Introduction: The Electricity Bill at the Edge of the Artificial Intelligence Revolution

In Sullivan County, Indiana, on the eastern bank of the Wabash River, a retired coal-fired power station may become the setting for one of the American artificial intelligence economy’s most consequential experiments in energy financing. The former Breed Generating Station belongs to an earlier chapter of American industrial history, the long period in which utilities built large centralized plants to serve households, factories, railways, and regional growth, and then recovered the cost of those plants over decades through regulated rates paid by everyone connected to the system. On October 19, 2026, American Electric Power and X-energy announced a joint development agreement to evaluate the construction of an advanced nuclear complex on that site, with Amazon supplying development funding and standing ready to purchase the output under a long-term contract if the project reaches a final investment decision.[1][2][3] The proposal calls for twelve Xe-100 high-temperature gas-cooled reactors, each rated at 80 megawatts and arranged in groups of four, for a combined potential capacity of 960 megawatts, which means that a property once devoted to burning coal for the general public could become part of the physical foundation of cloud computing and of increasingly demanding artificial intelligence systems.[1][2]

The announcement mattered well beyond the nuclear industry, because it suggested that the technology companies whose products depend on enormous quantities of electricity are becoming direct financial participants in the production of that electricity and are no longer content to be customers of whatever the grid happens to offer. Amazon was not asking for permission to consume existing utility resources; it was exploring a contractual structure that could justify the construction of new generating capacity that would not otherwise be built. That distinction carries real economic weight, because the American electricity system has traditionally recovered most of its infrastructure investment through regulated charges spread across a broad population of customers, and so whenever a new industrial consumer requires substantial generation or transmission, the decisive question is never only whether the utility can deliver the additional megawatts but which customers ultimately pay for the associated assets, which customers enjoy their benefits, and who absorbs the losses if the commercial assumptions that justified construction turn out to be wrong. Amazon’s representative on the project put the company’s position in a single phrase.

“pay our full costs without passing them to others”

— Josh Weissman, Vice President of Data Center Infrastructure Capacity Delivery, Amazon [1]

A separate development had already been unfolding inside the same utility system. On August 26, 2026, Indiana Michigan Power, the AEP subsidiary that owns the Breed site, announced a proposed rate plan that it said would deliver approximately $59 million in customer bill reductions during 2027, an amount that the company translated into savings of roughly $100 a year for an Indiana residential customer using 1,000 kilowatt-hours each month, together with a three-year freeze on every rate that appears on a residential bill.[4] The company did not present this as generosity, and it did not attribute it to falling fuel prices or to an accounting adjustment; it said plainly that the plan was made possible by load growth and by increased revenue from large customers, datacenters among them.[4][5] As of November 21, 2026, the proposed reductions remain before the Indiana Utility Regulatory Commission, whose decision is expected in June 2027, and they are therefore prospective savings that a regulator may approve, modify, or reject, and not reductions that any household has yet been guaranteed.[4][6]

The two announcements should not be confused with one another, and a good deal of this paper is devoted to keeping them apart. The proposed Breed nuclear facility did not cause the proposed $59 million reduction: the rate filing preceded the nuclear announcement by almost two months, and the nuclear plant has been neither approved for construction nor budgeted, and it rests on a reactor design that has not yet operated commercially anywhere in the world.[7] What the two developments share is a place in a single economic argument. The rate plan concerns how revenue from large electricity consumers can reduce the share of existing infrastructure costs borne by everyone else, while the nuclear agreement concerns whether those same consumers can finance additional generation without transferring unreasonable development and operating risk to households, and so the link between them is conceptual and institutional and should not be mistaken for an established causal relationship.

A third development placed this debate on the national legislative agenda. On November 19, 2026, the United States Senate approved the Bipartisan American Affordability and Jobs Act by a vote of 62 to 33, with roughly fifteen members of the Democratic caucus joining most Republicans.[8] The bill, introduced on September 30 by Senators Shelley Moore Capito, Mike Lee, Sheldon Whitehouse, and Martin Heinrich, couples reforms to federal permitting and transmission planning with a new framework for allocating the infrastructure costs created by large computational loads, and its Section 2107 would add a provision to the Federal Power Act requiring datacenters and high-density computing facilities of 20 megawatts or more to pay the incremental cost of the generation, storage, and delivery infrastructure built to serve them.[9][10] The bill awaits action in the House of Representatives during a lame-duck session and cannot be treated as enacted federal law, but its passage through the Senate by a margin of that size indicates how far the politics of the question have travelled in a single year.

These developments form a striking contrast with the public debate that dominated the preceding eighteen months. Communities from northern Virginia to central Ohio had grown increasingly concerned that datacenters would absorb scarce generation, drive up capacity-market prices, require expensive transmission upgrades, consume water, and leave ordinary customers paying for infrastructure designed to serve a handful of exceptionally profitable corporations, and those concerns were not imaginary: the independent market monitor for the PJM Interconnection has attributed $29.4 billion of capacity-market charges across four consecutive auctions to existing and forecast datacenter load.[11] The Indiana proposal introduces a different possibility, which is that a properly regulated large-load customer might contribute enough incremental revenue to spread existing fixed costs across a broader paying base while separately financing the infrastructure necessary to support its own operations. The Governor of Indiana, endorsing the I&M plan, stated the condition on which the whole idea depends.

“Economic growth should not come at the expense of existing customers.”

— Mike Braun, Governor of Indiana [4]

I have chosen the title Load Dividend because it names an economic possibility that the prevailing discussion of datacenters as liabilities of the electricity system has largely obscured. A large industrial customer creates more than an engineering obligation for the utility that serves it; under appropriate contractual, regulatory, and financial conditions, that customer can also produce measurable benefits for households and smaller businesses by contributing toward shared infrastructure expenses that existing customers would otherwise carry alone. The word dividend is deliberate, because a dividend is something that must be distributed, measured, and sustained before anyone is entitled to say that it exists, and a promise of future benefit does not qualify. Lower household bills, reduced base rates, stronger financial protection, and avoided future cost increases are different forms of benefit, and one of the recurring errors in public discussion is to treat them as interchangeable. The title therefore sets a demanding research question: under what conditions can artificial intelligence infrastructure create an electricity benefit that reaches beyond hyperscalers and utilities to the communities that host the physical base of what I have elsewhere called the Five-Layer AI Economy?

The proposition this paper advances can be stated in one sentence. The Load Dividend is the financial benefit that existing electricity customers receive when carefully structured contracts with new, high-volume industrial users spread fixed system costs across a larger paying customer base, while ensuring that those new users bear the incremental expenses and risks attributable to their own demand. The concept requires far more than the arrival of a large customer: it requires an identifiable economic surplus, enforceable commitments, transparent allocation of costs, independent regulatory verification, and a mechanism by which households actually receive the resulting benefit on their bills.

The analysis proceeds in seven steps. Section 1 sets out the economics of fixed-cost recovery and defines the Load Dividend mechanism with enough precision to distinguish it from its weaker relatives. Section 2 reviews the empirical research of 2020 through 2026, which has produced findings that appear contradictory until one sees which institutional settings each study is measuring. Section 3 examines Indiana as the experimental ground. Section 4 follows the federal transformation from voluntary corporate pledges to proposed statutory requirements. Section 5 compares the regulatory approaches of Pennsylvania, Virginia, and Texas. Section 6 develops a prospective investment and policy model for 2027 through 2030, and Section 7 gathers what has been learned into seven pillars before the conclusion returns to Sullivan County.


Section 1: The Economics of the Load Dividend — When Electricity Demand Becomes a Shared Financial Asset

Every argument about whether datacenters raise or lower household electricity bills is, underneath its political surface, an argument about how the fixed costs of a network are divided among the people who use it. Electricity is unusual among major commodities in that a very large share of what customers pay has little to do with the quantity of energy they consume in a given month and a great deal to do with the cost of owning, financing, and maintaining a system of wires, substations, transformers, and power plants that must exist whether or not anyone happens to be drawing on it at a particular hour. Once that is understood, the central puzzle of this paper stops looking like a paradox. A new customer who uses a great deal of electricity can make everyone else’s bill smaller if that customer pays toward costs that would have existed anyway, and the same customer can make everyone else’s bill larger if serving it requires new assets whose cost is then averaged across the whole customer base. The physics is identical in both cases; what differs is the accounting, the contract, and the regulator’s decision, which is why this section begins with economics and ends with a test that regulators can apply.


1.1. Why the Five-Layer AI Economy Begins With the Household Electricity Bill

Artificial intelligence is usually discussed through the performance of models, the benchmarks of semiconductors, the adoption of enterprise software, and the financial strategies of hyperscale technology corporations, and each of these indicators describes something real about the industry’s technological and commercial development. None of them, however, describes the physical and economic system required to sustain it, and that system begins with energy. In the framework I use throughout this research, Layer 1 provides the electricity and infrastructure required for computation; Layer 2 supplies the chips and supporting semiconductor systems; Layer 3 comprises the datacenters, electrical equipment, cooling, and networking in which computation takes place; Layer 4 consists of foundation models and related intelligence systems; and Layer 5 encompasses the applications, autonomous agents, robotics, and software services through which artificial intelligence produces economic value. Each layer depends on the reliability and affordability of the layers beneath it, and the dependence is most consequential at the boundary between energy and datacenters, because semiconductor manufacturers may deliver ever more efficient accelerators and hyperscalers may deploy racks of extraordinary density, yet higher aggregate deployment can still increase total demand even as each processor improves in performance per watt.

The scale of what is being asked of Layer 1 is now documented by institutions that have no commercial interest in exaggerating it. The International Energy Agency estimated that datacenters consumed roughly 415 terawatt-hours in 2024, about 1.5 percent of global electricity, and projected that consumption would more than double to around 945 terawatt-hours by 2030, with the United States accounting for the largest share of the increase and with datacenters responsible for nearly half of American demand growth over the period.[12][13] Its 2026 update reported that global datacenter consumption grew 17 percent in 2025 to about 485 terawatt-hours and kept the 2030 projection near 950.[14] The Agency’s executive director framed the stakes in language that energy ministries have since repeated.

“AI is one of the biggest stories in the energy world today”

— Fatih Birol, Executive Director, International Energy Agency [12]

For the United States specifically, the June 2026 update of Lawrence Berkeley National Laboratory’s datacenter energy report, prepared for the Department of Energy by Arman Shehabi, Sarah Smith, and colleagues, places 2030 consumption at 649 terawatt-hours in its reference case, equivalent to 11.8 percent of forecast national electricity use, within a scenario range of 9.5 to 15.3 percent.[15] An industry approaching an eighth of national consumption does not purchase electricity at the margin of the system in the manner of an ordinary customer; it reshapes generation investment, transmission planning, equipment procurement, and market prices, and the economic consequences therefore extend well beyond the corporations buying the computing equipment. The International Monetary Fund, examining the same phenomenon from the perspective of macroeconomic stability, reached a conclusion that explains why finance ministries have joined energy regulators in paying attention.

“The resulting strain on power grids has major implications for global electricity demand.”

— International Monetary Fund, IMF Blog [16]

When a utility invests in new substations, transmission lines, transformers, generating facilities, and control systems, some of those expenditures are dedicated to particular customers and others strengthen the shared network, and the central analytical challenge is to distinguish the benefits and costs that belong in each category. A transmission project built for one datacenter may also improve regional service for other users, while a project presented as a general reliability improvement may have become necessary mainly because one extraordinary load was added to the system. Neither situation can be understood from consumption figures alone, since both require cost-of-service analysis, engineering studies, transparent contracts, regulatory accounting, and a careful judgment about who would have paid for the investment had the datacenter never arrived. This is the reason the Load Dividend belongs within the Five-Layer AI Economy and not solely within conventional utility finance: the development of artificial intelligence changes the scale, speed, and concentration of electricity demand, and the allocation of electricity-system costs determines how the economic consequences of that expansion are distributed among people who will never write a line of code or own a share of a hyperscaler.


1.2. The Fixed-Cost Structure of Modern Electricity Systems

A utility’s expenses cannot be understood by multiplying the electricity it sells by the cost of generating each kilowatt-hour, because electricity service consists of several distinct categories of expenditure that behave differently as sales change. Generation requires construction, fuel, maintenance, staffing, financing, insurance, and eventual retirement; transmission requires towers, conductors, transformers, substations, land rights, and operating reserves; and distribution connects these higher-voltage systems to individual homes, schools, hospitals, and businesses. Some of these costs move with consumption, but many do not, at least over the horizon that matters for a monthly bill: a distribution feeder is no cheaper to own in a month when customers use less, a substation needs inspection whether or not demand fluctuates, and the utility continues to recover depreciation and an authorized return on prudent capital through regulated rates regardless of how many kilowatt-hours pass through its meters.

Severin Borenstein and James Bushnell of the University of California established the academic foundation for this point in a 2022 article showing that American retail prices recover a large volume of fixed and residual cost through volumetric charges, so that the price of an additional kilowatt-hour to a household routinely exceeds the private marginal cost of supplying it.[17] Christopher Knittel of MIT returned to the theme in a June 2026 research commentary arguing that tariffs designed for an era of one-way power flows and analog meters have become steadily more inefficient and regressive now that rooftop solar, batteries, electric vehicles, and datacenters are changing when and where load appears, and that the efficient structure would price energy by time, charge for capacity according to contribution to system peaks, and recover residual costs through transparent fixed charges.[18] The implication for this paper follows directly, and it is the hinge on which the whole argument turns: when fixed costs are embedded in per-kilowatt-hour rates, any increase in total sales that is not accompanied by a proportional increase in fixed costs reduces the fixed-cost burden carried by each unit sold.

An additional large customer therefore produces two financial effects at once. It increases the revenue available to the utility, and it may create additional expenses required to serve it, and the net consequence for everyone else depends entirely on the relationship between the two. Suppose a utility already has generation and network capacity available in most hours, as nearly every American utility does, given that the Duke University study led by Tyler Norris found the nation’s balancing authorities operating at an average load factor of roughly 53 percent.[19] A new industrial customer that uses some of this existing capacity contributes toward infrastructure costs that were previously allocated almost entirely to incumbent customers, and if the additional revenue exceeds the avoidable cost of supplying it, a surplus becomes available for allocation. Now consider the opposite situation, in which a large datacenter requires a dedicated substation, significant transmission reinforcement, new generating capacity, and expensive reliability services, and pays for its electricity under a tariff that does not adequately reflect those new costs. Its presence then increases total system expenses by more than its incremental revenue covers, and what was advertised as an economic benefit turns out to be a subsidy flowing from households to one of the largest corporations in the world.

The difference between these outcomes is fundamental, and it is the first discipline that the Load Dividend framework imposes. A load increase does not automatically create a dividend; it creates the opportunity for one, and that opportunity is realized only when the customer’s contribution exceeds the fully attributed incremental costs and risks that its service imposes on the system. Ryan Wiser of Lawrence Berkeley National Laboratory, whose team has produced the most comprehensive recent analysis of American retail price trends, compressed the point into the title of a December 2025 presentation to state policymakers.

“It could go either way: it’s not preordained!”

— Ryan Wiser, Senior Scientist, Lawrence Berkeley National Laboratory [20]


1.3. Defining the Load Dividend Mechanism

The framework can be expressed as a simplified economic identity, which I offer as an analytical device and not as a substitute for the rate calculations that utility regulation requires.


Potential Load Dividend = Incremental Customer Revenue − Fully Attributable Incremental Costs − Required Risk Protection − Retained Utility Obligations


Incremental customer revenue comprises the payments a new large user makes through energy charges, demand charges, minimum-billing obligations, dedicated infrastructure contributions, and other contractual arrangements. Fully attributable incremental costs comprise the generation, transmission, distribution, storage, capacity, operating, and maintenance expenses that would have been avoided had the load not materialized. Required risk protection accounts for exposure to customer default, early termination, construction delay, uncertain load growth, and the possibility that expensive assets remain in service long after the customer for whom they were built has departed. Retained utility obligations recognize that a utility may need to preserve reserves, maintain service quality, finance replacement assets, and satisfy lawful financial requirements before any surplus is distributed. The amount that remains must then pass through the applicable regulatory framework before it can become a verified customer benefit, because a surplus that stays on the utility’s income statement is a gain to shareholders and nothing more.

The value of the identity lies in the separation it forces among three outcomes that public discussion routinely merges, and which Table 1 sets side by side. The first is cost neutrality, in which a large customer pays the full incremental cost of its service and leaves other customers no worse off than they would have been in its absence. The second is cost avoidance, in which a large customer finances infrastructure that would otherwise have generated costs for existing customers, or removes the need for some alternative expenditure. The third is positive dividend distribution, in which an identifiable benefit is transferred to existing customers through reduced rates, bill credits, or another verified mechanism. Only the third is the strong form of the Load Dividend. Cost neutrality is a necessary protection, but it does not establish that any household bill has fallen, and the distinction matters because a utility, a hyperscaler, or an elected official may describe a large-load agreement as beneficial when the agreement does no more than prevent certain costs from being shifted onto households. Preventing harm is valuable, and it should be recognized as such, but it is a different thing from creating a financial return.


Table 1. Three outcomes that are routinely conflated, and the evidence each requires

OutcomeWhat it meansEvidence required2026 example
Cost neutralityThe new customer pays the full incremental cost of its service, leaving other customers no worse off than without it.Cost-of-service study; published large-load tariff; minimum bills and security.Texas large-load interconnection rule; Virginia GS-5 minimum-demand charges.[21][22]
Cost avoidanceThe new customer funds infrastructure, or removes a cost, that existing customers would otherwise have carried.A documented prior allocation and a regulator’s order changing it.Virginia transmission order of August 2026, cutting the projected residential increase from $2.90 to $0.94 a month.[23]
Positive dividendAn identifiable surplus is transferred to existing customers through lower rates or bill credits.An approved rate reduction or credit, followed by a realized change in household bills.Indiana Michigan Power’s proposed $59 million reduction for 2027, pending a decision expected in June 2027.[4]

1.4. The Counterfactual Test

Every credible Load Dividend analysis rests on a counterfactual, which is to say on a disciplined answer to the question of what would have happened without the new customer. The comparison begins with two scenarios: a baseline describing the system without the proposed datacenter, including expected household consumption, commercial and industrial demand, generation retirements, fuel prices, planned grid investment, financing costs, and ordinary reliability requirements; and a large-load scenario that adds the datacenter’s expected consumption together with the investments, contractual payments, and operational changes required to serve it. The difference between the two reveals the incremental economic impact, but only if several complications are handled honestly. An investment may serve multiple purposes, as when a new transmission line accommodates datacenter growth while also relieving congestion that predated it, and the cost-allocation analysis must then avoid charging the datacenter for benefits that belong mainly to others while equally avoiding the transfer of datacenter-specific costs to households. A lower residential rate may also reflect conditions that have nothing to do with large-load growth, among them reduced fuel expense, changes in depreciation, regulatory settlements, or revisions to the overall revenue requirement, and the analyst must isolate the contribution of large customers instead of claiming every observed saving as an artificial intelligence dividend.

The Indiana case makes this requirement especially important. Indiana Michigan Power explicitly identified large-customer revenue as a contributor to its proposed reduction, which supplies a relevant causal mechanism, but that statement does not mean that every dollar of the $59 million can be attributed to datacenters, and it does not establish that the later nuclear feasibility agreement had any part in generating those savings. A rigorous framework should therefore distinguish between the existence of a plausible mechanism and the verified measurement of its financial contribution, and Section 3 will argue that the pending rate case is valuable chiefly because it obliges a regulator to perform exactly that measurement in public.


1.5. Why Artificial Intelligence Changes the Conventional Industrial-Load Calculation

Large industrial electricity customers are not a novelty, since steel mills, aluminum smelters, chemical plants, and automotive factories have consumed substantial quantities of power for a century and have long been served under special industrial tariffs. The artificial intelligence economy nevertheless introduces five characteristics that the conventional analysis of industrial load was not designed to handle. Datacenter growth is exceptionally concentrated, so that a single campus may request hundreds of megawatts and a cluster of campuses may add more demand to a region in five years than the region added in the previous twenty; Indiana Michigan Power told its regulator that it expected peak load in its Indiana territory to rise from about 2,800 megawatts to more than 7,000 megawatts by 2030.[24] Deployment expectations can change rapidly, because announced demand depends on future financing, on the availability of advanced accelerators, on shifting model architectures, and on contracts that remain commercially contingent. Datacenters also combine long-lived electrical infrastructure with short-lived information technology, since a transformer has an operating life measured in decades while a generation of accelerators may be commercially superseded within a few years.

Operators differ, moreover, in their technical requirements, because an inference facility serving interactive applications has stringent latency and availability needs while a training cluster may be able to shift some workloads in time, subject to engineering and commercial constraints that vary from one operator to the next. Finally, and most importantly for the argument of this paper, the developers of these facilities command financial resources far beyond those of the communities in which they build. The third-quarter 2026 earnings season made the asymmetry concrete: Amazon reported capital expenditures of $77.3 billion in a single quarter, more than double the $35.1 billion of a year earlier, and held its 2026 guidance at roughly $220 billion, while one tabulation drawn from quarterly filings put the combined third-quarter spending of Amazon, Alphabet, Microsoft, and Meta on property and equipment at $201.9 billion, an increase of 108 percent over the same quarter of 2025.[25][26] Companies operating at that scale can negotiate long-duration energy agreements, post financial guarantees, fund generating assets, and absorb development risks that no smelter or assembly plant could have accepted, and that capacity creates an opening for new contractual arrangements at the same moment that it magnifies the consequences of inadequate cost allocation.

A utility that builds on speculative datacenter commitments exposes its existing customers to substantial stranded-asset risk, whereas a utility that obtains sufficiently strong commitments from financially capable customers may be able to distribute costs more efficiently while improving reliability for everyone. The result is an economic question that conventional industrial-development analysis cannot answer on its own terms: can the financial scale of artificial intelligence be converted into a durable public electricity benefit without requiring households to underwrite the industry’s uncertainties? That is the central question of the Load Dividend, and the next section asks what the evidence gathered so far can and cannot say about it.


Section 2: What the Evidence Shows — The Research Record From 2020 to 2026

The theoretical ambiguity described in the previous section is not a gap that better theory could close, because the theory is complete and it says, correctly, that the answer depends on the facts. What has changed over the past two years is the quantity and quality of the facts. Between early 2025 and November 2026, economists and engineers at national laboratories, research institutes, business schools, law schools, and a regional Federal Reserve bank published a remarkable sequence of empirical studies on the relationship between datacenter growth and electricity prices, and a reader who took them one at a time could be forgiven for concluding that the profession had reached no conclusion at all, since some of these studies find that datacenters lowered residential rates, others find that they raised them, and still others find that the effect is too small to matter. I want to argue in this section that the studies are less contradictory than they appear, and that when they are arranged according to what each one measures, in which institutional setting, and over which period, they converge on a proposition that is close to the thesis of this paper: the sign of the effect is determined by tariff design, market structure, and spare capacity, and can be chosen by policy.


2.1. The Studies That Find Load Growth Lowering Rates

The most influential body of evidence on the favorable side comes from Lawrence Berkeley National Laboratory and The Brattle Group. Ryan Wiser, Galen Barbose, Peter Cappers, and their colleagues examined state-level retail prices from 2019 through 2024 and found that the principal drivers of price increases were replacement of aging distribution and transmission equipment, storm and wildfire costs, natural gas price volatility, and state policy mandates, and that load growth was not among them; indeed, the states with the highest load growth tended to experience declining inflation-adjusted prices, while states with contracting sales generally saw prices rise.[27] Their 2026 edition extended the analysis through 2025 and reported that, adjusted for inflation, average American retail prices in 2025 stood only about three percent above their 2019 level and below their level of 2010, a finding that sits uncomfortably beside the popular narrative of a datacenter-driven affordability crisis.[27] A Columbia University white paper by Robin Millican, Douglas Arent, and David Sandalow, published in June 2026, reached a compatible result, observing that the states with the fastest demand growth from 2019 to 2025 generally saw all-sector real prices fall by more than one cent per kilowatt-hour as fixed infrastructure costs were spread over a larger base of consumption.[28]

Correlation of this kind invites an obvious objection, which is that datacenters choose to locate where electricity is already cheap and capacity already abundant, so that low prices might be the cause of datacenter growth and not its consequence. The Electric Power Research Institute addressed the objection directly. A January 2026 white paper by John Bistline set out the conditions under which new load lowers prices, namely that the long-run incremental cost of serving it must lie below the system’s current average cost, and summarized its answer to the title question in a sentence that deserves to be quoted because of its careful modesty.[29]

“Despite widespread concerns, the answer is sometimes yes.”

— John Bistline, Electric Power Research Institute [29]

A companion study by Asa Watten, Bistline, and Geoff Blanford then used an instrumental-variable strategy to control for location selection and estimated that a doubling of datacenter capacity caused residential retail prices to fall by roughly four percent, with the average effect of datacenter growth between 2019 and 2024 being a reduction in average rates of about six percent.[30] The authors explained the mechanism in the terms this paper has already used, as durable new demand spreading sales-independent costs over more kilowatt-hours and drawing in newer and cheaper capacity. From the University of Southern California, Angela Ryu and Shon Hiatt of the Marshall School’s Zage Business of Energy Initiative traced datacenter entry from the wholesale node to the retail bill and estimated congestion effects of roughly $1.45 to $2.30 per megawatt-hour at the local node, a retail impact below one dollar a month for a typical household, and no statistically significant effect among the investor-owned utilities that serve about two-thirds of American customers.[31]


2.2. The Studies That Find Datacenters Raising Rates

The contrary evidence is equally serious and comes from equally credible institutions. Robyn Meeks of Duke, Jacquelyn Pless of MIT Sloan, and their coauthors Zhiyuan Qi and Zhenxuan Wang linked facility-level data on datacenter entry between 2010 and 2024 to utility prices, investments, and regulatory outcomes, and found that entry raised average retail rates by 2.7 percent and residential rates by 2.1 percent, with the effect concentrated among investor-owned utilities, where average prices rose 5.6 percent, and in restructured states, where residential prices rose about four percent; in traditionally regulated states the estimated increase was small and statistically insignificant, and cooperatives showed no systematic response at all.[32] In household terms the average residential effect amounted to roughly three dollars a month, and about seven dollars a month for customers of investor-owned utilities. The authors summarized the pattern in a sentence that is, in effect, a statement of the research agenda this paper pursues.

“Data center entry increases retail electricity rates, but the burden is not distributed uniformly.”

— Robyn Meeks, Jacquelyn Pless, Zhiyuan Qi, and Zhenxuan Wang, MIT Center for Energy and Environmental Policy Research [32]

At the wholesale level, Owen Kay, Robert Reaser, and Reid Taylor of the Federal Reserve Bank of Dallas built an hourly, unit-level dispatch model and estimated that datacenter demand had lifted the load-weighted national wholesale price by 6.9 percent by 2025, an effect they project to reach 12.3 percent by 2028, with increases near 15 percent in northern Virginia and the Pacific Northwest and with spillovers into neighboring regions that host few datacenters of their own.[33] Perhaps their most consequential finding for the present argument is distributional: of an estimated $186.7 billion in additional payments to generators between 2021 and 2028, roughly half represents inframarginal rent, that is, higher prices paid to plants that would have operated anyway, which is a transfer from all electricity buyers to existing generation owners that no datacenter tariff can recover. The International Monetary Fund’s modeling, published in its April 2025 working paper by Christian Bogmans and colleagues, pointed in the same direction, documenting that electricity costs for vertically integrated artificial intelligence firms nearly doubled between 2019 and 2023 and simulating meaningful upward pressure on American electricity prices when supply cannot respond quickly to demand.[34]

Between these poles sits a September 2026 working paper by Vishrut Rana, Christian Kaps, and Serguei Netessine of Harvard Business School and the Wharton School, which analyzed 385 datacenters and the 42 utilities serving them from 2020 to 2025 and concluded that the facilities raised regional residential prices by an average of about $1.70 per megawatt-hour, accounting for roughly four percent of the observed residential increase over the period, and industrial prices by about $5.40 per megawatt-hour, accounting for roughly fifteen percent of the industrial increase.[35] Their result is a useful corrective in both directions: datacenters did raise prices in their sample, and they explain only a small fraction of what households experienced, the rest being attributable to inflation, fuel, and grid investment. Liang Min, who directs the Bits and Watts Initiative at Stanford University, made the same point about the composition of rising bills when asked whether making datacenters pay their own way would bring prices down.

“Bills would go up even without data centers”

— Liang Min, Managing Director, Bits and Watts Initiative, Stanford University [36]

The World Bank, surveying the question for developing economies in its World Development Report 2026, sided with the more cautious reading of the evidence, warning that where grids are shared and institutions are weak, households tend to

“bear a substantial proportion of the costs incurred to serve these data centers”

— World Bank, World Development Report 2026: The Promise of Artificial Intelligence [37]

and the United Nations Secretary-General, speaking in July 2025 about the collision between artificial intelligence and energy systems, offered a formulation that applies as well to affordability as to the emissions he had in mind.

“This is not sustainable — unless we make it so.”

— António Guterres, Secretary-General of the United Nations [38]


2.3. Reconciling the Record: Institutions Decide the Sign

Table 2 arranges the principal studies by what they measure, and once they are arranged in this way the apparent contradiction largely dissolves. The studies that find falling prices are measuring average retail prices at the state level, in a period when most datacenter load was being absorbed by systems with spare capacity and when the dominant population of datacenters in the sample was served by vertically integrated utilities under cost-of-service regulation. The studies that find rising prices are measuring either wholesale markets, where additional demand moves prices along a steep short-run supply curve and every buyer pays the higher clearing price, or retail prices in restructured states, where those wholesale movements pass through to households with little regulatory buffer. The MIT and Duke result that regulated states show no significant residential increase while restructured states show an increase of about four percent is therefore fully consistent with the EPRI and Berkeley Lab result that load growth has tended to lower average rates, because each set of authors is describing a different institutional channel.


Table 2. Principal empirical studies of datacenters and electricity prices, 2025–2026

StudyInstitutionWhat it measuresPrincipal finding
Wiser et al.[27]Lawrence Berkeley National Laboratory; BrattleState-level retail prices, 2019–2025States with the highest load growth saw real prices fall; price increases were driven by grid replacement, storms, wildfire, and fuel.
Watten, Bistline, and Blanford[30]Electric Power Research InstituteRetail rates, 2015–2024, instrumental-variable designA doubling of datacenter capacity lowered residential prices by about 4 percent.
Ryu and Hiatt[31]University of Southern CaliforniaWholesale node to retail billRetail effect under $1 a month; no significant effect at investor-owned utilities.
Rana, Kaps, and Netessine[35]Harvard Business School; Wharton385 datacenters, 42 utilities, 2020–2025Datacenters explain about 4 percent of the residential price rise and about 15 percent of the industrial rise.
Meeks, Pless, Qi, and Wang[32]Duke; MIT; North Carolina StateDatacenter entry and utility prices, 2010–2024Residential rates up 2.1 percent on average; 5.2 percent at investor-owned utilities; no significant effect in regulated states.
Kay, Reaser, and Taylor[33]Federal Reserve Bank of DallasHourly wholesale dispatch modelWholesale prices up 6.9 percent nationally by 2025, projected 12.3 percent by 2028.
Monitoring Analytics[11]Independent Market Monitor for PJMFour PJM capacity auctions$29.4 billion, or 46.2 percent of capacity charges, attributed to datacenter load.
Martin and Peskoe[39]Harvard Law SchoolNearly 50 regulatory proceedings in 10 statesSpecial contracts and socialized transmission costs can shift datacenter costs to households.

The legal scholarship supplies the remaining piece of the reconciliation. Eliza Martin and Ari Peskoe of Harvard Law School’s Electricity Law Initiative reviewed nearly fifty regulatory proceedings across ten states and showed that the mechanisms by which datacenter costs can reach households are embedded in ordinary ratemaking: special contracts negotiated in confidence and reviewed through opaque processes, cost-allocation formulas that lag behind rapid changes in the customer mix, and transmission charges that are socialized across all customers even when a specific customer triggered the investment.[39] Their paper is frequently read as an indictment of datacenters, but it is better understood as an indictment of rate design, and its recommendations, which include published large-load tariffs in place of secret contracts and direct assignment of customer-specific costs, are exactly the features that distinguish the Indiana and Virginia frameworks discussed later in this paper. Peskoe’s summary of the status quo before those reforms was blunt.

“We’re all paying for the energy costs of the world’s wealthiest corporations”

— Ari Peskoe, Director, Electricity Law Initiative, Harvard Law School [40]

Severin Borenstein, faculty director of the Energy Institute at the Haas School of Business at Berkeley, has put the opposing possibility with equal economy. Speaking to National Public Radio in November 2025, he observed that new datacenters ought in principle to cover their own power costs and somewhat more, which would lower bills for nearby residential customers, and he added that the obstacle is political and not technical, since regulators and legislators are often tempted to discount the rate in the name of economic development.[41]

“That doesn’t seem that hard to do”

— Severin Borenstein, Faculty Director, Energy Institute at Haas, University of California, Berkeley [41]

The engineering literature adds one more variable, which is the shape of the load and not merely its size. The Duke study led by Tyler Norris, with Tim Profeta, Dalia Patino-Echeverri, and Adam Cowie-Haskell, estimated that the existing American power system could integrate 76 gigawatts of new load if that load accepted curtailment in 0.25 percent of its hours, 98 gigawatts at 0.5 percent, and 126 gigawatts at one percent, because systems built to meet rare peaks carry unused capacity through most of the year.[19]

“could accommodate significant load additions with modest flexibility measures”

— Tyler Norris, James B. Duke Fellow, Nicholas School of the Environment, Duke University [19]

A 2026 study by Senga, Wang, and Knittel of MIT quantified the corresponding savings, finding that flexible operation of datacenters, compared with inflexible operation, would reduce power-system costs by as much as five percent in Texas, four percent in the Mid-Atlantic, and two percent in the western states, while cautioning that shifting load toward cheaper hours can raise emissions where the cheap marginal generator is fossil-fired.[42] Knittel described the two decisions that determine how much of this value a given facility can deliver.

“One is how much of their load in any one time period is flexible.”

— Christopher Knittel, George P. Shultz Professor, MIT Sloan School of Management [42]

Three conclusions follow from the research record, and they organize the remainder of this paper. A positive Load Dividend is empirically possible, and may even have been the average outcome through 2024, in regulated systems with spare capacity. The same demand growth has demonstrably raised costs through wholesale energy and capacity markets, and that channel is beyond the reach of any retail tariff. And every one of these studies describes a period that ended before the largest campuses now under contract began drawing power, which means that the historical average is a poor guide to the future and that the design of contracts being signed today will determine which body of evidence the next decade resembles. A Stanford assessment published in June 2026 by Michael Wara and Michael Mastrandrea made the forward-looking version of the argument, suggesting that if datacenter demand arrives as projected it could give utilities the opportunity to modernize aging infrastructure and add generation with the financial support of the technology industry, provided that the rules are written to capture that support.[43]


Section 3: Indiana as the Experimental Ground — Large-Load Tariffs, Household Savings, and Nuclear Investment

If the research record tells us that institutions determine whether load growth helps or harms households, then the most useful thing a researcher can do is to find a jurisdiction in which the institutions have been deliberately redesigned and observe what follows. Indiana is such a jurisdiction, and it is useful for a reason that has little to do with the size of its datacenter market, which is smaller than Virginia’s or Texas’s. Its value lies in the fact that within a span of roughly twenty months, a single utility operating under a single regulator produced three separable events that correspond to the three terms of the Load Dividend identity: an approved tariff that converts load forecasts into binding financial obligations, a proposed rate plan that would pass a surplus through to households, and an exploratory agreement under which a hyperscaler would help finance new generation. The first is a regulatory mechanism, the second is a proposed financial outcome, and the third is a potential future infrastructure investment, and each must be evaluated on its own evidence before anyone considers how they might fit together.


3.1. Indiana’s Convergence of Industrial Demand and Electricity Affordability

Indiana occupies a consequential position in the changing American energy economy, with an industrial history built on steel, automobiles, agriculture, and electricity-intensive manufacturing, and with a utility sector that sits inside regional power markets while remaining vertically integrated and subject to cost-of-service regulation. That combination places it squarely in the institutional category for which the MIT and Duke study found no significant residential price increase from datacenter entry. The state has also been the site of some of the largest datacenter announcements in the Midwest, including an Amazon Web Services campus near New Carlisle announced in 2024 at $11 billion, a $2 billion Google facility in Fort Wayne, and a Microsoft project in LaPorte, all of them in the service territory of Indiana Michigan Power, which serves more than 600,000 customers across Indiana and Michigan.[44][24]

During 2026, Governor Mike Braun’s administration placed electricity affordability at the center of its agenda. House Enrolled Act 1002, which the Governor signed on February 26, restructures Indiana rate cases around multi-year rate plans with performance incentive mechanisms tied to affordability and reliability, adds protections against summer disconnection, and requires utilities to enroll eligible low-income customers in levelized billing and assistance programs, with the first required multi-year filings due in 2029.[45] The law does not mention datacenters in its title, but it changes the setting in which datacenter revenue is treated, because a utility whose authorized return now depends in part on affordability metrics has a direct financial reason to show that load growth is lowering bills. Skeptics in Indiana are numerous and well organized, and their skepticism is part of the evidence. The Citizens Action Coalition, the state’s principal consumer advocate, called in 2024 for a moratorium on hyperscale facilities and has continued to track proposals across the state, and its program director described the pace of development in terms that explain why affordability became a political priority.

“Indiana is being inundated with AI data center proposals”

— Ben Inskeep, Program Director, Citizens Action Coalition of Indiana [46]

States competing for datacenters usually emphasize construction spending, technology investment, and new taxable property, but electricity infrastructure raises a different set of considerations, since a major campus may require hundreds of millions of dollars in electrical equipment and network upgrades without creating a workforce remotely proportionate to its energy consumption. Indiana’s emerging approach places the electricity contract at the center of that discussion, so that the relevant question is no longer whether datacenter investment is beneficial in general and becomes instead whether these particular customers make contributions adequate to protect, and if possible to benefit, the customers who were there first. Table 3 summarizes the three components of the Indiana case and the status of each as of this writing.


Table 3. The three components of the Indiana case, as of November 21, 2026

ComponentStatusKey termsWhat it establishes
Large-load tariff (Indiana Michigan Power)Approved by the IURC, February 2025.[47]Applies at 70 MW per site or 150 MW in aggregate; 12-year minimum term after a ramp of up to 5 years; 80 percent minimum billing demand; collateral; regulatory review of reductions above 20 percent.[44]A regulatory mechanism that converts load forecasts into binding obligations.
Customer Benefits Plan (rate reduction)Proposed August 26, 2026; decision expected June 2027.[4][6]About $59 million in 2027 bill reductions; about $100 a year at 1,000 kWh a month; three-year freeze on residential rates.A proposed financial outcome: a surplus passed through to households, if approved.
Breed nuclear joint development agreementFeasibility stage; announced October 19, 2026.[1][2]Twelve Xe-100 reactors, 960 MW; Amazon development funding and intended long-term purchase; no schedule or budget disclosed.[7]A potential future investment whose risk allocation is not yet contractually defined.

3.2. Indiana Michigan Power’s Large-Load Tariff

The foundation was laid in November 2024, when Indiana Michigan Power filed a settlement negotiated with Amazon Web Services, Google, Microsoft, the Data Center Coalition, the Indiana Office of Utility Consumer Counselor, and the Citizens Action Coalition, an unusual alliance in which the hyperscalers and their most persistent critics signed the same document.[44] The Indiana Utility Regulatory Commission approved the settlement in February 2025 with a single modification.[47] The resulting terms apply to new or expanded facilities with contract capacity of at least 70 megawatts at a single site or 150 megawatts aggregated across a company, and they require an initial contract of at least twelve years, which may follow a load-ramp period of up to five years; a monthly billing demand that cannot fall below 80 percent of contract capacity, an increase from the 60 percent floor that applied previously; collateral requirements scaled to the customer’s expected bills; and exit provisions that limit fee-free reductions in contract capacity to 20 percent after the fifth year on 42 months’ notice.[44] The Commission’s modification added that any mutually agreed reduction exceeding 20 percent of a customer’s contracted peak capacity must be submitted for regulatory review before it takes effect, which closes the door through which a utility and a large customer might otherwise have quietly renegotiated obligations on which other customers were relying.[47]

These provisions are central to the Load Dividend thesis because of the order in which money moves in the utility business. A utility generally commits capital before it receives the revenue expected from a new customer, and transformers, substations, transmission facilities, and generating units cannot be purchased in small increments as consumption rises. Without contractual protection, the utility could build to a datacenter’s forecast and later discover that the customer had delayed construction, reduced its requirements, or cancelled, in which case the cost of the unused assets would fall, by default, on everyone else. A minimum-billing arrangement changes the economics by obliging the customer to pay for a contracted level of service whether or not it uses it; a long contract term extends that obligation across a meaningful portion of the financing period; collateral protects against default; and exit provisions address the risk that infrastructure outlives the customer’s interest in it. The Citizens Action Coalition’s analyst estimated at the time that under the settlement’s formula a representative large customer would owe a minimum of about $332 million over the contract regardless of usage, compared with roughly $173 million under the previous terms.[44]

Taken together, these mechanisms transform a forecast of electricity demand into a financial commitment that a regulator can rely on, although they do not eliminate every risk. The credit quality of the counterparty still matters, asset lives can exceed twelve years, and some categories of cost fall outside the minimum bill. The economic significance of the tariff does not lie in the existence of a twelve-year contract as such; it lies in the transfer of a substantial portion of demand risk away from existing customers and toward the party that asked for the infrastructure. AEP has since applied the same logic across its system, and its third-quarter 2026 disclosures show the scale at which it now operates: incremental contracted load of 84 gigawatts through 2031, up from 69 gigawatts a quarter earlier, of which 35 gigawatts in PJM and the Southwest Power Pool are backed by firm take-or-pay electric service agreements, supporting a five-year capital plan of $100 billion.[48] The company’s chairman described the principle on which those agreements are written.

“new large-load customers bear the incremental system costs they create”

— Bill Fehrman, Chairman, President and Chief Executive Officer, American Electric Power [48]


3.3. The Proposed $59 Million Customer Benefit

On August 26, 2026, Indiana Michigan Power announced what it called one of the nation’s largest base-rate reduction plans. The company stated that it expected approximately $59 million in customer bill reductions in 2027, that a residential customer using 1,000 kilowatt-hours a month would save an estimated $100 a year, and that it would freeze all rates on residential monthly bills for three consecutive years, and it noted that its Indiana rates had already been reduced twice in the first six months of 2026.[4] Company representatives have described the proposal as a reduction of roughly five percent in the base-rate portion of the residential bill, with smaller reductions for business customers.[5] The explanation the utility offered is an unusually clear statement of the mechanism this paper describes: when large customers contribute toward fixed infrastructure costs, a smaller share of those costs needs to be recovered from everyone else. At the parent-company level, AEP told investors in October that take-or-pay agreements in its vertically integrated states are now expected to generate up to $21 billion in fixed-cost offsets for existing residential customers over the lives of the contracts, an estimate raised from $16 billion only three months earlier.[48]

The proposal therefore represents more than a promise that datacenters will pay for themselves, since it would transfer part of the financial benefit of customer growth into lower bills, and that step is what separates cost neutrality from a dividend. A utility could collect significant new revenue from hyperscalers without reducing residential charges at all, because the disposition of the revenue depends on tariffs, cost-of-service calculations, and approved rate design. The reaction in the service territory captured the political appeal of the idea, and the mayor of South Bend gave the proposal a description that could serve as an epigraph for this paper.

“the power of growth and getting more for less”

— James Mueller, Mayor of South Bend, Indiana [49]

Several qualifications remain necessary, and the local press has stated them as well as any economist could. The $59 million figure describes proposed bill reductions and is not a separately audited measure of surplus created by artificial intelligence consumption. The $100 figure describes a representative customer at a specified level of usage and should not be read as a uniform payment to every household. And the reduction is not final, because the Commission is expected to rule in June 2027, with any approved change reaching bills that summer.[4][6] An editorial in the Fort Wayne Journal Gazette, published ahead of a public field hearing in November 2026, welcomed the prospect of savings and then asked the questions on which the durability of the dividend depends: how well the plan holds up if a large customer opens late or uses less power than expected, and who pays for the infrastructure in that event.[50] Those are the questions the counterfactual test of Section 1 requires, and the fact that they are being asked in a docketed proceeding, with intervenors and discovery, is the reason a scholarly evaluation should treat the announced amount as a proposed regulatory outcome whose final form, billing implementation, and subsequent performance will supply the stronger evidence.


3.4. The Difference Between a Base-Rate Reduction and a Total-Bill Reduction

The structure of an electricity bill matters a great deal when claims about affordability are evaluated, because household charges include energy costs, fuel adjustments, generation and transmission charges, distribution costs, taxes, and a variety of approved riders, and a reduction in the base rate does not guarantee that every component of the total bill will decline. Fuel prices may rise, purchased-power expenses may change, weather may alter consumption, and new riders may offset part of the improvement. The national backdrop makes the caution concrete: the Energy Information Administration’s November 2026 outlook places the average American residential price at 18.17 cents per kilowatt-hour for 2026, up from 17.24 cents in 2025 and 16.48 cents in 2024, with a further rise to 18.68 cents forecast for 2027.[51] A household in a state where base rates fall can still experience a higher bill if the other components move against it, which is one reason the three-year freeze on all residential rates in the I&M proposal is arguably as significant as the headline reduction.

Consider a hypothetical residential customer whose annual electricity expense is $1,800. If approved base-rate changes reduce that customer’s annual charges by $100, the benefit is real, but if fuel-related adjustments raise the same customer’s costs by $75 over the same period, the observable net reduction is only $25, all else being equal. The example is illustrative and is not a forecast of Indiana rates, and its purpose is to show why analysts should distinguish the performance of a specific regulatory mechanism from broader movements in electricity costs. The strongest evidence would consist of standardized comparisons at consistent annual usage, with transparent bill components and actual charges, and it would report how benefits are distributed among renters, low-income customers, small businesses, and households with different consumption patterns, since an average reduction can conceal large differences among groups and a regulatory framework concerned with public benefit should disclose those differences instead of presenting one representative number as universal.


3.5. The Former Breed Generating Station: From Coal Infrastructure to Proposed Advanced Nuclear Capacity

The October 19 agreement introduces a separate dimension of Indiana’s strategy. Under the joint development agreement, AEP and X-energy will carry out feasibility work on a twelve-unit Xe-100 plant near the retired Breed station, a brownfield site that AEP already owns, and Amazon is supporting development through a funding agreement whose amount has not been disclosed.[1][2] If the project reaches a final investment decision and proceeds through construction to commercial operation, Amazon would purchase the output under a long-term contract that AEP says would cover the facility’s costs and operations for the contract’s duration, and the plant would be expected to operate for more than sixty years.[1] X-energy describes the project as the second under its collaboration with Amazon, which targets more than five gigawatts of new nuclear capacity on American grids by 2039, and as the fourth announced deployment within a commercial pipeline of roughly 11.5 gigawatts.[2]

The choice of location has an economic logic. Redeveloping a former generating property can offer access to land already zoned for industrial energy use, to transmission corridors, to cooling and transport infrastructure, and to a community with a memory of hosting a power plant, although none of this removes the need for engineering studies, licensing, environmental review, equipment procurement, financing, and construction. The technology’s regulatory position also improved materially in the weeks around the announcement: on November 9, 2026, the Nuclear Regulatory Commission approved a construction permit for the Long Mott Generating Station, the four-unit Xe-100 project that X-energy is developing with Dow at Seadrift, Texas, the first such permit for an advanced reactor at an American industrial site, and X-energy’s fuel subsidiary received in February 2026 the first Category II fuel-fabrication license issued in the United States.[52][53] These milestones reduce licensing risk for subsequent projects, but they do not alter the fact that no Xe-100 has yet been completed or operated commercially, and that the Indiana partners have disclosed neither a construction schedule nor a budget.[7]

The financial architecture is nonetheless instructive, because it separates responsibilities that conventional utility construction bundles together. The technology provider carries the design and commercialization of the reactor; the utility evaluates the project, would operate it, and integrates it with the grid; and the hyperscale customer supplies development capital and the commercial demand that justifies the investment. For households, the decisive issue is whether the plant’s financing and operating structure protects them against costs that arise primarily from serving that customer, and AEP’s chairman addressed the point by stating that the agreements contain the capital protections and financial safeguards the company regards as essential, an assurance whose content will become testable only when the definitive agreements are filed.[1] A press release can describe the intended allocation of risk, but the final financing documents, the regulatory orders, and the contracts themselves determine who actually bears it.


3.6. Nuclear Economics and the Limits of an Anchor-Customer Agreement

Advanced nuclear generation has economic characteristics that make long-term commercial commitments especially valuable. Nuclear projects require very large capital outlays before any electricity is sold, together with licensing, engineering, site preparation, fuel supply, specialized procurement, quality assurance, and construction periods long enough for financing costs to compound, and a large committed purchaser reduces uncertainty about future revenue in a way that can lower the cost of capital and make the investment easier to evaluate. Amazon’s capacity to play that role is not in doubt, given that its chief executive told investors in October that the backlog at Amazon Web Services had reached $578 billion and that customers were reserving capacity in blocks of hundreds of megawatts years ahead of delivery.[54]

“almost as fast as we install capacity, customers are gobbling it up”

— Andy Jassy, President and Chief Executive Officer, Amazon [54]

A power purchase agreement is not, however, a universal guarantee against nuclear development risk, and the questions that matter are specific: who owns the project, when the customer’s payment obligations begin, who is responsible for cost overruns and for delay, what performance guarantees and decommissioning obligations apply, how fuel costs are treated, and what happens on early termination. A project may also raise questions about whether unused output can serve other customers and how the revenue from such sales is shared. When a regulated utility is involved, the regulator must examine whether any cost allocated to customers is prudent and whether the contractual arrangements are consistent with law, and a hyperscaler’s public assurance that it will bear the project’s costs should be weighed against enforceable terms.

The Load Dividend framework therefore distinguishes two possible outcomes for a project such as Breed. A nuclear project that protects existing customers against financial exposure would qualify as a strong example of cost neutrality. A nuclear project that also lowers the long-run cost of serving existing customers, improves the economics of shared infrastructure, or produces verifiable customer credits could contribute to a positive dividend. The second outcome should never be assumed from the first, and for the period through 2030 the influence of the Breed proposal is likely to be felt far more in investment planning and contract design than in delivered electricity.


3.7. What Indiana Must Demonstrate Before the Load Dividend Can Be Considered Proven

Indiana’s announcements offer a promising research opportunity, and the evidence they have produced so far is incomplete. A credible evaluation will follow the outcome of the pending rate case, determine which proposed reductions are approved, and examine how the approved rates affect actual bills, and it will then assess how much of the benefit can reasonably be attributed to new large-load revenue as distinct from other changes in the revenue requirement. It will identify the infrastructure obligations associated with that revenue, the protections established by the large-load contracts, and any costs that remain recoverable from other customer classes. For the nuclear project, researchers should monitor feasibility findings, licensing milestones, financing terms, the final investment decision, and the allocation of construction and operating risk, and they should examine with particular care what happens if the project is delayed, reduced in size, redesigned, or abandoned.

Those possibilities matter because the benefits of customer growth can materialize on a different timeline from the investments required to support that growth, so that a household may receive a rate reduction in 2027 while the new nuclear facility remains years from producing a kilowatt-hour, and the two outcomes can coexist without either depending on the other. Indiana’s potential contribution is therefore not proof that every major datacenter lowers electricity bills. It is evidence that large-load tariff design and revenue allocation can be organized around a different objective, which is to make customer growth financially beneficial to the existing customer base while controlling the associated infrastructure risk, and the stronger the regulatory record becomes, the more useful Indiana’s experience will be to every other state now writing its own rules.


Section 4: The Federal Transformation — From Voluntary Commitments to Proposed Cost-Allocation Requirements

For most of the history of American electricity regulation, the question of who pays to serve a large new customer was answered almost entirely at the state level, in rate cases that attracted little attention outside the utility bar. The year 2026 changed that, and it changed it in a sequence that is worth following closely because each step used a different legal instrument with a different degree of enforceability. In March the executive branch obtained a voluntary pledge from the largest technology companies. In June the Federal Energy Regulatory Commission opened formal proceedings against every regional grid operator under its jurisdiction. In November the Senate passed legislation that would write a cost-allocation rule for computational loads into the Federal Power Act. A pledge states an intention, a tariff approved by a regulator establishes legally operative terms of service, a contract creates obligations between identified counterparties, and a statute creates general legal duties subject to implementation and judicial review, and the durability of any protection for households depends on how many of these instruments point in the same direction. This section examines each in turn and then turns to the problem that none of them resolves.


4.1. The White House Ratepayer Protection Pledge

On March 4, 2026, the White House announced the Ratepayer Protection Pledge, under which Amazon, Google, Meta, Microsoft, OpenAI, Oracle, and xAI committed to build, bring, or buy the new generation needed to meet their demand and to pay the full cost of it, to pay for all new delivery infrastructure required to serve their datacenters, to negotiate separate rate structures with utilities and state governments, and to pay those rates for the power and infrastructure brought online on their behalf whether or not they ultimately use the electricity.[55] The accompanying presidential statement declared that the buildout should be turned toward affordability for all households and businesses and should not be financed at their expense, and the administration later condensed the principle into a slogan displayed on the pledge’s official page.

“Data centers pay their way. Ratepayers pay theirs.”

— The White House, Ratepayer Protection Pledge [56]

In July the administration announced an expansion that brought more than 200 additional utilities, cooperatives, developers, and state governments into the commitment, so that by its own account the pledge now spans more than 300 organizations, covers about 80 percent of the power delivered to American homes and businesses, and applies in areas where some 263 million people live.[57] The pledge reflects a recognition that electricity affordability has become part of the political environment in which artificial intelligence must operate, and it marks a change in the relationship between the technology industry and the physical infrastructure supporting it, since hyperscalers that once bought electricity as ordinary customers of established systems now fund generation, sign long-dated purchase agreements, support transmission upgrades, and negotiate bespoke tariffs.

The limitation of the instrument is equally plain. Analysts at Harvard’s Salata Institute observed shortly after the announcement that the commitments leave considerable room for interpretation, since whether datacenters pay all of the relevant costs or only a portion of them has very different consequences for consumers, and the pledge itself contains no mechanism for verification or enforcement.[58] Its long-term importance therefore depends on the degree to which the commitments are reflected in actual tariffs, contracts, investment decisions, and documented customer outcomes, which is to say on whether the voluntary instrument is followed by binding ones.


4.2. The Federal Energy Regulatory Commission’s June 2026 Show-Cause Orders

The first binding instrument arrived on June 18, 2026, when the Federal Energy Regulatory Commission voted unanimously to issue six orders under Section 206 of the Federal Power Act, one to each of the regional transmission organizations and independent system operators it regulates, directing PJM, the Midcontinent ISO, the Southwest Power Pool, the California ISO, ISO New England, and the New York ISO either to demonstrate that their tariffs adequately address the interconnection of large loads or to propose reforms.[59][60] The Commission’s preliminary view was that existing frameworks may be inadequate to the reliability, cost-allocation, and timing problems that large loads present, and its directives fell into several categories, among them standardized interconnection procedures, rules for co-located generation and flexible service, and, most relevant here, protections against cost shifting that include public information on network-upgrade costs, standard cost-recovery agreements under which a large customer commits to a minimum financial contribution toward the transmission owner’s revenue requirement, and credit support to secure that commitment.[60]

The procedural posture should be described with care. The orders opened proceedings; they did not themselves make final findings or impose new rates, and they left each region room to define large loads and to tailor operational requirements to its own conditions. The issues the Commission identified nonetheless correspond closely to the terms of the Load Dividend identity. Cost transparency helps determine which upgrades are attributable to a particular customer. Cost-recovery agreements reduce the risk that infrastructure costs are shifted after a customer withdraws. Clear procedures reduce delay and improve the reliability of investment planning, and flexible service allows a customer to accept operational limits in exchange for different pricing or faster interconnection. Each of these mechanisms alters the relationship between a customer’s technical requirements and its financial responsibilities, and together they point toward a market in which large computational loads are differentiated by their demand characteristics and contractual commitments instead of being treated as interchangeable sources of load growth.


4.3. The Bipartisan American Affordability and Jobs Act of 2026

The Senate’s passage of the Bipartisan American Affordability and Jobs Act on November 19 introduced the possibility of a statutory framework. The bill is principally a permitting and transmission measure: it would shorten environmental reviews, limit the window for legal challenges, broaden federal siting authority for interstate transmission, and require regional planners to evaluate transmission needs jointly against demand forecasts over a twenty-year horizon.[9][61] Its treatment of computational loads appears in Section 2107, which the Congressional Research Service describes as directing changes to federal transmission pricing policy and affirming state authority over datacenters and high-density computing facilities.[9] As analyzed by the Bipartisan Policy Center and by Columbia University’s Center on Global Energy Policy, the section would add a new provision to the Federal Power Act establishing a ratepayer-protection floor for loads of 20 megawatts or more, with the threshold measured across commonly owned sites and phased projects so that a campus cannot be divided to escape it, and it would require such loads to bear the incremental cost of the generation, storage, transmission, and network upgrades built to serve them, to provide financial assurance against the possibility that they never materialize, and to hold other customers harmless if they depart.[61][10]

The projected effects are substantial by the standards of energy legislation. Rhodium Group’s modeling of the bill’s transmission and datacenter provisions projects average household electricity bills $50 to $78 a year lower in 2040 than they would otherwise be, with the transmission reforms allowing lower-cost resources to reach the grid and the datacenter provisions preventing the assignment of datacenter costs to other customers.[62] Supporters have cited estimates that the measure would move roughly $55 billion of transmission cost onto datacenters over a decade.[63] The ranking Democrat on the Senate Energy and Natural Resources Committee, one of the bill’s four sponsors, stated its purpose in terms that track the first two outcomes of Table 1.

“make data centers pay their fair share of electricity and grid costs”

— Senator Martin Heinrich, Ranking Member, U.S. Senate Committee on Energy and Natural Resources [64]

Three features make the bill economically important for the argument of this paper. It would establish a federal floor beneath the patchwork of state tariffs, so that a hyperscaler could no longer shop among jurisdictions for the weakest cost-allocation rule. It recognizes that the financial consequences of interconnecting a large load extend beyond the immediate utility and customer into regional transmission. And it addresses directly the possibility that the consumption used to justify infrastructure might never arrive. The bill has not become law as of November 21, 2026; its implementation would depend on House action, on the final statutory text, on rulemaking at the Commission, and on coordination with state and regional rules, and it faces opposition from environmental organizations and tribal governments who argue that its permitting provisions weaken review procedures and consultation requirements.[8] This paper accordingly treats it as a significant proposed development and not as an operative mandate.


4.4. The 20-Megawatt Threshold and the Problem of Project Aggregation

The Senate’s attention to aggregation deserves emphasis, because developers routinely organize facilities through multiple phases, corporate entities, and adjacent parcels, and a single hyperscale development may consist of several buildings constructed at different times and connected through separate electrical services held by separate subsidiaries. A rule that looked only at the nominal demand of an individual building would invite developers to divide projects in order to remain beneath the threshold, even though the cost of serving a campus arises from its combined demand and a transmission upgrade does not become cheaper because ownership of the buildings it serves has been distributed among affiliates. Indiana’s tariff confronts the same issue with its alternative trigger of 150 megawatts aggregated across a company, and Virginia’s 25-megawatt and Texas’s 75-megawatt thresholds raise it as well.

The Load Dividend framework therefore requires careful identification of the ultimate beneficiary of new electricity service. Regulators should examine the relationships among project owners, operating companies, cloud tenants, electricity purchasers, and financial guarantors, and contractual responsibility should be assigned to entities capable of honoring commitments that last as long as the infrastructure does. At the same time, rules should avoid treating independent businesses as a single customer merely because they operate in the same region, and the appropriate test looks to the underlying electrical, contractual, and economic relationships.


4.5. The Unresolved Problem of Wholesale Electricity Prices

Electricity regulation in the United States is divided among institutions, with the federal Commission overseeing interstate transmission and wholesale markets, state commissions overseeing retail service and most rate design, and regional grid operators coordinating planning and market operations, and a large datacenter can affect all three levels at once without those effects ever appearing in the same proceeding. A state commission may approve a tariff that makes a hyperscaler pay every dollar of its local infrastructure, and regional capacity prices may still rise because aggregate demand has grown faster than available generation, so that a household can be fully protected against the cost of a dedicated substation while remaining exposed to a market-wide price increase.

This is the most important limitation of a narrowly defined cost-allocation policy, and the events of 2026 have made it impossible to ignore. PJM’s base residual auction for the 2028/2029 delivery year, announced on July 14, procured 138,318 megawatts of unforced capacity at $325 per megawatt-day, a price set by a cap negotiated among the governors of the PJM states and accepted by the Commission, and it still fell 6,831 megawatts short of the reliability requirement, the second consecutive auction to miss the target; the capacity cleared was valued at $16.4 billion.[65] The independent market monitor attributed $6.3 billion of that amount, or 38.2 percent, to datacenter load, and calculated that datacenter load had accounted for $29.4 billion, or 46.2 percent, of total capacity-market charges over the last four auctions, a finding that led the monitor to recommend removing datacenter load from the general capacity market and procuring its capacity through a dedicated mechanism.[11] Figure 1 shows the monitor’s estimate by delivery year.


Figure 1. PJM capacity-market cost attributed to datacenter load by the Independent Market Monitor, by delivery year

Four-auction total: $29.4 billion, or 46.2 percent of capacity-market charges. The 2028/29 value is derived as the remainder of the reported total. Source: Monitoring Analytics, note [11].


The Columbia analysis of the Senate bill states the consequence with a precision that should temper any expectation that Section 2107 will settle the affordability question in restructured markets.

“these protections would not reach higher wholesale capacity and energy prices”

— Robin Millican, Jack Andreasen Cavanaugh, Noah Kaufman, and coauthors, Center on Global Energy Policy, Columbia University [61]

The same authors identify a second-order risk that bears directly on the Load Dividend: if grid connection becomes sufficiently expensive, datacenters may elect to supply themselves with on-site gas generation, which they estimate at roughly $101 to $107 per megawatt-hour compared with an average industrial grid price near $86, and in that case the load never contributes toward the fixed costs of the shared system and the dividend is forgone before it can be earned.[61] A cost-allocation rule can therefore fail in two opposite directions, by charging too little, which shifts costs onto households, or by charging so much relative to the alternatives that the customer leaves the grid and takes its contribution with it. A credible evaluation must measure two financial effects that operate through separate channels, one inside the utility’s retail rate design and one in the wider wholesale market, and it must accept that a datacenter may reduce household costs through the first while increasing them through the second, an outcome that is economically mixed and should be reported as such.


4.6. The Legislative Tradeoff and the Missing Measurement System

The Senate bill’s permitting provisions raise a tradeoff that belongs in any honest account of the Load Dividend. Supporters argue that lengthy approval processes delay generation and transmission, make it harder to meet demand, and thereby raise costs, while environmental and community organizations respond that the proposed changes would weaken established review, reduce public participation, and favor infrastructure with substantial environmental consequences, and the split within the Democratic caucus on final passage reflected that disagreement.[8] A project that produces measurable bill reductions may still impose costs through water consumption, emissions, land-use conflict, and construction disturbance, and a framework concerned with public benefit cannot exclude those consequences merely because they do not appear on an electricity bill. The appropriate analytical response is to evaluate the actual costs, benefits, risks, and safeguards of particular projects, without assuming that either faster permitting or longer review is preferable in every case.

A further obstacle is the poor quality of information about how much datacenter demand is real. Public announcements describe projects in gigawatts without distinguishing operating facilities from those under construction, from approved developments, from interconnection requests, and from proposals that are little more than options on land. Research published in November 2026 by Latitude Media found that utilities across 53 service territories reported 537 gigawatts of datacenter demand in their pipelines while holding signed contracts for only about 120 gigawatts, or 22 percent, against roughly 22 gigawatts actually operating, and that one large utility reporting nearly 55 gigawatts of contracted capacity had only about 12 gigawatts under electric service agreements that commit a customer to take power.[66] Speakers at an industry forum the same month noted that actual demand in PJM during 2026 had run about 21 percent below the grid operator’s own forecast from the previous year.[67] Counting every announced project as inevitable demand encourages overinvestment whose cost falls on existing customers, while ignoring credible projects invites shortages, and the Senate bill’s provisions on demand reporting and grid data, if enacted, would begin to give regulators the means of telling the two apart.[9] The federal opportunity thus extends beyond imposing financial obligations to building the information system that would show whether those obligations are appropriate and whether promised benefits arrive.


Section 5: The States as Competing Laboratories — Pennsylvania, Virginia, Texas, and the Emerging Geography of Electricity Accountability

While Washington debated, the states acted, and they acted differently from one another in ways that follow from their different market structures. Indiana, as the previous section showed, is a vertically integrated state whose utility owns generation and whose regulator sets rates on the basis of cost, which makes it the natural setting for an experiment in passing surplus revenue through to households. Pennsylvania is a restructured state inside PJM, where households buy generation at market-determined prices and where the utility’s tariff can protect them from wires costs but not from the capacity auction. Virginia hosts the largest concentration of datacenters in the world inside a vertically integrated utility that nonetheless participates in PJM’s markets. Texas runs its own grid under its own rules and faces a queue of proposed load larger than its existing peak demand. Reading these four jurisdictions together is the closest thing available to a controlled comparison, and what it reveals is that each state has concentrated on a different term of the Load Dividend identity.


5.1. Pennsylvania: Cost Attribution Inside a Wholesale Market

Pennsylvania provides the sharpest contrast with Indiana. Governor Josh Shapiro signed Executive Order 2026-05 in August 2026, formalizing what his administration calls the Governor’s Responsible Infrastructure Development requirements, a set of conditions on datacenter development that address infrastructure financing, community approval, environmental responsibility, and the priority of datacenter curtailment when the grid is under stress.[68][69] On October 15 he petitioned the Pennsylvania Public Utility Commission for party status in a complaint brought by the state’s Office of Consumer Advocate against PPL Electric Utilities, a step his office described as the first intervention by a Pennsylvania governor in a Commission proceeding in at least fifteen years.[68] The complaint contends that PPL’s existing rates, service classifications, and procedures do not adequately protect existing customers from costs caused by large datacenters, and it asks the Commission to order a separate rate and service class for such customers, to allocate to that class the transmission upgrades its members require, and to place those customers on interruptible service so that their supply can be reduced in an emergency before other customers lose theirs. The administration stated its objective in the announcement of the filing.

“data centers are held responsible for the costs and reliability impacts they create”

— Office of Governor Josh Shapiro, Commonwealth of Pennsylvania [68]

PPL responded that it had already taken steps to protect customers and that its LP-6 rate class requires qualifying large customers to make long-term financial and usage commitments, which means that the proceeding will turn on whether existing protections are sufficient and not on whether any exist.[70] Pennsylvania’s approach highlights a distinction of sequence. Indiana’s case emphasizes the potential for new customer revenue to support reductions for existing customers, whereas Pennsylvania’s proceeding emphasizes the prior question of preventing large-load costs from being transferred to them in the first place, and the two concerns are compatible but represent different stages of the same problem, since a utility cannot credibly distribute a dividend until it has established which costs belong to the new customer.

Pennsylvania’s membership in PJM complicates even that prior question, because the cost of serving a large datacenter is not confined to the local distribution system. The facility affects regional generation adequacy and transmission planning, and those effects reach customers who live far from it; a municipality might negotiate strong local agreements and still face higher capacity charges set in a regional auction that no state regulator controls. The framework therefore suggests that analysts distinguish local effects, such as construction, tax receipts, employment, and direct community obligations, from utility-wide effects on the revenue requirement and rate design, and both of these from regional effects on wholesale prices, capacity procurement, congestion, and reliability obligations. A project may benefit its host county while imposing costs elsewhere, or reduce a utility’s revenue requirement while increasing regional capacity exposure, or, if it finances new generation and strengthens the regional network, benefit customers well beyond its host community, and a sophisticated state policy will account for each layer separately instead of relying on a single statewide estimate.


5.2. Virginia: Rate Classes, Minimum Bills, and Direct Assignment

Virginia has the most developed regulatory record on large-datacenter electricity service, as one would expect of the state where the industry is largest. In its November 25, 2025 order in Dominion Energy Virginia’s biennial review, the State Corporation Commission created a new GS-5 rate class, effective January 1, 2027, for customers with demand of 25 megawatts or more, and required those customers to enter fourteen-year contracts under which they pay at least 85 percent of their contracted transmission and distribution demand and at least 60 percent of their contracted generation demand, whether or not they use it, and to post collateral against the upgrades their service requires.[21][71] These provisions tie the financial obligation of a large customer to the infrastructure reserved on its behalf, which is the essential move, since a datacenter that contracts for substantial capacity creates obligations for the utility even if its consumption later disappoints.

Virginia’s framework is therefore an example of structured cost protection, and it is instructive precisely because it shows what cost protection does not accomplish. In the same order, the Commission approved increases in residential base rates, smaller than Dominion had requested but increases nonetheless, amounting to an average of about thirteen dollars a month over two years.[71] A large-load rate class and the overall trajectory of residential rates are separate matters, and the existence of the first did not reverse the second. The Commission went further in August 2026, after a filing by Governor Abigail Spanberger’s administration, by ordering Dominion to develop a tariff that assigns directly to datacenters and other large-load customers the cost of transmission infrastructure built exclusively for their facilities; the order reduced the projected increase in the residential transmission charge from $2.90 a month to $0.94, and the administration described the saving to other customers as amounting to hundreds of millions of dollars.[72][23] That outcome is a clear case of cost avoidance, the second row of Table 1: residential bills still rise, but by less than they would have risen under the prior allocation.

On September 18, 2026, the Governor announced a broader Data Center Accountability Framework and signed Executive Order 22 to begin implementing it. The framework is organized around transparency, environmental protection, energy costs, clean energy, and workers, and it includes a ban on non-disclosure agreements between state agencies and datacenter developers, a requirement of local approval for facilities above 25 megawatts, and standards for efficiency, water use, land impact, and backup generation, together with a legislative agenda for the 2027 General Assembly.[73] The Governor described its purpose in language that makes the distributional claim explicit.

“Virginia’s Data Center Accountability Framework is about delivering a better deal for Virginians”

— Abigail Spanberger, Governor of Virginia [73]

Virginia has also enacted an electricity consumption tax applicable to datacenters for a specified period.[74] Such a tax is financially distinct from a utility rate reduction, because the revenue may support state priorities without lowering any household’s electricity bill, and the framework proposed here accordingly separates bill benefits from fiscal revenues. Both may matter for economic development, and they should not be combined into a single affordability claim without an explanation of how the money is used.


5.3. Texas: Interconnection Discipline in a Rapidly Expanding Market

Texas presents a third regulatory environment, in which most of the state operates within the Electric Reliability Council of Texas under institutional arrangements unlike those of the other three states. Senate Bill 6, which took effect in June 2025, directed the Public Utility Commission of Texas to adopt standards for interconnecting loads of 75 megawatts or more that would support economic development while limiting stranded-infrastructure risk and protecting reliability, and it included provisions on curtailment and demand management during emergencies.[75] The Commission adopted its implementing rule on September 18, 2026, effective October 8. The final rule requires a large-load customer, before it can be included in an interconnection study, to execute an intermediate agreement, make officer-attested disclosures, pay a flat study fee of $100,000, and post financial security of $50,000 per megawatt of requested peak demand, so that a 300-megawatt request requires $15 million in security before the study begins; it dropped a non-refundable per-megawatt interconnection fee that had appeared in the March proposal and relaxed several milestone and refund terms.[22][76]

The difference between the proposed and adopted rules is a reminder that a financial commitment discussed in an initial proposal cannot be assumed to survive into the final text, and Texas illustrates the difficulty of balancing financial discipline against the desire to attract investment. Requirements that are too weak allow developers to reserve scarce grid capacity with little at stake, while requirements that are too burdensome impose unnecessary cost and delay on viable projects, and the appropriate balance depends on the infrastructure risk each project creates. AEP reported in October that 46 gigawatts of its contracted load in the Texas market sits under agreements that comply with the new law, which suggests that the largest counterparties have found the terms acceptable.[48]

The Texas queue is best understood as a problem in financial forecasting. Reporting during 2026 described requested large-load interconnections in ERCOT reaching into the hundreds of gigawatts, a figure that reflects proposals and not a forecast that all of the requested load will arrive.[22] Utilities and transmission providers must prepare for credible demand, yet electricity infrastructure involves long construction periods and large capital commitments, so that building every proposed upgrade before knowing which projects will proceed would burden customers with unnecessary cost, while refusing to invest until uncertainty disappears would create shortages. Future datacenter requirements depend on semiconductor supply, capital budgets, model architectures, financing conditions, and enterprise demand, any of which can change the size or timing of a campus. The economic solution lies in connecting investment approval and construction commitments to evidence of commercial readiness, including binding customer agreements, financial guarantees, equipment orders, verified construction milestones, and payment obligations, and Texas’s experience is therefore a test of whether regulation can distinguish demand that is commercially credible from demand that remains speculative.


5.4. Comparing the State Approaches

The four jurisdictions, summarized in Table 4, demonstrate different dimensions of one economic problem. Indiana combines approved contractual protections with a proposed rate reduction that would distribute part of the benefit of increased customer revenue. Pennsylvania emphasizes direct assignment of datacenter-related costs and the protection of households within a regional wholesale market. Virginia demonstrates large-load rate classes, minimum billing commitments, direct assignment of dedicated transmission, and a broader accountability agenda. Texas emphasizes interconnection procedure, financial security, demand management, and the screening of very large volumes of proposed load.


Table 4. Four state approaches to large-load cost allocation

StateMarket structureSize thresholdPrincipal instrumentTerm of the identity addressed
IndianaVertically integrated utility within PJM70 MW per site; 150 MW aggregate12-year take-or-pay contracts with 80 percent minimum demand; proposed rate reduction.[4][44]Revenue and distribution of surplus
VirginiaVertically integrated utility within PJM25 MWGS-5 class with 14-year contracts and 85/60 percent minimums; direct assignment of dedicated transmission.[21][72]Attributable cost and risk protection
PennsylvaniaRestructured; retail choice within PJMUnder litigationExecutive Order 2026-05; complaint seeking a separate datacenter class and interruptible service.[68]Attributable cost; wholesale exposure
TexasERCOT energy-only market75 MW$100,000 study fee and $50,000 per MW security before study; curtailment provisions.[22][76]Risk protection and demand credibility

These approaches are not mutually exclusive, and a state could combine long-term contracts, transparent interconnection requirements, direct infrastructure contributions, minimum billing, and customer-benefit mechanisms. The relevant policy question is which combination fits the conditions of a particular system, since a state with abundant generation faces different challenges from one with severe transmission congestion, a vertically integrated utility has different obligations from one operating within a competitive wholesale market, and a region with substantial nuclear generation will evaluate long-term procurement differently from one that depends on natural gas or variable renewables. A single national tariff formula would therefore miss much of what matters, and the Load Dividend framework is better understood as a set of principles that can be translated into different regional arrangements: large customers should bear their attributable costs, contractual commitments should protect against abandoned infrastructure, shared benefits should be measured transparently, positive benefits should be distributed through identifiable mechanisms, and economic development should be judged by its full long-term consequences. An analysis by the consultancy E3, published in 2026, counted about thirty large-load tariffs implemented across the country in 2025 and 2026 alone and concluded that while there is little evidence so far of systematic cost shifting under the new designs, it remains to be seen how well these risk-mitigation tools will perform over time, which is a fair summary of what the state laboratories have and have not yet shown.[77]


Section 6: The Load Dividend Economy, 2027–2030 —Investment Models, Regulatory Design, and the Future Social Contract

The preceding sections have described what the Load Dividend is, what the research says about when it appears, and how one state and several others are trying to produce or protect it. This section looks forward. The years from 2027 through 2030 are the period in which the contracts now being signed will begin to bind, in which the largest campuses now under construction will draw power at scale, in which the Indiana rate case and the Pennsylvania complaint will be decided, and in which the Senate bill will either become law or be overtaken by something else, and they are therefore the years in which the proposition of this paper will be tested against bills that households actually receive. The purpose of what follows is to give that test a structure: a quantitative model simple enough to be audited, an account of the contractual variables that drive it, a regulatory architecture that would make its results visible, and three scenarios against which the outcome can be compared.


6.1. The Scale of the Challenge and the Capital Behind It

The economic importance of the Load Dividend will depend in part on the scale of datacenter development, and here the Berkeley Lab reference case of 11.8 percent of national electricity consumption by 2030 is best treated as a scenario and not as a prediction, since the laboratory’s own sensitivity cases range from 9.5 to 15.3 percent depending on equipment shipments, the mix of specialized chips, and the operating life of accelerators.[15] A national percentage also conceals the local problem, because some regions will see modest growth and others extraordinary concentration; the International Energy Agency has observed that nearly half of American datacenter capacity sits in five regional clusters and that half of the capacity under development is being added to existing clusters.[13]

The Five-Layer framework explains why forecasts are so unstable. Semiconductor development determines the electricity required for a given computing task, datacenter construction determines where those chips are concentrated, foundation-model development determines the scale and timing of training and inference workloads, and applications and agentic systems determine how intensively the resulting infrastructure is used, so that energy demand is the product of four uncertain quantities and cannot be forecast by counting announced square footage. What is not uncertain is the financial weight of the companies involved. The four largest hyperscalers have guided to combined 2026 capital expenditures in the neighborhood of $730 billion, with Amazon near $220 billion, Alphabet at $195 to $205 billion, Microsoft near $170 billion, and Meta at $135 to $145 billion, and Table 5 sets out the third-quarter figures.[25][26] A hyperscaler in this position is not merely buying a commodity when it contracts for power; it is buying the availability of an input without which several hundred billion dollars of chips, buildings, and networking equipment cannot earn a return, and that is why such a company may rationally accept financing obligations for energy infrastructure that would be unthinkable for an ordinary industrial customer. For utilities and infrastructure investors the opportunity is to obtain long-duration commitments from counterparties whose strategic need for power is acute, and the corresponding discipline is to remember that not every company attached to a prominent artificial intelligence project has the balance sheet of the largest publicly traded hyperscalers, and that a capital-intensive startup financed by outside investors presents a different credit risk from a diversified cloud company with substantial free cash flow.


Table 5. Hyperscaler capital expenditure, third quarter of 2026

CompanyQ3 2026 purchases of property and equipmentFull-year 2026 guidanceChange in Q3 from a year earlier
Amazon$77.3 billionAbout $220 billionUp from $35.1 billion
Alphabet$51.4 billion$195–205 billionRoughly doubled
Microsoft$39.4 billionAbout $170 billion (calendar year, including finance leases)Higher
Meta$33.9 billion$135–145 billion (including finance-lease principal)Higher
Four-company total$201.9 billionAbout $730 billionUp 108 percent from $97.3 billion

Sources: company third-quarter 2026 reports as compiled in notes [25] and [26]. Definitions of capital expenditure differ among the companies.


6.2. Constructing a Quantitative Load Dividend Model

A practical application of the framework requires numbers, and the simplest way to show how the identity of Section 1 behaves is to run it through three cases for a hypothetical utility, as in Table 6. Suppose the utility recovers $300 million a year in fixed system costs from an established customer base of 300,000 accounts, and that a hyperscale customer enters a long-term agreement under which it pays $80 million a year in energy charges, demand charges, and minimum-billing obligations. In the first case, the fully attributable incremental cost of serving that customer is $55 million a year, and a further $10 million is set aside for risk protection, reserves, and other retained obligations, which leaves an illustrative net contribution of $15 million. That amount is not automatically a customer dividend, because the regulator must still determine how it enters the revenue requirement, whether any portion must be retained, and how the remainder is allocated; but if all of it were distributed to the 300,000 existing accounts, the average benefit would be $50 per account per year.

In the second case the customer pays the same $80 million, but the attributable cost of serving it rises to $85 million, perhaps because serving it requires new generation at a cost above the system average or transmission that the first case did not need. The contribution is negative before any risk reserve is considered, and a claim that this customer benefits existing households would require evidence of additional avoided costs or other valid offsets, since the size of the customer’s bill establishes nothing by itself. In the third case the arithmetic of the first case holds for four years, after which the customer closes a facility for which the utility built a dedicated substation and transmission connection in the expectation of fifteen years of service. If the contract provides insufficient recovery, the unrecovered cost of that infrastructure is transferred to other customers, and a dividend that was reported, and perhaps celebrated, in the early years reverses in the later ones.


Table 6. An illustrative Load Dividend calculation for a hypothetical utility with 300,000 existing accounts

Annual amountCase A: surplusCase B: subsidyCase C: early exit
Incremental revenue from the large customer$80 million$80 million$80 million for four years, then zero
Fully attributable incremental cost$55 million$85 million$55 million, plus unrecovered dedicated assets after exit
Risk protection and retained obligations$10 million$10 million$10 million
Net contribution available for allocation+$15 million−$15 million+$15 million for four years, then negative
Average effect per existing accountAbout $50 a year lowerAbout $50 a year higherEarly saving reversed unless the contract recovers the stranded cost

Illustrative figures only; not drawn from any regulatory filing.


The example is not drawn from Indiana’s regulatory filing and makes no claim about it. Its purpose is to demonstrate the structure of the calculation and to make two points that follow from it. The sign of the dividend is determined by the relationship between revenue and attributable cost, which is an empirical matter that only a cost-of-service study can settle, and the durability of the dividend is determined by contract terms, which is a legal matter that only the contract can settle. Short-term savings are therefore insufficient evidence if their existence depends on risks that have been deferred into future decades, and the framework must be applied over the full period across which infrastructure costs are recovered.


6.3. Contract Duration, Financial Security, and the Risk of Stranded Assets

Electricity infrastructure and computing equipment have very different economic lives, and the mismatch creates the central contractual challenge of the coming decade. A transmission line may remain useful for half a century and a substation for several decades, whereas the commercial attractiveness of a particular accelerator can change within a few years as new hardware arrives, and although operators often continue to use older equipment, the economics of their operations can shift quickly. If infrastructure has a thirty-year life and the customer signs for twelve or fourteen, the utility faces uncertainty over the remainder, and the problem is sharper when the infrastructure is specialized, since an upgrade serving a broad industrial district may remain useful after one datacenter leaves while a dedicated connection to an isolated campus may have no alternative user.

A stranded asset arises when an investment cannot recover its expected value from the customers or uses originally anticipated, and datacenter projects create this risk in two forms that contracts must address separately. The first is abandonment, in which a project is cancelled or closed, and the second, which receives less attention and may prove more common, is underutilization, in which a project is completed and operates while using substantially less electricity than was reserved for it. Minimum-billing provisions are the principal defense against the second, because they reduce the financial consequence of any gap between actual consumption and reserved capacity, and the differences among the 80 percent floor in Indiana, the 85 and 60 percent floors in Virginia, and the security-based approach in Texas reflect different judgments about how much of that gap the customer should carry. The appropriate level should reflect the cost of the infrastructure and the risk of the project, and it should not be chosen because a particular percentage happens to be convenient for either party.

Financial guarantees should likewise be assessed by their enforceability and by the creditworthiness of the entity that stands behind them, since a guarantee from a highly capitalized parent has different economic content from an obligation issued by a newly created project subsidiary, and neither should be evaluated by its label. The evidence reviewed in Section 4, that signed contracts cover only about a fifth of the datacenter demand that utilities report, suggests that the distance between announced load and committed load remains the largest single source of stranded-asset risk in the system.[66] The broader principle is that financial responsibility should correspond to the infrastructure commitments that the customer’s own decisions create.


6.4. Electricity Flexibility as an Economic Resource

The size of the dividend also depends on how a datacenter uses electricity, because a facility that can reduce consumption during periods of extreme system stress imposes lower reliability costs than an otherwise identical facility that requires uninterrupted maximum service. The Duke estimates cited in Section 2, under which acceptance of curtailment in a quarter of one percent of hours would allow 76 gigawatts of new load to be served by existing capacity, indicate how large the avoided investment could be, and the MIT estimates of system cost savings of two to five percent indicate what that is worth.[42][19] Not every workload is suitable, since interactive inference, critical enterprise applications, and real-time systems have stringent availability requirements, while certain training, testing, data-processing, and batch activities can be shifted in time if technical and commercial constraints permit. This creates the opportunity for differentiated service, under which a customer accepting limited or non-firm service under an approved tariff pays a different price from one requiring fully firm capacity, and the customer-specific contract between I&M and Google that was the subject of an April 2026 order of the Indiana commission, which pairs a demand-response arrangement with a transfer of accredited clean capacity to the utility, is an early example of the form such arrangements may take.[78]

Flexibility must, however, be verified before it is credited. A promise to curtail is not equivalent to a demonstrated capability to reduce load during a grid emergency, and regulators and grid operators should require metering, controls, testing, and enforcement, and should weigh the benefits of flexibility against any costs it creates, including the emissions of the backup generation that may run when grid supply is curtailed. The framework recognizes flexibility as a potentially valuable contribution without assuming that all datacenters are flexible or that every reduction in consumption produces a corresponding household saving.


6.5. Nuclear Power and the Long-Term Financing of AI Infrastructure

Advanced nuclear generation could become an important component of the energy system supporting artificial intelligence, since it offers high availability, compact land use, and electricity without direct carbon emissions, and its commercial prospects depend on technological readiness, licensing, construction cost, supply chains, fuel availability, and financing. The Indiana partnership is relevant because it joins a large prospective purchaser to a utility and a technology developer, an arrangement that could reduce revenue uncertainty if the project advances. A new nuclear development should not, however, be assumed to deliver cheaper electricity because a hyperscaler is involved, for capital cost and financing conditions remain decisive, construction delay increases cost, specialized equipment creates procurement difficulty, and regulatory approval takes time, and any honest comparison must include alternative generation and storage technologies, transmission investment, and the reliability characteristics of the whole system.

The strongest Load Dividend argument for nuclear investment is therefore not that nuclear generation is inexpensive. It is that a financially capable industrial customer may be willing to assume obligations that remove from existing utility customers some of the risks they would otherwise bear, and whether that produces savings beyond risk protection depends on final contracts and realized costs. The history of American nuclear construction, in which cost overruns were recovered from captive ratepayers through regulated rates, is the reason this allocation of risk matters so much, and the arrangement proposed for Breed, in which the anchor customer’s contract is intended to cover the plant’s costs and operations, would if realized invert the traditional allocation.[1] For the period through 2030, the influence of projects such as Breed is likely to be greater in investment planning and financial structuring than in delivered electricity.


6.6. The Social Contract of Datacenter Growth

The phrase social contract refers here to the practical allocation of responsibilities, costs, and benefits among technology companies, utilities, regulators, and communities, and it does not imply that every project requires an identical agreement, since communities differ in electricity supply, economic opportunity, environmental conditions, land use, and fiscal resources. A shared set of questions is nonetheless emerging, and the public now asks them at zoning hearings and utility commission field hearings across the country: who pays for the infrastructure necessary to serve the facility, who receives the benefit of the new utility revenue, what happens if the customer abandons the project, whether household bills will decline, remain stable, or increase, how environmental and local infrastructure impacts will be addressed, and what information the public will receive before long-term commitments are approved.

These questions extend beyond conventional corporate investment analysis, because a hyperscaler evaluates a project by the expected return from cloud and artificial intelligence services, a utility by capital investment, revenue requirement, and financial stability, a state by employment, taxes, and competitiveness, and residents by affordability, reliability, and environmental conditions. Lucy Qiu of the University of Maryland has pointed out that how much of any cost increase reaches residential customers, and when, depends on utilities’ power-purchasing contracts, on retail rate structures, and on regulatory decisions about who pays, which is another way of saying that the social contract is written in documents most residents never see.[79] The Load Dividend framework offers a common financial language in which these interests can be compared. It does not eliminate disagreement among them, but it does make their competing claims measurable.


6.7. A Proposed Regulatory Architecture for 2027–2030

A practical policy framework could incorporate twelve complementary mechanisms, each of which corresponds to a weakness identified in the evidence reviewed above. Regulators should require credible demand evidence, so that large projects document construction schedules, capacity requirements, and financial readiness before infrastructure is committed on their behalf. They should establish transparent cost attribution, under which utilities identify the infrastructure required specifically for major new customers and distinguish it from broader system improvement. They should require long-term financial commitments where the investment justifies them, with contract duration, minimum billing, guarantees, and termination obligations calibrated to project-specific risk. They should protect against double recovery, so that no customer is charged twice for the same cost through overlapping rates and contractual mechanisms, a concern that has been raised about proposals to charge large loads both embedded and incremental transmission costs.

Regulators should also recognize shared infrastructure benefits, allocating the costs and benefits of multi-purpose projects through defensible methods, and they should incorporate wholesale-market effects into their analysis, since retail protections alone do not address changes in capacity and energy prices. They should evaluate operational flexibility and credit only what has been verified. They should require disclosure of customer impacts in a form that separates proposed savings from approved rate reductions and from realized changes in household bills, which are three different things that are routinely reported as one. They should assess environmental consequences, because affordability cannot be evaluated independently of material effects on air, water, land, and local infrastructure. They should establish mechanisms for distributing benefits, so that when the analysis identifies a positive contribution available to existing customers the rate framework specifies how it is recognized. They should review performance over time, since arrangements made in 2026 will operate in conditions that cannot now be foreseen. And they should preserve commercial viability, so that protective measures reflect actual costs and risks and do not impose arbitrary obligations that push large loads off the shared grid and forfeit the contribution those loads would have made.

The purpose of this architecture is not to guarantee that every new datacenter lowers household bills, which no architecture could do. Its purpose is to make the possibility measurable and to prevent unsupported claims in either direction.


6.8. Three Scenarios for 2030

The future of the Load Dividend can be examined through three illustrative scenarios, which are offered as conditions to monitor and not as predictions. In the first, which might be called contractual protection without positive savings, utilities adopt stronger tariffs and large customers bear their attributable infrastructure costs, and financial guarantees reduce stranded-asset exposure, but overall electricity prices continue to rise because of fuel costs, regional capacity shortages, or other system-wide factors. Customer protection improves and lower household bills do not materialize, so that the result is cost neutrality or cost avoidance and not a verified positive dividend. Virginia after the 2025 biennial review and the 2026 transmission order is the nearest existing example.

In the second scenario, a measurable household dividend, large datacenter customers enter binding long-term arrangements that cover their incremental costs, their additional revenue contributes toward shared fixed expenses, regulators recognize the resulting net benefit through approved rate reductions or credits, and wholesale-market effects remain sufficiently contained that the benefit is not offset. In this scenario the Load Dividend becomes an observable economic outcome, and Indiana’s proposed 2027 reduction is an early example of the kind of regulatory result that would need to be tested against actual performance.

In the third scenario, delayed cost transfer, utilities build infrastructure on aggressive demand forecasts, some projects are delayed, downsized, or abandoned, contractual protections prove insufficient or fail to cover all relevant costs, and wholesale prices rise because supply investment does not keep pace with regional demand. Existing customers eventually absorb expenses not recovered from the original large-load customers, early claims of affordability prove unsustainable, and the long-term result is the opposite of a dividend. The gap between reported and contracted load, and the repeated capacity shortfalls in PJM, are the leading indicators of this scenario, and they are already visible.


6.9. Implications for Corporations and Public Leaders

For hyperscalers, the framework is becoming relevant to site selection and contractual strategy, since a location with transparent tariff rules, credible generation plans, and predictable infrastructure obligations may be preferable to one with a lower headline price and an unstable political settlement. For utilities, strong large-load contracts support investment planning and improve the distribution of fixed costs, and they are increasingly the basis on which equity investors value growth. For semiconductor manufacturers, reliable electricity infrastructure governs the timing and scale of accelerator deployment. For nuclear developers and other generation providers, financially credible long-term customers improve the commercial basis for new projects. For state governments, the framework provides a way to examine whether major corporate investments deliver observable benefits instead of relying on announced capital expenditures, and for federal policymakers it offers an empirical test of whether cost-allocation legislation achieves its consumer-protection purpose. For households the measure is more direct, since the benefits should eventually appear in forms they can read on a bill or verify in a regulator’s order: lower charges, documented avoided expenses, or stronger protection against risks they would otherwise have carried. The credibility of the entire concept depends on that last test.


Section 7: What Have We Learned? Seven Pillars

The argument of this paper has moved from economic theory through empirical research to the regulatory experience of four states and the legislative activity of one extraordinary year, and it is worth gathering what that movement has established before returning to the place where it began. The seven pillars that follow are not a summary of the sections in order. They are the propositions that I believe a regulator, an executive, an investor, or a resident could take from this research and apply to a specific project in a specific place, and each is stated in a form that the events of 2027 through 2030 could confirm or refute.


Pillar 1. Electricity Demand Is Not Automatically an Economic Burden

Growing electricity demand creates both costs and benefits, and the prevailing debate, which treats large datacenters primarily as sources of infrastructure pressure, captures only half of the ledger. The concern is justified wherever major new loads require expensive generation, transmission, or distribution without adequate financial commitments, but the same demand creates revenue that contributes toward the cost of maintaining an existing network, and whether the outcome is beneficial depends on the relationship among incremental revenue, attributable cost, and financial risk. The Berkeley Lab and EPRI findings that load growth has tended to accompany, and in the causal estimates to produce, lower average retail prices, and Indiana Michigan Power’s explicit attribution of its proposed reduction to large-customer revenue, do not prove a universal benefit. They demonstrate that the relationship between large electricity demand and residential affordability is more complicated than the assumption that additional demand always raises household bills, and that utilities and regulators should evaluate large loads on economic evidence instead of treating all demand growth as inherently good or inherently harmful.


Pillar 2. A Dividend Requires More Than a Promise to Pay

A hyperscaler can promise to pay for its electricity, and seven of them did so at the White House in March 2026, but that promise does not establish whether the payments will cover the infrastructure required to deliver the electricity or what happens if the facility is never built. Minimum-billing commitments, financial guarantees, cost-recovery provisions, and early-termination obligations are therefore essential elements of the framework, because their function is to ensure that the customers creating substantial infrastructure requirements carry appropriate responsibility for them. The proposed Breed nuclear facility illustrates why the principle matters: the project remains exploratory, the eventual allocation of cost and risk will depend on agreements not yet written, and a successful financing structure could protect households against exposure without producing a dollar of savings. Cost neutrality and customer dividends are related outcomes, and the second cannot be inferred from the first.


Pillar 3. Institutions Determine the Sign of the Effect

The most important lesson of the research record is that the divergent findings of serious scholars reflect differences in institutional setting more than differences in method. Datacenter entry shows no significant effect on residential prices in traditionally regulated states and a measurable increase in restructured ones; it lowers average retail rates where fixed costs can be spread and raises wholesale prices where supply is tight. A policymaker who wants to know whether datacenters will raise bills in a given state should therefore ask what kind of utility serves the load, how capacity is procured, and what the tariff requires, before consulting any national average.


Pillar 4. State Regulatory Design Determines Whether Benefits Reach Households

Corporate investment does not translate into household affordability without a regulatory mechanism that carries it there. Utility revenues enter complex arrangements, and the distribution of any financial benefit depends on the approved revenue requirement, the rate structure, and the method of cost allocation. Indiana’s pending rate proposal places this issue squarely before its regulator, Virginia’s customer classification and direct-assignment order show how obligations can be established for major users, Pennsylvania’s proceeding emphasizes the assignment of datacenter-related expenses within a wholesale market, and Texas’s interconnection rule highlights financial security and credible development commitments. Each offers a different view of the same challenge, and all of them confirm that the presence of hyperscale investment is insufficient, since the financial structure must protect existing customers and any claimed dividend must be translated into an observable regulatory or billing outcome.


Pillar 5. Federal and State Policy Must Address Different Parts of the Same System

State regulators set retail tariffs and oversee local service, federal regulators oversee interstate transmission and wholesale markets, and regional grid organizations manage planning and operations, and large datacenters affect all of these at once. The Senate-approved bill attempts to address important elements of this problem, and its cost-allocation provisions would strengthen financial responsibility for covered loads if enacted and implemented, but its ultimate effect will depend on its final form, on rulemaking, and on its interaction with state authority. Consumer protection requires attention to both local infrastructure cost and regional market consequence, and a positive utility-level contribution is not enough if changes in the wider market impose larger costs on the same customers.


Pillar 6. Committed Load, Not Announced Load, Should Drive Investment

The sixth lesson concerns the quality of the demand signal itself. When signed contracts cover roughly a fifth of the datacenter load that utilities report, and when a grid operator’s realized demand runs a fifth below its year-old forecast, the greatest threat to households is not that datacenters will arrive and fail to pay but that infrastructure will be built for datacenters that never arrive at all. Take-or-pay agreements, collateral, aggregation rules, and regulatory review of contract reductions exist to convert forecasts into obligations, and capacity markets and transmission plans that rely on unscreened forecasts transmit speculative demand directly into household bills. The discipline of planning on committed load is therefore a precondition of the dividend and not a refinement of it.


Pillar 7. The Five-Layer AI Economy Needs a Measurable Public Benefit

The final lesson extends beyond utility regulation. Artificial intelligence expansion involves a chain of investments connecting energy, semiconductors, datacenters, models, and applications, and the economic value generated at the upper layers depends on physical infrastructure at the lower ones. The communities hosting that infrastructure participate in the economic process whether or not they own any part of the companies that capture the revenue, and the Load Dividend concept asks whether an identifiable portion of the financial benefit of expanding electricity demand can flow back to them through their electricity bills. It does not assume that every community must receive the same result, but it does require a credible method for distinguishing actual benefits from projected benefits, avoided costs, and marketing claims, and its strongest form would allow policymakers, utilities, corporations, and residents to examine a common set of financial facts, so that the public value of artificial intelligence infrastructure becomes something that can be evaluated and need no longer be taken on assertion.


Conclusion: Why the Load Dividend Defines a Different Economic Future for Artificial Intelligence Infrastructure

The proposed transformation of the former Breed Generating Station introduces a consequential possibility for the future of American artificial intelligence, which is that a major technology company may become more than a purchaser of electricity and may participate in financing new generation, support long-term procurement, and assume commercial obligations that shape how energy projects are developed. Amazon’s collaboration with AEP and X-energy demonstrates that direction, and it does so at an early stage: the facility’s contribution to the regional system, its financial performance, and its ultimate effect on customers depend on decisions and events that have not yet occurred. Indiana Michigan Power’s proposed $59 million customer benefit addresses a different part of the electricity economy, offering a direct test of whether growth in large customers can reduce the share of fixed costs recovered from everyone else. That result remains subject to regulatory approval, and it should never be attributed to nuclear generation that has not been built.

Together these developments form a useful research foundation because they connect two separate financial questions. The first asks whether large industrial customers can finance the infrastructure they require without imposing unreasonable risk on ordinary households, and the second asks whether their additional revenue can produce measurable benefits for those households. The Senate’s passage of the Bipartisan American Affordability and Jobs Act places both questions within a national policy discussion whose legal outcome is still uncertain, and the convergence of all three developments suggests that the political economy of artificial intelligence infrastructure is entering a more demanding phase. Technology companies can no longer assume that the announcement of a multibillion-dollar investment establishes that a project serves the public interest, since communities and regulators now ask for evidence about electricity costs, infrastructure responsibilities, environmental effects, and long-term financial risk. Utilities must show that their investment plans rest on credible demand and enforceable commitments, governments must distinguish infrastructure that serves the public broadly from infrastructure attributable to particular customers, and investors must judge whether the demand assumptions beneath new generation and transmission are commercially sustainable.

These requirements do not conflict with continued growth, and stronger financial discipline may well provide a more reliable foundation for the expansion of the Five-Layer AI Economy, in which advanced semiconductors depend on reliable electricity, datacenters depend on available energy infrastructure, foundation models depend on those facilities, and applications and agents depend on continuing access to computation. A sustainable system requires the costs supporting that chain to be allocated transparently. The contribution of the Load Dividend is to pose a further question that conventional discussions of cost allocation leave unasked: after a hyperscaler has paid the full costs and assumed the appropriate risks of its electricity service, can its participation also create an identifiable financial benefit for existing customers? If the answer is yes, communities would have reason to assess a datacenter not only by the electricity it consumes or the capital invested in it but by the net financial contribution it makes to the wider electricity system, whether in the form of reduced base rates, customer credits, shared infrastructure savings, or other independently verified benefits.

The terminology must nonetheless remain disciplined, and the discipline can be stated as a series of refusals. A corporate pledge is not a dividend, and a proposed rate reduction is not an implemented one. A nuclear development agreement is not proof of lower electricity costs, an avoided infrastructure charge is not necessarily a positive household saving, and a temporary reduction that depends on transferring costs into future years is not durable. The concept becomes meaningful only when the financial outcome is supported by transparent evidence, properly attributed costs, enforceable protections, and an observable benefit to existing customers. That is why the title fits the paper: Load names the large and growing electricity demand created by the infrastructure of artificial intelligence, and Dividend names the financial return to the existing customer base when that demand is converted into a properly structured contribution toward the system, so that the term connects the industrial expansion of the artificial intelligence economy to the distribution of its consequences and supplies a standard against which future projects can be judged.

The fundamental question for 2030 is therefore not how many gigawatts of new datacenter demand American utilities will accommodate. It is whether the financial arrangements used to accommodate those gigawatts leave existing customers better protected and, where possible, measurably better off. Indiana may become an important case through which that question is studied, Pennsylvania, Virginia, and Texas are developing regulatory responses that will provide comparisons, and federal legislation could add obligations and reporting standards if it is enacted. As artificial intelligence expands across energy, chips, datacenters, models, and autonomous applications, the economics of its physical foundation will increasingly shape the economics of the whole. A future in which large artificial intelligence customers pay their full attributable costs would represent a more disciplined allocation of infrastructure responsibility, and a future in which those customers also contribute to verified reductions in what existing households pay for electricity would represent something more. That additional, measurable public benefit, arriving on the monthly bill of a household in Sullivan County that will never use the computing power its county helps to produce, is the Load Dividend.


Endnotes and Sources:

[1]     American Electric Power. October 19, 2026. “AEP Partners with Amazon and X-energy to Explore Advanced Nuclear Energy in Indiana.”
https://www.aep.com/news/stories/view/12330/

[2]     X-energy. October 19, 2026. “X-energy, Amazon, American Electric Power Partner to Bring SMR Project to Indiana.”
https://investors.x-energy.com/news-releases/news-release-details/x-energy-amazon-american-electric-power-partner-bring-smr

[3]     Amazon. October 19, 2026. “Amazon Invests in Nuclear Energy Project Targeting Nearly 1 Gigawatt of Carbon-Free Energy for Indiana.”
https://www.aboutamazon.com/news/sustainability/amazon-nuclear-energy-indiana-x-energy-aep

[4]     Indiana Michigan Power. August 26, 2026. “Hoosiers Expected to See Savings as I&M Advances One of the Nation’s Largest Rate Reduction Plans.”
https://www.indianamichiganpower.com/company/news/view?releaseID=12207

[5]     Indiana Michigan Power. 2026. “Savings and Stability for Customers.” Public explanation of the proposed Customer Benefits Plan and existing large-load arrangements.
https://www.indianamichiganpower.com/cp/lower-rates/

[6]     Indiana Utility Regulatory Commission. 2026. “Pending Base Rate Cases,” including Indiana Michigan Power Cause No. 46454.
https://www.in.gov/iurc/docketed-cases/find-a-docketed-case/pending-base-rate-cases

[7]     Yahoo Finance. October 19, 2026. “Amazon Backs 12-Reactor Nuclear Plant at Indiana Coal Site.”
https://finance.yahoo.com/energy/articles/amazon-backs-12-reactor-nuclear-185738852.html

[8]     NPR. November 20, 2026. “Senate Passes Bipartisan Energy Bill Amid Environmental Concerns.”
https://www.npr.org/2026/11/20/g-s1-154579/senate-bipartisan-energy-bill

[9]     Congressional Research Service. November 2026. “Bipartisan American Affordability and Jobs Act of 2026: Summary of All Sections.” Report R49506.
https://www.everycrsreport.com/reports/R49506.html

[10]   Cowie-Haskell, Adam. Bipartisan Policy Center. November 2026. “Explaining the ‘Affordability’ in the Bipartisan American Affordability and Jobs Act.”
https://bipartisanpolicy.org/explainer/explaining-the-affordability-in-the-bipartisan-american-affordability-and-jobs-act/

[11]    Monitoring Analytics, LLC (Independent Market Monitor for PJM). July 28, 2026. “Market Monitor Report” to the PJM Members Committee; and “Analysis of the 2028/2029 RPM Base Residual Auction,” 2026.
https://www.monitoringanalytics.com/reports/Presentations/2026/IMM_MC_Market_Monitor_Report_20260728.pdf

[12]   International Energy Agency (Fatih Birol). April 10, 2025. “AI Is Set to Drive Surging Electricity Demand from Data Centres While Offering the Potential to Transform How the Energy Sector Works.”
https://www.iea.org/news/ai-is-set-to-drive-surging-electricity-demand-from-data-centres-while-offering-the-potential-to-transform-how-the-energy-sector-works

[13]   International Energy Agency. April 2025. “Energy and AI: Executive Summary.”
https://www.iea.org/reports/energy-and-ai/executive-summary

[14]   International Energy Agency. 2026. “Key Questions on Energy and AI: Executive Summary.”
https://www.iea.org/reports/key-questions-on-energy-and-ai/executive-summary

[15]   Smith, Sarah Josephine; Hubbard, Alex; Newkirk, Alexander; Ganeshalingam, Mohan; Holecek, Billie; Sartor, Dale A.; Mills, Michael; Shehabi, Arman. Lawrence Berkeley National Laboratory. June 2026. “United States Data Center Energy Usage Report: 2025 Update.”
https://eta-publications.lbl.gov/publications/united-states-data-center-energy-2025

[16]   International Monetary Fund. May 13, 2025. “AI Needs More Abundant Power Supplies to Keep Driving Economic Growth.” IMF Blog.
https://www.imf.org/en/Blogs/Articles/2025/05/13/ai-needs-more-abundant-power-supplies-to-keep-driving-economic-growth

[17]   Borenstein, Severin; Bushnell, James B. 2022. “Do Two Electricity Pricing Wrongs Make a Right? Cost Recovery, Externalities, and Efficiency.” American Economic Journal: Economic Policy 14(4): 80–110.
https://www.aeaweb.org/articles?id=10.1257/pol.20190758

[18]   Knittel, Christopher R. MIT Center for Energy and Environmental Policy Research. June 2026. “Smart Meters, Dumb Prices: Why Electricity Tariffs Are Becoming (More) Inefficient, Regressive, and Outdated.” CEEPR Research Commentary 2026-05.
https://ceepr.mit.edu/wp-content/uploads/2026/06/MIT-CEEPR-RC-2026-05.pdf

[19]   Norris, Tyler H.; Profeta, Tim; Patino-Echeverri, Dalia; Cowie-Haskell, Adam. Nicholas Institute for Energy, Environment & Sustainability, Duke University. February 2025. “Rethinking Load Growth: Assessing the Potential for Integration of Large Flexible Loads in US Power Systems.”
https://nicholasinstitute.duke.edu/publications/rethinking-load-growth

[20]  Wiser, Ryan. Lawrence Berkeley National Laboratory. December 2025. “Impact of Data Center Load Growth on Retail Electricity Prices.” Presentation, Michigan State University Institute for Public Policy and Social Research.
https://ippsr.msu.edu/sites/default/files/policy/presentations/wiser1225.pdf

[21]   Virginia State Corporation Commission. November 25, 2025. “In Biennial Review Ruling, SCC Creates New Class for Large-Scale Energy Users.”
https://www.scc.virginia.gov/about-the-scc/newsreleases/release/scc-issues-order-on-dev-biennial-review-2025/scc-rules-in-dev-biennial-review-case.html

[22]  Utility Dive. September 21, 2026. “Texas PUC Adopts Softened Rules on Data Center Interconnection.”
https://www.utilitydive.com/news/texas-puc-adopts-data-center-interconnection-rules/830899/

[23]  WRIC ABC 8News. August 6, 2026. “SCC Orders Virginia’s Data Centers to Cover the Costs of Transmission Infrastructure; Here’s What It Means for You.”
https://www.wric.com/news/virginia-news/scc-order-data-centers-pay-infrastructure/

[24]  Utility Dive. February 2025. Report on the IURC’s approval of Indiana Michigan Power’s large-load interconnection settlement.
https://utilitydive.com/news/indiana-iurc-large-load-interconnection-data-center-aep-amazon-google/740452

[25]   CNBC. October 29, 2026. “Amazon (AMZN) Q3 Earnings Report 2026.”
https://www.cnbc.com/2026/10/29/amazon-amzn-q3-earnings-report-2026.html

[26]  Walnut. November 2026. “AI Capex Tracker: Hyperscaler Capital Spending by Quarter (2026).” Compiled from company Forms 10-Q and earnings releases.
https://walnutinvest.com/stats/ai-capex-statistics

[27]   Wiser, Ryan; O’Shaughnessy, Eric; Barbose, Galen; Cappers, Peter; Gorman, Will. Lawrence Berkeley National Laboratory. 2025. “Factors Influencing Recent Trends in Retail Electricity Prices in the United States.” The Electricity Journal 38(4); with “Retail Electricity Price and Cost Trends: 2026 Edition” (with The Brattle Group).
https://doi.org/10.1016/j.tej.2025.107516

[28]  Millican, Robin; Arent, Douglas J.; Sandalow, David. Center on Global Energy Policy, Columbia University SIPA. June 23, 2026. “Electricity Affordability and Load Growth: Diagnosing and Fixing the Problem.”
https://www.energypolicy.columbia.edu/publications/electricity-affordability-and-load-growth-diagnosing-and-fixing-the-problem/

[29]  Bistline, John. Electric Power Research Institute. January 2026. “Win-Win Watts: When Can Data Centers, Efficient Electrification, and New Loads Lower Electricity Prices?”
https://esca.epri.com/products/3002034619/

[30]  Watten, Asa; Bistline, John; Blanford, Geoff. Electric Power Research Institute. June 2026. “Have Data Centers Raised Your Electric Bill? Causal Evidence from the United States.”
https://esca.epri.com/_astro/2026-jun-have-data-centers-raised-your-electric-bill.CVr2j5ls.pdf

[31]   Ryu, Angela; Hiatt, Shon R. University of Southern California, Marshall School of Business, Zage Business of Energy Initiative. 2026. “Data Center Entry and U.S. Electricity Prices: From the Wholesale Market to the Power Bill,” as summarized by the National Center for Energy Analytics.
https://energyanalytics.org/research/dont-blame-the-data-centers

[32]  Meeks, Robyn; Pless, Jacquelyn; Qi, Zhiyuan; Wang, Zhenxuan. MIT Center for Energy and Environmental Policy Research. June 2026 (revised October 2026). “Who Pays for Growth? Evidence from Data Centers and the Grid.” CEEPR Working Paper 2026-13.
https://ceepr.mit.edu/workingpaper/who-pays-for-growth-evidence-from-datacenters-and-the-grid/

[33]  Kay, Owen; Reaser, Robert; Taylor, Reid. Federal Reserve Bank of Dallas. March 2026 (revised October 2026). “Processing Power: The Effect of Data Centers on Wholesale Electricity Markets.” Working Paper 2606.
https://www.dallasfed.org/~/media/documents/research/papers/2026/wp2606r1.pdf

[34]  Bogmans, Christian; Gomez-Gonzalez, Patricia; Ganpurev, Ganchimeg; Melina, Giovanni; Pescatori, Andrea; Thube, Sneha D. International Monetary Fund. April 2025. “Power Hungry: How AI Will Drive Energy Demand.” IMF Working Paper 25/81.
https://www.imf.org/en/publications/wp/issues/2025/04/21/power-hungry-how-ai-will-drive-energy-demand-566304

[35]   Rana, Vishrut; Kaps, Christian; Netessine, Serguei. Harvard Business School and The Wharton School. September 2026. “The Cost of Intelligence: Impacts of AI Data Centers on U.S. Electricity Prices,” as summarized in HBS Working Knowledge, November 12, 2026.
https://www.library.hbs.edu/working-knowledge/blaming-data-centers-for-higher-electric-bills-its-not-that-simple

[36]  E&E News by POLITICO. 2026. “Trump Wants Data Centers to Pay for Power. Will That Lower Electricity Prices?” Interview with Liang Min, Stanford University.
https://www.eenews.net/articles/trump-wants-data-centers-to-pay-for-power-will-that-lower-electricity-prices/

[37]   World Bank Group. August 2026. “World Development Report 2026: The Promise of Artificial Intelligence.”
https://documents1.worldbank.org/curated/en/099090426145529770/pdf/P512385-5636baf0-fd26-410d-90c1-cf08125fe298.pdf

[38]  Guterres, António. United Nations Secretary-General. July 22, 2025. Remarks on the energy transition and AI data centers, as reported by Bloomberg / Insurance Journal, “AI Should Run on 100% Renewable Energy by 2030, UN Chief Says.”
https://www.insurancejournal.com/news/international/2025/07/23/832790.htm

[39]  Martin, Eliza; Peskoe, Ari. Harvard Law School Electricity Law Initiative. March 2025. “Extracting Profits from the Public: How Utility Ratepayers Are Paying for Big Tech’s Power.”
https://eelp.law.harvard.edu/wp-content/uploads/2025/03/Harvard-ELI-Extracting-Profits-from-the-Public.pdf

[40]  Louisiana Illuminator / States Newsroom. March 6, 2025. “Power for Data Centers Could Come at ‘Staggering’ Cost to Consumers.” Interview with Ari Peskoe.
https://lailluminator.com/2025/03/06/data-centers-2/

[41]   Borenstein, Severin. Energy Institute at Haas, University of California, Berkeley / NPR. November 6, 2025. “Is Your Electric Bill Going Up? AI Is Partly to Blame.”
https://haas.berkeley.edu/energy-institute/about/in-the-media/severin-borenstein-on-the-public-cost-of-ai/

[42]  Dizikes, Peter. MIT News. June 26, 2026. “How Data Centers Can Better Manage Energy Use.” Reporting on Senga, Wang, and Knittel, “Flexible Data Centers Reduce Power System Costs but Can Increase Emissions,” iScience (2026).
https://sustainability.mit.edu/article/how-data-centers-can-better-manage-energy-use

[43]  Stanford Report. June 2026. “Why Are Electricity Costs Rising?” Featuring Michael Wara and Michael Mastrandrea, Stanford University.
https://news.stanford.edu/stories/2026/06/electricity-costs-rising-experts-solutions

[44]  Indiana Capital Chronicle. November 26, 2024. “Ratepayer Advocates Hail ‘Landmark’ Settlement with Data Centers, Utility Company.”
https://indianacapitalchronicle.com/2024/11/26/ratepayer-advocates-hail-landmark-settlement-with-data-centers-utility-company/

[45]   Taft Stettinius & Hollister LLP. 2026. “Overview of Indiana House Enrolled Act 1002 (2026): Concerning Electric Utility Affordability.”
https://www.taftlaw.com/news-events/law-bulletins/overview-of-indiana-house-enrolled-act-1002-2026-concerning-electric-utility-affordability/

[46]  WSBT 22 Investigates. 2026. “Who Pays for Indiana’s Data Centers? The Fight Over Your Electric Bill.” Interview with Ben Inskeep, Citizens Action Coalition.
https://wsbt.com/news/22-investigates/22-investigates-data-center-costs-and-your-bill-consumer-watchdog-group-scale-boom-concern-communities-local-facilities-state-electricity-utilities-citizens-state-scale-paying-indiana

[47]   American Electric Power / Indiana Michigan Power. February 19, 2025. “Indiana Michigan Power Receives Order in Large Load Settlement.”
https://www.aep.com/news/stories/view/10037/

[48]  American Electric Power. October 28, 2026. “AEP Reports Third-Quarter 2026 Earnings, Expands Capital Plan and Increases Growth Outlook.” Form 8-K, Exhibit 99.1.
https://www.sec.gov/Archives/edgar/data/0000004904/000000490426000069/a3q20268kpressreleaseex991.htm

[49]  The Journal Gazette (Fort Wayne). October 2026. “Data Center Boom’s Benefits Many.” Guest column quoting South Bend Mayor James Mueller.
https://www.journalgazette.net/opinion/columnists/data-center-booms-benefits-many/article_698bd8ad-3b78-4388-8f3f-8553bb9bfaab.html

[50]  The Journal Gazette (Fort Wayne), Editorial Board. November 2026. “Consumers Have a Chance This Monday to Get Questions in Front of IURC.”
https://journalgazette.net/opinion/consumers-have-a-chance-this-monday-to-get-questions-in-front-of-iurc/article_2f562e9c-8f80-4893-a27b-986de0b639f9.html

[51]   U.S. Energy Information Administration. November 10, 2026. “Short-Term Energy Outlook: Electricity, Coal, and Renewables.”
https://www.eia.gov/outlooks/steo/report/elec_coal_renew.php

[52]   X-Energy, Inc. November 2026. “X-energy Reports Third Quarter 2026 Results.” Form 8-K, Exhibit 99.1.
https://www.sec.gov/Archives/edgar/data/0002088896/000119312526456633/xe-ex99_1.htm

[53]   U.S. Department of Energy, Office of Nuclear Energy. November 2026. “Dow, X-energy Receive Construction Permit for Advanced Nuclear Project in Texas.”
https://www.energy.gov/ne/articles/dow-x-energy-receive-construction-permit-advanced-nuclear-project-texas

[54]   Amazon (Andy Jassy). October 29, 2026. “Q3 Earnings: Andy Jassy on Why AWS Growth Keeps Accelerating.”
https://www.aboutamazon.com/news/company-news/aws-growth-ai-demand-q3-2026-earnings

[55]   The White House. March 4, 2026. “Fact Sheet: President Donald J. Trump Advances Energy Affordability with the Ratepayer Protection Pledge.”
https://www.whitehouse.gov/fact-sheets/2026/03/fact-sheet-president-donald-j-trump-advances-energy-affordability-with-the-ratepayer-protection-pledge/

[56]   The White House. 2026. “Ratepayer Protection Pledge.” Official pledge page.
https://www.whitehouse.gov/ratepayer-protection-pledge/

[57]   The White House. July 2026. “President Trump’s Ratepayer Protection Pledge Secures American AI Dominance, Protects Consumers.”
https://www.whitehouse.gov/releases/2026/07/president-trumps-ratepayer-protection-pledge-secures-american-ai-dominance-protects-consumers/

[58]  Salata Institute for Climate and Sustainability, Harvard University. March 17, 2026. Analysis of the Ratepayer Protection Pledge and data center grid costs.
https://salatainstitute.harvard.edu/data-centers-ai-artificial-intelligence-grid-permitting-transmission-electricity-energy

[59]   Federal Energy Regulatory Commission. 2026. “Cost Transparency and Customer Protection Initiatives in FERC’s June 2026 Show Cause Orders for Large Load Interconnections.”
https://www.ferc.gov/cost-transparency-customer-protection

[60]  Akin Gump Strauss Hauer & Feld LLP. June 2026. “FERC Issues Landmark Show Cause Orders on Large Load Interconnection.”
https://www.akingump.com/en/insights/blogs/speaking-energy/ferc-issues-landmark-show-cause-orders-on-large-load-interconnection

[61]   Millican, Robin; Cavanaugh, Jack Andreasen; Kaufman, Noah; et al. Center on Global Energy Policy, Columbia University SIPA. November 2026. “What the Bipartisan American Affordability and Jobs Act Means for the U.S. Energy System.”
https://www.energypolicy.columbia.edu/publications/what-the-bipartisan-american-affordability-and-jobs-act-means-for-the-u-s-energy-system/

[62]  King, Ben; Bower, Galen; Case, Parker; Kolus, Hannah; Larsen, John; Pastorek, Nathan; Wiens, Nathan. Rhodium Group. November 2026. “Estimating the US Energy System Impacts of the Bipartisan American Affordability and Jobs Act.”
https://rhg.com/research/baaja-us-energy-system-impacts/

[63]  Deseret News. November 20, 2026. “Senate Passes Sen. Lee’s Permitting Bill in Bipartisan Win.”
https://www.deseret.com/politics/2026/11/20/permitting-bill-passes-senate-with-support-from-sen-lee-gov-cox/

[64]  Heinrich, Martin. U.S. Senate Committee on Energy and Natural Resources (Democratic News). November 2026. Statement on Senate passage of the Bipartisan American Affordability and Jobs Act.
https://www.energy.senate.gov/press/dem/heinrich-senate-passage-of-landmark-bipartisan-permitting-deal-will-lower-electricity-bills-hold-data-centers-accountable-create-over-one-hundred-thousand-jobs-reduce-carbon-emissions/

[65]   PJM Interconnection. July 14, 2026. “PJM Capacity Auction Procures 138,318 MW of Generation Resources as Work Continues To Address Growing Electricity Demand.”
https://insidelines.pjm.com/pjm-capacity-auction-procures-138318-mw-of-generation-resources-as-work-continues-to-address-growing-electricity-demand/

[66]  Latitude Media / Latitude Intelligence. November 2026. “Quantifying the State of AI Power Development Post-Midterms.”
https://www.latitudemedia.com/news/quantifying-the-state-of-ai-power-development-post-midterms/

[67]   Latitude Media. November 2026. “The US Power Market’s Forecasting Problem Comes at a Cost to Ratepayers.”
https://www.latitudemedia.com/news/the-us-power-markets-forecasting-problem-comes-at-a-cost-to-ratepayers/

[68]  Commonwealth of Pennsylvania, Office of Governor Josh Shapiro. October 16, 2026. “Governor Shapiro Files Intervention Petition in PUC Case to Protect PPL Customers from Higher Electricity Costs and Hold Data Center Developers Accountable.”
https://www.pa.gov/governor/newsroom/2026-press-releases/gov–shapiro-protects-customers-and-holds-data-center-developers

[69]  Pennsylvania Department of Community and Economic Development. 2026. “The Governor’s Responsible Infrastructure Development (GRID) Requirements.”
https://dced.pa.gov/business-assistance/data-center-resources/grid-requirements/

[70]  Riley, Kim. Daily Energy Insider. October 20, 2026. “PPL Says Customer Protections Already Exist as Shapiro Seeks Stronger Data Center Rules.”
https://dailyenergyinsider.com/news/54157-ppl-says-customer-protections-already-exist-as-shapiro-seeks-stronger-data-center-rules/

[71]   Cardinal News. November 26, 2025. “Regulators Approve Dominion Energy Rate Increase.”
https://cardinalnews.org/2025/11/26/regulators-approve-dominion-energy-rate-increase/

[72]   Office of Governor Abigail Spanberger. August 5, 2026. “Following Governor Spanberger’s Urging, SCC Requires Data Centers to Pay for New Transmission Infrastructure.”
https://www.governor.virginia.gov/newsroom/news-releases/2026/august-releases/name-1122057-en.html

[73]   Office of Governor Abigail Spanberger. September 18, 2026. “Full Remarks: Governor Spanberger Unveils Nation’s Most Comprehensive & Aggressive Data Center Accountability Standards, Signs Executive Order.”
https://www.governor.virginia.gov/newsroom/news-releases/2026/september-releases/name-1123705-en.html

[74]   Virginia Department of Taxation. 2026. “2026 Legislative Summary,” including the Data Center Electricity Consumption Tax.
https://www.tax.virginia.gov/sites/default/files/inline-files/2026-legislative-summary.pdf

[75]   Texas Legislature. 2025. Senate Bill 6, 89th Regular Session. Enrolled bill summary.
https://capitol.texas.gov/billlookup/BillSummary.aspx?Bill=SB6&LegSess=89R

[76]   Texas AI Docket. October 2026. “PUCT Adopts the SB 6 Large Load Interconnection Rule, Requires $50,000 per Megawatt of Security and Drops the Interconnection Fee” (16 TAC §25.194, Project 58481).
https://texasaidocket.com/item/tx-2026-0187/

[77]   Energy and Environmental Economics, Inc. (E3). May 2026. “Understanding the Drivers of Rising Electricity Rates and the Role of Data Centers.”
https://www.ethree.com/wp-content/uploads/2026/05/Understanding-the-Drivers-of-Rising-Electricity-Rates-and-the-Role-of-Data-Centers_E3-2026.pdf

[78]  Indiana Utility Regulatory Commission. April 29, 2026. Order in Cause No. 46276 (Indiana Michigan Power customer-specific contract with Google LLC).
https://secure.in.gov/iurc/files/ord_46276_042926.pdf

[79]   NPR. October 1, 2026. “AI Data Centers: How Much Are Ratepayers on the Hook For?” Featuring Lucy Qiu, University of Maryland.
https://www.npr.org/2026/10/01/nx-s1-5984949/data-centers-ai-ratepayers-congress