Introduction: The Morning After the Datacenter Bidding War
The Script That Governed a Decade
For most of the first half of the 2020s, the economic-development conversation surrounding hyperscale datacenters followed a script so familiar that it had become nearly ceremonial. A governor would appear at a podium beside an executive from Amazon, Microsoft, Google, Meta, Oracle, or a specialized infrastructure developer. A number would be announced — hundreds of millions, then billions, then tens of billions of dollars. State legislatures would confirm or extend sales-and-use-tax exemptions on servers, networking gear, transformers, switchgear, chillers, and replacement hardware. Counties would negotiate property-tax abatements running fifteen, twenty, or fifty years. Economic-development agencies would accelerate permits, assemble land, extend roads, upsize water mains, and advertise electricity that was inexpensive by the standards of the industrialized world.
Underneath the ceremony sat a single, rarely examined economic assumption: that datacenter capital was mobile, that states were competing against one another for a finite pool of it, and that the jurisdiction offering the most attractive package would capture the investment. Everything followed from that premise. If capital was scarce and locations were fungible, then the rational move for any individual state was to bid. The bidding was not irrational; given the assumption, it was close to optimal. Virginia bid harder than anyone and became, by some measures, the largest concentration of datacenter capacity on earth, hosting roughly a third of the world’s facilities. Georgia, Ohio, Indiana, Texas, Arizona, Iowa, and Nebraska followed with their own variants. The exemptions were real money: Virginia’s datacenter sales-and-use-tax exemption alone accounted for fifty-three percent of all state economic-development spending across ten fiscal years, some $2.73 billion between 2021 and 2024, eleven times larger than the second-largest incentive on the books. [34]
By August 2026, the assumption underneath the ceremony had begun to invert.
Harrisburg, August 18, 2026
On a Tuesday morning in the Pennsylvania Capitol, Governor Josh Shapiro signed Executive Order 2026-05 and said something that would have been politically unthinkable eighteen months earlier. He began, as he almost always does, with the familiar assurance:
“Yes, Pennsylvania is open for business.”
— Governor Josh Shapiro, Commonwealth of Pennsylvania [5]
Then he added the twist. The Executive Order directed the Department of Environmental Protection to review datacenter permit applications only where the developer had made a legally binding commitment to the Governor’s Responsible Infrastructure Development — GRID — Requirements, and only where the developer had already obtained local land-use approval. It removed every AI datacenter proposal from Pennsylvania’s Permit Fast Track Program and barred the program from considering datacenter projects going forward. It prohibited Commonwealth agencies from entering nondisclosure agreements with datacenter developers. It required developers to secure or finance the incremental electricity their projects consume, to pay the interconnection, transmission, distribution, and network-upgrade costs their projects cause rather than socializing those costs across the ratepayer base, and to source an increasing share of that power from clean generation — solar, advanced nuclear, or battery storage. It required robust community benefits agreements and local hiring processes. [1] [2] [4]
Shapiro’s framing was unusually blunt for an economic-development announcement. He described more than one hundred datacenter proposals under discussion in the Commonwealth, of which only fifteen had applied for even one DEP permit and only five had obtained all permits required for a first phase of development. Most of the remainder, he argued, were speculative — thinly financed, lacking anchor tenants, and unlikely ever to be built — yet were nonetheless generating real friction in real communities:
“these speculators are nevertheless scaring our communities”
— Governor Josh Shapiro, at the August 18, 2026 signing [2]
And he described the intended effect of the order in the language of a filter rather than a welcome mat:
“Taken together, these steps will effectively block objectionable, unwarranted, and costly projects”
— Governor Josh Shapiro, Executive Order 2026-05 announcement [3]
The political symbolism was difficult to miss. Only a year earlier, Pennsylvania had celebrated Amazon’s announced $20 billion investment in Luzerne and Bucks counties — described at the time as the largest private-sector investment in the Commonwealth’s history — and had done so through the Fast Track permitting program that the August 2026 order now closed to the entire industry. [3] Pennsylvania still wants AI investment. It still wants electricity investment, advanced nuclear development, construction employment, and the industrial supply chain that follows digital infrastructure. What changed was not the state’s appetite. What changed was the state’s negotiating position.
The old sentence was: Come here, and we will help lower your costs.
The new sentence is: Come here, if you can pay the costs your project creates.
The Same Reversal, From the Opposite Political Direction
Fifteen days earlier and 1,500 miles southwest, a Republican governor had arrived at a structurally identical conclusion by an entirely different political route. On August 3, 2026, Texas Governor Greg Abbott sent a letter to Public Utility Commission of Texas Chairman Thomas Gleeson and ERCOT President and Chief Executive Officer Pablo Vegas directing both institutions to conduct a comprehensive verification and audit of every datacenter project advancing through ERCOT’s interconnection process — and directing that no additional datacenter be approved to move forward until that audit was complete. Projects failing the audit, the letter said, must be denied connection to the Texas grid. [8] [9] [10]
The numbers in Abbott’s letter explain the intervention better than any rhetoric could. ERCOT was then evaluating approximately 474 gigawatts of requests to connect to the Texas grid — more than five times ERCOT’s record peak electricity demand — and datacenters accounted for roughly ninety percent of those new power requests. [8] Abbott framed the directive in the plainest possible terms:
“Our top priority is to protect Texans’ safety and quality of life”
— Governor Greg Abbott, letter to PUCT and ERCOT, August 3, 2026 [11]
ERCOT responded the same day with Market Notice M-A080326-01, announcing that it would not deliver Batch Zero Large Load classification notifications by the previously scheduled August 7 deadline and would instead seek a good-cause exception from the Commission. [10] At the August 14 open meeting, ERCOT told the Commission it expected the verification process to take several months — but less than nine — and that it would also issue requests for information to loads between 25 and 75 megawatts outside Batch Zero, roughly 8,766 megawatts’ worth. [9] By mid-August, Texas regulators were describing an audit universe of roughly 250 to 300 projects. [12] Abbott had already, in June, directed the Commission to ensure that datacenter interconnections produced reduced rates for residential customers and that datacenters pay the cost of any new transmission infrastructure built to serve them. [12]
Texas did not ban datacenters. Texas placed a verification gate in front of the scarcest asset it controls — a position in the interconnection queue of the fastest-growing large grid in North America.
Five States, One Physical Reality
Pennsylvania and Texas were not isolated. They were the two most conspicuous entries in a sequence that ran through the summer of 2026 across four time zones and both political parties.
On June 25, Massachusetts Governor Maura Healey halted acceptance of applications for the Commonwealth’s newly launched twenty-year Qualified Data Center Sales and Use Tax Exemption — a program she had herself signed into law as part of the 2024 economic development act, and whose final implementing regulations had taken effect only the previous month. [23] [25] On July 14, New York Governor Kathy Hochul signed Executive Order 62, creating what her office described as the nation’s first statewide moratorium on new hyperscale datacenters, pausing discretionary state environmental permits for up to one year while the Department of Public Service develops a Generic Environmental Impact Statement and runs the “Energize NY” proceeding. [15] [19] On July 20, Nebraska Governor Jim Pillen signed an executive order barring new datacenter applications from approval under the ImagiNE Nebraska Act, the state’s flagship performance-based incentive program, and establishing a task force to assess impacts on land, water, and electricity. [28] [30]
Their political philosophies could hardly be more different. Shapiro and Hochul are Democrats governing large, unionized, high-cost northeastern states. Healey is a Democrat governing a state with among the highest electricity prices in the continental United States. Abbott and Pillen are Republicans governing energy-abundant states whose entire economic-development brand rests on low taxes and light regulation. Yet each confronted the same physical arithmetic, and the arithmetic does not read party registration.
Electric grids do not recognize political ideology. Neither do transformer lead times, transmission queues, aquifer recharge rates, substation construction schedules, or gigawatt-scale demand arriving faster than gigawatt-scale supply.
What These Decisions Actually Represent
It would be a serious analytical error to read this sequence as a national turn against artificial intelligence. It is not. Pennsylvania’s order explicitly contemplates continued datacenter development under conditions. Nebraska’s governor said directly:
“This is not a moratorium on data centers”
— Governor Jim Pillen, Nebraska, July 2026 [31]
New York’s moratorium is time-limited and paired with a rate proceeding whose explicit purpose is to define the terms on which hyperscale load can be served. Texas’s audit is a verification exercise, not a prohibition, and it exempts projects that bring their own generation or sit outside ERCOT.
What the sequence represents is something more economically consequential than a ban, and considerably more durable: a renegotiation of who pays for scarcity.
For fifteen years, states worried that datacenter companies would leave if another jurisdiction offered cheaper electricity, faster permits, or larger tax incentives. The generative-AI infrastructure cycle has produced a different environment. Hyperscalers, frontier model developers, sovereign investors, cloud providers, and specialized infrastructure firms are now pursuing enormous quantities of physical compute capacity simultaneously and on overlapping timelines. Capital is abundant — the four largest American hyperscalers alone raised their combined 2026 capital-expenditure guidance to roughly $725 billion after second-quarter earnings, up approximately seventy-seven percent from about $410 billion in 2025. [101] [102] But suitable substations, energized transmission corridors, dispatchable generation, cooling water, large-frame transformers, skilled electrical labor, developable land, and — critically — politically acceptable host communities are not abundant.
The scarce asset in this transaction is therefore no longer the hyperscaler’s money.
The scarce asset is the state’s capacity to accommodate the hyperscaler’s money.
That inversion is the central subject of this paper.
The Reversal Has an Audience
None of this is happening in a technocratic vacuum. It is happening in an election year, in front of an electorate whose views moved faster in eight months than almost any measured public opinion on an infrastructure question in recent memory. Emerson College Polling found that opposition to datacenters being built in or near one’s own community rose from 42 percent in December 2025 to 63 percent in July 2026 — a twenty-one point swing — with support falling to 27 percent and the neutral share collapsing from 25 percent to 10 percent. [80] [84] Opposition rose in every educational cohort and was strongest among the most educated: 79 percent among postgraduate degree holders, up twenty-seven points in eight months. [80] In the same survey, voters described themselves as concerned rather than excited about artificial intelligence by 63 percent to 14 percent. [81]
The market consequences were already visible. Data Center Watch, a project of the AI intelligence firm 10a Labs, found that at least 75 projects worth approximately $130 billion were blocked or delayed by local opposition in the first quarter of 2026 alone — roughly matching the total for all of 2025 in three months — with active opposition groups more than doubling to 833 across 49 states. Their assessment of what that meant was unusually direct:
“The quarter reflected a structural shift rather than a cyclical spike”
— Data Center Watch (10a Labs), Q1 2026 Report [87]
More than three hundred datacenter-related bills were introduced in state legislatures in the first six weeks of 2026, which the same report characterized as a decisive move away from incentive-focused policy toward regulatory oversight. [88]
This is the environment in which five governors acted. The remainder of this paper is an attempt to explain what they did, why the economics support it, where the economics do not support it, and what it means for the companies, investors, communities, and public officials who now have to operate inside it.
Why “Subsidy Retrenchment”: A Note on the Title
Why Not “Regulation,” “Backlash,” or “Moratorium”
Naming matters in policy analysis because the name determines the comparison set. Call this phenomenon “datacenter regulation” and it joins a category populated by zoning ordinances, noise limits, and setback requirements — important, but local and procedural. Call it “AI backlash” and it becomes a story about sentiment, which implies it will pass when sentiment passes. Call it “the politics of electricity prices” and it becomes a story about a single input cost, which understates the range of resources actually at issue. Call it a “moratorium” and the analysis collapses into a binary — build or don’t build — that describes only one of the five instruments actually in use and the least durable one at that.
None of those names captures what is economically distinctive about 2026. What is distinctive is that governments are reversing direction on a specific class of fiscal and administrative concessions that they had previously treated as the mandatory admission fee for participation in a national competition. That reversal deserves its own name, and the name should be precise about both halves of the motion.
“Subsidy,” Defined Broadly
Subsidy, in this paper, does not mean only an explicit check written by a government to a company. It means the full set of mechanisms through which a government reduces the private cost of locating a large industrial facility within its jurisdiction. That set includes:
- sales-and-use-tax exemptions on servers, networking equipment, power systems, cooling systems, and replacement hardware;
- property-tax abatements, payment-in-lieu-of-tax agreements, and millage preferences;
- corporate income-tax credits and performance-based incentive packages;
- expedited, prioritized, or otherwise preferential permitting;
- publicly financed roads, water mains, sewer capacity, and site preparation;
- favorable electric-rate treatment, including confidential special contracts negotiated outside the ordinary rate case;
- socialized transmission, distribution, and network-upgrade costs recovered from the general ratepayer base rather than from the causing customer;
- and the administrative confidentiality — nondisclosure agreements, code-named applicants, redacted filings — that made the true value of the foregoing difficult for legislatures, regulators, and the public to observe.
The last item on that list is the one most often omitted from subsidy accounting and the one Pennsylvania’s Executive Order attacked most directly. Opacity is itself a subsidy: it lowers the political cost of a concession by preventing anyone from pricing it.
The breadth of that definition matters because the retrenchment now underway operates across all eight categories simultaneously. Nebraska withdrew a performance-based tax credit. Massachusetts froze a sales-tax exemption. Pennsylvania withdrew expedited permitting and prohibited nondisclosure agreements and attacked cost socialization. New York attacked cost socialization and proposed eliminating tax subsidies outright. Texas attacked queue access. Virginia, uniquely, preserved its exemption while imposing an entirely new consumption tax. A framework that examined only tax incentives would miss most of the action.
“Retrenchment,” Defined Precisely
Retrenchment is chosen over “elimination,” “prohibition,” or “reversal” because it is the only word that accurately describes a partial, conditional, and negotiated pullback.
States are not, in the main, eliminating datacenters. They are not abandoning AI economic development. Several of them are actively courting the generation and transmission investment that AI load makes financeable. What they are doing is withdrawing, narrowing, conditioning, or repricing benefits that were previously offered with fewer demands attached. Nebraska terminated access to a program while explicitly denying that it had imposed a moratorium. Massachusetts froze applications while publishing a framework describing exactly what a compliant applicant would look like. Pennsylvania conditioned access to its exemption and its permitting process on compliance with enumerated standards. Texas conditioned queue advancement on passing an audit. These are different legal instruments operating on different margins, but they point in the same economic direction, and the direction is not “no.” The direction is “yes, at a price that reflects what you are consuming.”
Retrenchment also carries the correct connotation of reversion to a prior baseline. Nothing in these orders is exotic. Cost causation — the principle that the customer who causes a cost should bear it — is among the oldest doctrines in public-utility ratemaking. Requiring local land-use approval before state environmental review is ordinary administrative sequencing. Demanding proof of financing before allocating scarce infrastructure capacity is standard commercial practice. What was exotic was the fifteen-year period in which these ordinary principles were suspended for one industry because states believed suspension was the price of admission.
The Subtitle and the Word “Charging”
The second half of the title — When States Stop Bidding for AI Datacenters, and Start Charging Them for Scarcity — is where the analytical claim actually lives, and the word charging should be read broadly.
The charge sometimes appears directly and in cash. Virginia’s new $0.011-per-kilowatt-hour electricity consumption tax on datacenter operators, effective July 1, 2026, is a literal excise on compute, expected to generate up to $1.2 billion over two years with annual collections capped at $600 million and pro-rata refunds above the cap. [35] [36] A grid-access premium of the kind New York’s Energize NY proceeding is designed to construct is a literal price. An interconnection payment, a network-upgrade contribution, a water-withdrawal charge, or a community-benefit fund contribution is a literal transfer.
But the charge just as often appears indirectly, as a withdrawn benefit or an imposed obligation whose cash-equivalent value is substantial even though no invoice is ever issued. The loss of a twenty-year sales-tax exemption is a charge. A requirement to self-generate is a charge, because it converts an operating expense into a capital obligation on the corporate balance sheet. A larger interconnection deposit, a take-or-pay minimum-billing provision at eighty-five percent of contracted capacity, a ten- or twenty-year minimum contract term, a construction-milestone bond, an environmental mitigation requirement, an interruptibility obligation during grid emergencies, or the loss of expedited permitting — each is a charge in economic substance. In every one of these cases, something that was previously treated as an inexpensive or socially shared input is being reclassified as a scarce economic resource with a price.
The Framework
The paper therefore defines its central framework as a five-term composite:
Subsidy Retrenchment = Incentive Withdrawal + Cost Causation + Scarcity Pricing + Conditional Grid Access + Community Return
Each term is a distinct policy instrument with distinct legal machinery, distinct distributional consequences, and distinct political durability. Incentive Withdrawal operates through tax and economic-development statute and is the most politically visible and the least technically demanding. Cost Causation operates through utility ratemaking and interconnection agreements and is the most technically demanding and the most economically consequential. Scarcity Pricing operates through tariffs, excises, and premiums and is the newest and least settled. Conditional Grid Access operates through interconnection queues and reliability standards and is the most powerful because it controls the binding constraint. Community Return operates through land use, community-benefit agreements, and host-locality funds and is the most locally variable and the most legally fragile.
The paper’s argument is that these five instruments are converging into a coherent successor regime to the incentive-auction model, and that this transition could become one of the defining state-level economic-policy changes of the AI infrastructure era.
Section 1: From Datacenter Recruitment to Scarcity Accounting
Before examining what changed, it is necessary to be precise about what the prior regime actually was, because the prior regime is frequently caricatured. It was not stupidity, and it was not simple capture. It was a rational response to a specific structure of scarcity that has since inverted. Understanding why the old bargain made sense is the only way to understand why its collapse is significant rather than merely fashionable, and it is also the only way to evaluate honestly the arguments of those who believe the collapse is a mistake.
1.1 The Old Datacenter Bargain
The cloud-computing era created a genuine interstate competition for a genuinely mobile asset. Between roughly 2010 and 2022, a hyperscale datacenter was a comparatively modest object by the standards of heavy industry: a few tens of megawatts, a large but not extraordinary land footprint, moderate water consumption, low permanent headcount, and — decisively — a wide range of technically acceptable locations. Fiber routes mattered, latency to population centers mattered, natural-disaster exposure mattered, and electricity price mattered. But almost every state could clear the technical bar. When many jurisdictions can host a facility and only a few facilities are being built, the facility captures the surplus. That is not a policy failure; that is Bertrand competition.
States evaluated these projects through five channels: announced capital investment, construction-phase employment, local property-tax revenue, technology-sector branding, and ancillary economic activity. Only two of the five were reliably large. Capital investment was enormous and real. Construction employment was substantial and temporary. Property-tax revenue was meaningful where abatements did not consume it. Branding was speculative. Ancillary activity was, as Good Jobs First and others documented repeatedly, thin — datacenters purchase very few goods and services that local businesses can supply. [58]
Sales-and-use-tax exemptions became the dominant instrument for a structural reason. Datacenter capital expenditure is unusually concentrated in taxable tangible personal property. Servers, accelerators, networking equipment, uninterruptible power supplies, switchgear, generators, chillers, and computer-room air handlers are almost entirely equipment purchases, and — critically — they are replaced on a three-to-six-year cycle. A sales-tax exemption on datacenter equipment is therefore not a one-time closing cost. It is a perpetual annuity indexed to the refresh cycle, and in a rapid capacity buildout it compounds. Virginia’s experience illustrates the magnitude: companies reported investing $80.6 billion in Virginia datacenters across fiscal years 2024 and 2025 while claiming $3.2 billion in sales-tax exemptions, compared with $23.2 billion and $903 million respectively in fiscal 2023, and $13.8 billion and $673 million in fiscal 2022. [40] The exemption grew roughly fivefold in three years.
The pro-incentive argument was, and remains, economically serious, and it should not be dismissed. Virginia’s Joint Legislative Audit and Review Commission estimated that ninety percent of the datacenter investment made by exemption beneficiaries would not have occurred in Virginia without the exemption — a “but-for” percentage that a more recent JLARC analysis put as high as one hundred percent — and calculated that applying that ninety-percent reduction implied Virginia and its localities could have forgone nearly $1.3 billion in net tax revenues from fiscal 2024–2025 investments absent the exemption, as total revenues fell from $2.1 billion to $0.8 billion. [38] JLARC also estimated that each dollar of forgone revenue attributable to the exemption generated approximately $0.48 in additional state revenue. [35] The Virginia Economic Development Partnership put the industry’s annual contribution at:
“74,000 jobs, $5.5 billion in labor income, and $9.1 billion in GDP”
— Jason El-Koubi, President and CEO, Virginia Economic Development Partnership [40]
A 2022 tax-incentive evaluation from the Carl Vinson Institute of Government at the University of Georgia reached a comparable conclusion for Georgia, attributing ninety percent of datacenter activity in the state to the incentive. [55]
These are not trivial findings, and any honest treatment of retrenchment has to sit with them. If the but-for percentage is genuinely ninety percent, then withdrawal of the exemption is not a revenue gain — it is a revenue loss plus an investment loss. The counter-literature, which the paper takes up in Section 3.7, argues that the but-for estimates are inflated by self-reported industry data and by a methodology that treats incentives as causal simply because recipients say they were. The University of Texas political economist Nathan Jensen put the analytical distinction sharply:
“The key is the fiscal impact: does the state make its money back?”
— Nathan M. Jensen, Professor of Government, University of Texas at Austin [55]
The counterweight to the JLARC-style analysis is the per-job accounting. Good Jobs First’s canonical study of eleven major datacenter subsidy deals found an average cost per job of $1.95 million; the largest single per-job subsidy identified, $6.4 million, went to Apple in North Carolina. [58] [59] A New York project promising 125 jobs in exchange for $1.4 billion in incentives works out to roughly $11 million per job. [56] The organization also found that sixteen of thirty-six state datacenter subsidy programs impose no job-creation requirement whatsoever, meaning the recipient incurs no legal obligation to produce the employment the announcement advertised. [56] Their recommended cap — $50,000 per job — is roughly two orders of magnitude below observed practice, and their model reform legislation would additionally require unredacted applications disclosing beneficial ownership, standby generation and emissions, water use and discharge, transmission requirements, and noise levels, together with clawback provisions tied to continued operation. [58] [57] The practical counterweight is that in the largest hyperscale markets, tax breaks constitute a small share of total project cost, and power and land availability drive the actual siting decision — which is why several states with strong grid capacity have been able to scale back subsidies without visibly losing builds. [60]
1.2 When Capital Stopped Being the Scarce Commodity
The generative-AI investment cycle changed the bargaining equation by changing which side of the transaction faces a binding constraint.
The older economic-development model can be written as:
Few Projects → Many Competing States → Surplus Accrues to Capital
The AI infrastructure cycle increasingly resembles:
Many Gigawatt Projects → Few Immediately Viable Power-and-Land Locations → Surplus Accrues to Infrastructure
Four things changed at once.
Scale. The unit of account moved from tens of megawatts to hundreds of megawatts and then to gigawatts. The USC Marshall research group observed the historical anomaly precisely: over the past sixty years, only twenty-six facilities individually larger than 300 megawatts were added to American grids, and twenty of those were pumped-storage hydroelectric plants — which both consume and deliver power, a detail that is itself a portent for flexible datacenter load. [50] The industry is now proposing facilities of that scale routinely, and proposing many of them at once.
Simultaneity. In the cloud era, projects arrived sequentially and grids absorbed them incrementally. In the AI era, they arrive in overlapping waves driven by a competitive dynamic among a small number of extremely well-capitalized buyers, all of whom are attempting to secure capacity for the same model-training and inference roadmaps on the same three-year horizon.
Velocity mismatch. This is the structural heart of the problem. A hyperscale campus is planned and built in two to three years. In many regions, connecting a new facility of that size to the power grid takes four to ten years. The World Economic Forum’s 2026 assessment of the misalignment was categorical:
“grid connectivity has become the binding constraint”
— World Economic Forum, May 2026 [97]
Physical supply chains. Large-frame transformers, high-voltage breakers, gas turbines, and skilled electrical labor all have multi-year lead times that no amount of capital can compress below the manufacturing constraint. The International Energy Agency’s 2026 assessment captured the general condition:
“the AI surge is increasingly coming up against physical constraints”
— International Energy Agency, Key Questions on Energy and AI, 2026 [89]
Hyperscalers may possess extraordinary balance sheets. They cannot manufacture an available transmission corridor, a completed substation, a permitted gas lateral, a suitable river basin, an experienced electrical workforce, or an accepting local community instantaneously. Those things exist in the physical world on physical timelines, and most of them are controlled — directly or indirectly — by state and local governments.
That is the moment at which the state’s scarce resources begin acquiring negotiating value.
1.3 The Five-Layer AI Economy Meets State Public Finance
To see why this matters fiscally rather than merely operationally, it helps to place the transition inside the Five-Layer AI Economy.
| Layer | Content | Principal Economic Actors | State Policy Exposure |
| Layer 1 — Energy | Generation, transmission, substations, gas supply, nuclear assets, electricity rates, capacity markets | Utilities, IPPs, RTOs, state PUCs, public power districts | Very high: rates, siting, interconnection, environmental permits |
| Layer 2 — Chips | GPUs, accelerators, custom silicon, networking, memory, replacement cycles | NVIDIA, AMD, hyperscaler silicon programs, foundries | Moderate: equipment sales-and-use-tax exemptions |
| Layer 3 — Datacenters | Land, buildings, cooling, water, permitting, interconnection, construction labor | Hyperscalers, colocation developers, REITs, EPC firms | Very high: zoning, water, air permits, property tax |
| Layer 4 — Models | Frontier training and inference; increasingly compute-intensive | Frontier labs, cloud AI divisions, sovereign programs | Low: mobile, national, often taxed elsewhere |
| Layer 5 — Applications and Agents | Consumer and enterprise demand that ultimately drives utilization | Software firms, enterprises, consumers, agentic systems | Low to moderate: general corporate and sales tax |
The table makes visible an asymmetry that will increasingly preoccupy governors and legislatures. State subsidies are concentrated almost entirely in Layers 1 through 3. Most of the economic value is ultimately captured in Layers 4 and 5, by corporations operating nationally or globally and booking income in jurisdictions the host state does not control.
A state that exempts servers from sales tax (Layer 2), abates property tax on the shell (Layer 3), and socializes network upgrades across its ratepayers (Layer 1) is making three separate transfers into the cost structure of a value chain whose profits are realized in Layers 4 and 5. The host state bears the physical footprint, the ratepayer exposure, the water withdrawal, the land conversion, and the political friction. It receives construction employment, a modest permanent payroll, and whatever property tax survives abatement.
This is not an argument that the transfer is never worth making. It is an argument that the transfer is structurally asymmetric in a way that compounds as the layers above grow, and that the asymmetry is now large enough to be politically legible. Virginia is the clearest case: five companies received eighty-two percent of the benefit of the Commonwealth’s largest economic-development expenditure in 2023, and fewer than one hundred companies qualified for it at all. [37]
1.4 The Scarcity Ledger
If the old question was how much is the company investing, the new question is what scarce public and quasi-public resources must the state commit for every dollar of private investment. Answering that question requires an accounting device that does not currently exist in most state economic-development practice.
This paper proposes the Scarcity Ledger: a standardized, publicly filed accounting of the physical and fiscal resources a proposed hyperscale project consumes and produces, structured so that projects can be compared against one another and against alternative uses of the same resources.
| Ledger Line | Unit of Measure | Why It Belongs on the Ledger |
| Electricity capacity consumed | MW contracted; MW coincident peak | The binding constraint in most markets |
| New generation required | MW by resource type; additionality status | Distinguishes new supply from displacement of existing load |
| Transmission and distribution investment | $ of network upgrades; miles of line; substation count | Determines cost-causation exposure |
| Water withdrawal and consumption | Gallons/day withdrawn, consumed, returned; source basin | Often the true limiting factor in arid and aquifer-dependent regions |
| Wastewater and treatment load | Gallons/day; contaminant profile | Municipal capacity is finite and capital-intensive |
| Land utilization | Acres; prior use; conversion of farmland or habitat | Irreversible on policy-relevant timescales |
| Road and transportation burden | Peak construction trips/day; permanent heavy-vehicle traffic | Local capital cost borne by county |
| Emergency-service requirements | Fire suppression class; hazmat exposure; staffing | Small localities frequently cannot absorb this |
| Air permits and generator emissions | Backup generator MW; permitted run-hours; NOx/PM | Central to public-health objections |
| Noise footprint | dBA at property line and at nearest residence | Most common source of sustained local complaint |
| Permanent employment | FTE; wage distribution; local-hire share | The value most often overstated in announcements |
| Construction employment | Peak FTE; duration; apprenticeship participation | The value most often understated in criticism |
| Tax incentives received | $ NPV by instrument and by year | Must be NPV, not headline |
| Tax revenue generated | $ by level of government; net of abatement | The actual fiscal return |
| Community investments | $ committed; enforceable? clawback provisions? | Distinguishes obligation from public relations |
| Ratepayer exposure | $ at risk if project is cancelled post-construction | The stranded-asset question |
| Opportunity cost | Alternative industrial loads foreclosed per MW | The question almost never asked |
The final line is the one that most distinguishes the Scarcity Ledger from a conventional cost-benefit analysis. Conventional analysis asks whether a project’s benefits exceed its costs. The Scarcity Ledger asks whether a project’s benefits exceed those of the best alternative use of the same scarce resource — which, in a capacity-constrained grid, is the only question that actually determines whether the state has allocated well.
1.5 Defining the Subsidy-Retrenchment Threshold
A state reaches the Subsidy-Retrenchment Threshold when the marginal economic value of attracting one additional datacenter becomes smaller than — or politically harder to justify than — the marginal scarcity cost of accommodating it.
Conceptually:
Net State Value = (Investment + Tax Revenue + Employment + Strategic Value) − (Incentives + Infrastructure Cost + Resource Scarcity + Ratepayer Exposure + Community Externalities)
Three features of this expression deserve emphasis.
First, the left-hand terms are front-loaded and highly visible; the right-hand terms are back-loaded and diffuse. Investment and construction employment appear at announcement. Ratepayer exposure appears three years later in a rate case. This asymmetry in timing is the single most important reason the incentive-auction equilibrium persisted as long as it did: the political system observes the benefits before it observes the costs.
Second, several right-hand terms were, until recently, not measured at all. No state maintained a systematic accounting of ratepayer exposure to datacenter-driven network upgrades. Few measured cumulative water withdrawal against basin recharge. Almost none measured opportunity cost per megawatt. The Harvard Electricity Law Initiative’s central methodological finding is precisely that the measurement gap is not accidental — it is produced by confidentiality:
“The public faces significant risks that utilities will … profit from new data centers”
— Ari Peskoe and Eliza Martin, Harvard Law School Electricity Law Initiative [44]
Third, the threshold is reached at different points in different states, which is why the 2026 sequence looks heterogeneous even though it is driven by a common mechanism. Massachusetts crossed it with essentially no hyperscale datacenters built, because its electricity prices and constrained grid meant the ratepayer-exposure term was large from the first project. Nebraska crossed it with nine Meta-owned facilities on the ground and $317 million in direct credit revenue foregone between January 2021 and June 2026, plus $15 billion in personal property exempted from taxation. [29] Virginia has not crossed it — it preserved its exemption — but it has repriced around it. Texas has not crossed it either; it has instead built a gate.
Once the expression becomes uncertain or negative, governments acquire an economic incentive to renegotiate the traditional datacenter bargain. That is what 2026 was.
1.6 The Global Frame: What the International Institutions and the Research Community See
American state policy is not being made in isolation, and the analytical frameworks now emerging in Harrisburg, Austin, Albany, Boston, Lincoln, and Richmond track closely with assessments produced by institutions that have no stake in any particular state’s competitive position. It is worth establishing that frame, because it clarifies which features of the 2026 reversal are local politics and which are responses to a genuinely global physical transition.
The International Energy Agency’s measurement of the underlying trend is the reference point. Data-centre electricity consumption stood at roughly 415 terawatt-hours in 2024 — about 1.5 percent of global electricity use — having grown at approximately 12 percent per year over the preceding five years, more than four times the growth rate of total global electricity consumption. [92] Under the Agency’s central case, that figure roughly doubles by 2030 to approach 945 to 950 terawatt-hours, equivalent to the present electricity demand of Japan, with the AI-focused share tripling toward 465 terawatt-hours. [98] The concentration is what makes this a subnational governance problem rather than a global one: in Virginia, datacenters already consume roughly 26 percent of state electricity, and in Dublin the figure reaches 79 percent. [98] Roughly half of American electricity-demand growth over the next five years is expected to come from datacenters. [98]
The Agency’s five-year outlook confirms that the constraint is durable rather than cyclical. Global electricity demand is forecast to grow 3.6 percent in 2026, accelerating to 3.8 percent in 2027, and the next five years will add on average 50 percent more electricity demand per year than the past decade — with United States consumption rising close to 2 percent annually, more than twice the rate of the previous ten years, driven substantially by datacenter expansion. [90] [91] Electricity demand is now expected to outpace economic growth globally through 2030, reversing a relationship that held for most of the last century. [90]
The International Monetary Fund has framed the investment side of the same phenomenon in terms that are unusually vivid for the institution. Managing Director Kristalina Georgieva, speaking at the India AI Impact Summit in February 2026, described the scale of capital moving into datacenters, power plants, and neural networks with a historical analogy that captures both the ambition and the risk:
“It is like when the world was building the railways”
— Kristalina Georgieva, Managing Director, International Monetary Fund [93]
The comparison is apt in both directions. Railways transformed economies permanently; they also produced repeated capital-market crises when construction outran demonstrated demand. That is precisely the tension the Fund’s April 2026 Global Financial Stability Report examines, and it is the tension underlying every state’s decision about who should bear the cost of infrastructure built against a forecast.
The World Bank Group’s International Finance Corporation has extended the analysis to emerging markets, examining the energy risks and policy prerequisites for datacenter and AI investment outside advanced economies. [96] The Brookings Institution’s regulatory survey, prepared for the Forum for Cooperation on AI and updated in April 2026, reaches the same structural conclusion from the governance side: baseline estimates for datacenter energy consumption have already outrun overall electricity growth, and a consensus among leading analytical bodies points to a doubling or more of demand by 2030. [99]
The research community has converged on a similar diagnosis while disagreeing sharply about the remedy. The MIT Energy Initiative launched a dedicated Data Center Power Forum in response to the projection that global datacenter power demand will more than double by 2030, noting that American datacenters consumed 4 percent of national electricity in 2023 with demand expected to reach 9 percent by 2030. [52] Its director framed the moment in terms that apply as much to state fiscal policy as to engineering:
“We’re at a cusp of potentially gigantic change throughout the economy”
— William H. Green, Director, MIT Energy Initiative, and Hoyt C. Hottel Professor, MIT [53]
The efficiency counterargument — that improving hardware and software will absorb the demand — has been examined and largely set aside by the researchers closest to it. Andrew A. Chien, presenting at the Stanford Energy Seminar hosted by the Precourt Institute for Energy, put the point in the language of energy economics:
“Jevon’s paradox has overcome 50 years of energy efficiency improvements”
— Andrew A. Chien, University of Chicago and Argonne National Laboratory, Stanford Energy Seminar [54]
Chien’s constructive proposal is directly relevant to Pillar 8 of this paper: redesigning the service model so that computing flexibility — the ability of datacenter load to move in time and to respond to grid conditions — becomes a cooperative resource for grid stability and decarbonization rather than a pure stressor. [54]
Finally, the University of California, Berkeley perspective identifies the specific property of AI load that makes it a state and local problem rather than merely a national energy statistic. Datacenters do not constitute an enormous share of global energy use compared with air conditioning or heavy industry. What distinguishes them is spatial concentration:
“they are the most concentrated energy demands we’ve ever created”
— Carl Boettiger, Associate Professor, University of California, Berkeley [51]
Because grids are local, concentrated demand produces concentrated price and reliability effects in the communities that host it — and, as Boettiger observes, those communities are frequently the ones least able to absorb them. [51] That observation is the bridge between the global data and the state policy: a global aggregate of 950 terawatt-hours is an abstraction, but a one-gigawatt campus in a single county is a specific set of substations, wells, roads, and rate cases belonging to specific people who vote.
Section 2: Six States, Six Instruments of Retrenchment
The most useful way to understand the 2026 reversal is not as a single policy but as a set of six distinct legal instruments, each operating on a different margin of the bargain, deployed by governors with different constitutional powers, different grid structures, and different political coalitions. Read together, they constitute something close to a complete menu of the tools available to a subnational government that wants AI investment on better terms without forfeiting it entirely. Read individually, each reveals something the others cannot.
2.1 Pennsylvania: From Fast Track to Pay-Your-Own-Way
Pennsylvania is the paper’s primary case study because it is the state where the reversal is most complete, most legally structured, and most instructive — and because it is the state where the reversal is most obviously not hostility to AI.
The Commonwealth possesses an unusual combination of advantages: the Marcellus and Utica shale formations, a substantial existing nuclear fleet, membership in PJM Interconnection, large population centers, an industrial land bank of brownfield and post-anthracite sites, a unionized construction workforce, and — as of 2025 — an announced $20 billion Amazon investment that Governor Shapiro described as the largest private-sector investment in Pennsylvania history. Pennsylvania is not a state that needs to beg.
Executive Order 2026-05 does eight things. It removes all AI datacenter proposals from the Permit Fast Track Program and forecloses the program to datacenters prospectively. It directs the Department of Environmental Protection to review a datacenter permit application only where the developer has filed a legally binding notice of intent to comply with the GRID Requirements. It conditions DEP review on prior local land-use approval — meaning the township, borough, or county acts first and the state acts second. It prohibits Commonwealth agencies from executing nondisclosure agreements with datacenter developers. It requires developers to pay for the electricity their facilities consume and for the grid infrastructure upgrades their facilities cause, rather than socializing those costs across all ratepayers as had been the prevailing practice. It requires an increasing share of that power to come from clean sources — solar, advanced nuclear, or battery storage. It requires local hiring processes and a robust community benefits agreement. And it makes clear that noncompliant developers may lose access to Pennsylvania’s existing sales-tax exemption for datacenter equipment. [1] [2] [4]
Two features distinguish it from the other five states.
The first is sequencing. By requiring local approval before state environmental review, Pennsylvania inverted the ordinary order of operations in which a developer accumulates state permits and then presents the local government with a fait accompli. The Department of Environmental Protection’s Secretary, Jessica Shirley, confirmed the operational consequence: developers who decline the standards will not have their applications reviewed until local land-use approval is obtained. [4] This converts local consent from a downstream risk into an upstream gate — and, as Section 5.3 argues, into a priced input.
The second is provenance. Shapiro did not begin with an executive order. He began with legislation. The GRID standards were originally introduced as voluntary standards tied to tax benefits and expedited permitting. The Pennsylvania House advanced them on a bipartisan vote in June 2026. The Republican-led Senate declined to bring the measure up, and the majority leader stated he had no intention of taking legislative action to regulate datacenter development. [1] Shapiro’s response was to convert the voluntary framework into a binding one using the executive authority available to him. [4]
That provenance is analytically important for two reasons. It means the substantive content of the Pennsylvania framework has already cleared a bipartisan legislative chamber, which is evidence of political durability. And it means the framework’s legal foundation is an executive order rather than a statute, which is evidence of legal fragility — an executive order can be rescinded by a successor governor in a single afternoon. Advocates recognized this immediately; even supportive commentary described the order as a strong start with more to do. Climate Power’s senior advisor framed the underlying principle in the language that has become the movement’s slogan:
“make powerful corporations pay their own way”
— Jesse Lee, Senior Advisor, Climate Power [6]
The important development in Pennsylvania is therefore not that the Commonwealth rejected AI. It is that Pennsylvania began repricing access to its own advantages — and did so while explicitly retaining the ambition to be the place where the generation gets built.
2.2 Texas: Grid Access Becomes the Negotiating Table
Texas is the most revealing case precisely because it is the least ideologically predisposed to intervene.
Texas markets itself through inexpensive energy, rapid permitting, limited regulation, abundant land, and no corporate or personal income tax. It operates its own interconnection — ERCOT — largely outside Federal Energy Regulatory Commission jurisdiction, which historically has been a competitive advantage in speed. It has been the single largest beneficiary of the AI buildout by announced capacity. When a state with that profile places a gate in front of grid access, the gate is not ideological. It is arithmetic.
The arithmetic: approximately 474 gigawatts of interconnection requests against a record peak demand roughly one-fifth that size, with datacenters representing about ninety percent of the new requests. [8] No grid planner can distinguish real demand from speculative reservation at that ratio, and the distinction matters enormously, because transmission planning, generation procurement, and reserve-margin calculation all take the queue as an input.
The Texas response has three layers, and they were built sequentially rather than all at once.
The statutory layer is Senate Bill 6, signed in June 2025, which established disclosure and curtailment obligations for large loads of 75 megawatts or greater and formed the regulatory foundation for everything that followed. [10] The reliability layer arrived on July 9, 2026, when the PUCT adopted Nodal Operating Guide Revision Request 282 and Nodal Protocol Revision Request 1308, establishing “Large Electronic Load” reliability standards for facilities of 75 megawatts or greater where a substantial portion of load consists of power-electronic-based computational equipment. [14] The verification layer is Abbott’s August 3 directive and ERCOT’s Market Notice M-A080326-01 suspending Batch Zero.
To pass the audit, developers must produce detailed information on financial assistance received — including all state and local tax incentives, grants, and abatements — projected peak and annual power consumption, water consumption and cooling technology, ownership, project maturity, and community impact. [13] Notably, Abbott’s letter cited the failure of some datacenters to comply with the PUCT’s existing survey measuring water and power usage under the General Appropriations Act as a proximate justification. [8] In other words: the state asked, the industry did not fully answer, and the state escalated from request to condition.
The market’s reading of the trajectory was captured, unusually candidly, in an SEC filing. Fermi Inc., a datacenter developer that sits in the Southwest Power Pool rather than ERCOT and relies largely on behind-the-meter generation, told investors in its FY2026 Form 10-Q that Senate Bill 6, NOGRR 282, and NPRR 1308 were “evidence of a changing, and more restrictive, regulatory regime in Texas with respect to the data center industry.” [14] When a company that is not directly subject to a regime describes it as a risk factor, the regime is real.
Texas thereby establishes a principle with implications far beyond Texas:
Interconnection approval is itself an economic-development incentive — and can be allocated as one.
Instead of granting every proposed project equal standing in the queue, governments and grid authorities can distinguish among credible projects, speculative reservations, self-powered campuses, flexible loads, and projects willing to finance the infrastructure they require. That is not deregulation and it is not prohibition. It is rationing by verification, and it is the single most powerful instrument on the menu because it operates directly on the binding constraint.
2.3 New York: From Tax Incentive to Scarcity Premium
New York provides the clearest literal illustration of the paper’s subtitle, because New York is the state that has most explicitly proposed to charge a premium for access rather than merely to withdraw a benefit.
Executive Order 62, signed July 14, 2026, pauses discretionary state environmental permits for new hyperscale datacenters for up to one year while the Department of Environmental Conservation declines to issue any discretionary permit not already deemed complete, and while the Department of Public Service develops a Generic Environmental Impact Statement establishing consistent statewide standards covering energy demand, water use and quality, and air quality. [15] [17] The Governor’s framing at the signing was explicit about the mechanism:
“data centers to either produce their own energy or pay a premium”
— Governor Kathy Hochul, Executive Order 62 signing, July 14, 2026 [16]
She had made the same argument earlier in her State of the State address and repeated it through a statewide roundtable tour across Long Island, Western New York, the Finger Lakes, and the Rochester region:
“these huge consumers of power, these data centers need to bring their own power”
— Governor Kathy Hochul, Long Island roundtable, July 2026 [20]
Three components of the New York framework are structurally novel.
The Energize NY proceeding. Hochul directed the Department of Public Service to begin a proceeding designed to require datacenters either to pay a premium for grid power or to supply their own energy, with the Public Service Commission considering options including a separate rate class for datacenters. [19] This is scarcity pricing in its purest regulatory form: not a tax, not a fee, but a tariff differential reflecting the cost and scarcity of serving a particular class of load.
The New York Grid Acceleration Fund. The Governor directed DPS to consider creating a fund requiring datacenters to invest in the state’s aging grid infrastructure. The fund could support procurement of new clean energy supply and — this is the genuinely innovative element — the establishment of an insurance pool to which developers may need to contribute, protecting against speculative large loads that create uncertainty and increase costs. [17] An insurance pool against speculative load is a direct market-based answer to the phantom-load problem discussed in Section 3.4, and it is the first proposal of its kind at state level.
The community-investment blueprint. New York’s framework contemplates host localities receiving substantial funding for roads, water systems, schools, emergency services, and parks, and gives smaller municipalities the benefit of statewide standards rather than requiring each village to fund its own study. Local officials in the Nassau County Village Officials Association and elsewhere endorsed the moratorium substantially on this ground. [18]
The New York bargain therefore reads:
Access to New York → Pay for Incremental Power + Protect Ratepayers + Insure Against Speculation + Compensate Host Communities
rather than the prior bargain:
Access to New York → Receive Tax Advantage
Two caveats belong in an honest account. First, New York was never a major hyperscale destination to begin with; the state’s leverage is partly a function of having less to lose. Second, the state legislature had passed its own moratorium bill, which the Governor’s office described as needing additional work, choosing instead an executive order effective immediately. [21] A separate bill, S9144, introduced by Senator Liz Krueger in February 2026, would impose a statewide moratorium of at least three years and ninety days. [22] The executive order is thus, among other things, a shorter and more controlled alternative to a longer legislative pause.
2.4 Massachusetts: Freezing the Incentive Before It Entrenched
Massachusetts is the instructive case for states that have not yet experienced clustering, because it demonstrates that retrenchment does not require saturation.
The Qualified Data Center Sales and Use Tax Exemption was enacted as part of the 2024 economic development law and offers a twenty-year sales-and-use-tax exemption to certified facilities meeting eligibility requirements including the maintenance or creation of at least one hundred jobs within five years of certification. [26] Final regulations launched the program in May 2026. On June 25, 2026 — approximately one month later, and before the administration had received a single application — Governor Maura Healey halted acceptance of applications. [23] [25]
“I am halting any tax incentives for data centers until we have strong protections”
— Governor Maura Healey, Commonwealth of Massachusetts, June 25, 2026 [23]
Simultaneously, the administration released what it characterized as one of the most comprehensive state frameworks in the nation, covering cost, energy, water, air, noise, jobs, and community impacts, and setting out expectations for applicants seeking the exemption. [24] The framework’s central energy expectation is a “bring your own clean energy” standard under which operators are expected to meet the entirety of their demand from clean supply while minimizing or mitigating air and noise pollution. [26] No facility will be certified as a qualified datacenter until the guardrails are in place. [24]
The industry response was immediate and worth quoting because it articulates the strongest version of the opposing case:
“Massachusetts will be an uncertain and hostile market for development”
— Dan Diorio, Vice President of State Policy, Data Center Coalition [26]
That criticism has force. Massachusetts has forty-two datacenters, high energy costs, limited undeveloped land, and stringent environmental regulation, and has not been a significant participant in the national buildout. [26] Freezing a one-month-old incentive is, from a developer’s perspective, evidence that the state’s commitments have a short half-life — and predictability, as Section 6 argues, is itself a competitive asset.
But the Massachusetts case demonstrates something the mature-market cases cannot: governments can modify incentive structures before large-scale clustering becomes irreversible, and the political economy of doing so is far easier ex ante than ex post. Once a Northern Virginia exists, the incumbent industry has the resources, the employment base, and the legislative relationships to defend its treatment. Before it exists, the state’s options are open. Holyoke, meanwhile, became the first Massachusetts community to enact a citywide ban on AI datacenters after a proposed 20-megawatt facility along the Connecticut River drew organized opposition — evidence that the local layer moves whether or not the state does. [26] By August, a statewide ballot initiative had been proposed that would require local voter approval and other conditions for datacenter permits, the second such statewide measure in the country after Ohio. [27]
2.5 Nebraska: When Resource Accounting Overrides Recruitment
Nebraska matters because it decisively broadens the geography and the politics of the phenomenon. Subsidy Retrenchment is not a coastal policy, not a Democratic policy, and not a response to Northern-Virginia-style congestion. A Republican governor in an agriculture- and resource-intensive Great Plains state with public power districts rather than investor-owned utilities arrived at the same destination by a different road.
Governor Jim Pillen’s July 20, 2026 executive order prohibits new datacenter applications from being approved for tax credits under the ImagiNE Nebraska Act; requires the Departments of Economic Development, Revenue, and Water, Energy and Environment to review project proposals collaboratively to determine whether they serve the state’s interest; and establishes a datacenter task force through the Department of Water, Energy and Environment to protect the state’s natural resources and to equip county officials with zoning and permitting tools. [28] [30]
Pillen’s own account of the reasoning is notable for its consistency and for the fact that it predates the national backlash:
“Early in my administration, I called for a pause in building new data centers”
— Governor Jim Pillen, Nebraska, July 2026 [28]
His argument against incentives is a straightforward market-power argument, made in the vocabulary of a Republican governor rather than that of a consumer advocate:
“we do not need to bring people together by incredible tax incentives”
— Governor Jim Pillen, Nebraska, July 20, 2026 [29]
The fiscal accounting behind it: the Governor’s office reported that Nebraska forwent $317 million in direct credit revenue to datacenters between January 1, 2021 and June 30, 2026; that the value of personal property exempted from taxation reached $15 billion; and that a two percent property-tax rate applied to that value would have been worth $519 million. [29] Good Jobs First’s Subsidy Tracker places Nebraska among the top five state subsidizers of Meta Platforms, which has received close to $311 million from the state and operates nine facilities there. [29] Pillen estimated the order would protect electric rates for two million Nebraskans by $1 billion to $2 billion. [33]
State Senator Mike Jacobson, who joined the signing and has indicated he will introduce codifying legislation, articulated the leverage argument in its simplest form:
“They have plenty of money to come and locate here”
— State Senator Mike Jacobson, Nebraska Legislature [32]
That sentence is the entire thesis of this paper compressed into ten words, delivered by a Republican state senator from North Platte. The proposition is not that datacenters are bad. The proposition is that the competitive pressure justifying the subsidy no longer exists, and that when the competitive pressure disappears, the subsidy becomes a pure transfer.
2.6 Virginia: The Sixth Instrument — Preserve the Exemption, Price the Consumption
Virginia belongs in this analysis even though it did not withdraw an incentive, because Virginia deployed the instrument that most literally embodies the paper’s title, and because Virginia is the state with the most at stake.
During the 2026 session, the Virginia General Assembly considered 61 datacenter-related bills; fifteen went to the Governor and forty-six carried over to 2027. [39] The Senate proposed eliminating the datacenter sales-and-use-tax exemption — worth roughly $1.6 billion annually — beginning in 2027; the House preferred conditioning it on environmental compliance. [37] [41] The disagreement was severe enough to delay the biennial budget and to collapse an April 23 special session in a matter of hours. [39]
The resolution, signed by Governor Abigail Spanberger on June 30, 2026, is the most economically interesting compromise of the year. Virginia preserved the sales-and-use-tax exemption for qualifying datacenter equipment and simultaneously created a new Data Center Electricity Consumption Tax of $0.011 per kilowatt-hour, effective July 1, 2026, applying to both utility-supplied and qualifying self-generated electricity, with annual collections above $600 million refunded pro rata and a sunset on June 30, 2028. [35] [36] The measure may generate up to $1.2 billion over the biennium. The budget also created a joint legislative subcommittee, required to report by December 15, 2026, to study the exemption, the approaches taken by other states, the 2024 JLARC findings, and mechanisms to provide direct revenue to the Commonwealth from the datacenter industry. [35]
Read structurally, Virginia did something no other state did: it decoupled the capital subsidy from the consumption charge. It kept the instrument that attracts investment — the equipment exemption, which lowers the cost of building — and added an instrument that prices the scarce resource — an excise on the electricity that operating consumes. In economic terms, Virginia lowered the marginal cost of capital and raised the marginal cost of the scarce input, which is very close to a textbook prescription for a jurisdiction that wants investment but wants it to be efficient with the constrained factor.
Whether the design survives is a separate question. It expires in 2028. The industry reacted negatively. And the applying of the tax to self-generated electricity partially blunts the incentive to bring one’s own power — arguably working against the direction New York and Massachusetts are pushing. But as a policy artifact, Virginia’s consumption tax is the clearest existing example of a state charging for scarcity rather than merely declining to subsidize it.
2.7 Comparative Summary
| State | Governor | Party | Date | Primary Instrument | Legal Vehicle | Core Demand |
| Massachusetts | Maura Healey | D | June 25, 2026 | Incentive freeze | Administrative pause + framework | Bring your own clean energy; protect ratepayers |
| Virginia | Abigail Spanberger | D | June 30, 2026 | Scarcity excise | Biennial budget (HB 30) | $0.011/kWh on datacenter electricity |
| New York | Kathy Hochul | D | July 14, 2026 | Permit moratorium + rate premium | Executive Order 62 + DPS proceeding | Self-generate or pay a premium; fund the grid |
| Nebraska | Jim Pillen | R | July 20, 2026 | Incentive termination | Executive order | No ImagiNE credits; resource review; task force |
| Texas | Greg Abbott | R | August 3, 2026 | Conditional queue access | Gubernatorial directive to PUCT/ERCOT | Verify demand, water, ownership, financing |
| Pennsylvania | Josh Shapiro | D | August 18, 2026 | Conditional permitting + cost causation | Executive Order 2026-05 (GRID) | Local approval first; pay your own way; no NDAs |
The instruments differ. The demand does not.
Section 3: The Economics of Charging Artificial Intelligence for Scarcity
Having established what six states did, this section asks whether the economics justify it. The answer is more textured than either the advocacy literature or the industry response tends to allow. Cost causation is a well-founded principle with a century of regulatory pedigree, and its application to gigawatt-scale load is straightforwardly correct. Scarcity pricing on a constrained factor is defensible microeconomics. The opportunity-cost framing is analytically sound and almost entirely absent from current practice. But the empirical premise on which much of the political argument rests — that datacenters are the principal cause of rising residential electricity bills — is contested by serious research, and a paper that ignored that contestation would be advocacy rather than analysis. This section presents the case for retrenchment in its strongest form and then presents the case against it in its strongest form.
3.1 Cost Causation Replaces Cost Socialization
The foundational principle of public-utility ratemaking is that rates should reflect the cost of service, and that customers who cause costs should bear them. In practice, that principle has always been implemented with substantial averaging, because averaging is administratively simpler and because most incremental load has historically been small relative to the system. A new subdivision, a new grocery store, a new light-manufacturing facility — none of these individually triggers a new substation, a transmission reinforcement, or a new generating unit. Spreading the cost of general system growth across the customer base is reasonable when growth is diffuse.
A 500-megawatt, one-gigawatt, or larger AI campus is not diffuse. It is the single most concentrated point load most utility systems have ever been asked to serve. When a project triggers a new substation, a transmission reinforcement, incremental generation, gas infrastructure, water-treatment expansion, road improvements, fire-protection capacity, and backup-generation permitting, the central question is no longer administrative convenience. It is:
Who caused the cost, and therefore who should pay it?
The Harvard Electricity Law Initiative’s 2025 paper Extracting Profits from the Public: How Utility Ratepayers Are Paying for Big Tech’s Power is the most rigorous treatment of how, in practice, the answer has often been “everyone but the causer.” Ari Peskoe and Eliza Martin identify three distinct mechanisms.
The first and most consequential is the confidential special contract. Their research uncovered dozens of contracts between datacenters and utilities providing bespoke rates negotiated outside the ordinary rate case. Where the special-contract rate is lower than the utility’s cost to serve that customer, the utility recovers the shortfall from other ratepayers in subsequent proceedings. Because utilities and datacenters routinely shield contract terms with claims of proprietary information, the extent of the cost shift is not observable to the public and often not meaningfully tested by regulators. Peskoe and Martin identified forty publicly available state commission proceedings in which regulators approved special contracts in short and conclusory orders, frequently without engaging the utility’s own analysis, because challenging that analysis is costly and time-intensive and commission staff may lack the resources. [47]
The second is transmission cost allocation across the federal-state seam. Interstate transmission development and wholesale power sales are regulated by the Federal Energy Regulatory Commission, which determines how costs are shared among utilities. State commissions then apply their own formulas for dividing FERC-allocated costs among retail rate classes. Datacenter-driven costs can enter general rates at that hand-off. And if the anticipated demand never materializes, ratepayers are left paying for transmission that serves no one. [43]
The third is colocation. Where a large load is sited directly at an existing generator — particularly a nuclear plant — and that generator’s output is effectively removed from the wholesale market for billing purposes, the consequences propagate through capacity markets and therefore to all ratepayers. The definitional boundaries remain unsettled, which is precisely why FERC issued an order in December 2025 requiring PJM to implement transparent rules for substantial loads co-located with generation. [67]
Peskoe’s characterization of the competitive dynamic driving all three is worth stating plainly:
“Attracting datacenters is now this big competition among utilities”
— Ari Peskoe, Director, Electricity Law Initiative, Harvard Law School [47]
His co-author framed the physical scale of what is being requested with an image that has been widely repeated since:
“energy demand for entire cities regularly materializing out of thin air”
— Eliza Martin, Legal Fellow, Harvard Law School Environmental and Energy Law Program [45]
Subsidy Retrenchment, at its analytical core, moves policy from socialization toward causation. Pennsylvania’s requirement that developers pay for infrastructure upgrades rather than socializing them among all ratepayers is a direct statement of the principle. [2] So is the Oregon Public Utility Commission’s guideline directing utilities to charge large-load customers for upgrades that would not have been needed but for that customer’s interconnection — explicitly irrespective of whether other customers happen to benefit from the resulting infrastructure. [109]
3.2 The Megawatt Becomes a Scarce Economic-Development Asset
The second economic argument is about allocation rather than cost recovery, and it is the argument least represented in current state practice.
In a capacity-constrained system, a megawatt allocated to one project cannot simultaneously serve another. This is elementary, but its implications for economic-development policy are profound and almost universally ignored. Governors and legislatures routinely evaluate a proposed datacenter against the counterfactual of no project. In a constrained grid, the correct counterfactual is the next-best project that will not now be served.
Consider one gigawatt of firm capacity in a constrained region. It could support:
- one hyperscale AI campus;
- several semiconductor fabrication or advanced-manufacturing facilities;
- a large electrified industrial cluster — steel, chemicals, cement;
- housing expansion and the associated residential and commercial load;
- existing households and businesses, at lower rates;
- electrolytic hydrogen production;
- electric-vehicle and battery manufacturing;
- or reserve headroom for future development not yet identified.
Each of these has a different profile of permanent employment per megawatt, tax revenue per megawatt, supply-chain linkage per megawatt, and strategic value per megawatt. A semiconductor fab employs thousands. A hyperscale datacenter of comparable load typically employs between thirty and fifty people, or perhaps twice that at a very large facility. [59]
The question is therefore no longer whether datacenters create economic value. They plainly do. The better question — the one the Scarcity Ledger’s final line is designed to force — is:
Do datacenters generate enough value per scarce megawatt, relative to the alternatives that megawatt forecloses?
This is not an argument that the answer is no. In many places, particularly where surplus generation exists and no competing industrial demand is queued, the answer is clearly yes. It is an argument that the question has almost never been asked, and that a state which does not ask it is not conducting economic development. It is conducting first-come, first-served.
3.3 Tax Exemptions Become Scarcity Discounts
The third argument reframes what a datacenter tax exemption actually does as market conditions change, and it produces a genuine paradox.
During an era of excess capacity — surplus generation, uncongested transmission, abundant developable land, low political salience — a tax exemption performs its intended function. It attracts incremental investment that would otherwise locate elsewhere. The state forgoes revenue it would never have collected and gains activity it would never have had. Under those conditions, the JLARC-style but-for analysis is not merely plausible; it is likely correct.
During an era of severe grid congestion, the same instrument does something different. If the binding constraint is interconnection capacity rather than capital, and if more projects are seeking that capacity than can be served, then the marginal effect of the exemption on whether investment occurs approaches zero. What the exemption now does is discount access to an already-scarce asset — it lowers the price of something that is being rationed. Under those conditions, the exemption is not attracting investment. It is transferring surplus from the state to the entity that would have won the rationing contest anyway.
This produces the central paradox of the current moment:
The more desperately corporations want a state’s electricity, land, and interconnection capacity, the weaker the economic case becomes for paying them to consume it.
Nebraska’s Senator Jacobson stated exactly this proposition when he observed that competition to entice datacenters is not currently an issue given the resources these companies command. [31] Nebraska’s governor said the same thing in different words. It is, notably, an argument that lands equally well on the political right — where it reads as opposition to corporate welfare — and on the political left — where it reads as opposition to regressive transfer.
The honest qualification is that “excess capacity” and “severe congestion” are regional and temporal conditions, not national ones. Wyoming, the Dakotas, and parts of the Southeast retain genuine surplus. Northern Virginia, the ERCOT Permian and Dallas load zones, and the PJM Mid-Atlantic do not. A single national verdict on datacenter tax exemptions is therefore analytically unsound. The correct verdict is jurisdiction-specific and time-varying, which is an argument for sunset provisions and periodic reauthorization — precisely the design Virginia adopted when it set its consumption tax to expire June 30, 2028 and commissioned a joint legislative subcommittee to report by December 15, 2026. [35] [42]
3.4 The Phantom-Load Problem
The fourth argument is the one with the most direct empirical support, and it concerns the integrity of the information on which every other decision depends.
Developers seek grid positions years before projects are fully financed, before end users are secured, and before construction is certain. Some do so because that is how responsible long-lead-time development works. Others do so because a queue position is a free option: it costs little to hold, it can be resold, and it preserves optionality against a demand forecast nobody can verify. When hundreds of enormous requests accumulate, grid planners cannot easily distinguish real demand from optionality, and the queue itself becomes a source of cost — distorting transmission planning, generation procurement, reserve-margin calculation, and capacity-market clearing.
The evidence that a large share of requested load is speculative is now substantial and, importantly, comes from the utilities themselves once financial commitments are attached.
The clearest natural experiment is Ohio. AEP Ohio initially projected approximately 30,000 megawatts of datacenter demand. After Ohio regulators approved a new large-load tariff requiring developers to make financial commitments, that figure fell to 5,700 megawatts — and the utility’s public forecast was revised to roughly 13,000 megawatts, a reduction of more than half. [104] [108] Ohio manufacturers subsequently argued that even the revised figure remained inflated. [108] Whatever the true number, the direction is unambiguous: requiring money to accompany a request removed between fifty-seven and eighty-one percent of the request.
Texas presents the same phenomenon at national scale: 474 gigawatts of interconnection requests against a peak demand roughly one-fifth as large, with datacenters comprising ninety percent of the new requests. [8] Shapiro’s characterization of Pennsylvania — more than one hundred proposals, fifteen with any DEP permit application, five with a complete first-phase permit set — describes an attrition rate of roughly ninety-five percent between announcement and permitting readiness. [1] [7]
The policy responses now emerging are correspondingly commercial:
- larger and non-refundable study deposits, such as AEP Ohio’s tiered structure of $10,000 / $50,000 / $100,000 based on the size of the capacity request; [107]
- proof of financing and executed customer agreements;
- construction milestones with forfeiture provisions;
- demonstrated land control;
- credible energization schedules;
- minimum-billing or take-or-pay obligations, which as of the DELTa database’s March 31, 2026 public update appeared in 33 of 77 tracked large-load filings at an average of approximately eighty percent of contracted capacity, with AEP Ohio at eighty-five percent, Consumers Energy at eighty percent, and Kentucky Power at ninety percent; [105] [107]
- long minimum contract terms, with the emerging archetype at ten years or more — PPL’s Schedule LP-6 at ten years, Xcel at fifteen, and Florida Power & Light and Kentucky Power at twenty; [107]
- collateral requirements including letters of credit and cash deposits to reduce default and stranded-asset risk; [105]
- and exit-refund caps, such as the twenty percent cap under Texas rules, which has drawn legislative pushback precisely because a low refund cap discourages speculative entrants from leaving the queue. [13]
New York’s proposed insurance pool for speculative large loads belongs on this list as the most conceptually elegant entry: rather than attempting to identify which projects are speculative, it prices the aggregate risk that some are and requires the class to fund it. [17]
The general principle that emerges is significant:
Scarcity pricing operates not only on electricity consumed, but on electricity capacity reserved.
3.5 Water, Land, and Infrastructure Join the Price
Electricity receives the most attention because AI demand is enormous and because electricity bills arrive monthly. But the same analytical structure applies to every other scarce input, and in several regions water is the true binding constraint.
Water is analytically harder than electricity for three reasons. It is basin-specific rather than grid-wide, so scarcity cannot be assessed at state level. Its scarcity is stochastic and seasonal rather than continuous. And withdrawal, consumption, and return are three different quantities that public discourse routinely conflates — a facility that withdraws large volumes and returns most of them has a very different impact from one that evaporates them. Nebraska’s executive order specifically directed its new task force to work with the state’s Natural Resources Districts, reflecting the fact that Nebraska’s water governance is fundamentally local. [28] Virginia’s SB 553 now requires water-consumption estimates as part of rezoning and special-use-permit applications. [39] Long Island officials cited their sole-source aquifer as a principal reason for supporting the New York moratorium. [18]
Land is analytically distinct again, because land conversion is effectively irreversible on policy-relevant timescales. Farmland converted to a datacenter campus does not revert. This is why the strongest local opposition frequently arises in agricultural and rural-historic areas rather than in industrial zones, and why land-use authority — which localities possess even where they lack rate-setting authority — has become the most immediately available tool for communities.
A comprehensive scarcity regime therefore prices more than power:
Total Scarcity Charge = Power Charge + Grid Upgrade Charge + Water Charge + Infrastructure Charge + Community Contribution + Environmental Mitigation
The sum of these charges is, in effect, an attempt to construct an accurate price for the physical footprint of compute — a price that the accounting conventions of the cloud era never had to produce because the footprint was small enough to ignore.
3.6 From Incentive Package to Reciprocity Package
The cumulative effect of the preceding arguments is a change in the fundamental question that a state asks at the negotiating table.
The old economic-development package asked: What can the state give the corporation to secure the investment?
The emerging model asks: What will the corporation build, finance, or guarantee in return for access to what the state controls?
The reciprocity provisions now appearing in state frameworks, tariffs, and community-benefit agreements include:
- new generation, whether developer-built, contracted, or financed;
- transmission upgrades and dedicated substations;
- battery storage and other flexibility resources;
- advanced nuclear and small modular reactor offtake commitments;
- water recycling, closed-loop cooling, and wastewater infrastructure;
- local road, bridge, and intersection improvements;
- workforce training programs and registered apprenticeship participation;
- fire, hazmat, and emergency-service capacity funding;
- host-community funds with enforceable disbursement schedules;
- local tax guarantees and payment floors independent of assessment appeals;
- and curtailment or interruptibility commitments during system emergencies.
That last item deserves particular emphasis because it converts a liability into an asset. PJM has a proposal before federal regulators under which datacenters that do not bring their own power would be subject to curtailment on days when electricity is short — and Pennsylvania’s governor stated directly that grid-connected datacenters would have to go offline in a shortage. [2] A facility that can shed several hundred megawatts within seconds is, from a system-operator’s perspective, functionally equivalent to a peaking resource. Flexibility is the one thing a datacenter can offer the grid that almost nothing else of comparable size can, and pricing it properly is the most underexploited opportunity in the entire policy space.
This reciprocity architecture is the fiscal heart of Subsidy Retrenchment. It is not the absence of a deal. It is a different deal.
3.7 The Counter-Case: Does the Empirical Record Support the Premise?
Intellectual honesty requires confronting the strongest argument against everything above, and that argument is not made by industry lobbyists. It is made by economists working with market data, and it deserves to be stated at full strength.
The political case for retrenchment rests substantially on a causal claim: that datacenters are driving up residential electricity bills. Research from the University of Southern California Marshall School of Business directly contests that claim. Professor Shon R. Hiatt, director of the Zage Business of Energy Initiative, summarized a study covering wholesale and retail electricity markets from 2015 through the end of 2025 with findings that are difficult to reconcile with the prevailing narrative.
At the local grid level, datacenter entry raises nearby congestion costs by approximately $1.45 to $2.30 per megawatt-hour. At the regional wholesale level, prices in an affected zone rise by roughly $3.44 per megawatt-hour in the year following entry — a predictable consequence of dispatching higher-cost generation to meet new demand. But the retail impact is smaller still: doubling a utility’s operating datacenter capacity is associated with only a 0.65 percent increase in residential electricity prices, roughly $0.001 per kilowatt-hour, or less than one dollar per month for a typical household. Most consequentially, that small effect is concentrated among electric cooperatives and municipal utilities. For investor-owned utilities, which serve sixty-eight percent of the American population, the study finds no statistically significant relationship between datacenter growth and retail rates. [49]
Hiatt attributes the broader price surge — more than thirty-three percent nationally since 2019, and exceeding sixty-six percent in California — to a different set of causes documented in Lawrence Berkeley National Laboratory’s 2026 retail price trends analysis and the 2026 NERC Long-Term Reliability Assessment: premature retirement of baseload thermal generation, state renewable-portfolio and carbon policies, and the capital costs of grid modernization driven by manufacturing reshoring and general electrification. His policy conclusion follows directly:
“The solution to rising energy prices is not to curtail demand”
— Shon R. Hiatt, Professor and Director, Zage Business of Energy Initiative, USC Marshall School of Business [49]
This finding has three serious implications for the retrenchment thesis, and they should not be minimized.
First, it suggests the political mechanism and the economic mechanism may be misaligned. If datacenters are not the principal driver of residential bill increases, then policies justified on that basis may deliver little relief while imposing real costs on investment. Voters may punish incumbents for electricity prices, incumbents may respond by restricting datacenters, and bills may continue rising for entirely unrelated reasons — a politically stable but economically unproductive equilibrium.
Second, it identifies where the harm actually concentrates. The cooperative-and-municipal finding is not a refutation of the cost-shifting concern; it is a localization of it. Small systems lack the customer base across which to absorb wholesale volatility. That maps precisely onto the Oregon situation, where legislators observed datacenters targeting the roughly twenty-five percent of the state served by cooperatives and municipal utilities in order to avoid the clean-energy mandate applicable to investor-owned utilities. [110] It also maps onto Nebraska, a public-power state. If the harm is concentrated in small systems, then blanket statewide instruments are poorly targeted and system-specific tariffs are better.
Third, it does not dispose of the forward-looking case. Hiatt is explicit that with reserve margins declining, rising demand from reshored manufacturing and datacenter growth will exert greater pressure on wholesale prices, and that those effects could flow through to retail prices within a few years. [49] A retrospective study covering 2015 through 2025 measures a period in which most AI-scale campuses were not yet operating. The retrenchment argument is fundamentally prospective: it concerns 474 gigawatts of requests, not the load already energized.
Two further considerations complicate the picture in the other direction.
The PJM capacity market provides evidence that is difficult to attribute to anything other than anticipated datacenter demand. Capacity prices rose from $28.92 per megawatt-day for the 2024/2025 delivery year to $269.92 for 2025/2026 — an 833 percent increase, the sharpest single-year move in the market’s twenty-seven-year history — and then to the FERC-approved cap of $329.17 for 2026/2027. [62] [61] The December 2025 auction for 2027/2028 hit the updated cap of $333.44 while falling 6,625 megawatts short of the reliability requirement for the first time ever. [66] The 2028/2029 auction in July 2026 cleared at the cap for the third consecutive time. [63] PJM’s Independent Market Monitor identified demand from new and proposed datacenters as the primary reason for the spike. [65] An independent analysis cited by advocacy groups estimated that annual customer capacity costs in PJM rose from $2.2 billion in 2024 to $14.7 billion the following year. [63] PJM itself, in coordination with the governors of all thirteen member states, imposed a price cap and floor and extended those parameters through the December 2026 auction to limit volatility. [64]
Notably, PJM’s own estimate of the retail effect is modest — the cap price was expected to translate into a year-over-year increase of 1.5 to 5 percent in some customers’ bills — which is directionally consistent with Hiatt’s finding that the retail pass-through of datacenter-driven wholesale effects is small. [61] Both things can be true: capacity markets can be dominated by anticipated datacenter load and the resulting retail effect can be a minority of the observed bill increase.
The second consideration is that public perception has already resolved the question in a way that policy must contend with regardless of the underlying econometrics. Pew Research found that forty-three percent of Americans consider datacenter energy use a major reason their home energy bills are rising, with another twenty-three percent calling it a minor reason. [82] Nationally, electricity rates measured by average retail revenue per kilowatt-hour rose 7.1 percent in 2025, with sharply higher increases in specific jurisdictions — 26.3 percent in the District of Columbia, 18.9 percent in Pennsylvania, and 16.3 percent in Rhode Island. [79] Seven Northeastern states ranked among the fifteen with the largest residential electricity cost increases from 2021 to 2025. [82]
Three in four Americans expect their utility bills to rise this year, and only twenty-nine percent believe their state government protects their interests in regulating utilities — a nearly ten-point decline from the prior year. [86] With roughly eighty million Americans struggling to pay utility bills, the founder of the consumer advocacy organization PowerLines described the stakes in terms no econometric coefficient can dislodge:
“it’s a life or death and ‘eat or heat’ type decision”
— Charles Hua, Founder, PowerLines [85]
The correct synthesis, in this paper’s assessment, is as follows. The cost-causation argument for retrenchment does not depend on datacenters being the principal cause of past bill increases. It depends only on the proposition that a customer who triggers a specific, identifiable, and large infrastructure cost should pay it. That proposition is sound whether or not the aggregate historical effect has been large, and it becomes more important as the aggregate effect grows. Where the empirical critique bites hardest is against the rhetoric of retrenchment — against governors who promise that restricting datacenters will lower bills. It bites much less against the architecture of retrenchment, which is chiefly about assigning costs correctly, verifying demand honestly, and pricing scarcity explicitly. A policy that survives only if the causal story is true is fragile. A policy grounded in cost causation survives either way.
Section 4: Consequences for Hyperscalers, Investors, and the Five-Layer AI Economy
If Section 3 established the logic of the transition, this section examines what it does to the companies on the other side of the table. The short answer is that Subsidy Retrenchment does not primarily reduce AI capital expenditure. It relocates cost from the public balance sheet to the private one, relocates activity from Layer 3 toward Layer 1, and relocates investment across the map. Each of those three relocations has strategic consequences that are only beginning to be priced.
4.1 The Datacenter Capital Stack Gets Larger
A hyperscaler’s project budget could once stop at a relatively short list:
Land + Building + Servers + Networking + Cooling
Increasingly it cannot. The budget that a state operating under a retrenchment framework will actually require looks like this:
Land + Building + Servers + Networking + Cooling + Generation + Transmission + Substation + Storage + Water Systems + Community Infrastructure + Environmental Mitigation + Long-Term Take-or-Pay Obligations + Collateral
Every item added to the right of “Cooling” represents a cost that governments previously absorbed, socialized, or waived, migrating onto the corporate balance sheet. This is the mechanical content of “pay your own way.” The White House’s Ratepayer Protection Pledge states the same principle at federal level: signatory companies commit to build, bring, or buy all of the energy needed to power their facilities and to pay the full cost of that energy and its supporting infrastructure — the pledge’s operative phrase being that they will pay no matter what, which is a take-or-pay commitment in plain language. [73] [76]
The financial consequence is not a smaller capital program. It is a capital program with a longer duration, a higher share of illiquid and non-redeployable assets, and materially different risk characteristics. A GPU cluster can be moved, resold, or repurposed. A 500-megawatt combined-cycle plant built to serve one campus cannot.
4.2 The Hyperscaler Becomes an Infrastructure Developer
The most consequential structural effect of retrenchment is that it pushes Layer 3 companies down into Layer 1.
Amazon, Microsoft, Google, Meta, Oracle, xAI, and the infrastructure partners of frontier labs are increasingly behaving less like large electricity customers and more like independent power producers and transmission developers who happen to own compute. They are financing or contracting for new nuclear capacity and small modular reactors, gas generation, solar and wind, battery storage, dedicated substations, microgrids, private transmission, and water-recycling systems.
Two developments in 2026 accelerated this. First, PJM began in June facilitating new bilateral long-term agreements between large load customers and generation providers — contracts often spanning ten years or more that allow buyers to secure supply from new generation, storage, or demand-side resources — and issued a request for proposals to support the matching process. [64] Second, the requirement in multiple state frameworks that a rising share of supply be clean and additional converts what might have been a procurement decision into a development decision, because the clean resources in question frequently do not yet exist.
The strategic implication is significant and under-discussed: Subsidy Retrenchment converts electricity from an operating expense into a strategic capability. Firms that can develop, finance, permit, and operate generation acquire a durable advantage over firms that can only buy power. That is a very different competitive landscape from the one that prevailed when the differentiating capability was chip procurement.
4.3 Site Selection Becomes a Total-Scarcity Calculation
The traditional site-selection model weighted electricity price, tax treatment, fiber connectivity, land cost, and latency. That model is now obsolete, because it optimizes on price while the binding constraint is availability.
The replacement calculation must be:
Effective AI Infrastructure Cost = Energy Price + Tax Burden + Grid Access Cost + Interconnection and Network Upgrades + Water + Permitting Time-Value + Community Obligations + Reliability + Curtailment Risk + Regulatory Volatility Premium
Two counterintuitive consequences follow.
A state with higher nominal electricity prices may be cheaper in effective terms if it offers faster interconnection, predictable rules, and a settled community-consent process. Time-to-energization is worth more than a rate differential when a delayed campus means a delayed model, and when the alternative use of the capital is earning nothing.
A state with low nominal prices may be expensive if its transmission system is saturated, its queue is subject to audit, its communities are organized in opposition, or its rules are subject to reversal at the next election. The last item — regulatory volatility — is the newest term in the function and the least well priced. Massachusetts froze a one-month-old incentive. Pennsylvania’s framework rests on an executive order that a successor can rescind. Virginia’s consumption tax expires in 2028 with no assurance about what replaces it. A rational developer must now carry a regulatory volatility premium in the underwriting, and the states that can credibly minimize that premium will win business from states that cannot, independent of headline rates.
4.4 Retrenchment Will Not Necessarily Reduce Aggregate AI Capital Expenditure
This distinction is essential and frequently missed by both advocates and critics.
The evidence from the second quarter of 2026 is that policy friction has not slowed the capital cycle at all — though it has begun to change how investors price it. When Alphabet raised its guidance alongside its second-quarter results, its shares fell roughly seven percent the following day and dragged the other hyperscalers with them, a reaction that reflected scrutiny of returns rather than scrutiny of regulation. [100] Alphabet raised its 2026 capital-expenditure guidance to a range topping out near $205 billion; Amazon raised its cash capital-expenditure guidance to approximately $220 billion, up from roughly $200 billion, citing higher memory costs; Meta lifted the floor of its guidance to a $130–145 billion range; and Microsoft tracked toward roughly $190 billion. [101] [102] The four companies’ combined 2026 guidance of roughly $725 billion represents an increase of about seventy-seven percent over 2025’s approximately $410 billion, with analysts projecting the combined figure to exceed $1 trillion in 2027. [102] The International Energy Agency observed that the capital expenditure of just five technology companies now exceeds global investment in oil and natural gas production. [89]
What the second-quarter earnings season revealed was not retrenchment-induced caution but a common risk-management playbook: commit early to long-lived assets — land, datacenter shells, power — and defer commitments on short-lived assets, principally chips, until demand is visible. [101] That playbook is, if anything, an argument for acquiring power and land positions faster in a tightening regulatory environment, not slower.
The policy shift therefore redistributes rather than eliminates investment. Capital should be expected to move toward:
- states and regions with genuine surplus generation;
- brownfield industrial sites with existing interconnection rights;
- retired or retiring thermal power station sites with intact switchyards;
- existing nuclear campuses and their adjacent land;
- natural-gas producing basins where behind-the-meter generation is straightforward;
- regions with expandable rather than saturated transmission;
- privately powered, off-grid, or islanded campuses;
- and smaller, more distributed datacenter architectures that fall below regulatory thresholds — a point worth watching, since FERC has proposed defining large loads at fifty megawatts or greater in some contexts and the Department of Energy’s initial framework used twenty megawatts, meaning threshold design will itself shape facility design. [70] [71]
The plausible result is a new American compute geography — one organized around the availability of power and consent rather than around the availability of tax abatement.
4.5 Corporate Bargaining Power Versus State Bargaining Power
None of this means states hold all the cards. Hyperscalers retain formidable leverage, and any framework that ignores it will fail.
A multibillion-dollar project brings construction employment in trades that are politically organized, property-tax revenue to counties that often desperately need it, supplier activity, and — increasingly — generation and transmission investment that the state wants anyway. The Data Center Coalition’s response to the Texas audit made the industry’s strongest argument explicitly: with billions in investment and hundreds of thousands of jobs at stake, regulators should move swiftly to distinguish speculative projects from serious, committed investors. [11] That is not an unreasonable position. It is, in fact, the same argument the retrenchment states are making, pointed in the opposite direction.
The emerging equilibrium is a negotiation between two genuinely valuable asset bundles:
| Corporate Asset Bundle | State Asset Bundle |
| Capital at unprecedented scale | Electricity capacity and interconnection rights |
| Technology and operational capability | Land, water, and basin authority |
| Compute demand that finances generation | Permitting bandwidth and administrative speed |
| Ability to build generation others cannot | Political legitimacy and community consent |
| Mobility across jurisdictions | Legal authority over all of the above |
Subsidy Retrenchment occurs when the right-hand column becomes sufficiently scarce to command a larger share of the joint surplus. States cannot impose unlimited costs without triggering migration; companies cannot ignore unlimited conditions without forfeiting access. The outcome is bargaining, and bargaining outcomes depend on outside options, credibility, and information — which is precisely why transparency provisions like Pennsylvania’s NDA prohibition and Texas’s audit disclosures are more strategically important than they first appear. Information asymmetry was the state’s principal weakness in the prior regime.
4.6 The Winners May Be the Companies That Can Pay Their Own Way
There is a significant and underappreciated irony in the retrenchment movement, and it deserves serious examination rather than a footnote.
The largest hyperscalers may gain competitive advantage from stricter rules. A company with hundreds of billions in annual capital expenditure, an in-house energy development team, investment-grade credit, and the capacity to sign twenty-year take-or-pay obligations can satisfy a bring-your-own-clean-power requirement, fund a substation, capitalize a community benefits fund, post letters of credit, and absorb a $0.011 per kilowatt-hour excise. A mid-market colocation developer, a regional data-center REIT, a sovereign-backed newcomer, or a startup building specialized inference capacity frequently cannot.
The consequence is that regulation designed to protect communities may simultaneously raise barriers to entry and increase infrastructure concentration in the hands of the five or six firms that already dominate the layer. Every instrument on the retrenchment menu — minimum-bill obligations, long contract terms, collateral requirements, self-generation mandates, community-benefit funds — is a fixed cost, and fixed costs favor scale.
This is not an argument against the instruments. Ratepayer protection is a legitimate objective and the alternative — allowing thinly capitalized speculators to reserve scarce capacity and default — is worse. But policymakers should recognize that they are making a structural choice about market concentration in the physical layer of the AI economy, and should recognize it explicitly rather than discovering it in five years. Mitigations exist: graduated requirements scaled to project size, pooled community-benefit vehicles that small developers can join, standardized rather than bespoke contract terms that reduce transaction costs, and public or cooperative interconnection entities. None is currently a prominent feature of any state framework reviewed for this paper.
4.7 The Financial-Stability Overlay
A final consideration links state-level retrenchment to a macro-financial concern that has moved from the margins to the center of institutional analysis in 2026.
The International Monetary Fund’s April 2026 Global Financial Stability Report drew an explicit comparison between the current AI investment cycle and the dot-com period, when massive investments were also made in fiber networks, datacenters, server farms, and undersea cables, and when revenues ultimately fell short of expectations because the business models had not reached commercial maturity. The Fund identified analogous channels through which expectations and fundamentals could become misaligned in the present cycle, including circular financing arrangements, power-generation capacity constraints, and cybersecurity risk, and assessed AI-value-chain balance sheets across leverage, liquidity, profitability, capital-expenditure intensity, and valuation. [94] [95] The report noted separately that surging corporate issuance, including AI-related debt, could put upward pressure on funding markets as dealers warehouse bond supply. [94] Datacenters, the Fund observed, have become a large part of global commercial real estate. [94] In the United States Senate, a group of legislators pressed the Financial Stability Oversight Council in January 2026 to investigate the risks associated with more than $1 trillion in projected AI infrastructure debt. [103]
The connection to state policy is direct and runs in both directions. A stranded gigawatt-scale campus is a state fiscal problem as well as a creditor problem. If a project is cancelled after dedicated transmission, substation, and generation assets have been constructed against it, someone pays for those assets, and under the pre-retrenchment allocation that someone was the ratepayer. This is precisely the risk that take-or-pay minimums, collateral requirements, long contract terms, and New York’s proposed insurance pool are designed to transfer back to the causing party.
Viewed this way, retrenchment is not only a distributional policy. It is ratepayer-side prudential regulation — an attempt by state governments to avoid becoming the residual claimant on a private capital cycle whose duration risk they do not control and whose revenue assumptions they cannot verify. The IMF’s diagnosis and Pennsylvania’s executive order are addressing the same underlying exposure at different altitudes.
Section 5: Governors, Communities, and the New Politics of AI Infrastructure
Economic analysis explains why retrenchment is defensible. It does not explain why it happened in the summer of 2026 rather than in 2024 or 2028. That requires attention to politics — specifically, to the way an abstract national debate about artificial intelligence became a concrete local debate about utility bills, farmland, water tables, transmission towers, generator exhaust, and construction traffic. The transformation of AI from a Silicon Valley subject into a county-commission subject is the proximate cause of the policy sequence this paper describes, and it is unlikely to reverse.
5.1 The Bipartisan Character of the Shift
The most politically significant fact about 2026 is that the movement crosses party lines without translation.
Josh Shapiro in Pennsylvania, Kathy Hochul in New York, Maura Healey in Massachusetts, and Abigail Spanberger in Virginia are Democrats. Greg Abbott in Texas and Jim Pillen in Nebraska are Republicans. Their governing philosophies diverge substantially on taxation, regulation, energy policy, climate policy, and the proper scope of executive authority. Yet each arrived at a version of the same conclusion within a ten-week window, and each used the vocabulary of their own coalition to describe it.
The Democratic framing emphasizes ratepayer protection, environmental justice, corporate accountability, and community consent. Healey’s halting of tax incentives until protections exist against higher gas and electric bills is the archetype. [23] The Republican framing emphasizes fiscal discipline, opposition to corporate welfare, resource stewardship, and local control. Pillen’s insistence that Nebraska should attract givers rather than takers, and Jacobson’s observation that the companies have ample resources of their own, are the archetype. [28] [31]
These are different arguments. They produce nearly identical instruments.
That convergence is what makes the phenomenon durable rather than cyclical. A policy adopted by one coalition is reversed when the other takes power. A policy adopted independently by both coalitions, for different stated reasons, tends to persist. The federal picture reinforces this: President Trump’s Ratepayer Protection Pledge, first announced in the February 24, 2026 State of the Union address and signed by Amazon, Google, Meta, Microsoft, OpenAI, Oracle, and xAI at the White House on March 4, commits signatories to the same core principle — pay for the energy and the infrastructure — that Shapiro and Hochul are imposing by order. [73] [75] The July 23 expansion brought in fifty-five utilities, one hundred six energy cooperatives, twenty-eight datacenter developers, and twenty-three governors, all Republican — with Energy Secretary Chris Wright and EPA Administrator Lee Zeldin framing the commitment as a condition of winning the AI competition with China, and with Democratic governors Wes Moore of Maryland and Josh Shapiro of Pennsylvania, both parties to an earlier White House agreement on PJM generation, conspicuously absent. [77] [78] The White House claimed the pledge covers eighty percent of power delivered to American homes and businesses. [74] The President’s own framing was notably continuous with the state-level logic:
“The right approach to data centers is not to stop progress altogether”
— President Donald J. Trump, White House signing, March 4, 2026 [75]
The disagreement between the federal pledge and the state orders is therefore not about the principle. It is about enforceability — and that disagreement is sharp. Harvard’s Ari Peskoe assessed the pledge’s legal character directly:
“The Ratepayer Protection Pledge appears to be an unenforceable corporate commitment”
— Ari Peskoe, Harvard Law School, March 2026 [48]
His point was structural rather than cynical: utilities hold the pen on rate filings and utility regulators make the final decisions, so a voluntary corporate commitment does not by itself change what a commission approves. [48] He has also noted that the pledge’s three substantive elements — covering the cost of new generation, paying for delivery infrastructure, and paying even if the facility never comes online — are not themselves novel, since companies have been making similar commitments for some time; what would be novel is a mechanism that makes them enforceable. [46] The state instruments described in this paper are attempts to make the same commitment binding through the mechanisms that actually determine outcomes — permits, tariffs, and interconnection agreements.
5.2 The Governor Becomes an Allocator of the AI Economy
The second political transformation concerns the role itself.
Governors have long been promoters of economic development. The 2026 sequence positions them as something different and considerably more consequential: allocators of scarce AI infrastructure capacity. Through executive orders, appointments to public utility commissions, directives to grid operators, environmental permitting authority, and control of state incentive programs, a governor now influences:
- which projects receive grid access and in what order;
- which projects receive tax benefits and on what conditions;
- who finances new generation and who owns it;
- who pays transmission and network-upgrade costs;
- how water is allocated among competing uses;
- what host communities receive and whether it is enforceable;
- which projects are accelerated;
- and which projects are never built.
Abbott’s directive to the PUCT and ERCOT is the clearest illustration: a single letter from a governor’s office suspended a grid operator’s classification process for the largest interconnection queue in North America, and the operator complied within hours. [8] [10] Hochul’s Executive Order 62 halted an entire category of discretionary state environmental permit. [15] Shapiro’s order made local land-use approval a precondition to state review. [1]
This is a substantial accretion of practical authority over industrial allocation, exercised largely through instruments — permits, queues, tariffs — that were not designed for the purpose. It carries a corresponding risk: allocation decisions made through executive discretion rather than transparent standards invite exactly the opacity and favoritism that retrenchment was meant to eliminate. The states that handle this well will convert discretion into published, appealable criteria. The states that handle it badly will replace a subsidy auction with a permitting auction.
5.3 Local Consent Becomes a Priced Economic Input
The third transformation is the elevation of local political acceptance from a downstream risk to an upstream input with measurable value.
Pennsylvania’s requirement that local land-use approval precede state environmental review, and New York’s emphasis on host-community participation and its blueprint for supporting localities, both formalize something that was already true in practice: a technically perfect datacenter site has little economic value if the host jurisdiction refuses it. [1] [17] The Data Center Watch findings quantify the cost of ignoring this — $130 billion in blocked or delayed projects in a single quarter, 833 active opposition groups across 49 states. [87] [88]
Municipalities occupy a peculiar and powerful position in this landscape. They generally lack authority over electricity rate-setting, which is a state and federal matter. But they possess robust authority over land use: where facilities can be built, how large they can be, what studies must precede groundbreaking, what infrastructure must be demonstrated as available, and what conditions must be satisfied for approval. [48] In the absence of adequate state and federal ratepayer protection, thoughtful local ordinances have become the most immediately available tool available to communities — and, increasingly, the ballot box. Holyoke, Massachusetts enacted a citywide ban on AI datacenters. [26] Brookhaven, New York passed its own moratorium before the state acted. [18] Statewide ballot initiatives requiring local voter approval for datacenter permits were proposed in Ohio and then Massachusetts. [27]
Community legitimacy should therefore be treated as a measurable component of infrastructure feasibility, with the same analytical seriousness applied to interconnection studies and geotechnical surveys. In underwriting terms, it is a real option that can expire worthless.
5.4 November 2026 and the Datacenter Kitchen-Table Issue
The fourth transformation is electoral, and it is accelerating everything else.
Electricity bills, farmland conversion, transmission lines, generator emissions, water consumption, construction traffic, noise, and property values convert an abstract debate about artificial intelligence into intensely local questions with immediate household consequences. That conversion is what moved public opinion twenty-one points in eight months.
The 2025 results provided the proof of concept. Analysts attributed Democratic gubernatorial victories in Virginia and New Jersey substantially to the positions candidates took against datacenter development, and Democratic candidates gained traction by connecting high electricity rates to datacenter growth. [79] Conservatives joined the critique from a different direction; Florida Governor Ron DeSantis proposed an AI bill of rights explicitly including protection from datacenter electricity costs. [79] By 2026, candidates in both parties were campaigning on the issue in congressional battlegrounds spanning California, Georgia, Michigan, Ohio, Pennsylvania, and Texas. [85]
Brookings’ assessment of the electoral dynamic was unambiguous:
“Public concern over electricity costs will likely dominate this year’s campaign dialogue”
— Darrell M. West, Senior Fellow, Governance Studies, The Brookings Institution [79]
There is, however, a structural warning embedded in the same analysis. Most candidates are folding electricity increases into a generalized “affordability” narrative alongside housing, food, and fuel. The difficulty with aggregating distinct problems under a single label is that it becomes harder to explain — or to enact — the specific remedies each requires. [79] Affordability politics can produce cost-causation reform. It can equally produce a moratorium that delays generation investment and raises prices. The distinction between those outcomes is precisely the distinction between the architecture of retrenchment and its rhetoric.
Pennsylvania is the clearest laboratory for this. In Lackawanna County, datacenters are rising on scarred post-anthracite ground, and residents have made electoral support conditional on candidates’ positions. [83] The state that produced the nation’s strictest guardrails did so in an election year in a state where electricity rates rose 18.9 percent in 2025. [79] Whether that is responsive democracy or reactive policymaking is a judgment this paper leaves to the reader; that it is causally connected is not in serious doubt.
5.5 A Durable Policy Framework for Governors
Rather than choosing between “welcome every datacenter” and “ban datacenters” — a binary that serves neither ratepayers nor investment — states can construct a middle path. Based on the instruments actually in use across the six states examined, a durable framework should address ten elements.
| # | Principle | Operative Requirement | Instrument |
| 1 | Additionality | Demonstrate how new demand will be supplied without displacing existing customers | Interconnection agreement; clean-supply condition |
| 2 | Cost Causation | Assign but-for infrastructure costs to the causing load | Large-load tariff; network-upgrade allocation |
| 3 | Queue Integrity | Require financial and construction milestones before reserving capacity | Deposits, collateral, milestone forfeiture |
| 4 | Transparency | Disclose owner, operator, load, water use, incentives, and timeline | NDA prohibition; mandatory filings |
| 5 | Local Consent | Require host-jurisdiction approval as a precondition, not an afterthought | Sequencing rule; zoning authority |
| 6 | Resource Pricing | Price electricity, water, and infrastructure at scarcity value | Consumption excise; grid premium; water charge |
| 7 | Community Return | Convert benefits into enforceable obligations with clawbacks | Community benefits agreement; host fund |
| 8 | Flexibility | Reward load reduction during system emergencies | Interruptible tariff; demand-response credit |
| 9 | Incentive Conditionality | Tie any remaining tax benefit to measured performance | Certification with recapture provisions |
| 10 | Long-Term Liability | Protect ratepayers from stranded assets if a project cancels | Take-or-pay; insurance pool; termination fees |
Three design cautions apply across all ten. Thresholds shape behavior: a 75-megawatt trigger invites 74-megawatt facilities, and a fifty-megawatt trigger invites 49-megawatt campuses in clusters. Sunset and review provisions are features, not weaknesses, because the underlying scarcity conditions are regional and time-varying. And standardization reduces the barrier-to-entry problem identified in Section 4.6: bespoke negotiation favors the largest incumbents, while published standard terms are navigable by mid-market developers.
5.6 From State Competition to State Differentiation
Not every state should adopt identical rules, and the emerging landscape suggests they will not.
Texas can compete through energy abundance, behind-the-meter generation, and speed, using the audit to protect the value of its queue rather than to shrink it. Pennsylvania can combine natural gas, an existing nuclear fleet, PJM connectivity, brownfield industrial land, and a skilled trades workforce, offering a demanding but coherent standard. New York can offer premium grid access with heavy community investment in a state where the alternative is no development at all. Virginia can offer unmatched existing density, fiber, and operational ecosystem at a metered price. Nebraska and other public-power states can offer low rates on the explicit condition that the public power districts remain protected. Other states will organize around hydroelectric surplus, geothermal potential, nuclear brownfields, renewable overbuild, or post-industrial redevelopment.
The future competition therefore shifts from:
“Which state gives the largest subsidy?”
to:
“Which state offers the most credible long-term infrastructure bargain?”
Twenty-three states had approved at least one large-load tariff as of May 2026, with another seven pending, according to Edison Electric Institute tracking. [105] [111] At least eighteen states had introduced bills creating special rate classes or infrastructure cost-sharing mandates for large energy users, while construction-moratorium proposals advanced in states including New York, South Dakota, Oklahoma, and Vermont, and rollback or restructuring efforts targeted tax credits in Georgia, Oklahoma, Indiana, and Virginia. [106] Oregon, Minnesota, Indiana, Texas, and Virginia have enacted statutes governing how utilities treat large-load customers, and Pennsylvania regulators are developing a model large-load tariff applicable to all utilities in the Commonwealth. [104] This is no longer a set of isolated experiments. It is the formation of a national regulatory architecture built state by state.
5.7 The Federal-State Collision
One further dimension will shape the next two years: the question of who decides.
The Department of Energy directed FERC in October 2025, under Section 403 of the DOE Organization Act, to initiate rulemaking to accelerate the interconnection of large loads to the interstate transmission system — defining large loads generally as demand exceeding twenty megawatts, proposing standardized study deposits, concurrent study of load and generation, and assignment of one hundred percent of network upgrade costs to the interconnecting load. [71] [68] FERC opened Docket RM26-4-000, accumulated an administrative record exceeding 3,500 pages from nearly 175 stakeholders, and committed to act by June 2026. [67] [68] Chairman Laura V. Swett framed the stakes broadly:
“Our nation stands at a pivotal moment as we face rapid growth in demand”
— Laura V. Swett, Chairman, Federal Energy Regulatory Commission [67]
On June 18, 2026, rather than issuing a single national rule, FERC took a targeted path: it issued Section 206 show-cause orders to all six jurisdictional RTOs and ISOs — PJM, MISO, SPP, CAISO, ISO New England, and NYISO — and their transmission-owning members, requiring each within sixty days either to justify why existing tariffs remain just and reasonable without provisions tailored to large loads, or to file tariff changes. Within thirty days, each was required to submit an informational report describing how it will ensure adequate generation is available to serve existing and new large loads. The orders also directed dedicated, expedited study processes for generating facilities serving electrically proximate and co-located large loads. [68] [69] [72]
The jurisdictional stakes are not incidental. FERC has historically regulated generation interconnection; load interconnection has been a state and local matter. The National Association of Regulatory Utility Commissioners, representing state regulators, argued in the docket that FERC asserting jurisdiction over load interconnection would interfere with state authority over retail ratemaking. FERC’s choice of region-specific show-cause orders over a generic rule can be read as a partial accommodation of that objection — but the boundary remains unsettled, and it is the boundary on which the entire retrenchment project rests. A state cannot condition grid access if grid access is federally allocated.
Texas is partially insulated by ERCOT’s non-jurisdictional status, which is one reason the Texas instrument is a gubernatorial directive rather than a tariff. New York and Pennsylvania are not. The 90th Texas Legislature convenes in January 2027, when the PUCT is expected to seek expanded statutory authority over the datacenter industry. [10] The next twenty-four months will determine whether Subsidy Retrenchment is a state-level regime with federal accommodation, or a state-level regime substantially preempted at the point where it bites hardest.
5.8 The 2027–2029 Outlook
Extrapolating from the instruments now in use, the policy debate through 2029 should be expected to concentrate on eleven fronts:
- special electricity tariffs and dedicated rate classes for hyperscale loads, extending the twenty-three-state base;
- minimum financial commitments — deposits, collateral, and take-or-pay floors — as the standard price of a queue position;
- mandatory self-generation or verified incremental generation, with additionality tests replacing accounting-based clean-energy claims;
- community-investment funds and host-locality trusts with statutory disbursement rules;
- stricter tax-exemption eligibility, including job-creation floors, clawback provisions, and periodic recertification;
- interruptible and flexible datacenter service priced as a grid resource rather than treated as a concession;
- dedicated water charges and basin-level withdrawal permits with seasonal terms;
- transmission cost allocation reform at both FERC and state level;
- site-specific environmental standards for generator emissions, noise, and thermal discharge;
- comprehensive disclosure requirements replacing NDAs and code-named applicants;
- and local referenda and enhanced zoning authority, following the Ohio and Massachusetts ballot models.
The datacenter incentive war is unlikely to disappear. States will continue to compete, and some will compete by offering better terms than their neighbors. But the unit of competition is changing from the tax dollar to the scarce megawatt — and the terms of trade are moving toward the party that controls the megawatt.
Section 6: What Have We Learned? Nine Pillars
Pillar 1 — AI Capital Is No Longer the Only Scarce Asset
The economic assumption underlying decades of technology recruitment was that companies controlled scarce capital while states possessed interchangeable land, electricity, and infrastructure. Under that assumption, bidding was rational and the surplus properly accrued to the mobile factor.
The AI infrastructure cycle has partially reversed the assumption. Capital is attempting to enter the physical economy faster than many regions can generate electricity, construct transmission, procure transformers, secure water, obtain permits, and win community acceptance. Roughly $725 billion of hyperscaler capital expenditure in a single year is chasing a set of physical inputs that cannot expand at anything like that rate. [102]
When infrastructure becomes scarcer than the demand for it, bargaining power moves toward infrastructure owners and host governments. That is the first principle of Subsidy Retrenchment, and everything else follows from it.
Pillar 2 — The Incentive Is Becoming Conditional, and Then Reciprocal
The future incentive will be contractual rather than automatic. A hyperscaler may still receive favorable tax treatment or expedited approval — but only if it supplies additional generation, finances upgrades, meets environmental standards, hires locally, protects ratepayers, discloses material facts, and delivers enforceable community benefits.
The trajectory is:
Subsidy → Conditional Subsidy → Reciprocity
Massachusetts and Nebraska have executed the first transition. Pennsylvania and New York are attempting the second. Virginia has attempted something more unusual: preserving the subsidy while separately pricing the scarce input. All four are moving away from the unconditional grant, and none is moving toward prohibition.
Pillar 3 — Grid Access Is Becoming a Form of Economic Capital
The ability to obtain 500 megawatts or one gigawatt of reliable electricity at a specific location on a specific date is now more valuable than a conventional tax incentive, and in constrained regions it is more valuable by a wide margin.
Texas demonstrates that verification can precede access. New York demonstrates that access can carry a premium. Pennsylvania demonstrates that access can be tied to incremental generation and full cost responsibility. Ohio demonstrates that requiring financial commitment reveals which requests were real. [104] [108]
For the AI economy, an approved megawatt is becoming a form of economic capital — an asset with a price, a queue, a counterparty risk, and increasingly a secondary market.
Pillar 4 — Communities Have Entered the AI Capital Stack
Datacenter finance has traditionally centered on developers, lenders, hyperscalers, utilities, and equipment suppliers. Communities appeared, if at all, as a public-relations line item.
That is no longer tenable. Communities now determine whether projects are politically and legally feasible, and they have demonstrated the capacity to block $130 billion of announced investment in a single quarter. [87] Community-benefit agreements, local infrastructure funds, tax guarantees, environmental protections, and host approval are not peripheral. They belong inside the project’s economic architecture, capitalized and underwritten like any other precondition to operation.
Pillar 5 — Cost Causation Is the Durable Core; the Bill-Blame Narrative Is Not
The most important analytical lesson of this paper is a distinction between two arguments that are frequently conflated.
The argument that datacenters caused the electricity price increases of 2021–2025 is empirically contested. The USC Marshall research finds no statistically significant relationship between datacenter growth and retail rates at investor-owned utilities serving sixty-eight percent of Americans, and attributes the broader surge to thermal retirements, state policy, and grid modernization. [49] Policies premised on that causal claim are vulnerable to being proven wrong and to failing to deliver the relief they promise.
The argument that a customer causing a specific, identifiable, gigawatt-scale infrastructure cost should bear that cost requires no causal claim about history at all. It is a principle of ratemaking, and it becomes more important as loads grow.
States that build their frameworks on the second argument will find them durable. States that build on the first will find them contested — and will find that bills continue rising after the datacenters are restricted.
Pillar 6 — Regulation Designed to Protect Communities May Concentrate the Industry
Every retrenchment instrument is a fixed cost, and fixed costs favor scale. Self-generation mandates, twenty-year take-or-pay commitments, collateral requirements, community-benefit funds, and bespoke negotiation all advantage firms with the largest balance sheets and in-house energy development capability.
The plausible unintended consequence of a movement motivated by concern about corporate power is greater concentration of the physical layer of the AI economy among the firms that already dominate it. This tension deserves explicit attention rather than discovery after the fact, and it has straightforward mitigations — graduated thresholds, standardized terms, pooled compliance vehicles — that no state framework reviewed here currently employs.
Pillar 7 — Transparency Is the Precondition for Everything Else
Cost causation cannot be enforced against contracts nobody can read. Queue integrity cannot be assessed against applicants identified only by code name. Opportunity cost cannot be calculated against load data that is not filed. Community consent cannot be meaningful when the community does not know who is applying or for how much power.
The Harvard finding that regulators frequently approved special contracts in short and conclusory orders, without engaging the underlying analysis, because challenge is costly and staff resources are limited, is not principally a finding about regulatory capture. [47] It is a finding about information economics: opacity shifts the cost of scrutiny onto the party least able to bear it.
Pennsylvania’s prohibition on nondisclosure agreements between Commonwealth agencies and datacenter developers is, for this reason, arguably the most consequential single provision in Executive Order 2026-05 — more consequential than the permitting changes, because it changes what every subsequent proceeding can see. [1]
Pillar 8 — Flexibility Is the Most Underpriced Asset in the System
The one thing a gigawatt-scale computational load can offer a grid that almost nothing else of comparable size can offer is the ability to stop consuming, quickly, on instruction.
Not all AI workloads are equally flexible — inference serving real-time users is not, while much training and batch processing is — but the flexible fraction is large, and it is currently compensated almost nowhere at anything approaching its system value. Pennsylvania’s requirement that grid-connected datacenters go offline in a shortage, and PJM’s parallel proposal at FERC, treat flexibility as an obligation. [2] Treating it instead as a product — procured, priced, and contracted — would convert the largest new source of grid stress into one of the largest new sources of grid reliability.
The states that solve this will be able to accommodate substantially more compute per firm megawatt than the states that do not. It is the single highest-leverage unexploited opportunity in the policy space.
Pillar 9 — The Winning State Will Offer the Most Credible Deal, Not the Cheapest
The long-term winner will not necessarily be the state that offers the largest subsidy. It will be the state that offers the most credible package:
credible electricity supply; credible transmission delivery; credible permitting timelines; credible community approval; credible environmental standards; credible cost allocation; and credible political durability across administrations.
Hyperscalers deploying tens of billions of dollars against multi-decade asset lives need predictability nearly as much as they need incentives — arguably more, since the incentives are a modest fraction of total project cost while a two-year delay can be existential to a model roadmap. Mortenson survey data cited in the subsidy literature found that only three percent of datacenter owners described tax credits and local incentives as the most important site-selection factor, with land availability and power cost dominating. [58]
This yields the paper’s most important practical conclusion, and it is genuinely counterintuitive:
Subsidy Retrenchment can be pro-investment. A state that replaces unpredictable backlash, ad hoc moratoria, and litigated local fights with transparent rules establishing in advance exactly what AI infrastructure will cost — in dollars, in obligations, and in time — may attract more capital than a state that offers a large exemption alongside an unresolved political conflict. The exemption is worth a few percent of project cost. The certainty is worth years.
Conclusion: From Paying for Compute to Making Compute Pay for Scarcity
The first era of American datacenter expansion was defined by recruitment. States wanted digital infrastructure and competed to attract it. Tax exemptions, expedited permits, infrastructure assistance, socialized network upgrades, and inexpensive electricity became instruments in an interstate bidding contest for cloud-computing investment. The bidding was not foolish. Given a world in which capital was scarce, sites were fungible, and facilities were modest, it was very close to the correct strategy.
The AI era is beginning under different physical conditions, and physical conditions determine bargaining power.
Frontier models require enormous computational systems. Those systems require accelerators, networking, cooling, land, and buildings. And beneath all of them sits the first layer of the AI economy: electricity — generation, transmission, substations, transformers, water for cooling, and a workforce capable of building all of it. As AI investment scales from hundreds of megawatts toward gigawatt campuses, and as roughly $725 billion of annual hyperscaler capital expenditure collides with grids that add capacity on decade-long timelines, the availability of that first layer becomes progressively more valuable relative to the capital seeking it. [102] The International Energy Agency’s observation that the AI surge is increasingly meeting physical constraints is not a technical footnote. It is the entire economic story. [89]
The events of summer 2026 are the political expression of that arithmetic.
Pennsylvania — a state that celebrated a historic $20 billion Amazon investment — now tells future developers that they must obtain local approval first, finance the electricity infrastructure they require, protect ratepayers from cost shifting, operate without nondisclosure agreements, and meet binding environmental and community standards or lose access to the Commonwealth’s tax exemption. [1] Texas — among the most aggressively pro-development states in the country — has placed new ERCOT-connected datacenter approvals behind a comprehensive audit designed to separate credible projects from the ninety percent of a 474-gigawatt queue that may not be real. [8] New York has proposed that hyperscalers provide their own energy or pay a premium, has moved to eliminate datacenter tax subsidies, and has proposed an insurance pool against speculative load. [15] [17] Massachusetts froze a twenty-year exemption one month after launching it. [23] Nebraska removed datacenters from its flagship incentive program while reconsidering the resource equation entirely. [28] And Virginia — the industry’s largest single market — preserved its exemption while imposing a per-kilowatt-hour excise on the compute itself. [35]
None of these developments proves that America is retreating from artificial intelligence. Taken together, they reveal something more nuanced and considerably more durable.
America is entering a period in which governments become more selective about which AI infrastructure receives scarce physical capacity, under what conditions, and at whose expense.
That is why the term Subsidy Retrenchment is the right one.
Subsidy identifies the regime being reconsidered: tax advantages, expedited permitting, infrastructure support, favorable rate treatment, administrative confidentiality, and the socialization of costs, all deployed to attract datacenter investment during a period when states believed those concessions were the price of admission.
Retrenchment describes the reversal without falsely implying prohibition. Governments are pulling benefits back, narrowing them, attaching conditions to them, or exchanging them for reciprocal investment. They are not, in the main, saying no to AI. Increasingly they are saying yes — at a price that reflects scarcity.
And that distinction is what makes the subtitle load-bearing rather than decorative: when states stop bidding for AI datacenters, and start charging them for scarcity.
The scarcity may be electricity. It may be a position in an interconnection queue. It may be transmission capacity, water, land, permitting bandwidth, political tolerance, or community consent. Sometimes the price appears as a direct tariff or excise — Virginia’s $0.011 per kilowatt-hour. Sometimes it appears as a forgone exemption, as privately financed generation, as a new substation, as a community-benefit agreement, as mandatory water recycling, as a twenty-year take-or-pay obligation, or as the requirement to shut down first when the grid is short. Economically, these mechanisms all belong to the same transition. They are attempts to construct an accurate price for a physical footprint that the accounting conventions of the cloud era never had to measure.
The Five-Layer AI Economy is therefore entering a phase in which Layer 3 cannot expand merely because Layers 4 and 5 demand more intelligence and Layer 2 can supply more chips. Layer 1 — and the governments and communities that control access to its physical resources — is beginning to send a price signal upward through the entire system. That signal will shape which models get trained, where inference gets served, which firms can afford to build, and which regions capture the industrial activity that follows.
For hyperscalers, this means tomorrow’s competitive advantage may depend less on acquiring accelerators than on becoming credible builders of electricity, water, and community infrastructure — a capability set almost entirely absent from the industry’s core competencies a decade ago.
For investors, it means datacenter economics must now internalize regulatory, infrastructure, and resource obligations that were previously external to the project, and must carry an explicit regulatory volatility premium in the underwriting.
For local governments, it means a gigawatt campus represents bargaining power as well as burden — and that the bargaining power is perishable, because it exists only while the queue is full.
And for governors, the defining AI infrastructure question of 2027 and beyond may no longer be:
“How much should we offer to convince this datacenter to come?”
It may instead become:
“If this company needs something from our state that is genuinely scarce, what should our citizens receive in return?”
That question is the essence of Subsidy Retrenchment. The states now answering it are writing the operating rules for the physical layer of the artificial-intelligence economy — and, in doing so, are deciding how the enormous surplus generated in Layers 4 and 5 will be shared with the communities that host Layer 3 and power Layer 1.
Endnotes and Sources:
[1] 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
[2] Susan Phillips, WHYY News Climate Desk — “Pennsylvania Gov. Shapiro signs wide-ranging executive order reining in data center development,” August 2026. https://whyy.org/articles/shapiro-data-centers-executive-order-pennsylvania/
[3] The Hill — “Josh Shapiro signs order restricting AI data centers in Pennsylvania,” August 19, 2026. https://thehill.com/policy/technology/6037158-shapiro-signs-ai-data-center-order/
[4] Whitney Downard, Pennsylvania Capital-Star — “Gov. Shapiro signs data center executive order. Critics say it falls short,” August 19, 2026. https://penncapital-star.com/technology-information/gov-shapiro-signs-data-center-executive-order-critics-say-it-falls-short/
[5] Times Leader (Wilkes-Barre) — “Gov. Shapiro signs executive order on data center development in Pennsylvania,” August 19, 2026. https://www.timesleader.com/news/1752416/gov-shapiro-signs-executive-order-on-data-center-development-in-pennsylvania
[6] Office of Governor Josh Shapiro, Commonwealth of Pennsylvania — “What People Are Saying About Governor Shapiro’s Executive Order on Data Centers,” August 2026. https://www.pa.gov/governor/newsroom/2026-press-releases/what-people-are-saying-about-governor-shapiro-s-executive-order-
[7] FOX 29 Philadelphia — “Gov. Shapiro signs executive order putting guardrails on Pennsylvania data centers,” August 18, 2026. https://www.fox29.com/news/gov-shapiro-signs-executive-order-places-new-guardrails-data-center-development
[8] Akin Gump Strauss Hauer & Feld LLP — “Texas Pauses Data Center Interconnections Pending Statewide Audit,” August 2026. https://www.akingump.com/en/insights/alerts/texas-pauses-data-center-interconnections-pending-statewide-audit
[9] 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/
[10] Troutman Pepper Locke — “Texas Hits Pause on Data Center Grid Connections Amid Growing Oversight Push,” August 2026. https://www.troutman.com/insights/texas-hits-pause-on-data-center-grid-connections-amid-growing-oversight-push/
[11] Natalie Weber, Houston Public Media — “Gov. Greg Abbott pauses new data centers until ERCOT, PUCT audit energy, water usage,” August 3, 2026. https://www.houstonpublicmedia.org/articles/news/energy-environment/2026/08/03/558529/gov-greg-abbott-pauses-new-data-centers-until-ercot-puct-audit-energy-water-usage/
[12] The Texas Tribune — “Texas will audit up to 300 projects, mostly data center proposals,” August 14, 2026. https://www.texastribune.org/2026/08/14/texas-data-center-approval-pause-ercot-power-grid/
[13] David G. Cabrales, Scott D. Johnson and Daniel Farris — “Governor Abbott Pauses Texas Data Center Interconnections and Calls for Verification and Audit,” August 2026. https://www.mondaq.com/unitedstates/trademark/1827218/governor-abbott-pauses-texas-data-center-interconnections-and-calls-for-verification-and-audit-what-data-center-developers-need-to-know-now
[14] Fermi Inc. — Quarterly Report on Form 10-Q for the period ended June 30, 2026, U.S. Securities and Exchange Commission. https://www.sec.gov/Archives/edgar/data/0002071778/000207177826000051/frmi-20260630.htm
[15] Office of Governor Kathy Hochul, State of New York — “First Statewide Moratorium on New Hyperscale Data Centers Launched by Governor Kathy Hochul,” July 14, 2026. https://www.governor.ny.gov/news/first-statewide-moratorium-new-hyperscale-data-centers-launched-governor-kathy-hochul
[16] Office of Governor Kathy Hochul, State of New York — Rush Transcript, “Governor Hochul Launches First Statewide Moratorium on New Hyperscale Data Centers,” July 14, 2026. https://www.governor.ny.gov/news/video-audio-photos-rush-transcript-governor-hochul-launches-first-statewide-moratorium-new
[17] Office of Governor Kathy Hochul, State of New York — “Governor Hochul Highlights First Statewide Moratorium on New Hyperscale Data Centers as Support Continues to Grow,” July 2026. https://www.governor.ny.gov/news/governor-hochul-highlights-first-statewide-moratorium-new-hyperscale-data-centers-support
[18] Office of Governor Kathy Hochul, State of New York — “On Long Island, Governor Hochul Highlights First Statewide Moratorium on New Hyperscale Data Centers,” July 2026. https://www.governor.ny.gov/news/long-island-governor-hochul-highlights-first-statewide-moratorium-new-hyperscale-data-centers
[19] Rochester Business Journal — “Hochul discusses data center moratorium in Henrietta,” July 29, 2026. https://rbj.net/2026/07/29/hochul-data-center-moratorium-power-concerns-henrietta/
[20] Spectrum News NY1 / Central NY — “Could the data center moratorium cost New York grid investments?” July 23, 2026. https://spectrumlocalnews.com/nys/central-ny/politics/2026/07/23/data-center-moratorium-grid-investments
[21] NBC News — “New York to impose the country’s first statewide moratorium on data centers,” July 14, 2026. https://www.nbcnews.com/news/us-news/new-york-impose-countrys-first-statewide-moratorium-data-centers-rcna587429
[22] Data Center Watch — Legislative Briefing, February 6, 2026 (New York S9144, Sen. Liz Krueger). https://datacenterwatch.substack.com/p/briefing-02062026
[23] Office of Governor Maura Healey and Lt. Governor Kim Driscoll, Commonwealth of Massachusetts — “Governor Healey Halts Data Center Tax Incentive and Calls for Strict Guardrails to Protect Ratepayers, Environment & Public Health,” June 25, 2026. https://www.mass.gov/news/governor-healey-halts-data-center-tax-incentive-and-calls-for-strict-guardrails-to-protect-ratepayers-environment-public-health
[24] Commonwealth of Massachusetts — “Massachusetts Qualified Data Center Sales and Use Tax Exemption,” Statement of Expectations and Framework, June 25, 2026. https://www.mass.gov/info-details/massachusetts-qualified-data-center-sales-and-use-tax-exemption
[25] Colin A. Young, State House News Service (via WBUR) — “Gov. Healey slams brakes on data center tax incentives,” June 26, 2026. https://www.wbur.org/news/2026/06/26/governor-healey-data-center-tax-incentives
[26] CBS News Boston — “Massachusetts pauses tax breaks for data centers and addresses energy, water and noise concerns,” June 26, 2026. https://www.cbsnews.com/boston/news/data-centers-massachusetts-tax-breaks-healey/
[27] Ballotpedia News — “Massachusetts initiative would require local voter approval and other conditions for data center permits,” August 13, 2026. https://news.ballotpedia.org/2026/08/13/massachusetts-initiative-would-require-local-voter-approval-and-other-conditions-for-data-center-permits-the-second-statewide-proposal-following-ohio/
[28] Office of Governor Jim Pillen, State of Nebraska — “Gov. Pillen Signs Executive Order on Data Centers,” July 20, 2026. https://governor.nebraska.gov/gov-pillen-signs-executive-order-data-centers
[29] Noelle Annonen, Nebraska Public Media — “Pillen ends tax incentives for data center developers building in Nebraska,” July 20, 2026. https://nebraskapublicmedia.org/en/news/news-articles/pillen-ends-tax-incentives-for-data-center-developers-building-in-nebraska/
[30] Erin Bamer, Nebraska Examiner — “Pillen rescinds ImagiNE Act tax incentives for new Nebraska data centers,” July 20, 2026. https://nebraskaexaminer.com/2026/07/20/pillen-rescinds-imagine-act-tax-incentives-for-new-nebraska-data-centers/
[31] The Summerland Advocate-Messenger — “Governor ends tax incentives for data centers,” July 29, 2026. https://www.summerlandadvocate.com/story/2026/07/29/news/governor-ends-tax-incentives-for-data-centers/9676.html
[32] KOLN/KGIN 10/11 News — “Gov. Pillen signs order to protect Nebraska’s public power from data center drain,” July 20, 2026. https://www.1011now.com/2026/07/20/gov-pillen-signs-order-protect-nebraskas-public-power-data-center-drain/
[33] Nebraska TV (KHGI) — “Nebraska Gov. Jim Pillen signs order ending taxpayer subsidies for large data centers,” July 20, 2026. https://nebraska.tv/news/local/nebraska-gov-jim-pillen-signs-order-ending-taxpayer-subsidies-for-large-data-centers
[34] Joint Legislative Audit and Review Commission (JLARC), reported by WRIC ABC 8News — “Data centers save $2.7 billion through tax exemptions, make up more than half of Virginia’s incentive spending,” December 2025. https://www.wric.com/news/virginia-news/data-centers-tax-exemption-virginias-incentive-spending-jlarc/amp/
[35] Holland & Knight LLP — “Virginia Preserves Data Center Tax Incentive, Adds New Electricity Consumption Tax,” August 2026. https://www.hklaw.com/en/insights/publications/2026/08/virginia-preserves-data-center-tax-incentive
[36] Williams Mullen — “Virginia Budget Creates New Electricity Consumption Tax for Data Centers,” June 30, 2026. https://www.williamsmullen.com/insights/news/legal-news/virginia-budget-creates-new-electricity-consumption-tax-data-centers
[37] The Commonwealth Institute for Fiscal Analysis — “Virginia’s Data Center Tax Debate: What Changed and Why It Matters,” July 21, 2026. https://thecommonwealthinstitute.org/tci_blog/virginia-data-center-tax-exemption/
[38] Virginia Department of Taxation and Virginia Economic Development Partnership — “Biennial Data Center Retail Sales and Use Tax Exemption Report,” Report Document RD40, January 2, 2026. https://rga.lis.virginia.gov/Published/2026/RD40/PDF
[39] Whiteford, Taylor & Preston LLP — “Virginia Data Center Tax Reform: Key Implications for Real Estate Developers,” 2026. https://www.whitefordlaw.com/news-events/client-alert-virginia-data-center-tax-reform-key-implications-for-real-estate-developers
[40] VPM News — “$2B data center tax break fight pushes Virginia budget negotiations,” March 12, 2026. https://www.vpm.org/generalassembly/2026-03-12/budget-data-center-tax-break-scott-lucas-spanberger-torian-rephann
[41] MultiState Associates — “Virginia Lawmakers Pass 15 Data Center Bills as Tax Exemption Fight Looms,” March 30, 2026. https://www.multistate.us/insider/2026/3/30/virginia-lawmakers-pass-15-data-center-bills-as-tax-exemption-fight-looms
[42] Virginia Department of Taxation — Legislative Summary 26-82, Joint Subcommittee on Tax Policy Data Center Study, 2026. https://www.tax.virginia.gov/laws-rules-decisions/legislative-summaries/26-82
[43] Eliza Martin and Ari Peskoe, Harvard Law School — “How You Subsidize Big Tech with Your Electricity Bill,” Salata Institute for Climate and Sustainability, Harvard University. https://salatainstitute.harvard.edu/how-you-subsidize-big-tech-with-your-electricity-bill
[44] Ethan Howland, Utility Dive — “Utilities may subsidize data center growth by shifting costs to other ratepayers: Harvard Law paper,” March 2025. https://www.utilitydive.com/news/utilities-subsidize-data-center-growth-ratepayer-cost-shif-harvard-peskoe/742001/
[45] Harvard Magazine — “How AI Could Be Raising Your Energy Bill,” July–August 2025 (updated 2026). https://www.harvardmagazine.com/2025/07/harvard-ai-increasing-energy-costs
[46] Harvard Climate Brief, Salata Institute, Harvard University — “The data center boom is colliding with the grid’s hardest problems,” interview with Ari Peskoe, March 17, 2026. https://salatainstitute.harvard.edu/data-centers-ai-artificial-intelligence-grid-permitting-transmission-electricity-energy
[47] S&P Global Market Intelligence — “Harvard paper tackles datacenter cost-shifting for US electric utility customers,” March 2025. https://spglobal.com/market-intelligence/en/news-insights/articles/2025/3/harvard-paper-tackles-datacenter-costshifting-for-us-electric-utility-customers-87921809
[48] NGLC Research — “Data Center Electricity Costs & Ratepayer Protections,” April 2026 (quoting Ari Peskoe, Harvard Law School, March 2026). https://nextgenlandco.com/research/data-center-electricity-costs-ratepayer-protections/
[49] Shon R. Hiatt, USC Marshall School of Business — “Data Centers Are Not Driving Up Your Electric Bill,” RealClearEnergy, June 8, 2026. https://www.realclearenergy.org/articles/2026/06/08/data_centers_are_not_driving_up_your_electric_bill_1186606.html
[50] Shon R. Hiatt and Angela Ryu, Zage Business of Energy Initiative, USC Marshall School of Business — “Data Center Energy Demand: Who, Where, and How Growth Is Emerging,” cited in National Center for Energy Analytics, “The Rise of AI: A Reality Check on Energy and Economic Impacts.” https://energyanalytics.org/the-rise-of-ai-a-reality-check-on-energy-and-economic-impacts
[51] Carl Boettiger, University of California, Berkeley — “Why we don’t need more data centers to build better AI,” Berkeley News, August 11, 2026. https://news.berkeley.edu/2026/08/11/why-we-dont-need-more-data-centers-to-build-better-ai/
[52] William H. Green, MIT Energy Initiative — “MIT Energy Initiative launches Data Center Power Forum,” Massachusetts Institute of Technology. https://energy.mit.edu/?p=47200
[53] Leda Zimmerman, MIT News — “Confronting the AI/energy conundrum,” Massachusetts Institute of Technology, July 2, 2025. https://lidspreview.mit.edu/news/confronting-the-ai-energy-conundrum
[54] Andrew A. Chien, University of Chicago and Argonne National Laboratory — Stanford Energy Seminar, Precourt Institute for Energy, Stanford University. https://events.stanford.edu/event/stanford-energy-seminar-andrew-chien
[55] Nathan M. Jensen, University of Texas at Austin, quoted in TIME — “Why Tax Breaks for AI Data Centers Could Backfire on States,” April 2026. https://time.com/7280058/data-centers-tax-breaks-ai/
[56] Good Jobs First — “Will data center job creation live up to hype? I have some concerns,” updated March 23, 2026. https://goodjobsfirst.org/will-data-center-job-creation-live-up-to-hype-i-have-some-concerns/
[57] Good Jobs First — “Data Centers: Key Reforms for State Subsidy Legislation.” https://goodjobsfirst.org/data-centers-best-reforms-for-state-subsidy-legislation/
[58] Data Center Dynamics — “Research: US state subsidies pay $2 million per data center job” (reporting Good Jobs First, *Money Lost to the Cloud*). https://www.datacenterdynamics.com/en/news/research-us-state-subsidies-pay-2-million-per-data-center-job/
[59] Pat Garofalo, American Economic Liberties Project — “How to Rein in Big Tech’s Secret Data Center Deals,” November 2025. https://www.economicliberties.us/wp-content/uploads/2025/11/data_center_brief_FINAL.pdf
[60] Construction Owners Association — “Data Center Tax Incentives by State: 2026 Update,” citing Brookings and Good Jobs First data. https://www.constructionowners.com/insights/data-center-tax-incentives-state-by-state-whos-still-giving-whos-taking-back
[61] PJM Interconnection — “PJM Auction Procures 134,311 MW of Generation Resources; Supply Responds to Price Signal,” PJM Inside Lines. https://insidelines.pjm.com/pjm-auction-procures-134311-mw-of-generation-resources-supply-responds-to-price-signal/
[62] Institute for Energy Economics and Financial Analysis (IEEFA) — “Projected data center growth spurs PJM capacity prices by factor of 10.” https://ieefa.org/resources/projected-data-center-growth-spurs-pjm-capacity-prices-factor-10
[63] Sierra Club — “PJM Capacity Auction Hits Price Cap for Third Consecutive Time,” July 15, 2026. https://www.sierraclub.org/press-releases/2026/07/pjm-capacity-auction-hits-price-cap-third-consecutive-time
[64] PJM Interconnection — “The Capacity Auction Is Coming. Here’s What We’re Doing Now,” PJM Inside Lines, June 25, 2026. https://insidelines.pjm.com/the-capacity-auction-is-coming-heres-what-were-doing-now/
[65] Citizens Utility Board — “Sustained High PJM Capacity Prices Ramp Up Urgency For Data Center Reform,” July 15, 2026. https://www.citizensutilityboard.org/blog/2026/07/15/cub-sustained-high-pjm-capacity-prices-ramp-up-urgency-for-data-center-reform/
[66] Introl — “PJM $100B Rate Shock: Data Centers vs Ratepayers,” February 2026 (compiling PJM BRA reports, IEEFA and Utility Dive data). https://introl.com/blog/pjm-rate-shock-100-billion-data-center-electricity-2026
[67] Federal Energy Regulatory Commission — “FERC to Act on Large Load Interconnection Docket by June 2026,” statement of Chairman Laura V. Swett, April 16, 2026. https://www.ferc.gov/news-events/news/ferc-act-large-load-interconnection-docket-june-2026
[68] White & Case LLP — “FERC orders grid operators to promptly revise or justify interconnection rules for data centers and large loads,” June 25, 2026. https://www.whitecase.com/insight-alert/ferc-orders-grid-operators-promptly-revise-or-justify-interconnection-rules-data
[69] McGuireWoods LLP — “FERC Issues Section 206 Show Cause Orders Directing All Six RTOs/ISOs to Justify or Reform Large Load Integration Rules,” June 2026. https://www.mcguirewoods.com/client-resources/alerts/2026/6/ferc-issues-section-206-show-cause-orders-directing-all-six-rtos-isos-to-justify-or-reform-large-load-integration-rules/
[70] Morgan, Lewis & Bockius LLP — “FERC Presses Grid Operators on Data Center, Large Load Interconnections,” Power & Pipes, July 2026. https://www.morganlewis.com/blogs/powerandpipes/2026/07/ferc-presses-grid-operators-on-data-center-large-load-interconnections
[71] Federal Energy Regulatory Commission — “Interconnection of Large Loads to the Interstate Transmission System,” Docket No. RM26-4-000. https://www.ferc.gov/rm26-4
[72] Duane Morris LLP — “FERC Acts to Advance Data Center and Large Load Integration in Six RTO Regions,” June 23, 2026. https://www.duanemorris.com/alerts/ferc_acts_advance_data_center_large_load_integration_six_rto_regions_0626.html
[73] The White House — “Ratepayer Protection Pledge,” Proclamation 11014 of March 4, 2026. https://www.whitehouse.gov/releases/2026/03/ratepayer-protection-pledge/
[74] The White House — “President Trump’s Ratepayer Protection Pledge Secures American AI Dominance, Protects Consumers,” July 2026. https://www.whitehouse.gov/releases/2026/07/president-trumps-ratepayer-protection-pledge-secures-american-ai-dominance-protects-consumers/
[75] The White House — “President Trump Secures Historic Commitment to Keep Electricity Costs Down Amid Data Center Boom,” March 4, 2026. https://www.whitehouse.gov/releases/2026/03/president-trump-secures-historic-commitment-to-keep-electricity-costs-down-amid-data-center-boom
[76] U.S. Environmental Protection Agency — “President Trump Expands Historic Ratepayer Protection Pledge to Protect American Ratepayers, Lower Electricity Prices,” July 23, 2026. https://www.epa.gov/newsreleases/president-trump-expands-historic-ratepayer-protection-pledge-protect-american
[77] NOTUS — “Trump Expands His Ratepayer Pledge With Republican Governors, Utilities,” July 23, 2026. https://www.notus.org/energy/trump-expands-ratepayer-pledge-republican-governors-utilities
[78] Amelia Twyman, Georgia Recorder (via The Current) — “Kemp, GOP governors join Trump pledge to shield ratepayers from data center costs,” July 24, 2026. https://thecurrentga.org/2026/07/24/kemp-gop-governors-join-trump-pledge-to-shield-ratepayers-from-data-center-costs/
[79] Darrell M. West, The Brookings Institution — “How rising electric rates could affect the 2026 midterms,” March 20, 2026. https://www.brookings.edu/articles/how-rising-electric-rates-could-affect-the-2026-midterms/
[80] Emerson College Polling — “July 2026 National Poll: Democrats with 11-Point Generic Ballot Advantage,” July 2026. https://emersoncollegepolling.com/july-2026-national-poll-democrats-with-11-point-generic-ballot-advantage/
[81] Newsweek — “Americans Are Increasingly Turning on Data Centers,” July 2026. https://www.newsweek.com/americans-are-increasingly-turning-data-centers-12234599
[82] Pew Research Center — “Many Americans hold utility companies responsible for their rising home energy bills,” May 6, 2026. https://www.pewresearch.org/short-reads/2026/05/06/many-americans-hold-utility-companies-responsible-for-their-rising-home-energy-bills/
[83] Marcus Baram, Capital & Main (via Pennsylvania Capital-Star) — “In Pennsylvania, the data center wars hit the ballot box,” August 6, 2026. https://penncapital-star.com/energy-environment/in-pennsylvania-the-data-center-wars-hit-the-ballot-box/
[84] Certus Insights — “The Data Center Backlash: Public Opinion Has Collapsed,” August 2026. https://certusinsights.com/the-data-center-backlash-public-opinion-has-collapsed/
[85] Marc Levy and Jesse Bedayn, Associated Press — “Voters’ anger at high electricity bills and data centers looms over 2026 midterms.” https://www.aol.com/articles/voters-anger-over-high-electricity-050526352.html
[86] Ipsos and PowerLines — “Three in four concerned their gas, electricity utility bills will increase this year,” May 19, 2026. https://www.ipsos.com/en-us/three-four-concerned-their-gas-electricity-utility-bills-will-increase-year
[87] Data Center Watch (10a Labs) — “Q1 2026 Data Center Watch Report,” June 2026. https://www.datacenterwatch.org/q1-2026
[88] NBC News — “Data center opponents have blocked or delayed projects worth nearly $130 billion in 2026, study finds,” June 12, 2026. https://www.nbcnews.com/tech/tech-news/data-center-opposition-sharply-rising-2026-study-finds-rcna349728
[89] International Energy Agency — “Key Questions on Energy and AI,” Executive Summary, IEA, Paris, 2026. https://www.iea.org/reports/key-questions-on-energy-and-ai/executive-summary
[90] International Energy Agency — “Electricity 2026: Demand,” IEA, Paris, 2026. https://www.iea.org/reports/electricity-2026/demand
[91] International Energy Agency — “Electricity Mid-Year Update 2026,” Executive Summary, IEA, Paris, 2026. https://www.iea.org/reports/electricity-mid-year-update-2026/executive-summary
[92] International Energy Agency — “Energy and AI: Energy demand from AI,” IEA, Paris. https://www.iea.org/reports/energy-and-ai/energy-demand-from-ai
[93] Kristalina Georgieva, Managing Director, International Monetary Fund — “A Global Vision for Indian AI,” Remarks at the India AI Impact Summit, New Delhi, February 19, 2026. https://www.imf.org/en/news/articles/2026/02/19/sp021926-a-global-vision-for-indian-ai
[94] International Monetary Fund — *Global Financial Stability Report: Global Financial Markets Confront the War in the Middle East and Amplification Risks*, Chapter 1, Washington, DC, April 2026. https://www.imf.org/-/media/files/publications/gfsr/2026/april/english/ch1.pdf
[95] International Monetary Fund — *Global Financial Stability Report*, April 2026, Chapter 1 Online Annex. https://www.imf.org/-/media/files/publications/gfsr/2026/april/english/ch1annex.pdf
[96] International Finance Corporation, World Bank Group — “Risks and Opportunities in Data Center and AI Investment in Emerging Markets,” 2026. https://www.ifc.org/en/insights-reports/2026/risks-and-opportunities-in-data-center-and-ai-investment-in-ems
[97] World Economic Forum — “Is power grid connectivity the strategic bottleneck for AI?” May 2026. https://www.weforum.org/stories/2026/05/electricity-data-grid-connectivity-strategic-bottleneck-ai-transformation/
[98] Carbon Brief — “AI: Five charts that put data-centre energy use — and emissions — into context.” https://www.carbonbrief.org/ai-five-charts-that-put-data-centre-energy-use-and-emissions-into-context
[99] The Brookings Institution — “Global energy demands within the AI regulatory landscape,” updated April 2, 2026. https://www.brookings.edu/articles/global-energy-demands-within-the-ai-regulatory-landscape/
[100] 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
[101] UncoverAlpha — “Amazon, Google, Microsoft, Meta Q2 2026 earnings: capital expenditure guidance and de-risking playbook,” August 2026. https://www.uncoveralpha.com/p/amazon-google-microsoft-meta-q2-earnings
[102] ValueAdd — “AI Spending Tracker 2026: $725B by Big Tech,” updated August 2026 (compiled from company earnings calls and guidance through Q2 2026). https://valueaddvc.com/ai-spending
[103] U.S. Senate Committee on Banking, Housing, and Urban Affairs, Ranking Member Elizabeth Warren — “Warren, Colleagues Press FSOC to Launch Probe into Financial Stability Risks of AI Debt Bubble,” January 22, 2026. https://www.banking.senate.gov/newsroom/minority/warren-colleagues-press-fsoc-to-launch-probe-into-financial-stability-risks-of-ai-debt-bubble
[104] Emma Penrod, Utility Dive — “Large load tariffs proliferate as states take more active role in data center regulation,” March 2026. https://www.utilitydive.com/news/large-load-tariffs-proliferate-as-states-take-more-active-role-in-data-cent/816184/
[105] Sabin Center for Climate Change Law, Columbia Law School — “Data Center Regulation: What Local Governments Should Know about Large-Load Tariffs and Clean Transition Tariffs,” Climate Law Blog, June 2, 2026. https://blogs.law.columbia.edu/climatechange/2026/06/02/data-center-regulation-what-local-governments-should-know-about-large-load-tariffs-and-clean-transition-tariffs/
[106] ArentFox Schiff LLP — “State Regulation of Data Centers in 2026 – A Shifting Landscape,” April 2026. https://www.afslaw.com/perspectives/alerts/state-regulation-data-centers-2026-shifting-landscape
[107] CoBank Knowledge Exchange — “The state of large load rate design: Insights from the DELTa database,” April 2026. https://www.cobank.com/knowledge-exchange/the-state-of-large-load-rate-design-insights-from-the-delta-database
[108] The Buckeye Institute — “Undermining Ohio’s Competitive Edge,” Policy Brief, March 16, 2026. https://www.buckeyeinstitute.org/library/docLib/2026-03-16-Undermining-Ohio-s-Competitive-Edge-policy-brief.pdf
[109] Utility Dive — “Oregon PUC approves PGE’s large-load tariff framework for data centers,” May 28, 2026. https://www.utilitydive.com/news/oregon-puc-approves-pges-large-load-tariff-framework-for-data-centers/821361/
[110] Data Center Dynamics — “Rate of play: How US states are changing the rules around data centers and power,” August 2026. https://www.datacenterdynamics.com/en/analysis/rate-of-play-how-us-states-are-changing-the-rules-around-data-centers-and-power/
[111] Edison Electric Institute — “List of Large Customer Projects and Tariffs,” August 2026. https://www.eei.org/-/media/Project/EEI/Documents/Issues%20and%20Policy/List%20of%20Large%20Customer%20Projects%20and%20Tariffs
