threatening to stymie the grid planning that is crucial to keeping the lights on.[1]
Texas illustrates the scale of the problem with almost theatrical clarity. Electricity requests from datacenters and other large users connected to the ERCOT grid increased from roughly 48 gigawatts in 2023 to more than 474 gigawatts by August 2026—a nearly tenfold expansion of the queue in barely three years.[1] Governor Greg Abbott consequently ordered the Public Utility Commission of Texas and ERCOT, in a letter dated August 3, 2026, to conduct a “comprehensive verification and audit” of all datacenters advancing through the interconnection process before any additional projects could move forward, effectively making Texas the first major American datacenter hub to freeze new grid connections while it investigates what is actually behind them.[3] The Governor’s office noted that the approximately 474 gigawatts then under consideration was more than five times ERCOT’s record peak electricity demand, and that approximately ninety percent of the new power requests were attributable to datacenters.[3][4]
The chairman of the Public Utility Commission of Texas, Thomas Gleeson, had already articulated the underlying dilemma months earlier, at an industry conference in March 2026, when he warned that when regulators cannot distinguish genuine projects from speculative ones,
“you really don’t know how to build the infrastructure for it.” [2]
— Thomas Gleeson, Chairman, Public Utility Commission of Texas
Something unusual has therefore happened to electricity planning. A grid operator can now watch hundreds of gigawatts appear in its interconnection pipeline without knowing which gigawatts represent future operating factories of intelligence and which represent merely aspirations, options, duplicate applications, speculative reservations, or projects that will eventually disappear without consuming a single kilowatt-hour.
That distinction matters because electricity infrastructure cannot be constructed as an abstraction. A utility responding to a supposed one-gigawatt datacenter may need substations, transformers, high-voltage transmission lines, new or contracted generation, gas pipelines, grid-scale batteries and potentially billions of dollars of supporting infrastructure. Those assets can remain in service—and in the rate base—for decades. If the datacenter arrives, they may prove essential to the American AI economy. If it does not, somebody must still pay for them, and that somebody is very often the ordinary residential and commercial ratepayer who never asked for the datacenter in the first place.
The error works in both directions, and this bidirectionality is the analytical heart of the present paper. Assume too much of the announced demand is genuine, and utilities risk overbuilding expensive, long-lived infrastructure for customers who never materialize, with the stranded costs migrating quietly onto household electricity bills. Assume too much is speculative, and America risks underbuilding, leaving genuine AI facilities without electricity, delaying investment, pushing computation offshore, and weakening the country’s ability to compete in the global artificial-intelligence economy at precisely the moment that competition is intensifying.
This is becoming especially consequential because artificial intelligence has changed the scale of the individual electricity customer. An ordinary commercial development might request several megawatts; a large factory might request tens of megawatts. A hyperscale AI campus can seek hundreds of megawatts or multiple gigawatts—the load of a mid-sized city arriving as a single customer with a single set of corporate intentions. A small number of siting decisions by OpenAI, Microsoft, Amazon Web Services, Google, Meta, xAI, Oracle, Anthropic’s infrastructure partners, or other major computing companies can therefore alter the long-range electricity forecast of an entire utility territory. When the marginal customer is the size of Houston, forecasting stops being statistics and starts being intelligence-gathering.
The financial stakes behind these requests are no longer hypothetical. Reuters noted that Big Tech’s planned AI datacenter spending would top 700 billion dollars in 2026 alone, and the second-quarter 2026 earnings season confirmed the trajectory: Amazon raised its 2026 capital-expenditure guidance to approximately 220 billion dollars, Alphabet lifted its range to 195–205 billion dollars, Meta raised the floor of its guidance to 130–145 billion dollars, and Microsoft’s calendar-year trajectory approached 190 billion dollars, bringing the four largest hyperscalers to roughly 725–760 billion dollars of planned 2026 capital expenditure, an increase of nearly eighty percent over the approximately 410 billion dollars they spent in 2025.[34][35][38] Goldman Sachs now projects a combined 5.3 trillion dollars of capital expenditure for the four largest hyperscalers between fiscal 2025 and fiscal 2030.[40] Much of that capital is real, contracted, and already being poured into concrete and copper. But the electricity requests surrounding it exceed even these staggering sums by a wide margin, and the gap between the two is where the ghosts live.
Reuters found direct evidence that financial screening can substantially change utility forecasts. Chicago-based Exelon reduced its estimate of high-probability datacenter demand by roughly forty percent, to about 11 gigawatts, after adopting stricter collateral requirements.[1] In Ohio, AEP Ohio’s prospective datacenter pipeline fell by more than half after new requirements introduced stronger financial commitments and connection-study fees of as much as 100,000 dollars.[1][2] Yet—and this is fundamental—those reductions do not prove that the AI electricity boom itself is fictitious. Reuters also noted that grids can remain overwhelmed even after questionable projects are removed, because the remaining substantiated demand is still enormous.[1] PJM, the nation’s largest grid operator, continues to maintain that genuine load is arriving faster than necessary generation in parts of its thirteen-state footprint.
That distinction is fundamental to this paper. The problem is not that hundreds of gigawatts of future AI demand are necessarily fake. The problem is that America’s existing institutions have enormous difficulty assigning probabilities to them.
Traditional load forecasting was largely built around demographic growth, industrial activity, weather patterns, household formation, appliance saturation and relatively gradual changes in electricity consumption—phenomena that evolve over years and decades and that reveal themselves through census data and historical consumption curves. The AI economy introduces a different creature entirely: enormous, geographically mobile, capital-intensive loads capable of entering several interconnection queues simultaneously while their developers negotiate land, tax incentives, GPUs, power contracts and anchor customers across competing jurisdictions. The forecasting tools of the twentieth-century utility were never designed to price the strategic behavior of twenty-first-century hyperscalers.
The electricity system is therefore confronting not merely a generation shortage but an information problem—and information problems, as economists have understood since the foundational work on markets with asymmetric information, can distort investment just as severely as physical scarcity can.
This paper calls the resulting phenomenon Ghost Megawatts.
A Ghost Megawatt is not simply an imaginary datacenter. It is the portion of announced or requested electrical demand that appears in infrastructure planning but has a low probability of becoming an energized, sustained and paying electricity load within the period for which the grid is being planned. Conceptually:
Ghost Megawatts = Requested Megawatts − Probability-Weighted Credible Megawatts
For an individual project:
Ghost MW(i) = Requested MW(i) × (1 − Probability of Commercial Energization(i))
A 1,000-megawatt request with an estimated ninety percent probability of reaching commercial operation would therefore contain approximately 100 Ghost Megawatts. A different 1,000-megawatt proposal with uncertain financing, no anchor tenant, no generation contract, no permits and only a twenty percent probability of completion would contain approximately 800 Ghost Megawatts. Both appear identical in a headline queue total; they are radically different objects in economic reality.
The purpose of this formalization is not to pretend that regulators can calculate these probabilities with actuarial precision. They cannot, and this paper will argue that false precision is itself a danger. The purpose is to replace the binary distinction between “real” and “fake” projects—a distinction that flatters neither the complexity of datacenter development nor the intelligence of the public debate—with a more useful spectrum of credibility that can be measured, disclosed, compared across jurisdictions, and improved over time.
That spectrum is becoming one of the most important planning problems in what this author has elsewhere described as the Five-Layer AI Economy: Layer 1, Energy; Layer 2, Chips; Layer 3, Datacenters; Layer 4, Models; and Layer 5, Applications and Agentic Systems. Increasing demand at Layers 4 and 5 causes companies to reserve accelerators at Layer 2, campuses at Layer 3, and ultimately electricity at Layer 1. But the physical infrastructure necessary for Layer 1 must often be planned five to fifteen years before anyone knows whether today’s forecasts of model usage, inference demand and agentic workloads will materialize exactly as projected. Uncertainty at the top of the stack is passed downward, layer by layer, until it lands on the entity least able to absorb it quickly: the electric grid, whose assets last half a century and whose costs are socialized across every household in a service territory.
The AI economy has consequently created an extraordinary paradox that will recur throughout this paper: America simultaneously faces a very real electricity shortage and a potentially enormous amount of electricity demand that may not be real at all. Understanding that paradox—measuring it, governing it, and ultimately pricing it—is the purpose of Ghost Megawatts.
Why I Chose the Title “Ghost Megawatts”
I chose Ghost Megawatts because “ghost demand,” the phrase Reuters and a growing chorus of regulators now use, describes the existence of the problem but does not yet adequately describe its scale, its structure, or its unit of account. Governments, utilities and grid operators do not construct infrastructure for abstract demand; they construct infrastructure for megawatts. Transmission lines are sized in megawatts, transformers are rated in megawatts, capacity auctions clear in megawatt-days, and the checks that ratepayers ultimately write are denominated in the cost of serving megawatts. The meaningful question is therefore not merely whether speculative projects exist—everyone now concedes that they do—but how many megawatts inside a given utility’s forecast have a realistic probability of ever becoming operating load. By translating uncertainty into Ghost Megawatts, the concept becomes something policymakers, utilities, investors and datacenter developers can identify, classify, price and systematically reduce, rather than merely lament.
I also chose the phrase because it captures one of the defining contradictions of the AI infrastructure era. The United States may urgently need hundreds of gigawatts of additional generation over the coming decade—the International Energy Agency projects that American datacenter electricity demand will grow by roughly 130 percent by 2030, and Lawrence Berkeley National Laboratory’s 2025 update estimates that datacenters could account for 11.8 percent of all United States electricity by 2030, within a scenario range of 9.5 to 15.3 percent—while simultaneously planning billions of dollars of infrastructure around some megawatts that will vanish before consuming anything at all.[28][29][30] Ghost Megawatts are electricity promises without electricity consumption. They occupy forecasts, influence transmission decisions, affect generation planning, compete for scarce interconnection capacity, inflate capacity-market prices, and alter political debates even when the underlying datacenter is never completed. A ghost, in the traditional telling, is a presence without a body; a Ghost Megawatt is a demand without a load. That contradiction makes the title not merely evocative but analytically precise, and it fits naturally within the Five-Layer AI Economy series, of which this paper forms the energy-layer installment.
Section 1: The Emergence of the Ghost-Megawatt Economy
The deepest transformations in infrastructure economics rarely announce themselves as transformations; they arrive disguised as paperwork. The emergence of the Ghost-Megawatt Economy is a story about how a routine administrative artifact—the interconnection request—quietly became a strategic financial instrument, and about how the institutions that receive those requests discovered, almost too late, that their planning models had no way to tell an instrument from an intention. This section establishes the conceptual foundations: the distinction between electricity consumption and electricity claims, the formal taxonomy of megawatt credibility, the Texas laboratory in which that taxonomy is now being stress-tested at unprecedented scale, and the game-theoretic logic that explains why rational individual behavior by developers produces collectively irrational forecasts.
1.1 From Electricity Demand to Electricity Claims
The first distinction this paper must establish is between electricity consumption and electricity claims, because the conflation of the two is the raw material from which every Ghost Megawatt is manufactured.
A running datacenter consumes electricity; its load is a physical, metered, billable fact. A datacenter under construction represents probable future demand, discounted only by construction risk and commissioning schedules. A signed electric service agreement with collateral posted represents something stronger still—a financial commitment that survives even if the servers never arrive. But an inquiry asking whether 500 megawatts might someday be available at a particular substation represents none of these things. It is a question wearing the costume of a commitment, and in an era of abundance it would have been treated as such: filed, studied at leisure, and forgotten if nothing came of it. In an era of scarcity, however, every inquiry enters a queue, every queue feeds a forecast, and every forecast justifies capital expenditure. The costume becomes indistinguishable from the customer.
Yet these categories have become thoroughly blurred inside public discussions of the AI boom. Reuters found no standardized national system governing how utilities report prospective datacenter demand: one company may disclose only signed contracts while another includes early-stage inquiries in its investor presentations, and the roughly 270 gigawatts of datacenter electricity requests that Reuters tallied across ten of the biggest utilities in the Midwest, Mid-Atlantic and South—including AEP Ohio, Southern Company and PPL—was drawn from quarterly earnings calls in which the definitional boundaries of a “request” varied from company to company.[1] Aggregating such figures produces national numbers whose components have radically different probabilities of becoming real loads, and those national numbers then circulate through capital markets, congressional testimony and press coverage as if they were forecasts rather than compilations of claims. Jonathan Koomey, the veteran researcher of computing energy use whose eponymous law describes the historical improvement of computation per unit of energy, has spent two years warning that the public conversation is saturated with
“a lot of hype” [22]
— Jonathan Koomey, Ph.D., research affiliate, and co-author of the Bipartisan Policy Center’s guide to electricity demand growth
and his February 2025 report with Zachary Schmidt and Tania Das, Electricity Demand Growth and Data Centers: A Guide for the Perplexed, remains the essential methodological caution: announced capacity, queued capacity, contracted capacity and consumed energy are four different quantities that the perplexed—which is to say, nearly everyone—routinely treat as one.[23] Koomey’s broader counsel to an industry gripped by dueling projections has been, in effect, to calm down: society is not helpless before the growth in AI demand, and efficiency, measurement and skepticism remain available in exactly the proportions the moment requires.[39]
The scale of the resulting distortion has now been quantified by the consultancy Wood Mackenzie, whose analysis of the disclosed United States datacenter pipeline concluded that roughly seventy-two percent of AI datacenter power requests constitute phantom load—requests that will never be built as specified—and whose April 2026 assessment found approximately 600 gigawatts of proposed capacity still searching for power agreements against only 183 gigawatts that had actually secured them, a firm-commitment rate of roughly twenty-three percent.[26] Ben Hertz-Shargel, Wood Mackenzie’s global head of grid edge, identified the structural signature of the problem, noting that much of the planned capacity belongs to new developers concentrated in
“a small number of massive, speculative projects” [26]
— Ben Hertz-Shargel, Global Head of Grid Edge, Wood Mackenzie
disproportionately targeting the South and Southwest—precisely the regions where land is cheap, permitting is fast, and interconnection queues were, until recently, open to anyone with a filing fee.
1.2 Defining the Ghost Megawatt: A Five-Category Taxonomy
If the boundary between consumption and claims is the problem, then taxonomy is the beginning of the solution. This paper proposes that every megawatt appearing in an interconnection queue or utility forecast be assigned to one of five categories, ordered by descending credibility.
Table 1. A five-category taxonomy of requested megawatts, ordered by descending credibility.
| Category | Definition | Typical Evidence | Planning Treatment |
| Committed Megawatts | Binding contracts, credible financing, land control, permits and defined customers. | Signed ESA with collateral; construction underway; anchor tenant disclosed. | Near-full weight; discount only for ordinary execution risk. |
| Probable Megawatts | Advanced projects missing one or two significant conditions but reasonably expected to proceed. | Land, permits and tenant secured; generation contract in final negotiation. | High probability weighting, reviewed at each milestone. |
| Option Megawatts | Positions reserved while a developer decides among several candidate locations. | Parallel discussions in multiple states; one campus intended. | Weight equal to (1 ÷ number of live alternatives), pending disclosure. |
| Duplicate Megawatts | Substantially identical demand appearing in more than one utility or regional queue. | Same ultimate owner and project profile filed across jurisdictions. | Consolidate to a single entry via registry cross-checks. |
| Ghost Megawatts | Requested capacity whose probability of becoming operating load is low enough to distort planning if treated as firm. | No customer, no financing, no permits; queue position held for option or resale value. | Exclude from firm forecasts; monitor; price the reservation. |
Committed Megawatts are projects with binding contracts, credible financing, land control, permits and defined customers—the megawatts a planner can treat as nearly certain, discounted only by ordinary execution risk. Probable Megawatts belong to advanced projects missing one or two significant conditions but reasonably expected to proceed; a campus with land, permits and an anchor tenant that is still finalizing its generation contract belongs here. Option Megawatts are electricity positions deliberately reserved while developers decide among several candidate locations; they are not deceptive, but only a fraction of them can ever convert, by construction. Duplicate Megawatts are substantially identical future demand appearing in more than one utility or regional queue—the statistical shadow of a single project shopping multiple jurisdictions. And Ghost Megawatts, in the narrow sense, are requested capacity whose probability of becoming operating demand is sufficiently low that treating it as firm load creates material planning distortion: the speculative reservations, the land plays without customers, the queue positions held for resale value rather than construction.
The categories jointly define what this paper calls the Ghost-Megawatt Ratio:
GMR = Estimated Ghost Megawatts ÷ Total Requested Megawatts
The higher a utility territory’s GMR, the less reliable its headline datacenter pipeline becomes, and the more dangerous it is to build against that pipeline without probability weighting. Wood Mackenzie’s seventy-two percent phantom-load estimate is, in effect, a first national GMR calculation—and a national GMR above two-thirds means that the headline numbers dominating American energy discourse are, at best, three times too large.[26]
1.3 Texas and the 474-Gigawatt Reality Test
Texas provides the natural first case study, because nowhere else has the gap between requested and credible megawatts grown so large so fast, and nowhere else has a government moved so abruptly from courtship to audit.
The trajectory bears restating. In 2023, requests from datacenters and other large users to connect to the ERCOT grid totaled approximately 48 gigawatts—already remarkable for a system whose all-time peak demand was roughly 86 gigawatts. By June 2026, ERCOT was tracking more than 438 gigawatts of large-load requests, of which nearly eighty-nine percent were associated with datacenters, and by early August the figure cited in Governor Abbott’s letter had reached approximately 474 gigawatts—more than five times the record peak demand ever recorded on the Texas grid.[1][3][4] No serious analyst believes Texas will energize 474 gigawatts of new load on any relevant timescale; the entire installed generating capacity of the United States is on the order of 1,300 gigawatts. The queue had ceased to be a forecast and had become a phenomenon requiring its own explanation.
ERCOT’s institutional response, even before the Governor’s intervention, was the June 2026 “Batch Zero” framework: rather than evaluating enormous projects sequentially, as legacy processes assumed, ERCOT moved to evaluate qualifying large loads of 75 megawatts or greater collectively, so that it could understand their combined effect on system reliability and the transmission investment they would jointly require.[5] Senate Bill 6, signed into law in June 2025, had already established disclosure and curtailment obligations for loads of 75 megawatts or more, creating the statutory foundation on which the current oversight escalation rests.[8]
Governor Abbott’s August 3 directive then transformed the question entirely. By ordering the Public Utility Commission and ERCOT to complete a comprehensive verification and audit of every datacenter in the interconnection process before any additional project may advance—and by declaring that any project found in violation of applicable requirements “must be denied connection to the Texas grid”—the state stopped asking how quickly can Texas connect these projects and began asking which of these projects should Texas believe.[5] The directive requires project-level information on public financial assistance, self-supplied versus grid-supplied power, projected annual and peak electricity use, and planned water consumption and cooling technology; ERCOT paused its Batch Zero classification notifications, sought good-cause exceptions from its own deadlines, and now aims to complete the audit—covering as many as 300 projects—by December 2026, while acknowledging that its major transmission study will not be finished by the original April 2027 target.[5][6][7] ERCOT’s general counsel, Chad Seely, described the change with admirable candor: the verification that once happened late in the process has now been moved to the front of the line, before interconnection study even begins.[6]
The Texas experience is this paper’s central laboratory for a distinction that will recur throughout: the difference between a queue and an economy. Four hundred seventy-four gigawatts in an interconnection process does not mean Texas is about to consume an additional 474 gigawatts; but neither does the existence of speculative requests mean the remaining demand is trivial. Texas is, by several measures, the fastest-growing region for AI and cloud infrastructure in the world, second only to Virginia in installed datacenter capacity and widely expected—before the audit introduced new uncertainty—to take the top position.[8] The audit’s task, and the nation’s, is to find the boundary between the two truths.
1.4 Why Artificial Intelligence Produces More Ghost Megawatts Than Any Previous Technology
It is tempting to treat ghost demand as a moral failing of developers, but the more rigorous explanation is structural: artificial intelligence makes speculative electricity reservation economically rational in a way that no previous industrial technology quite did.
Consider the developer’s problem from the inside. A hyperscale campus requires three to seven years from site selection to energization under favorable conditions; grid interconnection in congested regions can take longer still. Meanwhile, the demand signals descending from the upper layers of the AI economy—model capabilities doubling on cadences measured in months, inference volumes growing triple digits year over year, enterprise adoption still in its early innings—arrive far faster than infrastructure can respond. A developer who waits until every permit, investor, GPU allocation and anchor tenant is finalized before requesting power may discover that the electricity it needs will not be available for another five or ten years, by which point the commercial opportunity has migrated elsewhere. The logical response is to reserve power early, in multiple places, at maximum plausible scale, and to treat the reservations as options to be exercised or abandoned as the future clarifies.
The result is a form of infrastructure game theory whose central proposition can be stated simply: when electricity becomes scarce, rational developers attempt to secure it before they know exactly how much they will ultimately need; and if every developer behaves rationally from its individual perspective, the collective forecast becomes irrational. Each individual request is a defensible hedge; the sum of the requests is a fiction. Economists will recognize the structure as a commons problem operating in reverse—not the overgrazing of a shared physical resource, but the overclaiming of a shared informational one, namely the credibility of the queue itself. Every speculative request slightly degrades the signal value of all requests, until the queue that was meant to communicate demand communicates mostly noise, and governments respond—as Texas, Pennsylvania and Ohio now have—by imposing the verification costs that the queue’s open design once spared everyone.
That is the beginning of the Ghost-Megawatt Economy: an equilibrium in which claims are cheap, verification is expensive, infrastructure is slow, and the gap among the three is measured in hundreds of gigawatts.
1.5 The Scholarly Context: What the 2020–2026 Literature Establishes
The academic and technical literature of the past six years frames the problem with increasing precision, and it is worth pausing to establish what is actually known, because the Ghost Megawatt concept sits deliberately at the intersection of three research streams.
The first stream measures real consumption. The Lawrence Berkeley National Laboratory’s landmark 2024 United States Data Center Energy Usage Report—the first federal assessment in nearly a decade, authored by Arman Shehabi, Sarah Smith, Jonathan Koomey, Eric Masanet and colleagues—found that datacenters consumed about 4.4 percent of total United States electricity in 2023, some 176 terawatt-hours, up from 58 terawatt-hours in 2014, and projected consumption of 325 to 580 terawatt-hours by 2028, or 6.7 to 12 percent of the national total.[27] The 2025 update pushed the horizon to 2030 with a reference case of 649 terawatt-hours—11.8 percent of projected national electricity—inside a compounded-uncertainty band of 521 to 843 terawatt-hours, or 9.5 to 15.3 percent.[28] The International Energy Agency’s Energy and AI report projects global datacenter consumption rising from roughly 415 terawatt-hours in 2024, about 1.5 percent of world electricity, to approximately 945 terawatt-hours by 2030 and 1,200 terawatt-hours by 2035, with the United States—which accounts for about forty-five percent of global datacenter consumption—growing around 130 percent by 2030.[29][30] These are large numbers, and they are the reason no serious version of this paper can claim the boom is illusory. But notice their magnitude relative to the claims: even the high end of LBNL’s 2030 band corresponds to well under 150 gigawatts of average national datacenter load, against 700-plus gigawatts of regional requests. The measured literature and the queued claims differ by a factor that no plausible forecast error can explain. Only ghosts can.
The second stream studies who pays. Eliza Martin and Ari Peskoe of Harvard Law School, in their widely cited 2025 paper Extracting Profits from the Public: How Utility Ratepayers Are Paying for Big Tech’s Power, examined roughly forty state regulatory proceedings and documented how special contracts between utilities and datacenters—often confidential, often approved with minimal scrutiny—can shift infrastructure costs onto captive ratepayers, warning that the industry’s current approach of
“luring data centers with discounted contracts or lopsided tariffs is unsustainable” [31]
— Eliza Martin and Ari Peskoe, Harvard Law School Electricity Law Initiative
and that without systematic ratemaking reform the public faces significant risks of subsidizing precisely the customers least in need of subsidy.[31] One study they cite estimates that Virginia-area ratepayers could see electricity costs rise by 150 to 450 dollars per year by 2040 under current structures.[33] The Ghost Megawatt extends their analysis one step earlier in the causal chain: before a discounted contract can shift costs, an unverified forecast must justify the infrastructure.
The third stream studies flexibility as a partial escape. Tyler Norris, Tim Profeta, Dalia Patiño-Echeverri and Adam Cowie-Haskell of Duke University, in their February 2025 study Rethinking Load Growth—arguably the most discussed energy paper of the period—showed that because power systems are built for extreme peaks that occur only a few hours a year, the existing United States grid could accommodate as much as 76 gigawatts of new load at 0.25 percent annual curtailment and up to 98 gigawatts at 0.5 percent, with ERCOT alone able to absorb roughly 15 gigawatts.[20][21] As Norris summarized the finding, the existing system, intentionally designed to handle extreme peak swings,
“could accommodate significant load additions with modest flexibility measures” [20]
— Tyler Norris, Nicholas School of the Environment, Duke University
The relevance to this paper is subtle but important: flexibility does not eliminate Ghost Megawatts, but it changes their cost, because a flexible megawatt requires less speculative infrastructure per unit of claimed demand—and, as Section 5 will argue, a developer’s willingness to accept curtailment is itself one of the strongest available signals that its megawatts are real.
Together, the three streams establish the boundary conditions of the Ghost-Megawatt problem: real consumption is growing fast but measurably; the costs of error fall on the public; and the physical system has more headroom than panic suggests, if—and only if—planners can tell which loads to believe.
Section 2: How Ghost Megawatts Are Manufactured
Ghost Megawatts are not conjured by fraudsters in the night; they are manufactured, in broad daylight, by the ordinary interaction of rational corporate strategy with fragmented institutional design. This section dissects the production process: the option value that scarcity confers on a grid position, the multiplication of demand through multi-jurisdiction site selection, the extraordinary length of the datacenter development chain, the historic reversal by which power now precedes customers rather than following them, and the crucial analytical discipline of distinguishing speculation from fraud. Understanding the manufacturing process matters because every policy instrument examined later in this paper—audits, permits, tariffs, deposits—works by raising the cost of one specific step in it.
2.1 The Option Value of a Grid Connection
A secured pathway to hundreds of megawatts of electricity has become, in the middle 2020s, an economically valuable asset in its own right—arguably one of the scarcest assets in the American economy. That fact alone reorganizes developer behavior. Where grid access is abundant, an interconnection request is a logistical formality undertaken when a project is ready; where grid access is scarce and the wait is measured in years, the request becomes something else entirely: an option contract written by the utility, priced at approximately zero, with no expiration date and no obligation to exercise.
Financial theory tells us exactly what happens to the demand for free options on a scarce underlying asset: it explodes. A power reservation functions like a call option on future electricity delivery—the developer secures the possibility of exercising future demand without committing immediately to build—and the scarcer and more volatile the underlying resource becomes, the more valuable the option grows, which in turn attracts holders who have no intention of ever building anything and every intention of selling their position, attaching it to land, or using it to court investors. The Financial Times, reporting on what it called phantom datacenters, quoted Tom Falcone, president of the Large Public Power Council, delivering the construction industry’s oldest wisdom against the new speculation:
“Nobody builds a 100-story tower in Manhattan without some anchor tenants” [25]
— Tom Falcone, President, Large Public Power Council
Yet gigawatt-scale power requests were being filed, in effect, for towers with no tenants, no financing and sometimes no architect. United States Energy Secretary Chris Wright went so far as to urge the Federal Energy Regulatory Commission to adopt rules deterring speculative projects in interconnection queues, and PJM—warning of unclear and uncertain datacenter projections—began pressing its member utilities to purge duplicative requests from their forecasts.[25] When the nation’s energy secretary, its largest grid operator, and its public power utilities all converge on the same diagnosis, the option-value dynamic has ceased to be a curiosity and become a systemic condition.
2.2 One Datacenter, Several States: The Multi-Queue Multiplier
Datacenter developers routinely evaluate several jurisdictions before selecting a final location, weighing power price and availability, land, water, fiber routes, tax incentives, permitting speed and political climate. This is prudent corporate practice, indistinguishable from how any manufacturer sites a factory. The distortion arises from what happens to the electricity system’s information while the evaluation proceeds.
Imagine a hyperscaler evaluating Texas, Ohio, Pennsylvania and Virginia for a future one-gigawatt AI campus. If serious discussions in all four states enter utility forecasts—and under pre-2025 practices they generally did, because utilities had every incentive to count prospective load and no mechanism to verify exclusivity—the national planning system temporarily observes four gigawatts of supposed future demand even though only one gigawatt will ultimately be constructed. Nothing fraudulent has occurred; no individual filing contains a false statement. The duplication emerges from rational corporate site selection interacting with fragmented utility forecasting, in which no registry exists to inform the utility in Columbus that the identical project is also being discussed in Amarillo, Harrisburg and Loudoun County.
This paper calls the phenomenon the Multi-Queue Multiplier: the ratio between the megawatts a project injects into the nation’s aggregate forecasts and the megawatts it will ever actually consume. A developer shopping four states with equal seriousness carries a Multiplier of four; sophisticated developers shopping eight or ten sites for a portfolio of projects can inflate national totals dramatically while behaving, at every step, with complete commercial propriety. Jonathan Koomey has emphasized that no national or regional tracking of this shopping exists in publicly available form—tech companies are secretive about scouting, and utilities do not share prospect lists with one another or the public—which means the Multiplier cannot currently be observed, only inferred from the wreckage it leaves in aggregate statistics—a flood of phantom projects that trade observers were already documenting in the load queues by the spring of 2025.[22][24] Mario Sawaya of the engineering firm AECOM described the resulting environment as a gold rush in which there is simply not enough power for all announced facilities, a candid admission from inside the industry that the queues contain multiples of any buildable reality.[22]
The Multi-Queue Multiplier explains why the 700-gigawatt figure should never have been read as a forecast, and also why merely dividing it by some assumed multiplier is insufficient: the multiplier varies by developer, by region and over time, and only disclosure regimes of the kind examined in Section 4 can compress it.
2.3 The Datacenter Development Chain: Fifteen Links, Any of Which Can Break
A credible AI datacenter requires an unusually long chain of commitments, and the length of that chain is itself a machine for producing Ghost Megawatts, because the electricity request typically enters the planning system when only the first two or three links exist.
The chain runs, in rough sequence, through land control; water rights and supply; local zoning and community acceptance; environmental permits; transmission access and interconnection study; secured generation or supply contracts; long-lead electrical equipment, above all transformers and switchgear, whose global order books now stretch years; backup power; networking and fiber; cooling infrastructure, increasingly liquid-based for AI densities; project financing; GPUs or custom accelerators, allocated by a handful of suppliers under their own scarcity; anchor customers; skilled construction labor in markets where electricians are the new bottleneck trade; and finally long-term operating economics that must survive electricity price volatility, chip depreciation cycles, and the possibility that model efficiency improves faster than demand grows. Failure or delay at any major link pushes expected energization backward, sometimes by years; failure at several can eliminate the project entirely. The cancellation of QTS’s massive Digital Gateway campus in Virginia amid record community pushback illustrates that even well-capitalized projects with famous sponsors can dissolve late in the chain.[37]
The asymmetry that matters for this paper is temporal: the power request is filed at the beginning of the chain, while the load arrives—if it arrives—only at the end. Every project that enters the queue and later breaks at link six or link eleven spends the intervening years as pure Ghost Megawatts inside somebody’s forecast, and under legacy practices nothing obliged the developer to report the breakage promptly, or at all.
2.4 Power Before Customer: The Great Reversal
The Pennsylvania numbers reveal how profound the reversal of traditional industrial logic has become. Reuters reported that of the more than 100 datacenters proposed in the state, only about 20 had applied for the permits necessary to advance, and a member of the governor’s office confirmed that most proposals had secured neither a power source nor a customer critical to funding the facilities.[1] The Commonwealth’s own Department of Environmental Protection counted 58 projects engaged in permitting discussions at some level of formality, of which only 15 had filed formal applications.[9]
Historically, companies built electricity infrastructure because economic activity required it: the aluminum smelter had customers for aluminum before it contracted for power; the rail yard existed because freight demanded it. In the emerging AI infrastructure economy, a meaningful population of developers inverts the sequence—first attempting to secure electricity, then marketing the secured electricity to find the economic activity that will justify consuming it. Power itself becomes part of the product. Land brokers advertise acreage by its megawatts rather than its acres; project marketing decks lead with substation proximity; and a signed—or even plausibly pending—interconnection position can add tens of millions of dollars to a parcel’s value before a single tenant conversation occurs. In such a market, filing power requests is not preparation for development; it is development, in the same way that acquiring mineral rights is mining before any ore is cut. The queue fills accordingly.
2.5 Speculation Does Not Mean Fraud: Four Species of Uncertainty
A serious paper must resist the rhetorical temptation to describe every uncertain project as deceptive, because the policy consequences of that error would be severe. There are at least four distinguishable species of uncertainty inside the ghost population, and they merit different regulatory treatment.
Strategic uncertainty describes the hyperscaler that genuinely possesses several viable locations and will build at one of them; its duplicate requests are real demand wearing multiple masks, and policy should aim to unify the masks, not kill the demand. Technological uncertainty reflects the honest unknowability of future power requirements: chip efficiency, model architectures, cooling technology and utilization patterns are all moving targets, and a developer’s 800-megawatt request may honestly become a 500-megawatt need—or a 1,200-megawatt one—by energization. Financial uncertainty attends projects whose completion depends on securing tenants or capital that may or may not appear; these are contingent megawatts, neither ghost nor firm. And speculative uncertainty, the narrowest and most troubling species, describes actors attempting to control scarce infrastructure without any realistic project at all—the pure option-harvesters of Section 2.1.
Policymakers must learn to distinguish among the four, because instruments calibrated for the fourth species will, if applied bluntly, destroy value in the first three. A collateral requirement that a speculator cannot meet is working as intended; the same requirement, sized wrongly, may exclude the credible startup developer whose project is real but whose balance sheet is young, thereby consolidating the datacenter industry into the handful of trillion-dollar incumbents who can post any collateral instantly. Regulation designed to eliminate Ghost Megawatts can, if crudely designed, eliminate legitimate megawatts as well—and with them, the competitive dynamism of Layer 3 of the AI economy. The credibility instruments surveyed in Sections 4 and 5 must therefore be judged not only by how many ghosts they exorcise but by how few real projects they wound in the exorcism.
Section 3: The Cost of Believing the Ghosts
If Ghost Megawatts were merely a bookkeeping curiosity, no governor would order audits and no legislature would convene hearings. They matter because believing them is expensive—expensive in stranded capital, in distorted market prices, in misallocated public money, and ultimately in political trust. This section quantifies the costs that are already visible, beginning with the classical danger of stranded infrastructure, proceeding through the striking evidence from PJM’s capacity markets that ghosts impose costs before they consume anything, confronting the equal and opposite danger of underbuilding, tracing the propagation of forecast error through all five layers of the AI economy, and ending with the politics—because in the end, every stranded dollar has a voter attached to it.
3.1 Transmission Built for Customers Who Never Arrive
The oldest and most intuitive economic danger is stranded infrastructure. Transmission systems, substations and generation facilities have service lives measured in decades—forty, fifty, sometimes seventy years—while a speculative datacenter can disappear in months, leaving nothing behind but a withdrawn filing. When a utility builds long-duration assets against short-duration expectations, the temporal mismatch does not vanish; it migrates onto the bills of the customers who remain. Under traditional cost-of-service regulation, prudently incurred investment enters the rate base and earns a return regardless of whether the anticipated load materializes, which means the ratepaying public functions, structurally, as the insurer of last resort for forecasting error—an insurer that collects no premium, writes no policy, and discovers its exposure only when the bill arrives.
The Harvard Electricity Law Initiative’s research makes the mechanism concrete: because utilities profit by building infrastructure, surging datacenter demand is a lucrative opportunity that incentivizes aggressive service commitments, discounted special contracts negotiated outside ordinary rate cases, and—where forecasts prove wrong—the socialization of the resulting costs across captive customers.[31][33] Ari Peskoe has noted that even the White House’s March 2026 announcement, in which technology companies pledged to cover the cost of new power plants and delivery infrastructure for their datacenters and to pay even if facilities never come online, largely restates commitments companies were already making, and leaves untouched the harder questions of enforceability and of the transmission costs that flow through regional formulas designed decades ago.[32] This explains why Ghost Megawatts are ultimately a ratepayer problem, not merely an industry forecasting problem: the entity best informed about a project’s credibility—the developer—bears the least cost when the forecast fails, and the entity least informed—the household—bears the most.
3.2 PJM and the Price of Forecast Growth: Ghosts That Bill Before They Exist
The most counterintuitive discovery of the past two years is that Ghost Megawatts do not need to consume electricity to cost money. The proof comes from PJM Interconnection, the grid operator spanning thirteen states and the District of Columbia, whose capacity market procures resources years in advance against forecast demand—which means that forecast demand, credible or not, clears directly into prices.
The numbers assembled by Joseph Bowring’s Monitoring Analytics, PJM’s independent market monitor, are remarkable. Across PJM’s last four base capacity auctions, existing and forecast datacenter demand drove 29.4 billion dollars of capacity charges—forty-six percent of the 63.6 billion dollars in total charges over that span.[16][17] In the December 2025 auction alone, datacenter load accounted for roughly 6.5 billion of 16.4 billion dollars in costs, and critically, about 6.2 billion of that was attributable to datacenters that had not yet been built but were forecast to arrive by the 2027/28 delivery year.[18] In the 2025/26 auction, the monitor calculated that datacenter load by itself raised revenues by more than 9.3 billion dollars, an increase of some 174 percent, making datacenters responsible for roughly sixty-three percent of that auction’s total cost.[19] The consequences have already reached kitchen tables: capacity costs contributed to residential increases on the order of 21 dollars per month for Pepco customers in Washington, 18 dollars in western Maryland, and 16 dollars in Ohio, and the price caps negotiated between PJM and Governor Shapiro—which suppressed the last auction by an estimated 3.2 billion dollars—expire with the 2028/29 auction, removing the ceiling just as datacenter forecasts peak.[18][19] Bowring’s own characterization of the moment was blunt: PJM is behaving as if this were ordinary growth, when in truth
“it is really a paradigm shift” [16]
— Joseph Bowring, President, Monitoring Analytics, Independent Market Monitor for PJM
and failing to recognize the shift imposes costs on every other customer in the footprint. His proposed remedy—removing datacenters and their forecast uncertainty into a separate capacity auction so that ordinary customers no longer underwrite speculative load—is precisely a Ghost-Megawatt containment mechanism by another name.[16] The Union of Concerned Scientists has separately estimated that PJM ratepayers will pay about 4.4 billion dollars for datacenter-related transmission projects approved in 2024 alone, and household electricity prices nationally have risen more than ten percent over two years, outpacing general inflation, with anticipated datacenter load among the drivers.[19][41]
The significance is profound and generalizes far beyond PJM: expected consumption alone can move procurement, investment and market prices. A megawatt can have an economic impact before it has a physical impact—and if the expectation later evaporates, the money is not refunded. The ghost, in other words, bills in advance.
3.3 The Opposite Risk: Underbuilding for the Boom That Is Real
There is an equally dangerous conclusion that policymakers must now be warned against, because it is rhetorically seductive and analytically wrong: if some demand is speculative, America’s electricity shortage is exaggerated. That does not follow, and the evidence is explicit on the point. Reuters specifically found that grids can remain overwhelmed even after less credible projects are eliminated, because the substantiated remainder is still enormous relative to the pace of new generation; PJM’s position remains that genuine load is arriving faster than the resources needed to serve it in parts of its system.[1] AEP Ohio’s experience, examined in detail in Section 4, makes the same point numerically: after the most aggressive financial screening in the country cut a 30-gigawatt pipeline by more than eighty percent, the surviving contracted load—17,861 megawatts against a historical statewide peak of 8,000 to 10,500 megawatts—still implies a near-tripling of the system.[14][15] The ghosts were real, and so is the boom they haunted.
The measured literature reinforces the conclusion from the demand side. LBNL’s 649-terawatt-hour reference case for 2030 represents nearly a fourfold increase over 2023 consumption; the IEA’s 130 percent growth projection for United States datacenter demand by 2030 would, on its own, absorb a large share of all planned generation additions; and the hyperscalers’ Q2-2026 disclosures—Amazon describing striking demand visibility as far out as 2028, and Andy Jassy telling investors that AWS could
“very possibly be $1 trillion annual revenue business for us in time” [38]
— Andy Jassy, Chief Executive Officer, Amazon
—indicate that the corporate commitments beneath the credible fraction of the queue are hardening, not softening.[28][29][38] America can therefore simultaneously have too many requested megawatts and too few available megawatts. This paradox is not a puzzle to be resolved in favor of one side; it is the operating condition of the Ghost-Megawatt Economy, and every sound policy must hold both truths at once. Overreaction to the ghosts starves the boom; credulity toward the ghosts taxes the public. The corridor between the two errors is narrow, and it is exactly the corridor that Sections 4 and 5 attempt to map.
3.4 Ghost Megawatts and the Five-Layer AI Economy: How Error Propagates Downward
The phenomenon can propagate through every layer of the AI economy, and tracing that propagation explains why an electricity-queue problem belongs in a series about artificial intelligence rather than merely in a utility trade journal.
At Layer 5, applications and agentic systems, forecasts anticipate massive inference adoption—every enterprise workflow instrumented, every consumer interaction mediated by models. Those expectations flow down to Layer 4, where laboratories project escalating training runs and inference fleets, and justify them with usage curves that are genuinely steep but inherently young. Layer 3 converts the model projections into campuses: hyperscalers and developers reserve land, shells and power on the assumption that the curves continue. Layer 2 converts campuses into silicon: GPU and accelerator purchase agreements—now routinely denominated in hundreds of billions of dollars and sometimes financed circularly, with suppliers investing in their own customers—determine rack densities and therefore power densities. And Layer 1 receives the sum of it all as electricity requests that utilities must translate into transformers, wires and generation years before the upper layers’ assumptions can be tested against reality.[41]
Every forecast therefore contains assumptions inherited from the layer above it, and errors compound as they descend because each layer adds its own hedge: the application company over-forecasts to be safe, the lab reserves compute against the over-forecast, the developer reserves power against the reservation, and the utility plans transmission against the reservation of the reservation. A mistaken assumption about future agent usage at Layer 5 can eventually become a mistaken transmission line at Layer 1—except that the agent forecast can be revised in a quarter, while the transmission line, once built, is a fifty-year fact. Ghost Megawatts are therefore best understood as the physical shadow of uncertainty higher in the AI stack: the place where the entire economy’s optimism, hedging and duplication finally condense into steel, and where the cost of being wrong stops being a writedown and becomes a rate case.
3.5 The Politics of Paying for the Wrong Forecast
Once household electricity bills rise, AI infrastructure stops being a technology-policy issue and becomes electoral politics, and the summer of 2026 marked the crossing of that threshold. Electricity prices became a campaign topic across the PJM footprint; New York enacted its own datacenter pause legislation; an Ohio constitutional amendment to ban datacenters was proposed, though it failed to reach the ballot; and the credit-rating agency Morningstar DBRS formally cautioned that mounting stakeholder resistance—new taxes, restrictions and moratoriums under consideration across the states—could begin to weigh materially on the credit quality of datacenter projects themselves, reducing development visibility and challenging assumptions about the pace and certainty of AI-driven capacity expansion.[16] When political backlash is priced into the cost of capital, the ghost has completed its journey from planning artifact to macro-financial variable.
Governors, utility commissioners, mayors, county officials and members of Congress now face three questions from constituents that admit no technical evasion: Who benefits? Who pays? And was the infrastructure actually necessary? The first two questions are distributional and can, in principle, be answered with tariffs and contracts. The third question is the Ghost Megawatt question, and it is unanswerable without credibility measurement—because “necessary” depends entirely on whether the load that justified the investment was ever real. Those questions will be asked with particular intensity as the November 2026 midterm elections approach, in states where the datacenter buildout, electricity prices and the AI economy’s local footprint have fused into a single political issue. Abbott’s audit and Shapiro’s executive order are, among other things, answers to that political moment—and the fact that a Republican governor of Texas and a Democratic governor of Pennsylvania converged within fifteen days on the same underlying instrument, the credibility gate, suggests that Ghost-Megawatt governance has become one of the rare genuinely bipartisan projects in American politics.
Section 4: Texas, Pennsylvania and Ohio Build the Credibility Gate
Between mid-2025 and September 2026, three states with very different political economies—deregulated, energy-exporting Texas; transmission-rich, PJM-embedded Pennsylvania; and industrial, utility-regulated Ohio—independently constructed mechanisms for the same purpose: forcing announced megawatts to prove themselves before they may consume public planning resources. This paper calls the generalized instrument the Credibility Gate: a checkpoint at which a developer must demonstrate, with information or money or both, that its proposed load deserves to be believed before it receives scarce grid capacity. This section examines the three state experiments in turn, extracts the common architecture emerging from them, and poses the national policy question their divergence creates.
4.1 Texas: Audit Before Energization
Governor Abbott’s August 2026 directive represents a philosophical shift whose importance exceeds its administrative details: the state that built its brand on frictionless connection is no longer treating a grid application as sufficient evidence of economic seriousness. The directive’s text is uncompromising—a comprehensive verification and audit of all datacenters in the interconnection process, completed before any additional projects advance, with any project violating PUCT, ERCOT or statutory requirements to be denied connection—and its stated trigger was itself an information failure: the refusal of some datacenters to comply with the state’s survey of water and power usage under the General Appropriations Act.[3][5] The audit apparatus that ERCOT then previewed operates on two tracks: a Batch Zero eligibility verification for large loads generally at or above 75 megawatts, and a community-impact information collection for unenergized computational loads at or above 25 megawatts, with project-level disclosure of ultimate ownership characteristics, public financial assistance, power sourcing, projected annual and peak consumption, and water and cooling technology.[5] As many as 300 projects will pass through the gate, with ERCOT targeting completion by December 2026 and conceding that its flagship transmission study will slip past April 2027.[6][7]
Texas is therefore creating the purest form of the Credibility Gate: verification moved, in Chad Seely’s formulation, to the front of the line.[6] The Data Center Coalition’s response—expressing hope that the audits would
“separate those who are responsible water and energy stewards from those who are not” [4]
— Data Center Coalition, statement on the Texas audit directive
—is notable precisely because it comes from the industry itself: serious developers have discovered that they benefit from a gate that distinguishes them from the speculators with whom they currently share a queue, a dynamic economists will recognize as the demand for signaling in a market degraded by adverse selection. The unresolved risks run in the other direction: an audit conducted slowly or politically becomes a de facto moratorium on real projects, Jefferies analysts have already warned that the 2027 legislative session could extend the delays, and financing parties are being advised to treat audit clearance as a condition precedent in project documents—evidence that the gate’s existence is already repricing Texas development risk.[7][8]
4.2 Pennsylvania: Permit the Project, Not the Press Release
Governor Josh Shapiro’s Executive Order 2026-05, signed August 18, 2026, attacks the same problem through the permitting system rather than the interconnection queue. The order directs Commonwealth agencies to review datacenter permit applications only where developers have made a legally binding commitment to the Governor’s Responsible Infrastructure Development requirements—standards spanning energy affordability, environmental protection, workforce and economic development, transparency and community engagement—and have first obtained local approval; it removes all AI datacenter projects from Pennsylvania’s Fast Track permitting program permanently; it prohibits the use of nondisclosure agreements in datacenter dealings; it requires community-benefit agreements including local hiring and investment; it applies its core requirements to projects above 25 megawatts; and it directs the pursuit of new rules on curtailment priority and cost allocation through the Public Utility Commission.[9][10][11][12] Shapiro, who had originally proposed the GRID standards as voluntary incentives before the legislature declined to act, framed the escalation in unambiguous terms, declaring that he had no choice but
“to protect the good people of Pennsylvania from these predatory developers” [11]
— Josh Shapiro, Governor of the Commonwealth of Pennsylvania
The empirically decisive number in Pennsylvania is again the gap between announcement and execution: more than 100 projects proposed in public databases, 58 engaged with the Department of Environmental Protection at some level, roughly 20 having sought permits, and 15 with formal applications filed.[1][9][12] A permitting-based gate converts that gap into information automatically, because a permit application is costly, project-specific, and impossible to duplicate casually across jurisdictions: the eighty-plus proposals that never file simply never encumber the system. The design’s vulnerability is the mirror image of its strength—an executive order can be litigated, revised or succeeded, and the Data Center Coalition’s Dan Diorio warned pointedly against rules being changed midstream for verified and responsible projects whose communities have already prepared for the investment—but as a mechanism for separating projects from press releases, Executive Order 2026-05 is the most comprehensive disclosure regime yet attempted by an American state.[11]
4.3 Ohio: Put Money Behind the Megawatt
Ohio provides the clearest financial experiment, and its results are the closest thing the Ghost-Megawatt hypothesis has to a controlled demonstration. The Public Utilities Commission of Ohio approved AEP Ohio’s dedicated datacenter tariff—Schedule DCT—in July 2025, following a period in which the utility, facing an unprecedented 30 gigawatts of datacenter requests against a statewide historical peak of 8,000 to 10,500 megawatts, had paused new datacenter connections entirely. The tariff’s provisions read like a syllabus in commitment devices: qualifying customers above 25 megawatts face substantial connection-study fees, long contractual terms, collateral requirements, demonstrated financial viability, exit fees for cancelled or defaulted projects, and minimum-demand obligations under which large customers must pay for the great majority of their contracted capacity even if actual usage falls short.[13][14][15]
The pipeline’s response was immediate and diagnostic. Of the roughly 30 gigawatts of pre-tariff interest, formal load-study requests under the new regime totaled 13,023 megawatts across 36 sites; and when the utility presented service plans and financial obligations in black and white, projects representing 5,642 megawatts signed binding electric service agreements with collateral, as reported to the commission on February 12, 2026.[13][14][15] Combined with 12,219 megawatts contracted before the tariff took effect, AEP Ohio now holds 17,861 megawatts of financially committed datacenter load scheduled to arrive progressively over the coming decade.[13][15] Marc Reitter, AEP Ohio’s president, explained the planning logic of the exercise: accurate estimates matter because clarity
“gives us the clarity to plan and align infrastructure investment” [14]
— Marc Reitter, President and Chief Operating Officer, AEP Ohio
Read through this paper’s framework, the Ohio experiment measured a Ghost-Megawatt Ratio directly: roughly eighty percent of the headline pipeline declined to put capital behind its claims, while the surviving twenty percent hardened into the most credible datacenter load book in the country. Critics remain—the Ohio Manufacturers’ Association has challenged the tariff before the Ohio Supreme Court, arguing that even the screened forecasts may be inflated and may transmit costs through regional capacity markets—and their objection is a sophisticated one, since a signed contract reduces but does not eliminate forecast risk.[15] But the tariff establishes the simple credibility principle on which all financial gates rest: if a developer wants the grid to believe its megawatts, the developer should be willing to place capital behind them. Exelon’s parallel experience—cutting its high-probability pipeline by roughly forty percent to about 11 gigawatts after imposing stricter collateral—confirms that the principle generalizes beyond a single utility.[1]
4.4 From Interconnection Queue to Financial Commitment: The Emerging Architecture
Viewed together, the three state experiments and their counterparts elsewhere—utilities serving eight of the ten largest datacenter markets have now overhauled pricing structures or imposed stringent project rules—outline the architecture of next-generation datacenter regulation.[25] The emerging Credibility Gate combines proof of land control; disclosure of ultimate ownership; identified customers; evidence of financing; power-source plans; deposits and study fees; collateral scaled to requested capacity; minimum-demand payments; exit fees; milestone-based queue retention under which positions expire if development milestones are missed; and, increasingly, curtailment obligations that both protect reliability and function as honesty tests. Together these mechanisms transform electricity capacity from a low-cost reservation into a financial obligation—which is to say, they reprice the free option of Section 2.1 at something approaching its true social cost, and thereby drain the speculative demand for it.
Table 2. Three state architectures of the Credibility Gate, 2025–2026.
| State | Primary Instrument | Key Mechanisms | Observed Effect (through September 2026) |
| Texas | Verification audit before interconnection (Abbott directive, Aug. 3, 2026) | Comprehensive audit of ~300 projects; Batch Zero collective study for loads ≥75 MW; SB6 disclosure and curtailment duties; community-impact data for computational loads ≥25 MW; ownership, power-sourcing and water disclosure. | New connections frozen pending audit; ERCOT targets completion by December 2026; ~474 GW queue under review, ~90% datacenters. |
| Pennsylvania | Binding permitting and transparency gate (Executive Order 2026-05, Aug. 18, 2026) | Legally binding GRID requirements; local approval before DEP review; removal from Fast Track permitting; NDA prohibition; community-benefit agreements; ≥25 MW threshold; curtailment-priority and cost-allocation rulemaking. | Of 100+ proposals, ~20 sought permits and 15 filed formal applications; speculative projects blocked from advancing. |
| Ohio | Financial commitment tariff (AEP Ohio Schedule DCT, approved July 2025) | Study fees up to $100,000; collateral; demonstrated financial viability; long contract terms; minimum-demand payments; exit fees for cancellation or default. | ~30 GW of claims reduced to 5,642 MW of signed, collateralized ESAs (plus 12,219 MW pre-tariff), yielding 17,861 MW of committed load. |
4.5 The Emerging National Policy Question
The United States now faces a structural choice between fifty-state experimentation and standardized national principles. Electricity regulation is, by constitutional design and historical accident, heavily state-based; different grids genuinely face different conditions, and the laboratory dynamics of 2025–2026 have been productive precisely because Texas, Pennsylvania and Ohio tried different instruments. But wildly inconsistent definitions of proposed datacenter demand make national forecasting nearly impossible—the 700-gigawatt figure is uninterpretable partly because its components were assembled under incompatible reporting conventions—and inconsistency invites regulatory arbitrage, in which speculative projects migrate to whichever jurisdiction has not yet built its gate, exporting the Ghost-Megawatt problem rather than solving it.[1]
A reasonable federal role therefore may not be to decide which datacenters receive electricity—a task for which no federal agency has the local knowledge and which would centralize exactly the discretion that should remain contestable. It may instead be to establish common disclosure and classification standards, so that a gigawatt of proposed AI demand means approximately the same thing in Texas, Pennsylvania, Ohio and Virginia; to support a national registry through which duplicate requests can be identified across balancing authorities without exposing commercially sensitive site selection; and to encourage, through FERC’s interconnection authority, the milestone-and-deposit structures that Energy Secretary Wright has already urged in order to deter speculative projects.[25] Measurement standardization is the classic minimal federal function—the Ghost-Megawatt equivalent of uniform accounting principles—and it is probably the highest-leverage intervention available at the national level.
Section 5: Building a Ghost-Megawatt Framework for the AI Age
Diagnosis without instrumentation is commentary, and the preceding sections would remain commentary if they did not culminate in tools. This section assembles the Ghost-Megawatt framework proper: a standardized credibility score for individual projects, probability-weighted load forecasting for system planners, the Ghost-Megawatt Ratio as a comparative indicator across jurisdictions, the allocation principle that should govern who bears forecast risk, a staged-commitment architecture that matches obligations to development maturity, and—because the framework would be worse than useless if it merely suppressed applications—an explicit optimization target that keeps the AI boom itself in view. The framework synthesizes what Texas, Pennsylvania and Ohio have improvised separately into a design that any jurisdiction, and eventually any national standard-setter, could adopt.
5.1 The Datacenter Credibility Score
The foundation of the framework is a standardized scoring system under which each major project receives points across ten dimensions of demonstrated commitment: land control; local approval; environmental permitting; anchor customer; financing; grid study completion; generation contract; equipment procurement; construction progress; and binding minimum-payment obligation. The dimensions are deliberately drawn from the development chain of Section 2.3, because each one corresponds to a link whose completion is observable, documentable and difficult to fake: a developer can announce anything, but it cannot casually possess a signed generation contract, a posted collateral instrument, or transformers on order in a market where transformer lead times are measured in years.
Table 3. The ten dimensions of the Datacenter Credibility Score.
| Dimension | What It Verifies | Example Evidence |
| 1. Land Control | The project has a physical site it can actually use. | Recorded deed, executed long-term lease, or exercised option. |
| 2. Local Approval | The host community has consented through its own processes. | Zoning approval; executed community-benefit agreement. |
| 3. Environmental Permitting | Regulatory feasibility at the site. | Filed and advancing air, water and land permits. |
| 4. Anchor Customer | Demand exists for the computing capacity. | Executed or binding-LOI tenancy with a creditworthy counterparty. |
| 5. Financing | Capital exists to complete construction. | Committed equity and debt; sponsor guarantees. |
| 6. Grid Study | The system impact has been engineered, not asserted. | Completed interconnection or Batch Zero-style study. |
| 7. Generation Contract | Electrons have been secured, not assumed. | PPA, self-supply plan, or utility service agreement. |
| 8. Equipment Procurement | Long-lead hardware is actually on order. | Transformer, switchgear and cooling purchase orders. |
| 9. Construction Progress | The project has left paper and entered the ground. | Notice to proceed; verified milestones. |
| 10. Binding Minimum-Payment Obligation | The developer bears the cost of its own forecast. | Minimum-demand charges, collateral and exit fees in force. |
Projects satisfying nearly every criterion would receive a high credibility score and full statistical weight in planning; early-stage projects would remain visible—visibility is essential, since today’s inquiry is sometimes tomorrow’s gigawatt—but would enter forecasts at the discounted weight their demonstrated commitments justify. The scoring must be dynamic, re-assessed at fixed intervals and at every milestone, because credibility is a trajectory rather than a state: a project that gains an anchor tenant and closes financing in the same quarter has changed category, and so has one that quietly lets its land option lapse. The score’s second function is communicative. A public, standardized score gives serious developers the signaling instrument they currently lack—the means to distinguish themselves from the speculators degrading their queue—which is why, as the Data Center Coalition’s reaction to the Texas audit suggests, sophisticated industry participants can be expected to support rather than resist well-designed scoring.[4]
5.2 Probability-Weighted Load Forecasting
The second instrument translates project scores into system planning. Instead of reporting that 100 gigawatts of datacenters have requested electricity—a sentence that has misled more state legislatures in two years than any statistic in the modern history of utility regulation—a grid operator applying the framework would report three numbers: 100 gigawatts of requested load; 62 gigawatts of probability-weighted load, computed by multiplying each project’s request by its score-derived probability of commercial energization; and 38 gigawatts of estimated Ghost Megawatts, the residual that should be planned around rather than planned for. Forecasts should further be presented as confidence bands rather than single gigantic numbers marching toward inevitability, because the honest representation of the AI era’s demand is a distribution, and every institution that has pretended otherwise—from utilities defending record capital plans to commentators declaring the boom fictitious—has been punished by events within quarters. Lawrence Berkeley National Laboratory’s practice of publishing reference cases inside explicit sensitivity ranges, 649 terawatt-hours within 521 to 843 for 2030, is the methodological model, and it is telling that the most credible numbers in this entire literature come from the institution most insistent about its own uncertainty.[28]
Probability weighting has a further, less obvious virtue: it makes forecast manipulation expensive in both directions. A utility that overweights speculative load to justify capital expenditure must now defend project-level probabilities in public proceedings; an operator that underweights credible load to slow-walk interconnection must defend the discounts. The argument moves from theology to evidence.
5.3 The Ghost-Megawatt Ratio as a Comparative Indicator
The third instrument, the Ghost-Megawatt Ratio introduced in Section 1.2, does for jurisdictions what the credibility score does for projects. A 100-gigawatt pipeline with a ten percent GMR and a 100-gigawatt pipeline with a seventy percent GMR are, for planning purposes, different objects separated by a factor of three in expected load, yet headline megawatts render them identical—and headline megawatts are what circulate in economic-development announcements, bond prospectuses and political speeches. Published GMRs would allow regulators to benchmark screening effectiveness across utilities; allow investors to discount pipelines with jurisdiction-appropriate haircuts rather than uniform skepticism; and allow researchers to test, over time, which gate designs actually compress the ratio. The early evidence already suggests an answer: Ohio’s financially gated pipeline carries a demonstrably low residual GMR, since every surviving megawatt is collateralized, while ungated queues elsewhere plausibly sit near Wood Mackenzie’s seventy-two percent national phantom estimate.[15][26] The ratio, in other words, is not merely descriptive; it is the outcome variable of the entire policy experiment now underway.
5.4 Make the Developer Carry the Forecast Risk
Beneath the instruments lies an allocation principle that public policy should adopt explicitly: the party making the extraordinary electricity forecast should bear a meaningful portion of the financial risk of that forecast proving wrong. The current default inverts this—developers assert, utilities build, ratepayers insure—and every pathology documented in this paper flows from the inversion. The principle does not require punitive regulation, and this paper does not endorse the confiscatory versions occasionally proposed; it requires only that the instruments of ordinary commercial seriousness—deposits, minimum-demand payments, collateral, milestones, transferable capacity agreements, and appropriately designed exit fees—be attached to grid claims in proportion to their scale. Ohio’s Schedule DCT provides one working model; the Harvard Electricity Law Initiative’s proposals for transparent, broadly applicable tariffs in place of confidential special contracts provide the ratemaking complement; and transferability deserves particular attention as the market-friendly refinement, since a developer whose plans genuinely change should be able to sell its committed position to a credible successor rather than abandon it, converting even failure into information and liquidity rather than stranded process.[13][31]
The deepest justification for the principle is informational rather than fiscal. Money placed at risk is a truth-revelation device: it forces the developer, who holds private knowledge about its own project’s reality, to disclose that knowledge through behavior. The 24 gigawatts that declined to sign in Ohio told regulators more about the state’s true demand curve than any survey could have, and they told it at zero cost to ratepayers.[14][15]
5.5 Different Rules for Different Levels of Certainty: The Staged-Commitment Architecture
A healthy policy system should not force every project to satisfy every requirement on day one, because the legitimate species of uncertainty catalogued in Section 2.5—strategic, technological and financial—all require room to resolve. The better approach is staged commitment, in which obligations escalate with development maturity.
Table 4. The staged-commitment architecture: obligations that escalate with development maturity.
| Stage | Developer Obligation | What the Grid Provides in Return |
| Exploration | Inexpensive preliminary inquiry; basic identity disclosure. | Information on availability; no queue position. |
| Reservation | Site identification and a modest, partially refundable deposit. | Provisional queue position with expiration milestones. |
| Engineering | Substantial study fee; technical documentation; ownership disclosure. | Formal interconnection study; indicative timeline. |
| Commitment | Financing evidence; customer disclosure; collateral posted. | Firm queue position; capacity allocation; accelerated review. |
| Construction | Binding minimum-demand obligations; milestone reporting. | Scheduled energization; infrastructure construction proceeds. |
| Energization | Full operational requirements, including any curtailment obligations. | Commercial service; durable, transferable capacity rights. |
At the Exploration Stage, inexpensive preliminary inquiry remains available to all, preserving the openness that lets young companies and novel projects enter the system. The Reservation Stage attaches site identification and a modest deposit. The Engineering Stage requires a substantial study fee and technical documentation—the point at which Ohio’s 100,000-dollar study fees and Texas’s disclosure requirements now operate. The Commitment Stage demands financing evidence, customer disclosure and collateral; the Construction Stage carries binding minimum-demand obligations; and the Energization Stage imposes full operational requirements, including whatever curtailment obligations the jurisdiction has adopted. The governing gradient is simple: the closer a developer gets to receiving scarce grid capacity, the more expensive it should become to walk away—and, symmetrically, the faster and more certain its path forward should become, since developers who have paid for credibility deserve queue priority over those who have not. Speed-for-commitment is the trade that aligns every legitimate interest in the system, and its absence—as Duke’s Tyler Norris observed of PJM’s early flexibility proposals, which offered mandatory curtailment with no defined path to faster interconnection—is precisely what makes gates feel like walls.[21]
5.6 Do Not Kill the AI Boom While Eliminating Its Ghosts
The framework requires, finally, an explicit statement of its own objective function, because a Ghost-Megawatt policy evaluated by the wrong metric will optimize for the wrong world. The objective cannot simply be to reduce interconnection requests: any sufficiently hostile regime can empty a queue, producing an impressive statistical improvement while destroying the country’s AI infrastructure capacity and handing the buildout to jurisdictions—and nations—with more welcoming grids. The objective is to maximize Credible Megawatts while minimizing Ghost Megawatts, and the two goals are separable: a well-designed gate raises the ghost’s cost of entry while lowering the credible project’s cost of proof, shrinking the queue and growing the buildout simultaneously, exactly as Ohio’s experience—an eighty percent smaller pipeline and a record contracted load book, at once—demonstrates is possible.[15]
The stakes of getting the balance right are geopolitical as much as economic. America’s strategic advantage in artificial intelligence will ultimately depend not on which country announces the most electricity demand—announcement is free everywhere—but on which country most efficiently transforms proposed gigawatts into productive intelligence infrastructure. A nation that builds against ghosts wastes capital it cannot spare; a nation that refuses to build until every ghost is exorcised cedes the decade to competitors; the nation that learns to tell the difference converts its planning institutions themselves into a competitive weapon. Forecasting credibility, in the AI era, is industrial policy.
Section 6: What Have We Learned? Seven Pillars
The argument of this paper can be consolidated into seven pillars—five drawn from the core analysis, and two that emerged in the course of the investigation, concerning flexibility as a credibility currency and the capital-market feedback loops that now amplify every error at Layer 1. Together they constitute the working doctrine of Ghost-Megawatt governance.
Pillar 1 — Electricity Has Become an Option Before It Becomes a Commodity
The AI infrastructure boom has changed the economic character of grid access itself. Developers increasingly seek to secure future electricity before every element of their business case is finalized, which means a grid position now carries option value independent of any facility—value that grows with every increment of scarcity and attracts holders with no intention of building. This behavior is rational individually and destabilizing collectively: when thousands of megawatts are reserved as strategic options, system planners mistake optionality for inevitability, and infrastructure gets built for exercises that never occur. The lesson is that grid-access scarcity, left unpriced, manufactures its own phantom demand—and that the free option must be repriced through deposits, milestones and collateral before any forecast built on the queue can be trusted.[25][26]
Pillar 2 — A Megawatt Request Is Not a Megawatt Forecast
The most important lesson is methodological. Requested capacity, contracted capacity, probability-weighted capacity and actual electricity consumption are four different quantities, and they should never be casually combined—yet the public conversation about AI and energy combines them daily, producing numbers like 700 gigawatts that are neither forecasts nor fictions but uninterpretable hybrids.[1][23] The AI era requires an entirely new vocabulary for electricity planning because a single gigawatt-scale project can now distort a regional outlook, and because the institutions reporting demand have heterogeneous incentives and no shared definitions. Governments must learn to measure the quality of demand, not merely its quantity; the taxonomy, credibility score and Ghost-Megawatt Ratio of this paper are offered as the beginning of that vocabulary.
Pillar 3 — Financial Commitment Is Becoming Infrastructure Verification
Texas is using audits; Pennsylvania is using permitting, transparency and community obligation; Ohio is using contracts, collateral and minimum-demand payments. The instruments differ, but the states are converging on a single principle: evidence must replace aspiration. Among all forms of evidence, money placed at risk has proven the strongest, because it converts a developer’s private knowledge into public signal—as Ohio’s collapse from 30 gigawatts of claims to 5,642 megawatts of collateralized contracts demonstrated more decisively than any survey, audit or hearing could have.[13][14][15] A developer willing to commit substantial capital to a proposed load provides more useful information to a grid planner than any press release announcing a future campus; the corollary, which regulators should embrace explicitly, is that capital commitment should purchase speed.
Pillar 4 — Ghost Megawatts Can Cost Money Even When They Consume No Power
One of this paper’s most counterintuitive conclusions is that a nonexistent future datacenter produces real present-day economic effects. Forecast demand influences generation planning, transmission proposals, capacity procurement, utility investment programs, land values, incentive packages, political debates and ultimately electricity prices—and PJM’s capacity markets have now quantified the channel, with 29.4 billion dollars of datacenter-driven capacity charges across four auctions, much of it attributable to facilities not yet built and some of it attributable to facilities that never will be.[16][17][18] The ghost does not need to become physical to become expensive. In the AI infrastructure economy, expectations themselves have infrastructure costs, and the governance of expectations—disclosure, classification, probability weighting—is therefore a form of consumer protection.
Pillar 5 — Flexibility Is Becoming a Credibility Currency
A pillar that emerged from the flexibility literature deserves independent statement: a load’s willingness to bend is among the most credible signals that it is real. The Duke findings—76 to 98 gigawatts of headroom at curtailment rates of a quarter to a half percent, with at least half the load retained in ninety percent of curtailment hours—mean that flexible loads can connect faster, at lower system cost, with less speculative infrastructure per megawatt; and Google’s demand-response agreements with Indiana Michigan Power and the Tennessee Valley Authority mark the first documented integration of AI datacenter flexibility into formal utility planning.[20][21] The signaling logic is the point here: a speculator holding an option has no operations to flex and no incentive to accept curtailment obligations that only bind real facilities, so curtailability functions as a self-selecting honesty test. Jurisdictions should therefore fold flexibility directly into their credibility gates—faster interconnection for curtailable load, following the direction of Texas’s Senate Bill 6 and the Southwest Power Pool’s expedited pathways—and treat a developer’s refusal of all flexibility as information.[8][21]
Pillar 6 — Capital Markets Now Amplify the Ghosts
The Ghost-Megawatt problem is no longer confined to utility planning; it has become entangled with the largest capital-allocation event in corporate history. The four biggest hyperscalers alone plan roughly 725 to 760 billion dollars of 2026 capital expenditure—up nearly eighty percent from 2025—against Goldman Sachs projections of 5.3 trillion dollars through 2030, with free cash flow compressing across the group and Alphabet’s shares falling seven percent on the very Q2-2026 report in which it raised its capex ceiling.[34][35][36][38][40] Layered atop the corporate spending are circular financing structures in which suppliers invest in customers who buy their products, and a datacenter development sector whose announced projects, in aggregate, exceed what can physically be built within their stated timeframes.[41] In such an environment, uncorrected Ghost Megawatts do double damage: they inflate the apparent addressable demand against which equity and credit are priced, and—when gates finally force the truth—they generate abrupt revisions of the kind that rating agencies now flag as material credit factors.[16] Ghost-Megawatt governance is therefore also financial-stability policy: the credibility gates that protect ratepayers simultaneously protect investors from underwriting the same illusion twice.
Pillar 7 — America’s AI Competition Is Becoming a Competition in Forecasting Credibility
The Five-Layer AI Economy ultimately rests on physical resources: applications create model demand, models create compute demand, compute creates datacenter demand, and datacenters create power demand. But if uncertainty at every higher layer is passed downward without adjustment, Layer 1 becomes responsible for constructing half-century infrastructure around the most speculative assumptions in the stack—and the compounding of hedges through the layers guarantees that the assumptions arriving at the bottom are the most inflated of all. The countries and states that solve this information problem will possess a durable advantage: they will build enough electricity for genuine AI expansion without wasting capital on projects that never arrive, and their planning institutions will attract exactly the credible developers that rival jurisdictions repel with chaos or credulity. The next stage of AI infrastructure leadership requires not merely more megawatts, but better intelligence about which megawatts are real—and better intelligence about intelligence infrastructure is a competition America is unusually well positioned to win, if it chooses to run.
Conclusion: The Megawatts That Haunt the AI Economy
Artificial intelligence has given America an electricity problem of extraordinary scale, but the problem is more complicated than the familiar argument that datacenters simply need too much power. The United States must simultaneously answer two apparently contradictory questions. Where will it find enough electricity for the AI infrastructure that is actually coming—the infrastructure attested by 725 billion dollars of hyperscaler capital expenditure, by contracted load books like Ohio’s 17,861 megawatts, by the measured trajectories of Lawrence Berkeley National Laboratory and the International Energy Agency? And how will it avoid constructing billions of dollars of infrastructure for the AI projects that are not coming—the duplicates, the options, the speculative reservations that inflate a 700-gigawatt queue above any buildable reality?[1][15][28][29][35]
The emerging evidence from Texas, Pennsylvania and Ohio suggests that American governments are beginning to understand the distinction. Texas has moved toward auditing enormous interconnection requests before allowing projects to advance, freezing the fastest-growing queue in the world while it verifies what stands behind it. Pennsylvania has imposed binding permitting, transparency and community requirements that convert announcements into applications or into silence. Ohio has constructed financial mechanisms requiring large customers to demonstrate commitment through contracts, collateral and minimum payments, and has watched eighty percent of its claimed pipeline decline the invitation. These approaches differ in instrument—audit, permit, tariff—but all are attempts to solve the same underlying information problem, and their bipartisan, cross-regional convergence within a span of months is itself evidence of how acute the problem has become.[3][9][13]
That information problem explains why I chose the title Ghost Megawatts. Reuters’ phrase “ghost demand” captures the existence of questionable demand; Ghost Megawatts extends the idea into an analytical unit. It asks how much electricity inside a forecast is supported by enough economic, financial, technical and regulatory evidence to justify treating it as future operating load—and it insists that the question be answered in megawatts, because electricity is measured in megawatts, transmission is constructed around megawatts, generation is financed around megawatts, and utility customers ultimately pay for infrastructure designed to deliver megawatts.
A Ghost Megawatt is therefore more than an exaggerated forecast. It is a claim on future infrastructure. It can occupy a position in an interconnection process, crowd a real project out of a study cluster, and delay the credible by the weight of the incredible. It can inflate a utility’s load forecast and the capital plan justified by it. It can move a capacity auction and add dollars to a pensioner’s electricity bill three states away. It can anchor a land valuation, a municipal tax abatement, a campaign speech, a bond offering. And it can do all of these things—as PJM’s 29.4 billion dollars of datacenter-driven capacity charges attest—without ever powering a single Nvidia GPU.[16]
This is why Ghost Megawatts fits naturally within the Five-Layer AI Economy. The upper layers of the system—applications, agents and models—expand at software speed, revising themselves in weeks. Chips and datacenters expand at industrial speed, in quarters and years. Electricity grids, transmission networks and power plants expand at infrastructure speed, in decades. Those clocks are profoundly different, and the Ghost Megawatt is what accumulates in the gaps between them: an AI company can revise a model architecture in months, a hyperscaler can change its preferred geography in weeks, a developer can withdraw a project in days—but a transmission line constructed around those expectations can remain in the ground for half a century, and its costs in the rate base for nearly as long.
The great infrastructure challenge of the next stage of artificial intelligence is therefore not simply to build faster. It is to determine what deserves to be built. That will require utilities to separate inquiries from commitments; governments to separate announcements from executable projects; financial markets to distinguish pipelines from realizable demand; grid operators to weight what they cannot verify; and policymakers to protect ordinary ratepayers without starving legitimate AI infrastructure of the electricity it genuinely needs. The scholars have supplied the foundations—Koomey’s taxonomy of demand claims, Norris’s arithmetic of flexibility, Martin and Peskoe’s anatomy of cost-shifting, Shehabi’s disciplined measurement—and the states have supplied the first working instruments. What remains is synthesis, standardization, and the institutional patience to run the experiment well.[20][23][27][31]
America should not respond to Ghost Megawatts by becoming afraid of megawatts. Nor should it respond to the AI race by believing every megawatt placed into an interconnection queue. The better objective is a more sophisticated one, and it can be stated as a program: find the ghosts, through disclosure and audit; price the ghosts, through deposits, collateral and minimum-demand obligations; remove the ghosts from infrastructure forecasts, through probability weighting and standardized classification; and then build aggressively—genuinely aggressively, at the pace the credible demand deserves—for the megawatts that remain.
If the artificial-intelligence economy is becoming one of the largest infrastructure transformations in modern American history—and the evidence assembled here suggests it is—then the difference between imagined electricity and consumed electricity may ultimately determine where trillions of dollars of capital are deployed, which regions prosper from the transformation and which merely pay for it, and whether the United States converts its energy institutions into an advantage or an anchor. That is the deeper meaning of Ghost Megawatts: the measurable distance between the electricity the AI economy says it will need and the electricity the AI economy will actually consume—and the proposition, defended throughout this paper, that a nation which learns to measure that distance can close it.
Footnotes and Endnotes:
[1] Laila Kearney, Reuters, “Analysis: Texas’ Halt on Powering Data Centers Reflects US Reckoning Over ‘Ghost’ Demand,” September 1, 2026. https://www.usnews.com/news/top-news/articles/2026-09-01/analysis-texas-halt-on-powering-data-centers-reflects-us-reckoning-over-ghost-demand
[2] Tech Startups, “AI Data Centers Request 700 GW of U.S. Power, But Much of the Demand May Be an Illusion,” September 1, 2026 (quoting PUCT Chairman Thomas Gleeson). https://techstartups.com/2026/09/01/ai-data-centers-request-700-gw-of-u-s-power-but-much-of-the-demand-may-be-an-illusion/
[3] Akin Gump Strauss Hauer & Feld LLP, “Texas Pauses Data Center Interconnections Pending Statewide Audit,” August 2026 (analysis of Governor Greg Abbott’s August 3, 2026 letter to the PUCT and ERCOT). https://www.akingump.com/en/insights/alerts/texas-pauses-data-center-interconnections-pending-statewide-audit
[4] Robert Walton, Utility Dive, “Facing an Estimated 474 GW of Interconnection Requests, Texas Hits Pause on Data Centers,” August 2026. https://www.utilitydive.com/news/texas-hits-pause-data-center-interconnections/827046/
[5] Gibson, Dunn & Crutcher LLP, “What Governor Abbott’s Data Center Audit Directive Means for ERCOT and the Batch Zero Study Process,” August 2026. https://www.gibsondunn.com/what-governor-abbotts-data-center-audit-directive-means-for-ercot-and-the-batch-zero-study-process/
[6] The Texas Tribune, “Texas Will Audit Up to 300 Projects, Mostly Data Center Proposals,” August 14, 2026 (quoting ERCOT general counsel Chad Seely). https://www.texastribune.org/2026/08/14/texas-data-center-approval-pause-ercot-power-grid/
[7] Ethan Howland, Utility Dive, “ERCOT Aims to Complete Texas Governor’s Data Center Audit by December,” August 2026. https://www.utilitydive.com/news/ercot-texas-puc-data-center-audit/828472/
[8] Troutman Pepper Locke LLP, “Texas Hits Pause on Data Center Grid Connections Amid Growing Oversight Push,” August 2026 (Senate Bill 6 and regulatory foundations). https://www.troutman.com/insights/texas-hits-pause-on-data-center-grid-connections-amid-growing-oversight-push/
[9] Office of Governor Josh Shapiro, Commonwealth of Pennsylvania, “Governor Shapiro Signs Executive Order on Data Center Development in PA,” August 18, 2026. https://www.pa.gov/governor/newsroom/2026-press-releases/governor-shapiro-signs-executive-order-on-data-center-developmen
[10] Commonwealth of Pennsylvania, Executive Order 2026-05, “Protecting Pennsylvania Consumers from Data Center Impacts,” August 18, 2026 (full text). https://www.pa.gov/content/dam/copapwp-pagov/en/governor/documents/eo2026_05_protecting%20pennsylvania%20consumers%20from%20data%20center%20impacts_final_executed.pdf
[11] The Philadelphia Inquirer, “Gov. Josh Shapiro Signs Executive Order Restricting Data Center Development in Pennsylvania,” August 18, 2026 (quoting Governor Shapiro and Dan Diorio, Data Center Coalition). https://www.inquirer.com/politics/pennsylvania/josh-shapiro-data-center-order-20260818.html
[12] Duane Morris Government Strategies, “Governor Shapiro’s Data Center Executive Order: What It Means for Pennsylvania,” August 2026. https://statecapitallobbyist.com/artificial-intelligence-ai/governor-shapiros-data-center-executive-order-what-it-means-for-pennsylvania/
[13] AEP Ohio, “AEP Ohio Updates PUCO on Data Center Load: Figures Show Tariff Is Working,” February 2026. https://www.aepohio.com/company/news/view?releaseID=10753
[14] Ohio Capital Journal, “AEP Ohio Says New Data Center Tariff Is Working, Critics Aren’t Buying It,” February 20, 2026 (quoting AEP Ohio President and COO Marc Reitter). https://ohiocapitaljournal.com/2026/02/20/aep-ohio-says-new-data-center-tariff-is-working-critics-arent-buying-it/
[15] New Project Media, “AEP Ohio Touts ‘Right-Sized Load’ as Data Center Pipeline Shrinks Further Under New Tariff,” March 2026. https://newprojectmedia.com/policy-aep-ohio-touts-right-sized-load-as-data-center-pipeline-shrinks-further-under-new-tariff/
[16] Ethan Howland, Utility Dive, “Data Centers Drove $6.3B in PJM Capacity Auction Costs: Market Monitor,” July 2026 (quoting Joseph Bowring, Monitoring Analytics; Morningstar DBRS commentary). https://www.utilitydive.com/news/pjm-data-centers-capacity-auction-imm-bowring/825626/
[17] The Hill, “Data Center Electricity Demand Is Expected to Drive Up Costs After Recent Power Auction,” July 2026. https://thehill.com/policy/technology/5970522-data-center-power-costs-pjm/
[18] Ethan Howland, Utility Dive, “Data Centers Were 40% of PJM Capacity Costs in Last Auction: Market Monitor,” January 7, 2026. https://www.utilitydive.com/news/data-centers-pjm-capacity-auction/808951/
[19] Cathy Kunkel, Institute for Energy Economics and Financial Analysis (IEEFA), “Projected Data Center Growth Spurs PJM Capacity Prices by Factor of 10,” 2025. https://ieefa.org/resources/projected-data-center-growth-spurs-pjm-capacity-prices-factor-10
[20] American Public Power Association, “Study Examines Potential for Integration of Large Flexible Loads in U.S. Power Systems,” February 2025 (on Tyler Norris, Tim Profeta, Dalia Patiño-Echeverri and Adam Cowie-Haskell, Rethinking Load Growth, Nicholas Institute, Duke University, 2025). https://www.publicpower.org/periodical/article/study-examines-potential-integration-large-flexible-loads-us-power-systems
[21] Latitude Media, “A Status Update on Data Center Flexibility,” October 2025 (interview with Tyler Norris, Duke University). https://www.latitudemedia.com/news/a-status-update-on-data-center-flexibility/
[22] Jeff St. John, Canary Media, “Utilities Are Flying Blind on Data Center Demand. That’s a Big Problem,” February 2025 (quoting Jonathan Koomey and Mario Sawaya, AECOM). https://www.canarymedia.com/articles/utilities/utilities-are-flying-blind-on-data-center-demand-thats-a-big-problem
[23] Jonathan Koomey, Zachary Schmidt and Tania Das, Bipartisan Policy Center and Koomey Analytics, “Electricity Demand Growth and Data Centers: A Guide for the Perplexed,” February 2025. https://bipartisanpolicy.org/download/?file=/wp-content/uploads/2025/02/BPC-Report-Electricity-Demand-Growth-and-Data-Centers-A-Guide-for-the-Perplexed.pdf
[24] Bianca Giacobone, Latitude Media, “Phantom Data Centers Are Flooding the Load Queue,” March 26, 2025. https://www.latitudemedia.com/news/phantom-data-centers-are-flooding-the-load-queue/
[25] Financial Times (republished by the Large Public Power Council), “‘Phantom’ Data Centres Muddy Forecasts for US Power Needs,” November 2025 (quoting Tom Falcone, LPPC, and U.S. Energy Secretary Chris Wright). https://www.lppc.org/news/phantom-data-centres-muddy-forecasts-for-us-power-needs
[26] Wood Mackenzie pipeline analysis, summarized in TFTC, “Wood Mackenzie: 72% of US AI Data Center Power Requests Are Phantom Load,” August 2026 (quoting Ben Hertz-Shargel, Wood Mackenzie). https://www.tftc.io/ai-data-center-phantom-power-demand-wood-mackenzie
[27] Arman Shehabi, Sarah J. Smith, Alex Hubbard, Alexander Newkirk, Nuoa Lei, Md Abu Bakar Siddik, Billie Holecek, Jonathan G. Koomey, Eric R. Masanet and Dale A. Sartor, Lawrence Berkeley National Laboratory, “2024 United States Data Center Energy Usage Report,” LBNL-2001637, December 2024. https://eta-publications.lbl.gov/sites/default/files/2024-12/lbnl-2024-united-states-data-center-energy-usage-report_1.pdf
[28] Sarah J. Smith, Alex Hubbard, Alexander Newkirk, Mohan Ganeshalingam, Billie Holecek, Dale A. Sartor, Michael Mills and Arman Shehabi, Lawrence Berkeley National Laboratory, “United States Data Center Energy Usage Report: 2025 Update,” 2026. https://eta.lbl.gov/publications/united-states-data-center-energy-2025
[29] International Energy Agency, “Energy and AI,” 2025. https://www.iea.org/reports/energy-and-ai
[30] The Brookings Institution, “Global Energy Demands Within the AI Regulatory Landscape,” June 2026. https://www.brookings.edu/articles/global-energy-demands-within-the-ai-regulatory-landscape/
[31] Ethan Howland, Utility Dive, “Utilities May Subsidize Data Center Growth by Shifting Costs to Other Ratepayers: Harvard Law Paper,” March 2025 (on Eliza Martin and Ari Peskoe, “Extracting Profits from the Public: How Utility Ratepayers Are Paying for Big Tech’s Power,” Harvard Law School Electricity Law Initiative, 2025). https://www.utilitydive.com/news/utilities-subsidize-data-center-growth-ratepayer-cost-shif-harvard-peskoe/742001/
[32] Harvard University, Salata Institute for Climate and Sustainability, “The Data Center Boom Is Colliding With the Grid’s Hardest Problems” (interview with Ari Peskoe, Harvard Law School), March 17, 2026. https://salatainstitute.harvard.edu/data-centers-ai-artificial-intelligence-grid-permitting-transmission-electricity-energy
[33] Harvard Magazine, “How AI Could Be Raising Your Energy Bill” (on research by Ari Peskoe and Eliza Martin), 2025. https://www.harvardmagazine.com/2025/07/harvard-ai-increasing-energy-costs
[34] CNBC, “Amazon, Meta and Microsoft Face Skeptical Investors This Week After Google Report Sparked Sell-Off,” July 28, 2026. https://www.cnbc.com/2026/07/28/hyperscalers-face-higher-capex-scrutiny-after-alphabet-report-panned.html
[35] Statista, “Big Tech’s AI Spending to Reach $760 Billion in 2026,” July 31, 2026 (based on Q2-2026 earnings of Microsoft, Alphabet, Meta and Amazon). https://www.statista.com/chart/35046/capital-expenditure-of-meta-alphabet-amazon-and-microsoft/
[36] CNBC, “Tech AI Spending Approaches $700 Billion in 2026, Cash Taking Big Hit,” February 6, 2026. https://www.cnbc.com/2026/02/06/google-microsoft-meta-amazon-ai-cash.html
[37] Vision Times, “Texas Halts New Data Center Grid Connections as ‘Ghost Demand’ Surges,” September 1, 2026 (noting the termination of the QTS Digital Gateway project in Virginia). https://www.visiontimes.com/2026/09/01/texas-halts-new-data-center-grid-connections-as-ghost-demand-surges.html
[38] Uncover Alpha, “Amazon, Google, Microsoft, Meta Q2 Earnings: Key Takeaways,” August 3, 2026 (Q2-2026 capital-expenditure guidance and Amazon CEO Andy Jassy remarks). https://www.uncoveralpha.com/p/amazon-google-microsoft-meta-q2-earnings
[39] Latitude Media, “Data Center Experts on Energy Use for AI: ‘Calm the Heck Down,’” March 2025 (quoting Jonathan Koomey). https://www.latitudemedia.com/news/data-center-experts-on-energy-use-for-ai-calm-the-heck-down/
[40] Yahoo Finance, “Meta, Microsoft, Amazon, and Alphabet Are About to Spend a Shocking Amount of Money to Dominate the AI Era” (Goldman Sachs Research estimates of $5.3 trillion hyperscaler capex, FY2025–FY2030), June 2026. https://finance.yahoo.com/sectors/technology/article/meta-microsoft-amazon-and-alphabet-are-about-to-spend-a-shocking-amount-of-money-to-dominate-the-ai-era-115359575.html
[41] Barchart, republished by Yahoo Finance, “How ‘Phantom’ Data Center Projects Are Making It Impossible to Forecast AI Power Demand,” August 2026. https://finance.yahoo.com/technology/ai/articles/phantom-data-center-projects-making-160554911.html
