Introduction: Two Electric Economies Begin to Separate
For most of the history of the American electric utility, an additional customer was something to celebrate. A new factory arriving on the edge of town, a subdivision rising out of former farmland, a hospital expanding its wings, a shopping center opening its doors, an office park filling with tenants — each of these represented more kilowatt-hours flowing across infrastructure whose costs could be spread across a progressively larger customer base, and therefore each of these made the entire electric system marginally more economical for everyone connected to it. Around that assumption, regulators and utilities constructed one of the most durable economic compacts in American industrial history: utilities would make prudent investments in generation, transmission, substations, and distribution networks; those investments would enter the regulated rate base; and customers would gradually repay the capital, the operating costs, and an authorized return on equity through their monthly bills, decade after decade, in payments so small and so predictable that most households never thought about the century of financial engineering embedded in the envelope that arrived from the power company. Growth, properly managed, made the whole system cheaper. That was the promise, and for roughly one hundred years, it was mostly true.
Artificial intelligence is now challenging that compact more profoundly than any development since electrification itself, because artificial intelligence has introduced a species of electricity customer the compact was never designed to accommodate: a single private campus that can request as much power as a mid-sized American city, that can arrive in a regulator’s queue alongside dozens of similar requests within a span of months, that may be built in phases or not built at all, and whose underlying computing hardware will be replaced several times before the transmission line constructed to reach it has recovered even half of its cost. The numbers involved are no longer hypothetical. The four largest hyperscalers — Microsoft, Amazon, Alphabet, and Meta — disclosed during their Q2-2026 earnings season that their combined capital expenditure for calendar 2026 would reach approximately $760 billion, up from roughly $413 billion in 2025, with the overwhelming share directed at AI datacenters, accelerators, networking, and power.[48] Goldman Sachs Research now projects that these four companies alone will deploy on the order of $5.3 trillion in capital expenditure between fiscal 2025 and fiscal 2030.[49] Every dollar of that spending eventually lands on a physical site, and every physical site eventually asks a utility, a grid operator, and a state regulator a single deceptively simple question: who will pay for the wires?
Virginia offers perhaps the clearest early glimpse of what comes next, because Virginia is already home to the world’s most concentrated datacenter market and therefore encounters every stage of this dilemma first. Yet the issue confronting its State Corporation Commission is no longer simply whether Virginia should permit more datacenters. The harder question — the question that will define the political economy of American electricity for the remainder of this decade — is becoming: who should financially guarantee the electricity infrastructure required to serve them?
The Virginia State Corporation Commission answered part of that question on November 25, 2025, when it issued its final order in Dominion Energy Virginia’s biennial review and created a distinct GS-5 rate class for customers demanding 25 megawatts or more — a category that captures hyperscale datacenters almost by definition, and that most of the roughly 450 datacenters already operating in Dominion’s territory are expected to fall within.[1, 6] Beginning January 1, 2027, qualifying large-load customers will be treated as a legally separate population from traditional customers. New large-load customers face minimum contractual commitments of fourteen years, with a load-ramp period of no more than four years embedded within that term. They must pay at least 85 percent of contracted transmission and distribution demand, and at least 60 percent of contracted generation demand, every month, regardless of whether their actual electricity consumption reaches those levels.[1, 2] Customers who cease operations or default during the contract term owe an exit fee covering the outstanding minimum charges across the remaining contract duration, and customers without sufficient credit can be required to post collateral — collateral that reporting in mid-2026 placed at approximately $1.5 million for every megawatt of contracted capacity, a figure that functions, in effect, as the regulator’s own written estimate of how much stranded-cost risk a failed hyperscale project would otherwise transfer to ordinary households.[4, 5] The Commission has explicitly described these safeguards as mechanisms for minimizing cost shifting from hyperscale facilities to other electricity customers.[2]
Virginia has gone further still. In a July 2026 proceeding concerning Dominion’s Rider T1 — the line-item charge through which the utility recovers roughly $1.5 billion in transmission investment — the Commission heard arguments from Governor Abigail Spanberger’s administration, from Meta, Google, Amazon, and Microsoft, and from its own staff about how transmission projects driven by datacenter development should be allocated.[8] On July 31, 2026, the SCC found that large-load datacenter customers must pay for transmission facilities constructed solely to serve them, and in early August it ordered Dominion to develop a tariff mechanism for directly assigning those costs rather than blending them into charges recovered broadly from all customers.[9, 10] The Governor’s office pressed for a “but for” cost-causation standard — the principle that infrastructure which would not exist but for a specific large-load customer should be financed by that customer — and the Commission’s own staff attorney, testifying in the Rider T1 case, characterized the status quo bluntly.
“There remains a glaring cross-class subsidization occurring to the benefit of new GS-5 customers.”
— Andrew Major, Attorney, Virginia State Corporation Commission Staff [8]
That seemingly technical accounting dispute has become important enough for one of the world’s largest technology companies to fight in court. In late August 2026, Microsoft filed a notice of appeal with the Supreme Court of Virginia challenging the SCC’s July order requiring datacenter developers to bear upfront costs for transmission infrastructure constructed specifically for their facilities — the first time a Big Tech company has formally contested a ruling by the regulator overseeing the world’s largest datacenter hub.[11, 12] The appeal is particularly striking because it arrived less than six months after Microsoft joined Amazon, Google, Meta, OpenAI, Oracle, and xAI at the White House to sign the Ratepayer Protection Pledge, publicly committing to build, bring, or buy the energy their facilities require and to pay the full cost of that energy and its supporting infrastructure.[28, 29] Microsoft insists that it continues to accept the principle that datacenters should pay the costs they cause; its stated concern is with the mechanics rather than the principle, framed in language that will recur throughout this paper.
“Our appeal seeks to ensure that cost allocation appropriately applies cost-causation principles.”
— Microsoft, statement on its Virginia Supreme Court appeal, September 2026 [11]
The disagreement exposes something much bigger than the allocation of a transmission line in Northern Virginia. It asks whether an electricity system developed around millions of relatively predictable households and businesses can continue operating under the same financial architecture when an individual AI campus can request hundreds of megawatts — in some cases more electricity than entire cities once required — and when the queue of such requests in a single utility territory can exceed the utility’s entire historical peak load several times over. Dominion’s own disclosures illustrate the scale: by the end of 2025 the company reported more than 48 gigawatts of datacenter capacity in various stages of contracting, against a system whose total peak for all customers is roughly 25 gigawatts, and a 2026 SCC hearing revealed that the combined load of approved and queued datacenter projects in its territory surpassed 70 gigawatts.[44, 46]
Ohio is already asking the same question through a different institutional design. AEP Ohio’s dedicated Data Center Tariff, adopted by the Public Utilities Commission of Ohio on July 9, 2025 and effective since July 23, 2025, places new datacenter developments above 25 megawatts into a specialized process: projects enter formal load-study tranches, sign binding commitments with minimum monthly demand charges of at least 85 percent of contracted capacity, demonstrate financial viability, provide collateral, and accept exit fees if their projects are canceled — rather than simply joining an undifferentiated utility planning forecast.[13, 14] The results have been extraordinary as a piece of revealed-preference economics: before the tariff, developers had submitted requests for roughly 30,000 megawatts of new load in AEP Ohio’s territory; once binding financial commitments were required, only about 13,000 megawatts of projects were willing to pay for a formal engineering study, and by February 12, 2026, binding contracts under the new tariff covered 5,642 megawatts, on top of 12,219 megawatts contracted before the tariff took effect.[15, 16] In August 2026, Ohio regulators added a further safeguard, requiring datacenter customers to give AEP Ohio 180 days’ notice before joining the grid on default service, explicitly so that the utility can procure adequate supply — at the datacenter’s cost — without burdening its other 1.5 million customers.[17]
Something fundamental is therefore changing. The United States is not merely constructing more electricity infrastructure for artificial intelligence. It is beginning to construct a separate financial architecture around the electricity consumed by artificial intelligence — a parallel system of rate classes, contract durations, collateral schedules, verification audits, exit fees, and direct cost assignments that exists alongside, but increasingly apart from, the traditional regulated compact serving homes and ordinary businesses.
I call this emerging phenomenon Ratebase Separation.
Ratebase Separation occurs when regulators conclude that the scale, concentration, uncertainty, infrastructure requirements, and abandonment risk of hyperscale AI loads are sufficiently different from conventional electricity consumption that the associated financial obligations must increasingly be isolated from the traditional utility rate base. Instead of asking households and small businesses to collectively absorb the risks of speculative transmission lines, oversized substations, new power plants, or abandoned capacity, regulators create distinct rate classes, minimum-load obligations, collateral requirements, exit fees, refundable deposits, queue deposits, direct assignment rules, and long-term take-or-pay structures for the customers creating those requirements — so that the entity that causes an extraordinary infrastructure obligation is also the entity that guarantees it.
This development belongs directly inside the Five-Layer AI Economy, the analytical framework that organizes my broader body of work. Layer 1 — energy — is not merely an upstream input supporting Layer 2 chips, Layer 3 datacenters, Layer 4 models, and Layer 5 applications and agents. Electricity regulation is becoming an economic governor on how quickly every subsequent layer can expand. The ability of Nvidia, Microsoft, OpenAI, Anthropic, Google, Meta, Amazon, xAI, and the next generation of AI companies to scale intelligence may increasingly depend upon something far less glamorous than GPUs or transformer architectures: whether regulators are willing to place billions of dollars of grid infrastructure into the public rate base, or whether they will require AI companies to finance it themselves, upfront, under contract, and against collateral. That distinction — socialized rate base versus privatized guarantee — could become one of the defining infrastructure questions of the 2027–2030 AI economy, and it is the subject of this paper.
Why I Chose the Title “Ratebase Separation”
I chose the phrase Ratebase Separation because the emerging conflict is no longer adequately described by the vocabulary that has dominated public discussion so far — phrases like “datacenter electricity consumption,” “the AI power crunch,” or “ratepayer backlash.” Those phrases describe symptoms. The deeper structural change concerns the financial boundary surrounding the regulated electric system itself. For more than a century, American utilities have pooled the great majority of their infrastructure costs through a common rate base and recovered those costs over decades from broad customer populations, on the theory that shared infrastructure serves shared interests and that diversification across millions of customers makes any single customer’s departure financially trivial. Hyperscale AI campuses break that theory at its foundation, because the investment triggered by a single customer can involve enormous transmission lines, dedicated substations, incremental generation capacity, and regional grid upgrades whose economic lives — forty years or more — can dramatically exceed the commercial life of the AI hardware inside the facility, and in some cases even the corporate commitment that justified the project in the first place. Virginia’s separate GS-5 classification, Ohio’s dedicated Data Center Tariff, Indiana’s large-load contractual protections, Pennsylvania’s statewide model tariff, and Texas’s increasingly formal verification and interconnection regime all point in the same direction: regulators are beginning to build a financial firewall around extraordinary loads.[1, 13, 18, 21, 25]
The word Separation therefore matters as much as the word ratebase. The states examined in this paper are not rejecting AI investment; in most cases their governors are actively courting it, and their utilities are eager to serve it because serving it is enormously profitable. What the states are separating is the benefit of attracting datacenters from the obligation of ordinary households to underwrite them. The emerging principle is simple to state even though it is difficult to implement: the economic actor creating an extraordinary infrastructure requirement should increasingly carry the financial risk associated with that requirement. If that principle spreads nationally — and the evidence assembled in Sections 3 through 5 suggests it is spreading with remarkable speed — then two overlapping electricity economies will emerge in the United States. One will serve conventional homes, businesses, and industry through the traditional regulatory compact of shared rate base and socialized cost recovery. The other will serve hyperscale compute through specialized tariffs, long-duration take-or-pay contracts, collateral, direct infrastructure contributions, flexible-load agreements, and private or quasi-private power arrangements that sit closer to project finance than to classical utility regulation. That transformation — the drawing of a legal and financial line between the household electricity bill and the AI factory — is precisely what the title Ratebase Separation is intended to capture.
There is also a scholarly reason for the title. The most influential academic work on this subject over the past two years — above all the Harvard Law School Electricity Law Initiative’s March 2025 study, Extracting Profits from the Public: How Utility Ratepayers Are Paying for Big Tech’s Power, by Eliza Martin and Ari Peskoe — documented in forensic detail how the absence of separation operates: through confidential special contracts approved in short and conclusory regulatory orders, through cost allocation formulas designed for a different era, and through utility incentives that reward capital deployment regardless of who ultimately pays for it.[38] Martin and Peskoe examined roughly forty state proceedings and warned that, without systematic changes to prevailing ratemaking practice, utilities would face powerful incentives to profit from datacenters by building aggressively and then shifting costs to captive ratepayers, concluding that the industry’s prevailing approach of attracting datacenters through discounted special deals could not endure.
“The industry’s current approaches of luring data centers with discounted contracts or lopsided tariffs is unsustainable.”
— Eliza Martin and Ari Peskoe, Harvard Law School Electricity Law Initiative [38]
Ratebase Separation names what comes after that unsustainable phase. It is the institutional response to the problem Harvard identified: the deliberate construction, state by state, of a second financial architecture in which hyperscale load is contracted, collateralized, verified, and directly assigned rather than quietly socialized. The remainder of this paper explains where that architecture came from, how five states are building it in five different ways, what financial instruments it consists of, what a national framework could look like, and what its emergence teaches us about the political economy of the Five-Layer AI Economy.
Section 1: The Traditional Utility Rate Base — The Economic Compact AI Is Disrupting
Before it is possible to understand why five American states are now constructing specialized financial regimes for AI datacenters, it is necessary to understand, with some precision, the machine those regimes are being carved out of. The regulated utility rate base is one of the least discussed and most consequential financial institutions in the American economy. It determines how roughly $3 trillion of long-lived electrical infrastructure is financed, who bears the risk when demand forecasts prove wrong, and how the cost of the physical grid is divided among more than 140 million customer accounts. The AI infrastructure boom is not colliding with a market; it is colliding with a century-old regulatory compact, and the nature of that compact explains both why the collision is happening and why the response — Ratebase Separation — is taking the specific institutional forms documented in this paper.
1.1 What the Rate Base Actually Is
The regulated utility model rests on a straightforward exchange. A utility receives an exclusive franchise to serve a territory, and in return it accepts an obligation to serve every customer within it and submits its prices to regulatory control. To meet its obligation, the utility constructs assets that regulators determine are prudent and useful: generation facilities, substations, transformers, high-voltage transmission infrastructure, distribution lines, meters, control systems, and the accumulating layers of equipment that convert fuel and sunlight into reliable voltage at the wall socket. Eligible investment — investment found to be prudently incurred and “used and useful” in serving customers — becomes part of the utility’s regulated rate base. The utility subsequently recovers depreciation on that capital, its operating expenses, and an authorized return on invested equity from customers over the useful life of the assets, which for transmission and generation can span four decades or longer. In Dominion’s November 2025 biennial review, for example, the Virginia SCC authorized a return on equity of 9.8 percent, modestly above the prior 9.7 percent and well below the 10.4 percent the company had requested — a reminder that the rate of return itself, multiplied across tens of billions of dollars of rate base, is among the most heavily litigated numbers in American administrative law.[7]
The rate base is therefore not merely an accounting category, and this point deserves emphasis because nearly everything in the AI-electricity debate follows from it. The rate base determines who finances infrastructure: customers do, collectively, through decades of small monthly increments, with the utility acting as a financing intermediary whose shareholders are compensated for deploying capital. The rate base determines who assumes long-duration risk: customers do, again collectively, because once an investment is found prudent, its costs remain recoverable even if the demand that justified it evolves in unexpected directions. And the rate base determines who ultimately pays when forecasts prove wrong — which, in a system built on forty-year assets and two-year rate cases, is the question that matters most.
1.2 Cost Socialization Worked When Loads Were Relatively Diversified
For most of the twentieth century, the socialization of infrastructure costs through a common rate base was not merely tolerable; it was genuinely efficient, and it is worth pausing to appreciate why, because the efficiency of the old arrangement explains why regulators are reluctant to abandon it wholesale even now. Historically, infrastructure investment served thousands or millions of customers simultaneously. A transmission improvement might strengthen reliability across multiple communities and unlock lower-cost generation for an entire region. A new generating unit might serve residential air conditioning in the afternoon, commercial lighting in the evening, and industrial processes overnight, its costs divided among customer classes whose consumption patterns conveniently interleaved. Diversification did the quiet work of insurance: because no single customer represented more than a sliver of the system, the disappearance of any one customer — a factory closing, a mill relocating — rarely eliminated the economic justification for regional infrastructure. Forecast errors at the level of individual customers washed out in the aggregate. The law of large numbers was, in a very real sense, the balance sheet of the American grid.
Hyperscale AI changes this relationship not by degree but by kind. When a single customer’s requested capacity rivals the peak demand of a major metropolitan area, the law of large numbers stops operating. The customer is no longer a statistical event inside a diversified portfolio; the customer is the portfolio. And when that customer’s future consumption depends on the trajectory of a technology whose economics are being rewritten every eighteen months — on chip efficiency, model architecture, inference demand, and the capital-markets appetite for AI itself — the utility’s forecast risk is no longer actuarial. It is speculative, concentrated, and enormous.
1.3 From Incremental Load to Lumpy Load
Traditional electricity planning assumes gradual change. Utility load forecasts were built for a world in which demand grew — or, for two decades after 2005, did not grow — in fractions of a percentage point per year, driven by population, appliance efficiency, and macroeconomic cycles that unfolded slowly enough for integrated resource plans to track them. AI campuses arrive differently. A datacenter developer can request hundreds of megawatts at a particular geographic node, sized not to any local economy but to the global compute plans of a handful of technology companies. Multiple developers can request service simultaneously, producing gigawatts of prospective demand within a relatively small area — often clustered around the same fiber routes, water resources, and substations. Dominion Energy told investors in early 2025 that typical individual campus requests had escalated from roughly 30 megawatts to between 300 megawatts and several gigawatts in the span of a few years, and by mid-2026 an SCC hearing revealed that the utility had approved 111 datacenter projects totaling roughly 27 gigawatts for connection through 2031 while 220 more waited in a queue whose combined load exceeded 70 gigawatts — against a system whose all-customer peak is approximately 25 gigawatts.[44, 46] The Lawrence Berkeley National Laboratory’s congressionally mandated assessments frame the national picture: datacenters consumed about 4.4 percent of total U.S. electricity in 2023, are projected to consume between 6.7 and 12 percent by 2028, and under LBNL’s 2025 update could account for approximately 11.8 percent of U.S. electricity by 2030, with scenarios ranging from 9.5 to 15.3 percent.[36, 37]
The relevant policy question that emerges from this lumpiness is deceptively simple: should speculative requested capacity be treated as though it were ordinary future electricity demand? Under the traditional compact, the answer was effectively yes — a request for service was a request for service, and the utility’s obligation to serve translated requests into planning assumptions and planning assumptions into rate-based investment. Under conditions of hyperscale AI development, that translation becomes dangerous, because the same developer may file duplicate requests across multiple utility territories, because financing and tenants may not yet exist, and because the difference between a request and a commitment can amount to billions of dollars of infrastructure that someone must eventually pay for. UC Berkeley energy economist Severin Borenstein, speaking at the Technology Policy Institute’s Aspen Forum in August 2026, warned regulators directly that interconnection queues have become unreliable measures of real demand, describing developers filing parallel requests across multiple territories in the hope of securing power wherever it becomes available first, and estimating that datacenter-driven cost pressures could raise retail electricity prices by two to three cents per kilowatt-hour — a 10 to 20 percent increase — across affected states if the buildout is financed carelessly.[43]
1.4 The Cost-Causation Principle
Utility regulation has long embraced some version of the idea that costs should follow causation — that the customers whose service requirements cause the system to incur an expense should be the customers who pay for it. In ordinary times, cost causation is an allocation doctrine applied at the level of broad customer classes, implemented through load studies and demand charges, and contested mainly by industrial intervenors seeking marginally better allocations at rate cases. Artificial intelligence transforms cost causation from an allocation doctrine into a constitutional question for the entire regulatory system, because the stakes of applying it — or failing to apply it — have grown by orders of magnitude. If a billion-dollar transmission project would not be needed absent a particular 500-megawatt campus, should millions of households contribute to financing it? If the datacenter subsequently delays construction, downsizes, changes its computing architecture, relocates its workloads, self-generates its electricity, or simply closes, who owns the stranded infrastructure? These are precisely the questions the Virginia SCC confronted in the Rider T1 proceeding, where Governor Spanberger’s office urged a “but for” standard under which infrastructure that exists only because of a large-load customer is paid for by that customer, and where the Commission ultimately ordered Dominion to develop a direct-assignment tariff for customer-specific transmission facilities.[8, 9] Ratebase Separation begins with the proposition that extraordinary load requires unusually clear cost causation — and, crucially, that clarity must be established contractually before the concrete is poured, not litigated afterward when the forecast has already failed.
1.5 Layer 1 Becomes the Governor of the Five-Layer AI Economy
This is where the present paper connects directly to the Five-Layer AI Economy framework that organizes my broader research. The five layers, and the role of each, can be summarized as follows.
| Layer | Domain | What It Contains | How Ratebase Separation Constrains It |
| Layer 1 | Energy | Generation, transmission, substations, distribution, grid reliability, and the regulatory institutions governing them | The layer where Ratebase Separation operates directly: rate classes, tariffs, collateral, verification, and direct assignment determine who finances the grid |
| Layer 2 | Chips | GPUs, custom accelerators, memory, and networking demand from Nvidia, AMD, Broadcom, and hyperscaler silicon programs | Chips cannot be energized without interconnection; verified, contracted megawatts become a precondition for deploying silicon at scale |
| Layer 3 | Datacenters | Physical campuses converting megawatts into compute; hyperscalers, neoclouds, and colocation operators | Tariff obligations, take-or-pay minimums, and exit fees now sit inside datacenter project finance models as first-order cost lines |
| Layer 4 | Models | Training and inference architectures; frontier laboratories and their scaling roadmaps | The pace and geography of training capacity increasingly follow the map of states where credible large-load frameworks exist |
| Layer 5 | Applications & Agents | Commercial services converting intelligence into economic activity | End-user AI economics inherit, through compute pricing, the cost of capital embedded in Layer 1 guarantees |
Table 1. The Five-Layer AI Economy and the constraint imposed by Layer 1 electricity regulation.
The crucial insight of this section — and the analytical spine of everything that follows — is that Layer 1 is beginning to impose a financial admission price on Layers 2 through 5. Electricity availability alone is no longer sufficient for AI expansion. Developers increasingly need electricity plus creditworthiness, contractual durability, and infrastructure guarantees: a fourteen-year contract in Virginia, an 85 percent take-or-pay minimum in Ohio, collateral scaled to megawatts, verification audits in Texas, and financial security sufficient to cover network improvements in Pennsylvania.[1, 13, 22, 26] The grid, in other words, has stopped asking only whether a customer wants power and started asking whether the customer can guarantee power — and that shift, from consumption to guarantee, is what elevates electricity regulation from background infrastructure into the governing constraint of the entire AI economy.
Section 2: Why a 500-MW AI Campus Breaks Traditional Cost-Allocation Assumptions
2.1 Five Hundred Megawatts Is Not Merely Another Customer
A 500-megawatt AI datacenter campus behaves differently from a conventional factory in nearly every dimension that matters to utility economics, and the differences compound one another. Its load is enormous: the Congressional Research Service notes that a single hyperscale facility drawing 100 megawatts consumes roughly as much electricity as 80,000 American households, which means a 500-megawatt campus is, in residential-equivalent terms, a city of several hundred thousand homes materializing at a single point on the grid.[53] It can operate continuously, with load factors above 75 percent — the very threshold Virginia uses to define its GS-5 class — which means it offers little of the daily diversity that historically allowed utilities to serve many customer types with shared capacity.[6] Its build-out may occur in phases spanning years, with contracted capacity far ahead of energized load, which means the infrastructure must be sized for a peak that may not arrive until the end of a four-year ramp, if it arrives at all. The computing hardware inside may be replaced several times during the economic life of the transmission infrastructure built to reach it. And its ultimate electricity requirement can change substantially as chip efficiency, liquid cooling, model architecture, and inference economics evolve — variables that no utility forecasting department, however sophisticated, has any comparative advantage in predicting.
The utility investment may last forty years. The GPU architecture may last four. That single mismatch — one order of magnitude between the life of the asset and the life of the technology it serves — is the foundation of the entire ratebase problem, and it deserves its own section.
2.2 Infrastructure Duration Versus Technology Duration
Transmission lines and substations are among the longest-lived productive assets in the economy; they are engineered, financed, and depreciated on horizons that assume the world around them changes slowly. AI hardware evolves at the opposite tempo. Blackwell gives way to Rubin; Rubin will eventually yield to another generation; custom accelerators from the hyperscalers reshape rack densities and power envelopes; model architectures shift the balance between training and inference; and more efficient inference can lower electricity intensity per token even as aggregate demand increases — or can, through the rebound dynamics familiar to energy economists since Jevons, increase total consumption precisely because each unit of intelligence has become cheaper. No one can say with confidence which of these forces will dominate in 2032, and yet the transmission line that serves the campus must be financed today on a schedule that runs to the 2060s. Ratebase Separation therefore confronts a fundamental duration mismatch: long-lived electricity infrastructure is being constructed around short-lived technological assumptions, and the regulatory system must decide who holds that mismatch on their balance sheet — the household, the utility shareholder, or the hyperscaler whose technology roadmap created it.
2.3 Forecast Error Becomes Financial Risk
Consider the mechanics concretely, because the abstraction of “stranded costs” conceals a very specific chain of events. Suppose a developer initially requests 500 megawatts. The utility, honoring its obligation to serve and its planning standards, begins constructing the substations, network upgrades, and transmission reinforcements that 500 megawatts of high-load-factor demand requires. Three years later, the developer — for reasons of chip supply, financing conditions, model economics, or corporate strategy — energizes only 250 megawatts. The physical infrastructure still exists in its entirety. The debt that financed it must still be serviced in its entirety. The utility still expects, and under prudence doctrine is generally entitled to, recovery of its investment. Without special protections, the unused half of that capacity migrates quietly into the broader rate base, where its costs are recovered from customers who neither requested it nor benefit from it. Households then become, in effect, the residual guarantors of someone else’s AI forecast — an outcome no one voted for, no one contracted for, and no regulator would have approved if the question had been posed explicitly. Virginia’s minimum demand charges answer exactly this scenario: under GS-5, the customer in this example would pay at least 85 percent of the transmission and distribution costs of its full contracted 500 megawatts every month, regardless of its actual draw, converting the forecast error from a public risk into a private contractual obligation.[1, 2]
2.4 The Queue Becomes an Economic Signaling Mechanism
Once forecast error carries this much financial weight, the interconnection queue can no longer remain a merely administrative artifact — a list of requests processed in the order received. It necessarily becomes an economic signaling mechanism, a test of project seriousness, and, increasingly, a capital-allocation system in its own right. Deposits, collateral, site-control requirements, engineering milestones, and binding contracts separate credible projects from speculative reservations, and the separation is dramatic wherever it has been tried. Ohio’s experience is the cleanest natural experiment available: 30,000 megawatts of requests became roughly 13,000 megawatts willing to fund an engineering study, which became 5,642 megawatts willing to sign binding contracts with collateral — a five-to-one ratio between claimed demand and financially committed demand.[15, 16] AEP Ohio’s own explanation of the winnowing was straightforward.
“More speculative or uncertain data center projects did not sign contracts.”
— AEP Ohio, filing update to the Public Utilities Commission of Ohio, February 2026 [16]
Texas has moved in the same direction on an even larger canvas. By mid-2026, large power users had requested more than 438 gigawatts of new connections to the ERCOT grid — roughly ninety percent of it datacenters — against an all-time statewide peak of about 85.5 gigawatts, and by early August the Governor’s office cited more than 474 gigawatts of requests.[22, 24] Following Governor Greg Abbott’s August 3, 2026 directive, ERCOT paused its new Batch Zero interconnection-study classifications and began a comprehensive verification of every datacenter advancing through the large-load process, sending requests for information to 231 large energy consumers with responses due September 23 and preparing a roughly 130-question survey of 461 datacenters covering financial assistance, power self-sufficiency, water and cooling, community impact, and ownership.[21, 22, 23] Grid queues, in short, are becoming capital-allocation systems: the right to hold a place in them is being priced, audited, and conditioned, because the alternative — planning a continental grid around unverified announcements — has become fiscally intolerable.
2.5 The Stranded-Asset Problem
The central question raised by all of this is not whether hyperscalers can pay today’s electricity bill. Microsoft, Amazon, Meta, and Google plainly can; these are among the most creditworthy enterprises in the history of capitalism, and their Q2-2026 results showed AI-related revenue scaling rapidly even as capital spending reached unprecedented levels.[48] The problem is determining who guarantees infrastructure costs over decades, because a wealthy counterparty today does not eliminate the underlying risks that accumulate across a forty-year asset life: project cancellation before energization; corporate restructuring that reassigns or abandons campuses; technology substitution that collapses the power intensity of compute; load migration toward states or countries with cheaper or faster power; behind-the-meter self-generation that removes load from the grid while leaving the wires in place; macroeconomic recession; AI-demand forecasting error at the industry level; and the simple possibility that a planned campus never reaches contracted capacity. It is telling that even Dominion’s own shareholders have begun to press the question: a shareholder resolution filed for the company’s 2026 proxy season asked the utility to disclose a stranded-asset risk assessment covering its 47-plus gigawatts of contracted datacenter capacity — roughly twice the company’s current peak system load — on the ground that inflated demand forecasts could leave Dominion overbuilding infrastructure for facilities that end up underutilized or uneconomic.[44] Ratebase Separation converts those uncertainties from public risks into contractual risks: it does not make them disappear, but it names the party that holds them, prices that holding through collateral and minimum charges, and thereby restores the correspondence between decision-making power and financial responsibility that the traditional compact, confronted with hyperscale load, had quietly lost.
Section 3: Five Emerging State Models of Ratebase Separation
No single state invented Ratebase Separation, and no two states are building it the same way. What has emerged instead, between mid-2025 and late 2026, is a set of five distinguishable institutional models — five different answers to the same underlying question of who guarantees the infrastructure of the compute boom. Virginia separates by rate class. Ohio separates by dedicated tariff. Indiana separates by contract. Texas separates by verification and grid admission. Pennsylvania separates by statewide model tariff. Examining each in turn, and then side by side, reveals both the diversity of American regulatory federalism and the striking convergence of its destination.
3.1 Virginia: The Rate-Class Separation Model
Virginia currently provides the clearest conceptual example of Ratebase Separation, which is fitting for the jurisdiction hosting the largest concentration of datacenters on Earth. The SCC’s GS-5 classification, established in the November 25, 2025 final order in Dominion’s biennial review, places qualifying large loads — defined by demand of 25 megawatts or more and a monthly load factor of roughly 75 percent or greater — into a legally separate rate class beginning January 1, 2027.[1, 6] The framework incorporates a layered set of protections that together read like a textbook of stranded-cost risk management: a minimum fourteen-year service commitment for new qualifying customers, with an optional load-ramp period of up to four years requiring a minimum annual ramp of 20 percent; minimum monthly payment obligations of at least 85 percent of contracted transmission and distribution demand and 60 percent of contracted generation demand; exit fees covering all outstanding minimum charges for the remaining contract term if a customer ceases operations or defaults; collateral requirements — reported at approximately $1.5 million per megawatt — for customers lacking sufficient credit; a requirement of three years’ notice for planned demand reductions, with limits on the magnitude of such reductions; alternative generation and transmission cost-allocation methodologies to be developed in future proceedings; and potential direct assignment of customer-specific transmission facilities.[1, 2, 4, 5] The Commission’s February 2026 fact sheet states the purpose plainly: these measures exist to ensure datacenters pay their own costs and to minimize shifting to other customer classes.[2]
Virginia thus represents Ratebase Separation through regulatory classification. Instead of pretending that a hyperscale campus is simply a very large commercial customer — the polite fiction embedded in most legacy tariff structures — Virginia has formally recognized it as a fundamentally different kind of electricity consumer, with a different risk profile deserving a different legal category. The direct-assignment strand of the framework goes further still: in its July 31, 2026 order in the Rider T1 proceeding, the SCC found that large-load customers must pay for transmission facilities constructed solely to serve them, ordered Dominion to file a new cost-assignment proposal within ninety days, and identified projects such as the planned Valley Link transmission line as candidates for direct assignment to the GS-5 class in the future.[9, 10] Governor Spanberger’s administration has reinforced the organizing principle throughout: infrastructure that would not exist “but for” large-load customers should be paid for by those customers rather than by residential ratepayers.[8] The Piedmont Environmental Council, an intervenor whose testimony the Commission cited, welcomed the July order while warning that it remains a first step — the group’s analysis of the earlier biennial-review decision argued that even a fourteen-year contract term could leave a substantial share of grid-upgrade costs with ordinary ratepayers after the contract expires, and its president, Chris Miller, criticized the Commission’s decision to leave interim cost allocation unchanged for two years.
“Still unfair and does not go far enough to protect the average Virginian ratepayer.”
— Chris Miller, President, The Piedmont Environmental Council, on the SCC’s interim cost-allocation decision [3]
Microsoft’s legal challenge makes Virginia particularly important as a national bellwether. When the company filed its notice of appeal with the Supreme Court of Virginia in late August 2026 — without specifying which provisions of the July order it seeks to overturn, while affirming its acceptance of cost-causation in principle and warning that Virginia’s approach could conflict with emerging federal efforts to standardize large-load cost allocation — the argument stopped being theoretical.[11, 12] The boundary of the AI electricity economy is now being litigated, in the state where that economy is most physically concentrated, by the company with at least eleven datacenter campuses operating, planned, or under construction there. However the Virginia Supreme Court rules, its decision will become the first appellate landmark of Ratebase Separation, and every utility commission in the country will read it.
3.2 Ohio: The Dedicated Datacenter Tariff Model
Ohio’s approach emphasizes a purpose-built tariff and a structured queue rather than a new rate class embedded in a general rate case. AEP Ohio’s Data Center Tariff became effective on July 23, 2025, following the Public Utilities Commission of Ohio’s unanimous July 9 approval of a settlement negotiated with commission staff, the Ohio Consumers’ Counsel, Ohio Partners for Affordable Energy, and other stakeholders.[13, 14] Under the tariff, new datacenter customers above 25 megawatts face minimum monthly bills based on at least 85 percent of contracted capacity or their highest recent billing demand; twelve-year contract terms; demonstration of financial viability; collateral from the customer or a co-signing financial sponsor; and exit fees if a project is canceled or obligations cannot be met.[14] Developers face a defined application and study process in which projects are evaluated together in tranches, so that the cumulative transmission consequences of clustered campuses are studied as a system rather than as a sequence of isolated requests.
Ohio therefore represents Ratebase Separation through customer-specific tariff architecture, and its most important innovation is not simply charging datacenters more. It is making contractual and financial certainty part of the right to obtain massive quantities of grid capacity — and, in doing so, generating the single most valuable dataset in the national debate. The tariff functioned as a truth serum for the queue: 30,000 megawatts of requests collapsed to 5,642 megawatts of binding post-tariff contracts (alongside 12,219 megawatts contracted before the tariff took effect, for 17,861 megawatts in total, scheduled to come online progressively through 2035 against a system whose recent peaks have ranged between roughly 8,000 and 10,500 megawatts).[15] Critics such as the Ohio Manufacturers’ Association argue that the original 30,000-megawatt figure was inflated from the beginning and that the tariff merely forced the utility to come clean about real demand — a critique that, notably, does not weaken the case for Ratebase Separation but strengthens it, since either interpretation demonstrates that unpriced queues produce fictitious planning numbers.[16] The August 2026 addition of a 180-day notice requirement before datacenters take default service, with the datacenter covering all associated procurement costs, extends the same logic from infrastructure to energy supply: PUCO granted the measure explicitly to protect AEP Ohio’s 1.5 million other customers, in a market where Columbus residential bills were already running more than 7 percent above the prior year.[17]
3.3 Indiana: The Contractual-Risk Model
Indiana presents a third variation, distinguished by who sat at the table. In February 2025, the Indiana Utility Regulatory Commission approved modifications to Indiana Michigan Power’s industrial tariff addressing large loads — and the settlement it approved had been negotiated and jointly filed not only by the utility and the Indiana Office of Utility Consumer Counselor and the Citizens Action Coalition, but by Amazon Web Services, Microsoft, Google, and the Data Center Coalition themselves.[18, 19] The context was transformational load growth: AWS had announced an $11 billion datacenter campus near New Carlisle and Google a $2 billion facility in Fort Wayne, among the largest economic development projects in the state’s history, and I&M projected that its Indiana peak load would jump from roughly 2,800 megawatts to more than 7,000 megawatts by 2030 — a two-and-a-half-fold increase driven by a handful of customers.[19]
The approved framework applies to new or expanded facilities with contract capacity of at least 70 megawatts, or 150 megawatts aggregated across a company, and incorporates mandatory long-term financial commitments proportionate to customer size, minimum billing requirements, collateral, exit fees, and restrictions on capacity reductions — with the IURC adding, on its own initiative, a modification requiring that any reduction exceeding 20 percent of a large-load customer’s contracted peak capacity be reviewed and approved by the Commission itself.[19, 20] The Indiana proceeding is revealing because the regulator explicitly recognized that the unexpected departure of a single major large-load customer could cause substantial financial harm to the utility and to remaining customers, and because it acknowledged the deeper structural shift: a small number of hyperscale customers could collectively become larger than entire traditional customer sectors within the utility’s system. That changes utility economics at the root. A diversified monopoly can become financially dependent upon a few enormous buyers — a concentration risk more familiar to corporate bond analysts than to utility regulators — and Indiana’s answer was to bind those buyers with contracts strong enough to carry the dependence. Indiana therefore represents Ratebase Separation through contractual risk transfer, and its settlement demonstrates something politically important: the hyperscalers themselves signed it, which suggests that sophisticated large-load customers prefer transparent, negotiated, durable frameworks over the reputational and regulatory hazards of quiet cost socialization.
3.4 Texas: The Verification and Grid-Admission Model
Texas faces a different version of the problem, at a scale that dwarfs every other jurisdiction. Its combination of low-cost land, abundant energy development, the distinctive ERCOT market structure, and aggressively business-friendly policy has attracted proposed AI load in quantities that have lost contact with physical reality: more than 438 gigawatts of large-load interconnection requests by mid-2026 — nearly ninety percent from datacenters — rising to more than 474 gigawatts by early August, more than five times the highest peak demand ERCOT has ever served.[22, 24] Requested megawatts, in Texas above all, are not real megawatts, and the state’s response has been to build the admission machinery that decides which ones are.
The machinery has two interlocking parts. The first is Batch Zero, approved by the Public Utility Commission of Texas in June 2026: a shift from one-project-at-a-time interconnection studies to a system-wide batch process for large loads of 75 megawatts and above, with defined submission deadlines, readiness evidence, dynamic-model requirements, and new ride-through performance rules for large electronic loads.[22] The second is verification. On August 3, 2026, Governor Greg Abbott directed ERCOT and the PUCT to conduct a comprehensive verification and audit of all datacenters advancing through the large-load interconnection process — covering grid and water usage, public financial assistance, onsite generation, community impacts, and project ownership — before any of them may proceed, and ERCOT immediately paused Batch Zero classifications and announced it would not authorize large computational loads to energize until verification is complete.[21, 22] By September, ERCOT had issued requests for information to 231 large energy consumers conditionally included in Batch Zero, with responses due September 23, and was preparing its roughly 130-question community-impact survey of 461 datacenters, with final audit results expected in December.[23] The Governor’s framing left no ambiguity about the stakes of admission.
“Any project that fails to comply … must be denied connection to the Texas grid.”
— Governor Greg Abbott of Texas, directive to ERCOT and the PUCT, August 3, 2026 [24]
Texas therefore highlights a precursor to Ratebase Separation rather than a rival to it: before determining who should pay for infrastructure, regulators must determine which projects are real enough to justify building it. The Texas model is consequently less about conventional rate classes and more about grid admission — and grid admission may prove equally important, because in a world of genuine electricity scarcity, the right to reserve grid capacity becomes economically valuable in itself. Texas has additionally pioneered the flexibility dimension of the new architecture: Senate Bill 6, signed in June 2025, requires new large loads above 75 megawatts connecting to ERCOT to demonstrate the ability to curtail on demand during declared grid emergencies, embedding load flexibility into the terms of admission rather than leaving it as a voluntary virtue.[52]
3.5 Pennsylvania: The Statewide Model-Tariff Approach
Pennsylvania has moved toward statewide standardization, and in doing so has produced the most generalizable template of the five. After more than a year of hearings, en banc proceedings, public comment, and stakeholder engagement, the Pennsylvania Public Utility Commission voted 5–0 on April 30, 2026 to adopt a modified model tariff framework for large-load customers, releasing its Final Order on May 12, 2026 — a first-of-its-kind statewide model applying to individual customers with maximum contract capacity above 50 megawatts, and to multiple closely located customers aggregating 100 megawatts or more.[25, 26] PUC Chairman Steve DeFrank framed the stakes in terms that could serve as an epigraph for this entire paper.
“This is one of the most important infrastructure and consumer protection issues facing utility regulators across the country.”
— Steve DeFrank, Chairman, Pennsylvania Public Utility Commission [25]
The model tariff’s substance maps almost one-to-one onto the conceptual architecture of Ratebase Separation. Network upgrade costs are allocated under an explicit “but for” cost-causation test — if an improvement would not have been needed but for the large-load customer’s interconnection, the customer pays for it as a contribution in aid of construction, regardless of incidental benefits to others, with the sole exception of upgrades already planned under an approved long-term infrastructure improvement plan.[26] Monthly billing demand is set at no less than 80 percent of contracted capacity, and a minimum monthly demand charge of at least 80 percent of contracted demand applies. Financial security must fully cover network improvement and interconnection facility costs, with collateral refundable or reducible as construction and load milestones are achieved. The minimum initial contract term is five years following the customer’s ramp-up period — shorter than Virginia’s fourteen years, but paired with a direction that utilities ensure contracts last long enough to recover the investments made specifically to serve the customer — and termination or major capacity reduction requires 48 months’ notice.[26, 27] The framework also encourages large-load customers to contribute to low-income assistance programs and creates rate options rewarding interruptible or flexible service during peak demand.[26]
The Pennsylvania approach therefore represents Ratebase Separation through statewide regulatory standardization, and its location gives it outsized significance. Pennsylvania sits inside PJM, the thirteen-state regional market where datacenter growth in Virginia and Ohio already shapes capacity prices, transmission plans, and electricity economics across state borders — PPL Electric alone reported an “advanced” datacenter pipeline of 28.3 gigawatts in Pennsylvania by May 2026, against a current summer peak of 7.5 gigawatts.[49] One state’s AI expansion can now raise another state’s electricity bill through the shared capacity market, a mechanism examined in detail in Section 4. That interdependence makes Ratebase Separation an inherently interstate issue, and it makes Pennsylvania’s decision to publish a portable model — explicitly intended as guidance that other utilities and, implicitly, other states can adapt — a quiet act of national institution-building.
3.6 The Five Models Side by Side
| Dimension | Virginia (GS-5) | Ohio (AEP DCT) | Indiana (I&M) | Texas (ERCOT) | Pennsylvania (Model Tariff) |
| Separation mechanism | Separate rate class in general rates | Dedicated datacenter tariff | Negotiated contractual settlement | Verification & batch grid admission | Statewide model tariff framework |
| Threshold | 25 MW; ~75% load factor | 25 MW | 70 MW (150 MW aggregated) | 75 MW (large loads / LCLs) | 50 MW (100 MW aggregated) |
| Minimum demand charge | 85% T&D; 60% generation | 85% of contracted capacity | Minimum billing per settlement | N/A (market-based energy) | 80% of contracted demand |
| Contract term | 14 years (≤4-year ramp) | 12 years | Long-term, size-proportionate | N/A; admission conditions instead | ≥5 years after ramp; cost-recovery-based |
| Collateral / security | Yes; ~$1.5M per MW reported | Yes; sponsor co-signing allowed | Yes | Study deposits; readiness evidence | Must cover network & interconnection costs; declines with milestones |
| Exit / reduction protection | Exit fee; 3-year notice; reduction limits | Exit fees on cancellation | IURC review of >20% reductions | Denial of connection for non-compliance | 48 months’ notice for exit or major reduction |
| Distinctive feature | Direct assignment of customer-specific transmission | 180-day notice before default service | Hyperscalers co-signed the settlement | Governor-ordered audit of 461 facilities | Portable statewide template; flexibility incentives |
| Effective | Jan 1, 2027 | Jul 23, 2025 | Feb 19, 2025 | Jun–Aug 2026 | Final Order May 12, 2026 |
Table 2. Five state models of Ratebase Separation compared. Sources: [1][2][13][14][18][19][21][22][25][26].
Read across the rows rather than down the columns, the table tells a single story five times. Every jurisdiction has concluded that hyperscale load requires a threshold definition, a take-or-pay floor or its admission-stage equivalent, duration matching between contract and investment, collateralization of forecast risk, and protection against abrupt exit. The parameters differ; the grammar is identical. That grammatical convergence, achieved without any federal mandate and in states with profoundly different politics and market structures, is the strongest available evidence that Ratebase Separation is not a regulatory fashion but a structural adaptation — the electricity system’s immune response to a category of customer it had never previously encountered.
Section 4: The New Financial Architecture of AI Electricity
Beneath the five state models lies a common toolkit — a set of financial instruments that, taken together, constitute the operating system of Ratebase Separation. Each instrument answers a specific failure mode of the traditional compact, and each transfers a specific risk from the public rate base to the party that created it. This section examines the six most important instruments in turn, because understanding them individually is the only way to understand why the overall architecture is coherent rather than punitive, and why sophisticated hyperscalers have, in several jurisdictions, chosen to negotiate within it rather than against it.
4.1 Upfront Contributions in Aid of Construction
The simplest protection is also the most controversial: if infrastructure would not exist without a datacenter, make the datacenter finance it — upfront, as a contribution in aid of construction, before the utility’s shareholders or its other customers advance a dollar. Pennsylvania’s model tariff makes this the default rule for network upgrades under its “but for” test, and Virginia’s July 2026 Rider T1 order pushes Dominion toward the same destination through direct assignment of customer-specific transmission facilities.[9, 26] Upfront contributions shift construction risk decisively away from ordinary ratepayers: if the project dies, the wires were never on the public’s account. But they also raise developers’ cost of capital, convert what was once a utility-financed operating expense into a developer-financed capital expense, and — in the argument Microsoft is now making to the Virginia Supreme Court — risk conflicting with federal efforts to standardize large-load cost allocation if fifty states adopt fifty incompatible prepayment mechanics.[11, 12] That tension, between the incontestable fairness of the principle and the contested engineering of the mechanism, sits at the exact center of the Virginia litigation, and its resolution will shape the cost of capital for the entire Layer 3 datacenter economy.
4.2 Take-or-Pay Electricity
Take-or-pay structures transform electricity capacity into something resembling long-duration industrial infrastructure finance — closer in spirit to a liquefied natural gas offtake agreement or a pipeline capacity contract than to a conventional retail electricity tariff. A hyperscaler agrees not merely to consume electricity when convenient but to pay for a minimum quantity of reserved capacity whether or not it is used: 85 percent of transmission and distribution demand in Virginia and Ohio, 80 percent in Pennsylvania’s model.[1, 14, 26] The take-or-pay minimum answers what may be the single most important question in the entire debate: if an AI company asks the grid to build for 500 megawatts, should it be allowed to pay as though it needed only 100 megawatts after construction is complete? Ratebase Separation increasingly, and in my judgment correctly, says no. The minimum charge does not punish underutilization; it prices it, and by pricing it, the minimum charge accomplishes something subtle and valuable — it forces the developer’s internal capacity request to converge toward the developer’s genuine expectation, because every megawatt of exaggeration now carries a fourteen-year annuity of cost. The take-or-pay floor is, in effect, a truth-telling mechanism disguised as a billing rule, and Ohio’s queue collapse from 30,000 to 5,642 binding megawatts is the empirical proof that it works.[15, 16]
4.3 Collateral as Grid Insurance
Collateral transforms corporate promises into financial protection. Letters of credit, parent guarantees, cash deposits, and co-signed sponsor obligations ensure that some portion of future obligations remains collectible even if a project fails, a subsidiary is abandoned, or a special-purpose development entity proves to be exactly as thinly capitalized as its name suggested. Virginia’s reported requirement of roughly $1.5 million per megawatt for customers lacking sufficient credit is best understood not as a fee but as a price signal: it is the regulator writing down, in dollars per megawatt, its own estimate of the stranded-cost exposure that a failed project would otherwise impose on the public.[5] Pennsylvania’s contribution to the design is the milestone-based release schedule: financial security must initially cover the full network improvement and interconnection costs, then declines or is refunded as construction and load milestones convert forecast into fact.[26, 27] This embodies a principle that any national framework should adopt explicitly: risk protection should be strongest when uncertainty is greatest and should decline as actual performance replaces forecasts. Collateral, structured this way, is not a tax on datacenters. It is grid insurance, underwritten by the party whose behavior determines whether the insured event occurs — which is precisely where insurance theory says the premium belongs.
4.4 Exit Fees and Duration Matching
A datacenter may eventually leave. Its transmission line cannot. Exit fees exist to recognize this asymmetry, and they are the instrument that most directly addresses the duration mismatch identified in Section 2.2. Under Virginia’s GS-5 rules, a customer that ceases operations or defaults during its contract term owes the outstanding minimum charges across the entire remaining duration; Ohio imposes exit fees on canceled projects; Indiana subjects capacity reductions above 20 percent to regulatory review; Pennsylvania requires 48 months’ notice before termination or major reduction.[4, 14, 20, 27] If customers could abandon contracts without compensating the system for infrastructure built around them, ordinary customers would become the ultimate insurer of every corporate strategy shift in the technology industry — which is exactly the position they occupied, silently, under the old compact. Exit fees therefore function as a form of infrastructure-duration matching: they extend the customer’s financial presence to the horizon of the assets its arrival caused, even if its physical presence ends sooner. The economist’s way of saying this is that exit fees internalize the externality of departure; the ratepayer’s way of saying it is simpler — you broke ground, you pay for it.
4.5 Queue Deposits and the Pricing of Speculation
Deposits serve a different and earlier function in the project lifecycle: they discourage developers from reserving grid capacity without a sufficient probability of construction. As electricity scarcity intensifies, speculative queue positions create real economic costs long before any physical infrastructure is built — they distort regional planning forecasts, inflate capacity-market demand curves, delay credible projects standing behind them, and, as the PJM experience examined below demonstrates, can transfer billions of dollars to ratepayers through market prices alone. Ohio prices queue seriousness through its tranche-study fees and binding-contract requirements; Texas prices it through Batch Zero’s submission deadlines, readiness evidence, and now the Governor’s verification audit; Dominion’s contracting ladder — from substation engineering letters of authorization to construction letters to full electric service agreements, with customer cost commitments increasing at each step — prices it through graduated contractual escalation.[15, 22, 44] A credible national framework should make queue access progressively expensive as a project approaches construction, for the same reason that options markets charge more for longer-dated certainty: the reservation of scarce capacity is itself a valuable position, and unpriced valuable positions attract exactly the speculation the system can least afford.
4.6 Stranded-Asset Risk and the PJM Warning
Stranded assets represent the heart of Ratebase Separation, and the PJM capacity market has already provided a preview — at continental scale — of what happens when speculative AI load is allowed to flow unpriced into shared cost structures. PJM’s independent market monitor, Monitoring Analytics, estimated that datacenters were responsible for 63 percent of the price increase in the 2025/2026 capacity auction, translating to roughly $9.3 billion in additional costs recovered from customers across the thirteen-state region in a single delivery year; capacity prices rose from $28.92 per megawatt-day for 2024/2025 to $269.92 for 2025/2026, then to the FERC-approved cap of $329.17 for 2026/2027 and $333.44 for 2027/2028 — an increase of roughly tenfold in three auction cycles.[32, 33] In the December 2025 auction, datacenter load accounted for $6.5 billion, or 40 percent, of $16.4 billion in total costs — and approximately $6.2 billion of that was attributable to datacenters that had not yet been built.[33] Households have felt the pass-through directly: Pepco customers in Washington, D.C. absorbed bill increases averaging $21 per month beginning June 2025, roughly half attributable to capacity prices; western Maryland households roughly $18 per month; Ohio households roughly $16; and the NRDC projects that if prices remain at the cap, cumulative regional capacity costs could reach $163 billion through 2033, translating to roughly $70 per month for the average household by 2028.[32, 34] The market monitor’s diagnosis was unequivocal.
“Data center load growth is the primary reason for recent and expected capacity market conditions.”
— Monitoring Analytics, Independent Market Monitor for PJM, January 2026 report [33]
Consider, against that backdrop, several entirely plausible 2030 outcomes: AI demand grows more slowly than the 2026 queue implies; a new accelerator generation dramatically improves performance per watt; a hyperscaler shifts workloads toward regions with faster interconnection; behind-the-meter generation removes contracted load from the grid; a startup-backed neocloud operator fails; a project loses financing in a capital-markets downturn; an environmental permit is revoked. In every one of those scenarios, transmission and generation infrastructure may nevertheless have been constructed — and the public-policy question is therefore not simply who pays when AI succeeds, but who pays when an AI infrastructure forecast is wrong. The PJM numbers demonstrate that the answer, absent Ratebase Separation, is already visible on 67 million people’s monthly bills. That question — who owns the downside of the forecast — is likely to become one of the most important questions any governor or utility commissioner faces between now and 2030, and it is the question every instrument in this section exists to answer in advance.
Section 5: A National Framework for an AI Large-Load Rate Class
5.1 The Case for a National Model
Electricity regulation remains primarily state-based, and nothing in this paper argues for federalizing retail ratemaking. Yet AI infrastructure is irreducibly national — and increasingly transnational — in its logic. Microsoft can compare Virginia against Ohio; Amazon can compare Indiana against Mississippi; Google can compare Pennsylvania against Texas; Meta can shift future development toward whichever state offers the fastest credible path to power. Without common principles, states risk sliding into a competition in which favorable electricity treatment becomes another form of economic-development subsidy — a race in which the “winning” state is the one whose households most quietly underwrite the infrastructure of trillion-dollar companies. The federal government has recognized the problem rhetorically: the Ratepayer Protection Pledge announced by President Trump on March 4, 2026 and signed at the White House by Amazon, Google, Meta, Microsoft, OpenAI, Oracle, and xAI commits signatories to build, bring, or buy the energy their datacenters need, to pay for delivery infrastructure upgrades, to negotiate separate rate structures with utilities and states, and to pay for power brought online for their facilities whether they use it or not; by July 2026 the White House reported that the expanded pledge — now including utilities, developers, and 23 governors — covered 80 percent of power delivered to U.S. homes and businesses.[28, 29, 51] But the pledge is voluntary and carries no enforcement mechanism, a limitation that Brookings scholars, industry observers, and Harvard’s Ari Peskoe have all emphasized: commitments of this kind are only as strong as the tariffs, contracts, and commission orders that operationalize them, and companies have been making similar promises for some time without the math of transmission costs and rising bills changing.[30, 31, 40] A national framework need not impose identical rates. It should establish common risk-allocation principles — the grammar of Table 2 — while leaving parameters to state judgment. The seven principles that follow, distilled from the five state experiments, are my proposal for that grammar.
5.2 Define the AI Large-Load Threshold in Tiers
The five states draw their thresholds at different levels — 25 megawatts in Virginia and Ohio, 50 in Pennsylvania, 70 in Indiana, 75 in Texas — and the differences are less important than the shared recognition that a threshold must exist. A national model could rationalize the landscape with graduated tiers, under which financial obligations increase with scale rather than switching on abruptly at a single line.
| Tier | Contracted Capacity | Indicative Obligations |
| Tier I | 25–99 MW | Standard large-load tariff; minimum demand charge; study deposits; standard collateral |
| Tier II | 100–299 MW | Extended contract term; milestone-based collateral covering customer-specific network costs; ramp schedules with minimum annual ramps |
| Tier III | 300–499 MW | But-for direct assignment of dedicated facilities; enhanced exit fees; long-notice reduction requirements; flexibility capability assessment |
| Tier IV | 500 MW and above | Full duration-matched contracting; contributions in aid of construction for dedicated transmission; regulator-reviewed capacity reductions; verified flexibility or firm self-supply commitments |
Table 3. A proposed four-tier national structure for AI large-load classification.
The objective of tiering is not to punish large projects. It is to recognize, in the architecture of the tariff itself, that a 25-megawatt facility and a 1-gigawatt AI campus create fundamentally different planning risks — different in the size of the dedicated infrastructure they trigger, different in the concentration risk they impose on the utility, and different in the harm their abandonment would cause — and that a single undifferentiated “large load” category repeats, one level up, the very error of undifferentiated classification that Ratebase Separation exists to correct.
5.3 Adopt the “But-For” Infrastructure Test
The single most powerful principle available for national adoption is the one Pennsylvania has already codified and Virginia’s Governor and Commission have embraced: if infrastructure would not reasonably be constructed but for a specific large-load project, the project should bear the attributable cost.[8, 26] The test does not require assigning every network improvement to one company — shared assets that genuinely strengthen regional reliability can and should still be allocated proportionally across beneficiaries, and Pennsylvania’s carve-out for upgrades already planned under approved infrastructure programs shows how to keep the boundary honest. What the test requires is transparency of demonstration: before costs enter the shared rate base, someone must show, on the record, who benefits and in what proportion. The but-for test is cost causation with a burden of proof attached, and the burden falls where the information lives — on the utility and the customer whose project created the question.
5.4 Require Graduated Financial Commitment
Projects should demonstrate seriousness progressively, through an escalating ladder of commitments that tracks the utility’s own escalating exposure: site control; application deposits; engineering deposits; interconnection studies; milestone payments; binding capacity commitments; collateral; minimum demand charges; and finally long-duration service agreements. Dominion’s three-stage contracting ladder and Ohio’s study-then-contract sequence already embody the idea, and the governing principle generalizes cleanly: the deeper a utility moves into irreversible infrastructure investment, the greater the customer’s financial commitment should become.[15, 44] Graduated commitment has a second virtue that deserves emphasis — it produces information. Each rung of the ladder is a revealed-preference test, and the aggregate of those tests is the only demand forecast worth planning a grid around.
5.5 Match Contract Length to Asset Risk
No single contract period will fit every project, and the spread between Pennsylvania’s five-year minimum and Virginia’s fourteen-year term — with the Piedmont Environmental Council arguing that even fourteen years leaves a substantial residual on ratepayers and should have been twenty — shows how unsettled this parameter remains.[3, 27] The governing principle, however, is settled by first principles: a utility should not finance a customer-specific asset over decades while possessing only a short-duration customer commitment. Contract design should therefore be derived from expected cost recovery rather than from arbitrary calendar conventions — the term should be whatever length is required for the minimum charges to amortize the customer-specific investment, given the ramp schedule and the collateral — and Pennsylvania’s directive that utilities ensure contracts last long enough to recover investments made specifically for the customer states exactly this rule in regulatory language.[26]
5.6 Reward Flexible AI Loads
Ratebase Separation should not consist only of obligations, because AI infrastructure may eventually provide the grid with something genuinely valuable: flexibility at unprecedented scale. The foundational research here comes from Duke University’s Nicholas Institute, where Tyler Norris and co-authors demonstrated in early 2025 that the existing U.S. power system could integrate nearly 100 gigawatts of new large load — within the current capacity of the 22 largest balancing authorities — provided that load can be curtailed for an average of roughly 0.5 percent of annual hours, with typical curtailment events lasting about two hours and usually involving partial rather than total reductions.[41] Norris described the discovered headroom as “significantly beyond expectations,” and the finding has since propagated through the industry with remarkable speed: Google signed roughly a gigawatt of demand-response contracts with utilities including Entergy Arkansas, Minnesota Power, and DTE Energy; EPRI’s DCFlex initiative expanded internationally; Texas wrote curtailment capability into statute through SB6; Pennsylvania’s model tariff created explicit rate options for interruptible and flexible service; and Norris himself was hired by Google in 2026 to lead market innovation for advanced energy — a career move that is itself a data point about where the industry believes the value lies.[26, 41, 52, 54] The academic literature has also matured enough to register the trade-offs: MIT’s Christopher Knittel and co-authors have shown in an NBER working paper that flexible datacenters can lower system costs while, under some dispatch patterns, increasing emissions — a reminder that flexibility incentives must be designed against the full objective function, not costs alone.[42] The design principle for a national framework follows directly: a datacenter that reliably and verifiably avoids peak load should receive materially different economic treatment from one requiring maximum grid availability twenty-four hours per day, because the two facilities impose materially different infrastructure requirements — and the difference should be priced symmetrically, as a reward, with the same seriousness that inflexibility is priced as an obligation.
5.7 Separate Speculative Load From Committed Load
Utilities and regional transmission organizations should be required to publish at least two forecasts: requested large-load demand, and financially committed large-load demand — the megawatts backed by binding contracts, collateral, and milestone payments. The distinction is not cosmetic; it may be the single most consequential piece of information in American energy planning today. Ohio’s five-to-one ratio between requested and committed load, ERCOT’s 474 requested gigawatts against an 85.5-gigawatt historical peak, Borenstein’s warning about duplicate filings across territories, and Monitoring Analytics’ finding that billions of dollars of PJM capacity costs were driven by datacenters not yet built all point to the same conclusion: the country does not currently know whether it faces a 50-gigawatt shortage or merely 50 gigawatts of competing development applications, and policy made under that ambiguity will systematically overbuild.[15, 22, 33, 43] Monitoring Analytics has proposed the logical endpoint — removing uncertain datacenter load from the base capacity auction entirely and procuring for it separately — and PJM has begun designing supplemental auction mechanisms along those lines.[35] Publishing the two forecasts side by side would not resolve every dispute, but it would relocate the burden of proof to where it belongs: on the claimed megawatts, not on the households asked to finance them.
5.8 Require Ratepayer Stress Tests
Before approving extraordinary infrastructure investments, regulators should model — and publish — scenario analyses covering full load realization; partial realization; delayed ramp; early customer departure; technology-driven efficiency improvements; and customer insolvency. The question each scenario must answer is concrete and singular: what happens to residential and small-business bills under this outcome? The practice would formalize what the best proceedings already do informally. Virginia’s SCC estimated that its cost-allocation changes could cut the projected Rider T1 increase for a representative residential customer from $2.90 per month to $0.94 — a published, falsifiable number that transformed an abstract allocation debate into a household arithmetic problem — and the shareholder resolution demanding that Dominion assess stranded-asset risk across its 47-gigawatt pipeline shows that investors, not only consumer advocates, now want the downside quantified.[2, 44] A stress-test requirement would also discipline the utilities’ own incentives, which — as Martin and Peskoe documented — currently reward capital deployment under essentially all demand scenarios, since the utility earns its return whether or not the load materializes.[38] What gets modeled gets governed; what remains unmodeled gets socialized.
5.9 Coordinate FERC, RTOs, and State Regulators
Transmission does not respect state borders, and neither does AI. PJM, MISO, ERCOT, SPP, and the other regional organizations are confronting hyperscale demand simultaneously, and the PJM capacity-market experience proves that one state’s admission decisions become another state’s bill increases through shared markets.[32, 33] FERC will therefore become structurally important even where retail-rate authority remains with the states — in large-load interconnection standards, in co-location rules for generation and datacenters, in capacity-market design for uncertain load, and in the very question Microsoft has raised in its Virginia appeal: whether and how state prepayment mechanisms should harmonize with federal cost-allocation standardization.[11, 12] The long-term architecture of Ratebase Separation will require sustained coordination among governors, public utility commissions, FERC, regional transmission organizations, utilities, hyperscalers, and consumer advocates — an unglamorous, committee-shaped answer, but the honest one. This is where Ratebase Separation could evolve from a collection of state experiments into a national policy doctrine: not through preemption, but through convergence on the shared grammar the states are already, independently, writing.
Section 6: What Have We Learned? Seven Pillars
Every long investigation should end by asking what it has actually established. The five state models, the six financial instruments, and the nine national principles examined above compress, in my reading, into seven durable lessons — seven pillars on which the analysis of AI electricity regulation can rest for the remainder of this decade. The first five concern the structure of the problem; the final two, added here because the evidence of 2026 demands them, concern the structure of the response.
Pillar 1 — AI Load Is Becoming Its Own Electricity Customer Class
The first lesson is conceptual, and it is the one from which all others follow. A hyperscale AI campus should no longer automatically be treated as simply another commercial or industrial electricity customer, because its defining characteristics — scale that rivals cities, concentration at single grid nodes, continuous load factors, phased and uncertain ramps, duration mismatch between its hardware and its infrastructure, and abandonment consequences measured in billions — make it structurally different from every customer category the twentieth-century tariff book contains. Virginia has formalized that conclusion in law through GS-5; Ohio, Indiana, Texas, and Pennsylvania are arriving at the same destination through tariff, contract, admission, and model-framework mechanisms respectively.[1, 13, 18, 22, 25] The Five-Layer AI Economy has therefore produced an institutional consequence beyond its five layers: a new regulated class of electricity consumer, created by compute itself. Once a legal category exists, everything else — data collection, differentiated obligations, targeted protections, specialized litigation — organizes around it, which is why the classification decision, dry as it sounds, is the constitutional moment of Ratebase Separation.
Pillar 2 — “Who Pays?” May Matter More Than “Can We Generate Enough?”
Much of the public AI-energy debate remains fixated on supply: how many gigawatts are available, how quickly natural-gas turbines can be manufactured, whether retired nuclear plants can restart, whether small modular reactors can arrive before 2030, whether renewables paired with storage can fill the gap. Those questions matter enormously, and nothing here diminishes them. But the evidence of 2025–2026 suggests that the binding political constraint may lie elsewhere: in the question of who finances the infrastructure necessary to connect power to AI. A state can possess theoretically sufficient generation and still face paralyzing political opposition if households believe their electricity bills are subsidizing trillion-dollar technology companies — and that belief is no longer speculative. Brookings reports that 35 percent of American voters agree that datacenter development drives up everyone else’s electricity costs, and that in Virginia itself, comfort with a new datacenter in one’s own community collapsed from 69 percent in 2023 to 35 percent in 2026; eleven states introduced datacenter moratorium legislation in early 2026, Maine’s passed before a gubernatorial veto, and a proposed constitutional amendment to ban datacenters circulated in Ohio.[16, 30] The PJM numbers — $9.3 billion in one year, a projected $70 per household per month by 2028 at the cap — explain the anger with arithmetic rather than ideology.[32, 34] Ratebase design, in short, determines the political durability of the AI boom. The industry that solves “who pays” will be allowed to keep building; the industry that does not will discover that transformers and turbines were never its scarcest input — public consent was.
Pillar 3 — Creditworthiness Becomes Part of Compute Infrastructure
The third lesson is that financial credit is becoming physically inseparable from compute. Historically, the key datacenter inputs were land, fiber, chips, cooling, and electricity. The new regimes add a sixth: creditworthiness. A company seeking hundreds of megawatts in Virginia after January 2027 needs a fourteen-year signature, an 85 percent minimum, and — absent investment-grade standing — collateral on the order of $1.5 million per megawatt; in Pennsylvania it needs security covering the full cost of its network improvements; in Ohio it needs a co-signing financial sponsor if its own balance sheet will not carry the obligation.[1, 5, 14, 26] This reshapes competitive structure across Layers 3 through 5 of the AI economy in a direction that deserves candid acknowledgment: it favors incumbents. Microsoft, Amazon, Alphabet, and Meta can write twenty-year guarantees against the strongest balance sheets in corporate history — even as FactSet reports that their capex has begun to outrun operating cash flow, with incremental debt rising from 9 percent of capex in fiscal 2024 to roughly 32 percent by mid-2026 — while venture-backed neoclouds and startup operators must rent that creditworthiness from sponsors, partners, or customers, at a price.[50] The future datacenter race may therefore reward not merely the company with the best AI model, but the company capable of writing the strongest twenty-year infrastructure guarantee — and any national framework that cares about competition in AI should think carefully, as it calibrates collateral schedules, about how heavy an admission price it is placing on the challengers.
Pillar 4 — States Are Becoming Laboratories for AI Utility Regulation
Virginia, Ohio, Indiana, Texas, and Pennsylvania are not merely competing to attract datacenters; they are experimenting, in the classic federalist sense, with different ways to govern them. Virginia emphasizes separate classification and direct cost assignment; Ohio a dedicated tariff and structured queue; Indiana long-duration contractual protection negotiated with the hyperscalers themselves; Texas verification and grid admission backed by gubernatorial directive; Pennsylvania a portable statewide model with flexibility incentives. These experiments are already cross-pollinating — Pennsylvania’s framework is being cited as a roadmap in North Carolina, Ohio’s revealed-demand data is quoted in every jurisdiction weighing a tariff, and the White House pledge machinery is recruiting governors as signatories — and they may ultimately produce a national template through convergence rather than command.[27, 29, 51] Governors should therefore understand electricity regulation as a component of AI industrial policy, with a counterintuitive corollary: the most successful state may not be the one offering the cheapest electricity. It may be the state offering the most credible, predictable, and politically sustainable allocation of AI infrastructure costs — because hyperscalers deploying $200 billion a year of capital value regulatory durability over marginal price, and because a bargain rate that collapses under ratepayer revolt in year three is worth less than a fair rate that survives for twenty.
Pillar 5 — The AI Boom Requires a New Social Contract for Electricity
The deepest lesson reaches beyond datacenters entirely. Electric grids are social infrastructure in the fullest sense: households depend on them for heat and safety, hospitals for life support, manufacturers for livelihoods, schools and military installations and water systems for basic function. The Five-Layer AI Economy is inserting customers of unprecedented scale into this shared system, and scale of that kind cannot be governed by the fine print of legacy tariffs; it requires a renegotiated social contract. The terms of that contract are becoming legible in the documents examined throughout this paper, and they can be stated in a single balanced sentence: artificial intelligence should pay for the extraordinary infrastructure artificial intelligence causes — while receiving genuine economic rewards when it demonstrably improves the grid. America should be able to build the electricity infrastructure necessary to compete in artificial intelligence; nothing in Ratebase Separation argues otherwise, and the framework’s flexibility provisions, its milestone-released collateral, and its shared-asset allocations all exist precisely so that legitimate AI investment proceeds. What the new contract refuses is the involuntary conscription of households as financiers of every speculative campus — and Harvard’s Peskoe has repeatedly identified the institutional precondition for holding that line.
“You really need vigilant regulators.”
— Ari Peskoe, Director, Electricity Law Initiative, Harvard Law School [39]
Pillar 6 — Speculative Megawatts Have Become a Financial Instrument — and Must Be Regulated Like One
The sixth pillar emerged from the data of 2026 rather than from theory. A position in an interconnection queue — costless to claim under the old rules — has become an asset with real option value and real externalities: it can secure scarce capacity ahead of rivals, inflate a developer’s prospects with lenders and tenants, and, when aggregated across hundreds of filings, move regional capacity prices by billions of dollars before a single facility exists. Monitoring Analytics attributed roughly $6.2 billion of a single PJM auction’s costs to datacenters not yet built; ERCOT’s queue reached five and a half times the state’s all-time peak; Borenstein described duplicate filings across territories as routine; and Ohio demonstrated that pricing the position eliminates most of it.[15, 22, 33, 43] The regulatory conclusion follows with the force of analogy: unpriced claims on scarce shared capacity are what securities regulators would call an unmargined position, and the remedy is the same in both domains — deposits, disclosure, verification, and escalating commitment. Texas’s audit of 461 facilities, Ohio’s tranche fees, Dominion’s contracting ladder, and PJM’s exploration of a separate procurement track for uncertain load are all, in this reading, the early machinery of margin requirements for megawatts.[15, 23, 35, 44] The queue has become a capital market. It should be governed with a capital market’s seriousness.
Pillar 7 — Voluntary Pledges Are Converging Into Enforceable Regulation
The final pillar concerns the trajectory from promise to law. The year 2026 opened with the grandest voluntary commitment in the history of the industry — seven companies signing the Ratepayer Protection Pledge at the White House, later joined by utilities, developers, and 23 governors, pledging to build, bring, or buy their power and pay its full cost “no matter what” — and closed its first half with one of those signatories appealing to a state supreme court against the most concrete regulatory implementation of exactly that promise.[11, 28, 51] The juxtaposition is not primarily an indictment of Microsoft, whose stated objection concerns mechanics rather than principle; it is a lesson about institutional physics. Voluntary pledges set expectations, but only tariffs, contracts, collateral, and commission orders allocate money — and when billions of dollars meet the difference between a press release and a payment obligation, the payment obligation is where the dispute will occur. Brookings’ assessment that the pledge “needs enforcement,” the energy industry’s shrug at its lack of any enforcement mechanism, and Peskoe’s observation that companies had been making similar promises for years without changing the underlying math all point the same way: the durable settlement of the AI-electricity question will be written in the regulatory instruments this paper has catalogued, not in pledge ceremonies.[30, 31, 40] The pledges matter — they establish the norm, recruit the governors, and make open cost-shifting reputationally expensive — but Ratebase Separation is what the norm looks like after it has been translated into enforceable text. The translation is the story of 2025–2027.
Conclusion: The Emergence of Ratebase Separation
The first phase of the AI infrastructure boom was measured in GPUs. The second was measured in datacenters. The third increasingly will be measured in gigawatts. But the next phase — the phase this paper has tried to bring into focus — may be measured in something less visible and more consequential than any of these: who signs the electricity guarantee.
Virginia’s GS-5 regime and its direct-assignment orders, Microsoft’s challenge to transmission-cost allocation before the state Supreme Court, Ohio’s dedicated datacenter tariff and its five-to-one collapse of speculative demand, Indiana’s hyperscaler-signed large-load settlement, Texas’s Batch Zero verification and audit of 461 facilities, and Pennsylvania’s first-of-its-kind statewide model tariff together reveal a regulatory transformation that is no longer prospective; it is underway, documented in final orders, effective tariffs, and pending litigation.[1, 9, 11, 13, 18, 22, 25] These states differ in mechanism, politics, and market structure, but they are confronting the same underlying problem: the traditional utility compact was not designed for individual customers capable of creating hundreds or thousands of megawatts of concentrated incremental demand, and no amount of nostalgia for the elegance of the old cost-socialization machinery can make it safe to run hyperscale load through it unmodified.
For more than a century, rate-base economics rested upon diversification: many customers jointly financed infrastructure expected to serve communities for decades, and the law of large numbers converted individual unpredictability into collective stability. Artificial intelligence introduces the opposite condition. A relatively small number of companies — Microsoft, Amazon, Google, Meta, Oracle, xAI, OpenAI’s infrastructure partners, and the emerging neocloud operators — can now individually determine the need for multibillion-dollar power investments, and their combined 2026 capital expenditure of roughly three-quarters of a trillion dollars means the determinations arrive not occasionally but continuously.[48, 49] That concentration changes the politics of electricity as thoroughly as it changes the finance. If hyperscale growth proceeds while residential electricity bills rise — and in PJM they are rising now, visibly, with datacenters identified by the market’s own independent monitor as the primary cause — the AI industry risks assembling, against itself, precisely the political coalition it can least afford: governors withdrawing incentives, commissions turning restrictive, communities blocking substations and corridors, legislators drafting moratoriums and special taxes.[30, 32, 33] Ratepayer politics could become a constraint on AI scaling as binding as chips, transformers, transmission equipment, or generation — and unlike those constraints, it cannot be solved with capital.
Ratebase Separation offers a different path, and its logic can now be stated in full. It does not require America to choose between AI leadership and affordable household electricity; it requires separating the risks appropriately, so that each is carried by the party best positioned to control it. Let AI developers finance customer-specific infrastructure, upfront where the but-for test is clean. Let utilities recover legitimate shared system costs through the traditional compact that still serves diversified load well. Let collateral, scaled to megawatts and released against milestones, protect the public against abandonment. Let binding contracts and verification audits distinguish real demand from speculative queues, and let the published gap between requested and committed load discipline every forecast that follows. Let flexible datacenters earn genuine economic rewards when they verifiably reduce grid stress, on the strength of the Duke findings that nearly 100 gigawatts of headroom exists for loads willing to bend.[41] Let shared infrastructure remain shared where benefits genuinely flow across the system. And let ordinary customers remain protected — structurally, contractually, and enforceably — when infrastructure exists primarily to serve extraordinary private loads.
This is why the phrase Ratebase Separation fits this paper so precisely. It describes more than a tariff, more than a rate class, more than any single state’s order. It identifies a structural transition in the political economy of electricity: the moment at which the Five-Layer AI Economy became so physically large that its first layer — energy — began to acquire its own regulatory boundaries, its own financial guarantees, and its own customer classification, distinct from the shared system that built modern American life. Virginia is where that boundary first became visible in law. Ohio, Indiana, Texas, and Pennsylvania demonstrate that it will not be the last, and the White House pledge machinery, whatever its enforceability, shows that even national politics now organizes itself around the boundary’s logic.[29, 51]
By 2030, the United States may no longer possess one electricity economy in the traditional sense. It may operate two overlapping systems: a conventional regulated economy designed around households, businesses, and traditional industry, financed through the century-old compact of shared rate base and gradual recovery; and an emerging compute electricity economy whose participants reserve hundreds of megawatts at a time, sign fourteen-year guarantees, post collateral by the megawatt, submit to verification audits, and increasingly finance outright the infrastructure needed to deliver their power. The line separating those two systems is what I call Ratebase Separation. And understanding where that line is drawn — between public obligation and private compute, between shared infrastructure and attributable infrastructure, and ultimately between the household electricity bill and the AI factory — may become one of the most consequential state-level economic policy questions of the entire artificial-intelligence era.
Footnotes and Endnotes:
[1] Virginia State Corporation Commission, News Release, “SCC Issues Order in Dominion Energy Virginia Biennial Review,” November 25, 2025. 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
[2] Virginia State Corporation Commission, “SCC Data Center Initiatives: Ensuring Data Centers Pay Their Own Costs,” Fact Sheet, February 2026. https://www.scc.virginia.gov/media/sccvirginiagov-home/about-the-scc/fact-sheets/scc-data-center-initiatives-02-2026.pdf
[3] Charles Paullin, “Virginia Regulators Approve New Dominion Rates, Assign More Costs to Data Centers,” Inside Climate News, January 7, 2026. https://insideclimatenews.org/news/07012026/virginia-regulators-approve-new-dominion-rates/
[4] Stephen D. Haner, “The SCC Decides: Dominion’s Rates and Profits Go Up, New Rules on Data Centers,” Thomas Jefferson Institute for Public Policy, December 3, 2025. https://www.thomasjeffersoninst.org/the-scc-decides-dominions-rates-and-profits-go-up-new-rules-on-data-centers/
[5] Dara Abasiita, “Virginia Now Makes Data Centers Post $1.5 Million a Megawatt,” Forbes, June 9, 2026. https://www.forbes.com/sites/daraabasiita/2026/06/09/virginia-now-makes-data-centers-post-15-million-a-megawatt/
[6] Data Center Dynamics, “Virginia Regulators Approve New Rate Class for Data Centers and Other Large Loads,” March 2026. https://www.datacenterdynamics.com/en/news/virginia-regulators-approve-new-rate-class-for-data-centers-and-other-large-loads/
[7] American Action Forum, “Virginia’s New Data Center Electricity Rate Class,” April 22, 2026. https://www.americanactionforum.org/insight/virginias-new-data-center-electricity-rate-class/
[8] Ethan Howland, “Virginia SCC Weighs Dominion Data Center Transmission Cost Allocation,” Utility Dive, July 15, 2026. https://www.utilitydive.com/news/virginia-scc-dominion-data-center-transmission-cost-allocation/825300/
[9] Charlie Paullin, “SCC Orders Dominion to Develop Tariff to Assign More Transmission Costs to Data Centers,” Virginia Mercury, August 5, 2026. https://virginiamercury.com/2026/08/05/scc-orders-dominion-to-develop-tariff-to-assign-more-transmission-costs-to-data-centers/
[10] The Piedmont Environmental Council, Press Release, “The Virginia State Corporation Commission Takes Important First Step of Requiring Large-Load Data Centers to Pay for Transmission Facilities Constructed Solely to Serve Them,” August 2026. https://www.pecva.org/resources/press/press-release-the-virginia-state-corporation-commission-takes-important-first-step-of-requiring-large-load-data-centers-to-pay-for-transmission-facilities-constructed-solely-to-serve-them/
[11] Tom Jowitt, “Microsoft Challenges Data Centre Cost Recovery, After Public Pledge,” Silicon UK (reporting on the Financial Times), September 2026. https://www.silicon.co.uk/cloud/ai/microsoft-virginia-appeal-631382
[12] BigGo Finance, “Microsoft Appeals Virginia SCC’s Data Center Transmission Cost Prepayment Rule to State Supreme Court,” September 2026. https://finance.biggo.com/news/052d8471-f6e1-4351-a6e7-7af55815584e
[13] AEP Ohio, “Data Center Tariff” (official tariff page; PUCO adoption July 9, 2025; effective July 23, 2025). https://www.aepohio.com/company/about/rates/data-center-tariff/
[14] Ethan Howland, “Ohio Regulators Approve AEP Data Center Interconnection Rules,” Utility Dive, July 10, 2025. https://www.utilitydive.com/news/Ohio-regulators-approve-aep-data-center-interconnection-rules/752690/
[15] The Ironton Tribune, “AEP Ohio Updates Commission on Data Center Load Under Contract,” March 1, 2026. https://irontontribune.com/2026/03/01/aep-ohio-updates-commission-on-data-center-load-under-contract-2/
[16] Nick Evans, “AEP Ohio Says New Data Center Tariff Is Working, Critics Aren’t Buying It,” Ohio Capital Journal, February 20, 2026. https://ohiocapitaljournal.com/2026/02/20/aep-ohio-says-new-data-center-tariff-is-working-critics-arent-buying-it/
[17] Data Center Dynamics, “Ohio Regulators Direct Data Center Firms to Provide Utility AEP Ohio With 180 Days Notice Prior to Grid Connection,” August 2026. https://www.datacenterdynamics.com/en/news/ohio-regulators-direct-data-center-firms-to-provide-utility-aep-ohio-with-180-days-notice-prior-to-grid-connection/
[18] Indiana Michigan Power / American Electric Power, “Indiana Michigan Power Receives Order in Large Load Settlement,” February 19, 2025. https://www.aep.com/news/stories/view/10037/
[19] Ethan Howland, “Indiana Regulators Approve ‘Large Load’ Interconnection Rules,” Utility Dive, February 20, 2025. https://www.utilitydive.com/news/indiana-iurc-large-load-interconnection-data-center-aep-amazon-google/740452/
[20] Power Technology, “IURC Approves I&M Large Load Tariff Settlement,” February 20, 2025. https://www.power-technology.com/news/iurc-american-electric-power-im-large-load-tariff/
[21] ERCOT, Market Notice M-A080326-01, “Update Regarding Batch Zero Timelines,” August 3, 2026. https://www.ercot.com/services/comm/mkt_notices/M-A080326-01
[22] Baker Botts LLP, “Texas Large Load Interconnection Update: ERCOT Batch Zero Pause and Verification Process,” August 2026. https://www.bakerbotts.com/thought-leadership/publications/2026/august/texas-large-load-interconnection-update—ercot-batch-zero-pause-and-verification-process
[23] Natalie Weber, “ERCOT to Survey More Than 400 Data Centers as Part of Gov. Abbott’s Audit,” Houston Public Media, September 11, 2026. https://www.houstonpublicmedia.org/articles/news/energy-environment/2026/09/11/561649/ercot-to-survey-more-400-data-centers-as-part-of-gov-abbott-audit/
[24] Blockspace / Yahoo Finance, “ERCOT Seeks to Suspend Batch Zero Deadlines for Texas Data-Center Audit,” August 11, 2026. https://www.yahoo.com/news/us/articles/ercot-seeks-suspend-batch-zero-134018194.html
[25] Pennsylvania Public Utility Commission, Press Release, “PUC Releases Final Order Establishing First-of-Its-Kind Large Load Model Tariff Framework,” May 13, 2026 (Chairman Steve DeFrank). https://www.puc.pa.gov/press-release/2026/puc-releases-final-order-establishing-first-of-its-kind-large-load-model-tariff-framework-05132026
[26] K&L Gates LLP, “Pennsylvania Public Utility Commission Adopts Model Interconnection Tariff for Large Load Customers,” May 29, 2026. https://www.klgates.com/thought-leadership/Pennsylvania-Public-Utility-Commission-Adopts-Model-Interconnection-Tariff-for-Large-Load-Customers-5-29-2026
[27] Womble Bond Dickinson, “Pennsylvania’s Large Load Model Tariff Offers a Road Map for North Carolina,” May 27, 2026. https://www.womblebonddickinson.com/us/insights/alerts/pennsylvanias-large-load-model-tariff-offers-road-map-north-carolina
[28] The White House, “Fact Sheet: President Donald J. Trump Advances Energy Affordability with the Ratepayer Protection Pledge,” March 4, 2026. https://www.whitehouse.gov/fact-sheets/2026/03/fact-sheet-president-donald-j-trump-advances-energy-affordability-with-the-ratepayer-protection-pledge/
[29] The White House, “Ratepayer Protection Pledge” (signed pledge and signatory tracker), updated July 2026. https://www.whitehouse.gov/ratepayer-protection-pledge/
[30] Sanjay Patnaik and Robert E. Litan, “The Pledge to Protect Ratepayers from AI Data Center Costs Needs Enforcement,” The Brookings Institution, July 9, 2026. https://www.brookings.edu/articles/the-pledge-to-protect-ratepayers-from-ai-data-center-costs-needs-enforcement/
[31] Lisa Martine Jenkins, “Energy Industry Greets White House Data Center Pledge With a Shrug,” Latitude Media, March 11, 2026. https://www.latitudemedia.com/news/energy-industry-greets-white-house-data-center-deal-with-a-shrug/
[32] Cathy Kunkel, “Projected Data Center Growth Spurs PJM Capacity Prices by Factor of 10,” Institute for Energy Economics and Financial Analysis (IEEFA), 2025. https://ieefa.org/resources/projected-data-center-growth-spurs-pjm-capacity-prices-factor-10
[33] Ethan Howland, “Data Centers Were 40% of PJM Capacity Costs in Last Auction: Market Monitor” (Monitoring Analytics report), Utility Dive, January 7, 2026. https://www.utilitydive.com/news/data-centers-pjm-capacity-auction/808951/
[34] Natural Resources Defense Council (NRDC), “Rising Demand from Data Centers Driving Reliability, Cost Concerns,” November 2025. https://www.nrdc.org/press-releases/rising-demand-data-centers-driving-reliability-cost-concerns
[35] Ethan Howland, “Data Centers Drove $6.3B in PJM Capacity Auction Costs: Market Monitor,” Utility Dive, July 20, 2026. https://www.utilitydive.com/news/pjm-data-centers-capacity-auction-imm-bowring/825626/
[36] U.S. Department of Energy / Lawrence Berkeley National Laboratory, “2024 Report on U.S. Data Center Energy Use” (Arman Shehabi et al.), December 20, 2024. https://www.energy.gov/articles/doe-releases-new-report-evaluating-increase-electricity-demand-data-centers
[37] Lawrence Berkeley National Laboratory, Center of Expertise for Data Center Energy, “Sector-Wide Modeling & Forecasting” (2025 Update: 11.8% of U.S. electricity by 2030). https://datacenters.lbl.gov/index.php/modeling-forecasting
[38] Eliza Martin and Ari Peskoe, “Extracting Profits from the Public: How Utility Ratepayers Are Paying for Big Tech’s Power,” Harvard Law School Electricity Law Initiative, March 2025; summarized in Ethan Howland, Utility Dive, March 10, 2025. https://www.utilitydive.com/news/utilities-subsidize-data-center-growth-ratepayer-cost-shif-harvard-peskoe/742001/
[39] Harvard Magazine, “How AI Could Be Raising Your Energy Bill” (interview with Ari Peskoe and Eliza Martin), 2025–2026. https://www.harvardmagazine.com/2025/07/harvard-ai-increasing-energy-costs
[40] Harvard Salata Institute / Harvard Climate Brief, “The Data Center Boom Is Colliding With the Grid’s Hardest Problems” (interview with Ari Peskoe), March 17, 2026. https://salatainstitute.harvard.edu/data-centers-ai-artificial-intelligence-grid-permitting-transmission-electricity-energy
[41] Tyler H. Norris et al., “Rethinking Load Growth: Assessing the Potential for Integration of Large Flexible Loads in US Power Systems,” Duke University Nicholas Institute for Energy, Environment & Sustainability, February 2025; coverage in Latitude Media. https://www.latitudemedia.com/news/the-us-grid-may-have-over-100-gw-of-load-to-spare/
[42] Christopher R. Knittel, Juan Ramon L. Senga, and Shen Wang, “Flexible Data Centers and the Grid: Lower Costs, Higher Emissions?,” NBER Working Paper 34065, MIT, 2025. https://www.nber.org/papers/w34065
[43] Broadband Breakfast, “Economist Warns Data Center Interconnection Requests May Be Inflated” (Severin Borenstein, UC Berkeley, at the Technology Policy Institute Aspen Forum), August 20, 2026. https://broadbandbreakfast.com/economist-warns-data-center-interconnection-requests-may-be-inflated/
[44] The Motley Fool, “Dominion Energy (D) Q4 2025 Earnings Call Transcript,” February 23, 2026 (CFO Steven Ridge; 48 GW contracted pipeline). https://www.fool.com/earnings/call-transcripts/2026/02/23/dominion-energy-d-q4-2025-earnings-transcript/
[45] Data Center Dynamics, “Dominion Energy Adds More Than 5GW of Contracted Data Center Load to Its Pipeline,” June 18, 2026. https://www.datacenterdynamics.com/en/news/dominion-energy-adds-more-than-5gw-of-contracted-data-center-load-to-its-pipeline/
[46] Prince William Times, “The Big Bottleneck: Power Crunch Keeps Data Centers in the Dark,” July 2026 (Dominion SCC disclosures; PJM CEO David Mills letter). https://www.princewilliamtimes.com/localnews/the-big-bottleneck-power-crunch-keeps-data-centers-in-the-dark/article_3b45f243-defa-573c-8ed1-c8709f73ce94.html
[47] Futurum Group, “AI Capex 2026: The $690B Infrastructure Sprint,” February 12, 2026. https://futurumgroup.com/insights/ai-capex-2026-the-690b-infrastructure-sprint/
[48] Statista, “Big Tech’s AI Spending to Reach $760 Billion in 2026” (Q2-2026 earnings of Microsoft, Alphabet, Meta, Amazon), July 31, 2026. https://www.statista.com/chart/35046/capital-expenditure-of-meta-alphabet-amazon-and-microsoft/
[49] Yahoo Finance / Goldman Sachs Research, “Meta, Microsoft, Amazon, and Alphabet Are About to Spend a Shocking Amount of Money to Dominate the AI Era” ($5.3 trillion hyperscaler capex, FY2025–FY2030), June 3, 2026. https://finance.yahoo.com/sectors/technology/article/meta-microsoft-amazon-and-alphabet-are-about-to-spend-a-shocking-amount-of-money-to-dominate-the-ai-era-115359575.html
[50] FactSet Insight, “Hyperscalers Tap External Financing as AI Capex Outruns Cash Flow,” July 23, 2026. https://insight.factset.com/hyperscalers-tap-external-financing-as-ai-capex-outruns-cash-flow
[51] U.S. Environmental Protection Agency, “President Trump Expands Historic Ratepayer Protection Pledge to Protect American Ratepayers, Lower Electricity Prices,” July 23, 2026. https://www.epa.gov/newsreleases/president-trump-expands-historic-ratepayer-protection-pledge-protect-american
[52] Carbon Direct, “The Billion-Dollar Case for Enabling Data Center Load Flexibility,” March 20, 2026. https://www.carbon-direct.com/insights/the-billion-dollar-case-for-enabling-data-center-load-flexibility
[53] Congressional Research Service, “Data Centers and Their Energy Consumption: Frequently Asked Questions,” Report R48646, 2025. https://www.congress.gov/crs-product/R48646
[54] Data Center Dynamics, “Duke Researcher Tyler Norris Named Google’s Head of Market Innovation, Advanced Energy,” July 27, 2026. https://www.datacenterdynamics.com/en/news/duke-researcher-tyler-norris-named-googles-head-of-market-innovation-advanced-energy/
