Introduction: The 96-Megawatt Signal
On September 21, 2026, a deceptively modest number entered America’s fast-expanding debate over the infrastructure required for artificial intelligence, and that number was 96 megawatts. Georgia Power and Google announced that Google would support uprates on Georgia Power’s owned portion of the nuclear units at Plants Vogtle and Hatch, work that is expected to add approximately 96 megawatts of new capacity to the grid to serve Georgia Power customers, while Google’s subscription is projected to enable roughly $900 million in customer benefits over the operating lives of those units [1]. The filings submitted to the Georgia Public Service Commission that same day, under Dockets 44280 and 56002, requested approval of a new Nuclear Uprate (NU-1) tariff structure and of a new extended power uprate for Hatch Units 1 and 2, while an extended uprate for Vogtle Units 1 and 2 had already been approved in the utility’s 2025 Integrated Resource Plan [1]. Under the arrangement, Google subscribes to the NU-1 tariff and receives the Zero-Emission Credits that represent the carbon-free attributes of the power generated by the uprated capacity, a structure that Georgia Power describes as designed to shield non-participating customers from the incremental costs of the uprate work [1].
“unlock the significant opportunity to bring online new nuclear power”
— Lucia Tian, Director of Advanced Energy Technologies, Google [1]
Ninety-six megawatts is small when placed beside the multigigawatt campuses now routinely discussed in AI infrastructure planning, it is small when compared with the nearly five gigawatts of total capacity at Plant Vogtle itself, and it is smaller still when measured against national aspirations involving fleets of advanced reactors, small modular reactors, new gas turbines, renewable generation, storage, and transmission corridors. Market commentators noticed the gap immediately, observing that the increment is modest relative to gigawatt-scale data-center demand and that its value still depends on regulatory treatment and successful engineering execution [5]. Yet the smallness of the number is exactly why it deserves careful attention, because it reveals a strategic pathway that is easy to overlook in a policy conversation dominated by dramatic construction targets. Instead of asking only how quickly the United States can construct entirely new power plants, utilities, hyperscalers, regulators, and investors are increasingly asking a different and more immediately actionable question: how much additional electricity can be extracted, safely and economically, from generating assets that already exist?
In nuclear engineering, the practical answer to that question carries a technical name, the power uprate, and it has a long regulatory history. When the NRC licenses a commercial reactor, it sets a maximum thermal power level that anchors many of the analyses demonstrating plant safety, so any increase in that level requires the Commission’s explicit approval through a license amendment, a process utilities have used since the 1970s to generate more electricity from their plants [6]. The NRC recognizes three categories of uprate. Measurement-uncertainty-recapture uprates, generally below two percent, rely on more precise instrumentation to recover margin previously held in reserve for uncertainty; stretch uprates, typically up to about seven percent, operate within the plant’s existing design capability; and extended power uprates go further, often requiring major modifications to turbines, pumps, motors, generators, transformers, and other balance-of-plant equipment [6]. The distinction matters because a nuclear power station is not merely its reactor vessel but an integrated industrial system of steam supply, turbine-generators, condensers, cooling infrastructure, electrical switchyards, instrumentation, safety systems, and transmission connections, and each of those subsystems may contain latent capability that can be unlocked without reproducing the full capital project represented by a greenfield station.
“U.S. commercial reactors are designed with excess capacity to allow for a potential uprate.”
— U.S. Nuclear Regulatory Commission [6]
America has done this before, and at meaningful scale. As of January 2022, the NRC had approved 171 uprates yielding a gain of roughly 24,089 megawatts thermal, or about 8,030 megawatts electric, which the agency characterizes as adding generating capacity equivalent to approximately eight new reactors [6]. What is new in 2026 is not the engineering practice but the economic environment surrounding it. The NRC’s schedule of expected applications, updated on September 8, 2026, now lists 31 potential uprates through 2032 representing about 7,336 megawatts thermal and approximately 2,421 megawatts electric, including extended uprates at Hatch Units 1 and 2, Vogtle Units 1 and 2, McGuire, Catawba, Columbia, Perry, Beaver Valley and Davis-Besse, alongside a substantial block of Constellation filings whose individual sites remain proprietary [2]. The schedule is a statement of expectations rather than a guarantee that each application will be filed or approved, but it demonstrates unmistakably that uprating has again become a live element of U.S. nuclear planning after years in which the practice had faded from strategic attention.
The timing is not accidental. Artificial intelligence is creating an electricity-demand problem that is unusual in both its speed and its geographic concentration. In the framework I have developed across this research series, the Five-Layer AI Economy begins with Layer 1, energy, before progressing through Layer 2 chips, Layer 3 datacenters, Layer 4 models, and Layer 5 applications and agents, and every accelerator installed at Layer 2 eventually expresses itself as a load at Layer 1. The International Energy Agency’s 2025 analysis projected that global data-center electricity consumption would more than double to around 945 terawatt-hours by 2030, with the United States accounting for nearly half of domestic electricity-demand growth over that period [7], and the agency’s April 2026 follow-up found that data-center electricity demand had already jumped 17 percent in 2025 while the capital expenditure of five large technology companies exceeded $400 billion and was expected to rise by a further 75 percent in 2026 [8]. As model training expands, inference becomes continuous, agents run persistently, and physical AI moves into factories, vehicles, and robots, electricity ceases to be a mere accompaniment to AI growth and becomes one of the principal determinants of how quickly every layer above it can expand.
“consuming as much electricity by 2030 as the whole of Japan does today”
— Fatih Birol, Executive Director, International Energy Agency [9]
This is why the nuclear debate surrounding artificial intelligence appears to be entering a second phase. The first phase, visible across 2023 through 2025, was largely about preservation and revival: preventing economically stressed reactors from closing, restarting retired facilities such as Palisades and the former Three Mile Island Unit 1 where technically and financially feasible, and recognizing firm, carbon-free nuclear generation as increasingly valuable in systems facing rapid data-center growth. The second phase, which the Georgia agreement crystallizes, is about extraction, meaning the disciplined search for additional safe output from reactors that are already connected to transmission systems, already operating with trained workforces, already occupying permitted sites with established cooling infrastructure, already supported by qualified supply chains, and already embedded within regional electricity markets and state regulatory relationships.
This paper calls that opportunity Reactor Yield, and it proposes that the concept changes the unit of analysis in a consequential way. For much of the nuclear-renaissance conversation, the most visible metric has been the number of reactors the country might build, whether ten new large reactors, dozens of small modular reactors, or a fleet of advanced designs sufficient to serve AI campuses and industrial electrification. Reactor Yield proposes a complementary metric, incremental megawatts per existing reactor, and therefore a different first question: before asking how many new plants America can build, how many additional safe megawatts can America’s existing fleet produce before the next generation of reactors reaches commercial operation? The distinction is especially pertinent because greenfield construction remains capital-intensive and slow; construction on Vogtle Units 3 and 4 began in 2009 against an original budget of $14 billion and planned operation in 2016 and 2017, yet the units entered commercial service in 2023 and 2024 with total project costs that Georgia Power estimates at more than $30 billion [10].
Power uprates are not substitutes for new nuclear construction, and this paper does not argue that they are. Every plant has unique technical, safety, cooling, transmission, licensing and economic constraints, uprates cannot deliver unlimited capacity, and meeting very large increases in electricity demand over the coming decades will almost certainly require new generation at scale. The argument is narrower and, I believe, more durable: time has become an economic variable in the AI race, and if several dozen or several hundred incremental megawatts can be produced at existing sites years sooner than multibillion-dollar greenfield plants can be completed, those megawatts carry disproportionate strategic value during the 2027–2030 window when AI load growth is steepest and new supply is scarcest. The central proposition of Reactor Yield therefore follows directly from that observation.
Before the AI economy waits for the next reactor, it should determine how much more electricity can be obtained, safely and economically, from the reactors it already has.
Why the Title “Reactor Yield”?
I chose the title Reactor Yield because the paper is not principally about whether nuclear power should be expanded, whether hyperscalers should purchase nuclear electricity, or whether new reactors will eventually be required, questions that a substantial literature already addresses. It is about the productive yield of an asset that has already been financed, constructed, licensed, interconnected and operated, and about the proposition that such an asset may contain additional electricity-producing capability that can be unlocked through uprates, equipment replacement, efficiency improvements and associated infrastructure investment. The word yield is borrowed deliberately from agriculture and finance, where it measures the output extracted from capital already in place rather than the quantity of new capital deployed, and that framing captures the analytical move this paper wants to make.
The word also reorganizes the analytical framework. Instead of counting reactors, the paper measures incremental megawatts extracted from each reactor, each nuclear site, and ultimately the entire U.S. fleet, which makes Reactor Yield a natural companion to the Five-Layer AI Economy: it converts Layer 1’s installed nuclear infrastructure from a fixed endowment into an expandable resource whose additional output can support growth across chips, datacenters, models, applications and agentic systems before an entirely new generation of plants becomes available. Earlier papers in this series examined adjacent questions, with Reactor Readiness asking whether the surrounding industrial system is prepared to deploy nuclear energy and Baseload Capture examining competition among large buyers for dependable long-duration electricity, whereas Reactor Yield asks a distinct question about how much more a given reactor can become capable of producing.
The remainder of the paper proceeds in seven sections. Section 1 develops the distinction between installed and extractable capacity and examines Georgia as a case study spanning three nuclear eras. Section 2 situates Reactor Yield within the AI demand shock and develops the economics of time, including the Reactor Yield Curve and a candid treatment of the limits and critiques of the concept. Section 3 analyzes hyperscaler-financed yield through the Google, Meta and Constellation transactions of 2026. Section 4 addresses the allocation question of who owns the incremental megawatt. Section 5 examines the federal policy and regulatory machinery now organized around uprates, Section 6 maps the 2027–2030 existing-fleet pipeline, and Section 7 distills eight pillars before the Conclusion returns to the 96-megawatt signal with which the paper began.
Section 1: The Hidden Megawatts Inside America’s Existing Nuclear Fleet
The conventional way to describe a nuclear fleet is through installed generating capacity, the figure that appears in utility filings, grid-operator databases, and national statistics, and by that measure the United States has for decades operated the largest commercial nuclear fleet in the world at roughly 97 gigawatts [10]. Installed capacity is a useful accounting convention, but it quietly embeds an assumption that is increasingly misleading in a period of electricity scarcity, namely that a plant’s licensed output is a fixed ceiling rather than a negotiable boundary defined by instrumentation accuracy, equipment ratings, safety analyses, cooling capability and regulatory permission. This section argues that Reactor Yield requires a second category of analysis alongside installed capacity, which I call extractable capacity, and that the gap between the two is where some of the fastest firm megawatts available to the AI economy are currently hidden.
1.1 From Installed Capacity to Extractable Capacity
Installed capacity describes what a plant is presently licensed and configured to produce, while extractable capacity asks how much additional safe electrical output might become available through engineering changes, instrumentation improvements, component replacement and regulatory approval. The difference may appear semantic, yet it transforms how a nuclear asset should be analyzed, because a 1,000-megawatt reactor ceases to be merely a fixed 1,000-megawatt asset and becomes, depending on its design, material condition, operating margins, cooling capability, turbine systems, electrical equipment and licensing basis, the anchor of a future portfolio of incremental megawatts. Some of those megawatts will be inexpensive, some will require major capital, some will be technically unavailable because of cooling-water limits or transmission constraints, and some may prove unusually valuable precisely because they can reach the grid years sooner than greenfield generating capacity.
The NRC’s own framing supports this reading. The agency notes that licensees sometimes modify or replace components to accommodate a higher power level, that depending on the desired increase and the original design this can involve major modifications such as replacing main turbines, and that every such change must be analyzed in a license amendment demonstrating that the new configuration remains safe [6]. Extractable capacity is therefore not a speculative engineering fantasy but a regulated, documented and historically exercised category of plant value, and the task of Reactor Yield analysis is to measure it systematically rather than discovering it project by project.
1.2 The Three Forms of Nuclear Uprating
The NRC’s three categories provide the natural foundation for a Reactor Yield framework, and each corresponds to a distinct economic logic that I label measurement megawatts, design-margin megawatts, and capital megawatts. The categories differ not only in the size of the increase they unlock but in the capital they require, the regulatory review they trigger, and the time they take to reach the grid, and the NRC has set explicit review targets for each: six months for a measurement-uncertainty-recapture application, nine months for a stretch uprate, and twelve months for an extended uprate, commitments communicated to the Nuclear Energy Institute in May 2024 as part of a broader effort to prepare dedicated reviewer cadres for an anticipated wave of applications [11].
Table 1. The three forms of Reactor Yield under the NRC uprate framework
| Category | Typical increase | Primary mechanism and equipment scope | NRC review target | Reactor Yield label |
| Measurement Uncertainty Recapture (MUR) | Less than 2% | More precise feedwater-flow instrumentation reduces the conservatism held in reserve for measurement error; little or no heavy equipment change [6] | 6 months [11] | Measurement megawatts |
| Stretch Power Uprate (SPU) | Up to about 7% | Uses margin within the original design envelope; instrumentation setpoint and operational changes, usually without major redesign of large components [6] | 9 months [11] | Design-margin megawatts |
| Extended Power Uprate (EPU) | Historically up to about 20% | Significant modifications to high-pressure turbines, condensate pumps and motors, generators, transformers and other balance-of-plant systems [6] | 12 months [11] | Capital megawatts |
Review targets are NRC goals for license amendment reviews, not guarantees; engineering, procurement and outage scheduling add time before and after the licensing step.
1.2.1 Measurement megawatts
Measurement-uncertainty-recapture uprates illustrate a subtle but powerful economic principle, which is that some additional electricity can come not from building more plant but from understanding the existing plant more precisely. Because reactor thermal power is inferred from measurements of feedwater flow and temperature, regulators historically required operators to hold a margin against the uncertainty in those measurements, and ultrasonic flow instrumentation that narrows the uncertainty allows part of that margin to be recovered as licensed output. The historical record also shows why measurement megawatts must be approached with rigor rather than enthusiasm: in 2004 the NRC approved a measurement-uncertainty-recapture uprate for Fort Calhoun, but after the vendor disclosed potential inaccuracies in the ultrasonic flow meter, the Omaha Public Power District asked to return the plant’s licensed thermal limit to its prior level before the uprate was ever implemented [12]. The lesson is that measurement megawatts are only as real as the measurement, and a credible Reactor Yield inventory must treat instrumentation quality as a first-order variable.
1.2.2 Design-margin megawatts
Stretch uprates capture electricity from capabilities that were built into the original engineering envelope but not used at the plant’s initially licensed operating level. Nuclear plants of the 1970s and 1980s were frequently designed with conservative margins in turbines, heat exchangers and electrical equipment, partly because designers were uncertain about long-term performance and partly because regulators and owners valued headroom. Stretch uprates convert a portion of that headroom into output, usually through setpoint and operational changes rather than wholesale equipment replacement, and for that reason they tend to sit on the lower and flatter part of the cost curve developed in Section 2.
1.2.3 Capital megawatts
Extended power uprates are the most strategically interesting category for the AI economy because they can deliver increases large enough to matter at the scale of a data-center campus. The NRC’s approved-applications record shows what such projects look like in practice: in 2012 the agency authorized a 13 percent extended uprate plus a 1.7 percent measurement recapture at Turkey Point Units 3 and 4, and 10 percent extended uprates plus 1.7 percent measurement recapture at St. Lucie Units 1 and 2 [12]. These are capital megawatts in the literal sense, requiring new or rewound generators, replaced high-pressure turbines, higher-capacity pumps and transformers, and sometimes cooling-system upgrades, and they are therefore not free capacity. Neither, however, are they equivalent to constructing a new station, and the economic question becomes whether modifying an existing nuclear platform can produce additional firm megawatts at lower cost, on a shorter schedule, or with lower execution risk than building equivalent firm capacity from scratch.
1.3 Georgia as the September 2026 Case Study
The Google–Georgia Power agreement makes Georgia an unusually instructive case because the state contains, within a few hundred miles, three distinct eras of American nuclear economics. Plant Hatch near Baxley, with two boiling water reactors that entered commercial operation in 1975 and 1979, represents the mature fleet; Vogtle Units 1 and 2 near Waynesboro, pressurized water reactors that entered service in the late 1980s and are licensed into the mid-2040s, represent the last great wave of large twentieth-century construction; and Vogtle Units 3 and 4, AP1000 units connected to the grid in 2023 and 2024, represent the newest large reactors in the national system [13][14]. The proposed uprates thus allow a single state to display the full arc from existing fleet to expanded existing fleet to newly built fleet, and to compare the economics of extraction directly against the economics of construction on the same site and under the same regulator.
Georgia is also the place where the demand side of the equation is most visible. In its 2025 Integrated Resource Plan, Georgia Power projected approximately 8,500 megawatts of electrical load growth over six years, driven substantially by data centers, an upward revision from the 8,200 megawatts in its earlier proposal [15]. Nuclear already supplies more than a quarter of Georgia Power’s energy, and Vogtle is described by the utility as the nation’s largest generator of clean energy [1]. The uprate program therefore sits at the intersection of an extraordinary load forecast, a nuclear-heavy generation mix, and a vertically integrated regulatory model in which the Public Service Commission reviews both resource plans and large-load contracts, which is exactly the combination in which incremental firm megawatts carry the highest option value.
It is also worth recognizing what the Google transaction is not. The 96 megawatts attributable to Google’s subscription sit within a larger uprate program at Southern Company’s plants: the Department of Energy reported in May 2026 that planned uprates at six existing reactors across Hatch, Vogtle and Farley would produce a combined 345 megawatts of additional baseload capacity, pursued alongside license renewals as part of Energy Dominance Financing loans to Southern Company totaling $26.5 billion across Alabama and Georgia [16][17]. The Google agreement is therefore best understood not as the origin of Georgia’s uprate program but as a financing and attribution layer placed on top of it, which is itself an important finding for the allocation questions discussed in Section 4.
1.4 Nuclear Life Extension Multiplies Reactor Yield
Uprating becomes considerably more valuable when combined with longer reactor life, because the economic worth of an incremental megawatt is a function not only of its size but of the number of years over which it will be produced. The NRC’s subsequent-license-renewal framework allows qualifying reactors to extend operation from 60 toward 80 years, subject to safety and environmental review, and 2026 brought a notable acceleration in that process. On June 12, 2026, the NRC renewed the licenses for Hatch Units 1 and 2 in under twelve months, making them the second and third units approved under a streamlined process after Duke Energy’s Robinson plant, whereas previous subsequent renewals had taken roughly two and a half years on average [14]. Hatch Unit 1 is now licensed through 2054 and Unit 2 through 2058, up to 80 years of operation for each reactor [18], and industry tracking data indicate that projects at Hatch now include power uprates and refueling outages representing roughly $380 million of investment [19].
The resulting multiplication effect is straightforward to state and easy to underestimate: additional megawatts multiplied by additional years of operation equals additional lifetime reactor yield. A 50-megawatt uprate that operates for five years before a plant retires is one economic asset, while the same 50 megawatts operating for three additional decades under an extended license is something categorically different, since the capital cost is amortized over many times more megawatt-hours and the risk premium demanded by financiers falls accordingly. The economic value of uprates should therefore never be evaluated independently of license extension, and the Hatch sequence, in which subsequent renewal was approved in June 2026 and an extended-uprate application followed with Google’s financial support in September, illustrates how the two decisions reinforce one another.
1.5 Building a National Reactor Yield Inventory
If extractable capacity is a real and regulated category of value, it follows that it should be measured systematically rather than discovered opportunistically. The United States could conceptualize each operating reactor through a standardized Reactor Yield profile, and the Department of Energy’s UPRISE initiative, which has stated that its near-term work will include assessing plant equipment for increased output, examining supply-chain readiness, and validating economic models to support investment decisions [20], is the natural institutional home for such an inventory. Table 2 sets out the fields that a credible profile would need to contain.
Table 2. Template for a plant-level Reactor Yield profile
| Profile field | Why it matters for extractable capacity |
| Existing licensed thermal and electrical output | Establishes the baseline from which all incremental yield is measured. |
| Measurement-recapture potential | Identifies the lowest-cost increment and tests instrumentation quality, as the Fort Calhoun reversal demonstrates [12]. |
| Stretch and extended uprate potential | Separates design-margin megawatts from capital megawatts and indicates the likely NRC review pathway [11]. |
| Turbine-generator, pump and condenser headroom | Determines whether major balance-of-plant equipment must be replaced and how long procurement will take. |
| Cooling-system and thermal-discharge constraints | Higher thermal power requires rejecting more heat; water availability and permit limits can cap yield. |
| Transformer, switchyard and transmission export capability | Incremental megawatts are worthless if they cannot be delivered; interconnection limits can delay even restarts [21]. |
| Fuel and core-design implications | Higher power levels affect fuel management and cycle economics. |
| Outage requirements and scheduling | Most EPU work is installed during refueling outages, which fixes the calendar of delivery. |
| Capital cost and approval timeline | Determines cost per incremental megawatt and time-to-megawatt. |
| Remaining licensed life and SLR status | Converts incremental megawatts into lifetime megawatt-hours, as in the Hatch case [18]. |
| Presence of a large-load counterparty | Indicates whether hyperscaler or industrial financing can de-risk the investment. |
The profile is a proposed analytical instrument; fields are illustrative and would be refined through the DOE UPRISE equipment-assessment work.
The result would no longer be merely a list of nuclear plants and their nameplate ratings but a national inventory of potentially extractable nuclear megawatts, ranked by cost, schedule, risk and remaining life. Such an inventory would give state commissions a common reference point when evaluating uprate proposals, give hyperscalers a map of where their capital could create firm supply rather than merely reallocating it, and give federal lenders a disciplined basis for prioritizing loans, and it would convert Reactor Yield from an intuition into an instrument of planning.
Section 2: The AI Demand Shock and the Economics of Time
Every infrastructure strategy is ultimately a response to a demand forecast, and the credibility of Reactor Yield as a strategy depends on whether the AI-driven increase in electricity demand is large, near-term and geographically concentrated enough to justify paying a premium for megawatts that arrive early. This section first assembles the evidence on the demand side from the 2024–2026 literature and from second-quarter 2026 corporate disclosures, then argues that the correct comparison between uprates and new construction is temporal rather than technological, develops time-to-megawatt as an explicit economic variable, introduces the Reactor Yield Curve, and closes with a candid treatment of the strongest criticisms of the concept, because a framework that cannot survive its critics does not deserve to guide capital.
2.1 The Scale and Speed of the Demand Signal
The most authoritative federal estimate of U.S. data-center electricity use comes from Lawrence Berkeley National Laboratory. Its December 2024 report, led by Arman Shehabi and co-authored with Jonathan Koomey, Eric Masanet and others, found that data centers consumed about 176 terawatt-hours in 2023, roughly 4.4 percent of U.S. electricity, and projected a range of 325 to 580 terawatt-hours by 2028, equivalent to between 6.7 and 12 percent of national consumption [22]. The laboratory’s 2025 update extended the horizon and raised the central case, estimating that data centers could account for 11.8 percent of total U.S. electricity by 2030 within a scenario range of 9.5 to 15.3 percent [23]. Even allowing for the wide uncertainty that both reports emphasize, the central estimates imply a scale of new load that the U.S. grid has not absorbed since the decades of post-war electrification.
International evidence points in the same direction. The IEA’s April 2026 analysis found that power consumption per AI task is falling at a rate unprecedented in energy history, yet that rising usage and energy-intensive applications such as AI agents mean total data-center consumption is still set to roughly double by 2030 from about 485 to 950 terawatt-hours, with AI-focused facilities tripling to around 465 terawatt-hours [24]. The agency’s executive director framed the shift not only as a burden but as a catalyst for new supply, observing that the technology sector is increasingly helping to bring forward innovative generation and storage solutions.
“while AI is still an energy taker, it is also becoming an energy maker”
— Fatih Birol, Executive Director, International Energy Agency [8]
Corporate disclosures through the second quarter of 2026 show that these forecasts are converting into contracted load, particularly in the Southeast where Georgia Power operates. Southern Company reported second-quarter 2026 adjusted earnings of $1.13 per share, above both its own estimate and consensus, and disclosed that system-wide data-center load had surpassed 1.2 gigawatts, an increase of more than 500 megawatts in a year, with data-center usage up 55 percent year over year [25]. Management reported that its prospective pipeline of large industrial and data-center projects remained well above 75 gigawatts, that 17 gigawatts were already contracted, and that another 8 gigawatts were in late stages of development [25]. During the quarter alone the company signed 6 gigawatts of new large-load contracts, including a 3.2-gigawatt, 25-year agreement with OpenAI near Savannah that incorporates one gigawatt of flexible demand response, and raised its full-year guidance to the top of its $4.50 to $4.60 range [26].
“we expect this trend to continue accelerating”
— David Poroch, Chief Financial Officer, Southern Company [25]
Two qualifications keep this evidence honest. First, contracted load is not operating load: Southern’s chief executive acknowledged that the company has learned to work closely with customers because load ramps may diverge from what was originally forecast, and analysts noted that seventeen gigawatts of contracts will take years to materialize while some prospective projects will never reach construction [27]. Second, efficiency gains are real and large, which is why forecasts span such wide ranges. Neither qualification, however, undermines the core point for Reactor Yield, which is that the demand uncertainty is overwhelmingly about how fast and how much load will arrive, not about whether substantial new firm supply will be needed in the late 2020s.
“may not be what was projected”
— Chris Womack, Chairman and CEO, Southern Company, on large-load ramp rates [27]
2.2 The Wrong Comparison and the Right One
It would be misleading to argue that uprates eliminate the need for new nuclear generation, and a great deal of public confusion arises from framing the choice as a technological contest between existing reactors and future reactors. The better comparison is temporal, because the AI infrastructure boom is creating electricity demand concentrated in the late 2020s, while most of the advanced reactors and greenfield large reactors being discussed today will arrive in the 2030s. The Department of Energy’s own assessment of nuclear-powered data centers acknowledges that widespread commercial advanced reactors are likely to arrive in the 2030s and that near-term data-center energy will come largely from natural gas, coal, wind, solar and existing nuclear plants [28]. The economic question therefore is not which technology is superior in the abstract but what supplies firm electricity during the interval, and Reactor Yield is one of the few nuclear answers available inside that interval.
“There are needs on different time scales”
— Patrick White, former Research Director, Nuclear Innovation Alliance [29]
The same analysis in MIT Technology Review noted that many technology companies will require large amounts of power within three to five years while new nuclear plants can take close to a decade to build, and it reported a Department of Energy estimate that uprates at existing sites could add between two and eight gigawatts without building new infrastructure [29]. Meta’s head of global energy made a parallel argument about relicensing, describing support for extending operating plants as one of the most powerful near-term tools for any buyer unwilling to wait a decade for new technology [29]. Uprates and license extension are thus two expressions of the same temporal logic: they convert existing assets into near-term supply while the long-duration construction program matures.
2.3 Time-to-Megawatt as an Economic Variable
Traditional levelized-cost comparisons emphasize dollars per megawatt-hour, implicitly treating a megawatt-hour delivered in 2035 as interchangeable with one delivered in 2028 once both are discounted at a cost of capital. AI infrastructure introduces a variable that levelized-cost analysis handles poorly, which I call time-to-megawatt: the elapsed time between a decision to procure firm capacity and the moment that capacity can actually serve a load. For a hyperscaler whose accelerators depreciate over a few years and whose competitive position depends on bringing training and inference capacity online ahead of rivals, a megawatt available in 2028 may carry far more strategic value than an otherwise cheaper megawatt available in 2035, because the idle chips, delayed model releases and forgone revenue associated with waiting are not captured in the utility’s cost of generation.
Princeton’s Jesse Jenkins, whose ZERO Lab models electricity-system expansion, has placed 60 to 80 gigawatts of additional load growth by 2030 within credible territory while warning that much of it may be deferred because supply cannot keep pace [30]. That diagnosis, from a researcher whose work has generally urged skepticism toward inflated demand projections, underscores that the binding constraint in the late 2020s is less the existence of demand than the speed at which firm supply can be added.
“we’re not able to build fast enough on the power side”
— Prof. Jesse Jenkins, Princeton University, ZERO Lab [30]
This suggests a broader infrastructure metric that developers and regulators should apply alongside levelized cost, namely a composite of capital cost, delivery time, reliability and operating life. Rather than optimizing only for the cheapest theoretical generation technology, developers increasingly need to optimize for the electricity that can actually reach their campus within the required construction window, and on that composite metric uprates at plants with long remaining licenses score unusually well, because they combine firm output, established interconnection, and a regulatory review target measured in months rather than years [11].
2.4 The Vogtle Lesson
Vogtle Units 3 and 4 demonstrate why schedule must sit at the center of any Reactor Yield discussion, and they do so in both directions. On one side, the project’s history is a cautionary tale: construction began in 2009 against an original expectation of $14 billion and commercial operation in 2016 and 2017, yet the units entered service in 2023 and 2024 with total costs that Georgia Power estimates at more than $30 billion [10], and the Center for Strategic and International Studies characterizes the undertaking as a fifteen-year effort costing over $36 billion when all participants’ costs are included [31]. On the other side, the completion of Vogtle despite a contractor bankruptcy and a pandemic demonstrated that the U.S. industry can still build supply chains, train a skilled workforce, and finish large reactors, and a 2025 technical report from Idaho National Laboratory, MIT and Oak Ridge argues that first-of-a-kind overruns can fall substantially through standardization, modularization and the transfer of lessons learned [32].
MIT’s Koroush Shirvan has quantified what a next pair of AP1000s might cost. His 2024 projection estimated the overnight capital cost of two additional AP1000 units at the Vogtle site at between $8,300 and $10,375 per kilowatt, with a construction schedule of roughly 80 to 96 months from first concrete to commercial operation [33]. Even on that optimistic learning-curve basis, a new large reactor ordered in 2026 would be unlikely to generate power before the early-to-mid 2030s, which means new nuclear plants and uprates solve different time problems: greenfield reactors address the long-duration capacity problem of the 2030s and 2040s, while Reactor Yield addresses a portion of the near-to-medium-term capacity problem of 2027 through 2031. A Stanford-based critic of the nuclear industry reaches a sharper conclusion about large construction, and his objection is worth stating in his own words because it strengthens rather than weakens the case for prioritizing extraction.
“big reactors with several times the promised cost and construction time”
— Amory B. Lovins, Stanford University, on the global record of large reactor projects [34]
2.5 Brownfield Nuclear Economics
Existing nuclear sites possess assets whose economic value is easy to underestimate because they rarely appear on a balance sheet at anything close to replacement cost. These include transmission interconnections and switchyards, trained and licensed operators, security infrastructure and emergency-planning systems, cooling systems and water rights, long regulatory histories with the NRC, maintenance organizations and nuclear-qualified supplier relationships, spent-fuel storage infrastructure, established local tax relationships, community familiarity with nuclear operations, and decades of plant-specific operational data. An uprate builds upon this installed institutional and physical capital, whereas a new reactor must create much of it, and that asymmetry is why cost per incremental megawatt at an existing site can be materially lower than the per-kilowatt cost of a greenfield plant even when the uprate requires major turbine and generator work.
The brownfield advantage is not absolute, and the Crane restart illustrates its limits. Even a plant that previously operated at the site, backed by a 20-year Microsoft power purchase agreement and a $1 billion federal loan for a $1.6 billion project, faced PJM studies indicating that transmission upgrades needed for full grid delivery might not be completed until the 2030s, prompting Constellation to seek permission from federal regulators to transfer interconnection rights from a fossil unit [21]. The Federal Energy Regulatory Commission granted that waiver in June 2026, allowing the transfer of 760 megawatts of capacity interconnection rights from the Eddystone plant and preserving a restart target in the second half of 2027 [35]. For Reactor Yield the implication is direct: incremental megawatts that exceed a site’s existing export capability can inherit exactly the interconnection delays that uprates are supposed to avoid, which is why transmission export capability belongs in every plant-level profile.
“we continue to expect to start this unit in 2027”
— Joseph Dominguez, President and CEO, Constellation Energy, on the Crane restart [36]
2.6 The Reactor Yield Curve
Each reactor can be understood as possessing its own Reactor Yield Curve, an ordering of available increments from cheapest to most expensive. The first increment is typically inexpensive measurement recapture, the next may require modest instrumentation and setpoint changes within the design envelope, a further increment may require high-pressure turbine replacement or generator rewinding, and beyond that additional output may require new pumps, condensers, main transformers, or cooling-system modifications. Eventually the curve reaches a point at which extracting another megawatt becomes uneconomic or technically impractical, whether because of reactor-core thermal limits, cooling-water constraints, or the cost of transmission reinforcement. Figure 1 presents the concept schematically.
Figure 1. Conceptual Reactor Yield Curve. Category boundaries follow NRC definitions; the cost index is illustrative and not specific to any plant.
The policy implication of the curve is that the objective should never be maximum uprating at any cost but identification of the economically rational portion of the curve for each unit, which will differ widely across the fleet. A boiling water reactor with generous turbine margins and abundant cooling water might rationally pursue a large extended uprate, while a pressurized water reactor constrained by thermal-discharge permits might stop after measurement recapture. The presence of a hyperscaler counterparty shifts the curve’s economic stopping point to the right, because a buyer willing to pay a premium for firm, carbon-free megawatts delivered early effectively raises the value of each increment, and this is precisely the mechanism through which the Google and Meta transactions discussed in Section 3 create capacity that might not otherwise be built.
2.7 Limits, Critiques and Alternatives
A framework intended to guide billions of dollars of investment must confront its strongest objections, and Reactor Yield faces at least three. The first is the scale objection: two or even five gigawatts of uprates are small relative to the tens of gigawatts of load now under contract nationally, and the IEA projects that renewables supported by storage, together with natural gas, will meet the largest shares of growth in data-center demand through 2035 [7]. The second is the sustainability objection raised by researchers who question whether the pace of data-center construction can be matched by clean supply at all.
“The demand for new data centers cannot be met in a sustainable way.”
— Noman Bashir, MIT Computer Science and Artificial Intelligence Laboratory and MIT Climate and Sustainability Consortium [37]
The third objection is that the fastest megawatts may not be generation at all but flexibility. A 2025 study by Tyler Norris, Tim Profeta, Dalia Patiño-Echeverri and Adam Cowie-Haskell at Duke University’s Nicholas Institute estimated that about 76 gigawatts of new load, roughly ten percent of national peak demand, could be integrated into existing U.S. power systems if those loads accepted curtailment averaging 0.25 percent of their maximum uptime, rising to 98 gigawatts at 0.5 percent and 126 gigawatts at 1 percent, with the Southern Company balancing authority alone able to accommodate about 8 gigawatts at the 0.5 percent level [38].
“could accommodate significant load additions with modest flexibility measures”
— Tyler Norris, Duke University, Nicholas School of the Environment [39]
Reactor Yield does not dismiss these objections but incorporates them. The scale objection is correct and is precisely why this paper frames uprates as a bridge rather than a destination. The sustainability objection strengthens the case for extracting carbon-free output from existing reactors, since each incremental nuclear megawatt at an existing site is one that need not come from new fossil capacity during the interval. And the flexibility finding is complementary rather than competitive, because flexible interconnection increases how much load a system can absorb while firm incremental supply determines how much energy that load can actually consume over a year; the OpenAI contract in Georgia, with its one gigawatt of flexible demand, shows the two approaches already being combined in practice [26]. The honest conclusion is that Reactor Yield is one necessary instrument within a portfolio, most valuable where it is cheapest per megawatt, fastest to deliver, and paired with long remaining operating life.
Section 3: Hyperscaler-Financed Reactor Yield
For most of the history of the U.S. nuclear fleet, the decision to invest in an uprate was made by a utility or merchant generator weighing electricity-system economics against the prospect of regulatory cost recovery or wholesale-market revenue, and many technically feasible uprates were never pursued because the incremental revenue could not justify the capital and outage risk. The AI era introduces a new source of capital into that calculation, namely large-load customers with very long planning horizons, investment-grade balance sheets and an unusually high valuation of firm, carbon-free electricity delivered on a predictable schedule. This section examines how that capital is entering the nuclear fleet through three distinct 2026 structures, the Google–Georgia Power subscription tariff, the Meta–Vistra long-term purchase agreements, and customer-enabled uprates within Constellation’s contracting program, and it argues that these transactions represent a shift from energy procurement to energy-capacity creation.
3.1 Google Changes the Capital Question
The significance of the Georgia agreement lies less in its 96 megawatts than in the financing question it answers. Georgia Power’s state regulators had approved a Vogtle Units 1 and 2 extended uprate in the 2025 resource plan, but the specifics of how to pay for the work had not been settled, and the September 2026 filing proposes a subscription program through which large customers such as Google can fund both the Vogtle and Hatch uprates [40]. The utility describes this subscription model as an extension of renewable-energy programs it has offered for years, and states that the structure, originally contemplated in the 2025 resource plan, is intended to protect non-participating customers from incremental uprate costs [1]. Georgia Power’s senior vice president for strategic growth characterized the program as an illustration of how large-load growth can benefit all customers, while aligning with the company’s Customer Protection Pledge [1].
Google’s broader nuclear portfolio shows that the uprate deal is part of a deliberate strategy rather than an isolated experiment. In the months preceding the Georgia agreement, the company signed a 22-year power purchase agreement with Fortum’s Loviisa plant in Finland, its first international nuclear deal, as well as agreements supporting the restart of NextEra’s 615-megawatt Duane Arnold Energy Center in Iowa, a 200-megawatt agreement with fusion developer Commonwealth Fusion Systems, an agreement with small-modular-reactor developer Kairos Power, and a strategic partnership with Elementl Power to develop three nuclear projects [41]. Seen against that portfolio, the uprate agreement fills a specific temporal niche: it is the Google nuclear commitment most likely to deliver incremental firm megawatts before the end of the decade.
3.2 From Power Purchase Agreement to Capacity Creation
Traditional hyperscaler energy procurement frequently involves contracting for electricity that was already expected to be produced, which can support a plant’s continued operation but does not necessarily add supply to a constrained grid. Reactor Yield introduces a different structure in which the customer helps finance new output from an old asset, and the distinction matters both economically and politically. When a large buyer merely contracts for existing output, other customers may perceive it as competing for a fixed pool of firm electricity; when the buyer’s capital expands the pool, the transaction can be framed, and in favorable regulatory designs can genuinely function, as a net addition to system capacity.
Vistra’s agreements with Meta, announced on January 9, 2026, are the clearest illustration of this shift at scale. The 20-year agreements cover 2,176 megawatts of existing output from the Perry and Davis-Besse plants in Ohio together with 433 megawatts of incremental capacity from uprates at Perry, Davis-Besse and Beaver Valley in Pennsylvania, which Vistra describes as the largest nuclear uprates ever supported by a corporate customer in the United States, with the electricity continuing to flow to the grid for all users [4]. The same announcement committed Vistra to begin planning subsequent license renewals at all three plants, and a company executive noted how far the fortunes of these assets had reversed.
“these plants were on a path to retirement”
— Stacey Doré, Chief Strategy and Sustainability Officer, Vistra, on Perry, Davis-Besse and Beaver Valley as recently as 2020 [4]
The Vistra transaction also shows that capacity creation takes time even at existing plants. Vistra’s chief executive indicated in February 2026 that delivery of Perry’s existing output would begin in December 2026 and Davis-Besse’s in December 2027, but that the uprate capacity was longer-dated, with Perry’s uprates expected online in 2031 and the remaining uprates following annually until all were complete by 2034 [42]. Those dates are consistent with the NRC’s schedule, which lists Perry’s extended-uprate application for the third quarter of 2029 and Beaver Valley and Davis-Besse filings in 2030 through 2032 [2]. Meta’s announcement frames the uprates as part of a broader portfolio of up to 6.6 gigawatts of new and existing nuclear capacity by 2035, including support for TerraPower and Oklo projects [43], which places Reactor Yield in its proper position as the existing-fleet layer of a multi-decade nuclear strategy.
3.3 Customer-Enabled Uprates in Constellation’s Fleet
Constellation Energy, the country’s largest nuclear operator, provides the third model, in which uprates are embedded within a broader long-term contracting program. The company’s second-quarter 2026 results showed adjusted operating earnings of $2.55 per share against $1.91 a year earlier, driven by the Calpine acquisition, higher PJM capacity prices and commercial performance [44], and management raised full-year guidance to $11.50 to $12.50 per share [45]. More revealing for Reactor Yield were the operational and contractual details: Constellation completed refueling outages in an average of 23 days against an industry average of 38 while simultaneously implementing a turbine uprate at Byron Unit 1, and it announced roughly 920 megawatts of new long-term nuclear agreements averaging 18.5 years, of which 890 megawatts came from existing generation while a customer commitment enabled a 30-megawatt uprate at Dresden [46]. Industry reporting identifies that Dresden-linked agreement as a power purchase arrangement with Walmart [47].
The Byron detail deserves emphasis because it demonstrates a practical feature of Reactor Yield that is invisible in national statistics: capital megawatts are frequently installed during refueling outages that must occur anyway, which means the marginal outage cost of an uprate can be small when the work is sequenced skillfully. An operator that can execute turbine replacements inside a 23-day outage window possesses a genuine competitive advantage in the extraction economy, because its time-to-megawatt is shorter and its lost-generation cost lower than those of a less efficient peer, and this is one reason why operational excellence, not only engineering headroom, belongs in any assessment of a fleet’s extractable capacity.
Table 3. Corporate-supported Reactor Yield transactions and programs, 2026
| Counterparties | Plants | Incremental MW | Structure and timing |
| Google – Georgia Power | Vogtle 1–2; Hatch 1–2 (Georgia Power share) | ~96 | NU-1 subscription tariff; Google receives Zero-Emission Credits; subject to Georgia PSC approval; NRC EPU filings expected 2027–2028 [1][2] |
| Meta – Vistra | Perry, Davis-Besse, Beaver Valley | 433 | 20-year PPAs alongside 2,176 MW existing output; uprates online 2031–2034 [4][42] |
| Customer – Constellation | Dresden | 30 | Long-term PPA within ~920 MW contracting program; customer commitment enables uprate [46][47] |
| DOE UPRISE / EDF – Southern Company | Hatch, Vogtle, Farley (six reactors) | 345 | Planned uprates pursued with license renewals under $26.5 billion in federal loans [16][17] |
The Southern Company programs overlap: the Google-supported Georgia increments form part of the broader Hatch–Vogtle–Farley uprate plan and are not additive.
Figure 2. Announced incremental-megawatt programs at existing U.S. reactors in 2026. Georgia programs overlap and should not be summed.
3.4 A Generalized Hyperscaler Uprate Model
Abstracting from these three cases, a generalized hyperscaler uprate model involves six parties with distinct roles. The utility or merchant owner holds the generating asset and the customer relationship; the nuclear operator, sometimes a sister company such as Southern Nuclear, engineers and executes the uprate; the hyperscaler provides subscription revenue, long-term contractual support, or other financing that de-risks the investment; the state regulator, in vertically integrated markets, evaluates the allocation of costs, risks and benefits among customer classes; the NRC independently reviews any change in licensed power level and the associated safety analyses; and grid customers potentially receive system benefits, depending on project structure and regulatory treatment. The result is a financial bridge between Layer 3 of the Five-Layer AI Economy, where datacenters are sited and financed, and Layer 1, where the electricity that powers them is produced.
The model’s institutional variants matter. In a vertically integrated state such as Georgia, the subscription tariff routes hyperscaler money through a regulated rate structure overseen by the Public Service Commission, which provides transparency but also subjects the arrangement to public contestation. In a restructured market such as PJM, where Vistra and Constellation operate, the same economic function is performed through bilateral power purchase agreements between a merchant generator and a corporate buyer, which is faster and more flexible but places greater weight on wholesale-market rules for large loads, an area in which Constellation expects regulatory clarity on PJM’s backstop procurement and interim resource-adequacy mechanisms by the end of 2026 [46].
3.5 Corporate Credit Meets Nuclear Infrastructure
Large technology companies possess financial characteristics that can materially alter the economics of nuclear upgrades. They can offer long contractual tenors, substantial balance sheets, predictable if rapidly growing capacity requirements, investment-grade credit, tolerance for premium pricing of firm clean electricity, and a willingness to commit years before the datacenters that will use the power are completed. Those characteristics reduce the revenue uncertainty surrounding infrastructure whose economic life may extend for decades, and in effect they allow the cost of capital for an uprate to be priced closer to the credit of the hyperscaler than to the merchant risk of the generator. Reporting in mid-2026 observed that every major AI company had signed at least one nuclear agreement during the year, together totaling nearly ten gigawatts of capacity, and argued that technology firms were increasingly financing plants directly rather than only purchasing their output [48].
The federal government has built a complementary channel. The DOE’s Office of Energy Dominance Financing reports more than $289 billion in available loan authority and the ability to finance up to 80 percent of eligible project costs, and the UPRISE initiative has committed to convening match-making workshops between plant owners and end users to facilitate exactly the kind of agreements examined in this section [20]. The combination of hyperscaler offtake and federal debt is potentially powerful, because the former reduces revenue risk while the latter reduces financing cost, and together they can move marginal uprate projects from uneconomic to financeable.
3.6 From Megawatt Buyer to Infrastructure Counterparty
The most important conceptual change revealed by these transactions is institutional. The hyperscaler ceases to be merely an electricity consumer and becomes an infrastructure counterparty whose capital, credit and contractual commitments shape which generating assets are extended, upgraded and ultimately built. Under Reactor Yield, the economic relationship runs from the AI company to the utility, from the utility to the nuclear asset, from the asset to incremental megawatts, and from those megawatts back to AI expansion, creating one of the clearest feedback loops within the Five-Layer AI Economy. The transition can be summarized as a progression from electricity buyer to infrastructure counterparty to capacity creator, and if the model spreads, AI companies may help finance uprates, life extensions, transmission reinforcement and eventually entirely new generation, profoundly altering the historical relationship between America’s technology and utility sectors.
Section 4: Who Owns the Incremental Megawatt?
Once a hyperscaler finances additional generation at an existing nuclear station, a deceptively simple question follows, and it is one that engineering analysis alone cannot answer: who receives the additional electricity, who bears the risk of producing it, and who captures the benefits it creates for the wider system? This section argues that Reactor Yield creates an allocation problem that sits at the intersection of utility economics, corporate finance, environmental-attribute accounting and public regulation, and that the durability of the hyperscaler uprate model will depend less on whether the engineering succeeds than on whether regulators, utilities and customers perceive the allocation of costs and benefits as fair. The Georgia filing offers an unusually clear window into one answer, while the PJM transactions offer another, and comparing them illuminates the design choices that every state will eventually confront.
4.1 Reactor Yield Creates an Allocation Question
There are several possible destinations for the output of a hyperscaler-financed uprate. The output could simply serve the general grid, with the financing customer receiving only environmental attributes; it could economically support a specific large-load customer through a bilateral contract; it could be delivered through a regulated tariff structure; it could offset system costs that would otherwise fall on all ratepayers; or an arrangement could combine several of these mechanisms. The Georgia structure is noteworthy because Georgia Power describes the uprated capacity as being added to the grid to serve Georgia Power customers generally, while Google provides economic support for the investment through its subscription and receives the associated Zero-Emission Credits [1]. The Meta–Vistra agreements take a similar stance in a different market, specifying that the electricity from the uprated plants will continue to flow to the grid for all users even as Meta purchases the incremental energy and capacity [4].
4.2 Physical Electricity Versus Economic Attribution
Electricity does not travel from one identified reactor directly to one identified server rack, because grid power is pooled and electrons follow physics rather than contracts. Reactor Yield therefore requires a careful distinction between physical delivery and economic attribution: a hyperscaler might financially enable 100 megawatts of new nuclear output without physically consuming precisely those 100 megawatts at every moment, and its claim to the output is expressed through contractual instruments such as power purchase agreements, capacity rights and environmental credits rather than through a dedicated wire. The distinction becomes more consequential as large-load tariffs and clean-energy contracting become more sophisticated, and as hyperscalers increasingly seek to demonstrate hourly rather than annual matching of their consumption with carbon-free supply, since an uprate at a high-capacity-factor nuclear unit produces attributes that are unusually well matched to a data center’s round-the-clock load profile.
Transparency about attribution is not a trivial concern. World Nuclear News observed that while the Georgia dockets are public, the document containing the subscription agreement itself was redacted in its entirety [13], which is common for commercially sensitive contracts but leaves outside observers unable to verify how the $900 million of projected customer benefits was calculated or how risk is shared if the uprates cost more or deliver less than expected. As hyperscaler-financed uprates multiply, the credibility of the model will depend on regulators ensuring that the economic substance of such agreements, even if not their commercial detail, is available for public scrutiny.
4.3 Large-Load Customers and Ordinary Ratepayers
AI infrastructure introduces a politically and economically important distribution problem into utility regulation. If a datacenter causes a utility to construct additional infrastructure, regulators must determine which costs properly belong to the datacenter operator, the utility’s shareholders, participating customers, the general rate base, future customers, or some combination of these. Reactor Yield sharpens the question because an uprate can produce both private and system-wide benefits: when a hyperscaler helps finance incremental generation that improves overall supply adequacy, ordinary customers can benefit through lower capacity costs and improved reliability, but when system costs are incurred primarily to accommodate one large customer, fairness suggests those costs should be assigned accordingly. The answer will differ by state, market structure and contract, and Georgia’s approach of pairing the uprate program to a subscription tariff paired with a public pledge to protect non-participants is only one of several defensible designs [1].
Southern Company’s leadership has argued that its vertically integrated, state-regulated structure is an advantage precisely because it allows large-load contracts to be designed with customer protections under direct commission oversight. The company reported that all new large-load contracts in Georgia are reviewed by the Public Service Commission and that its large-load agreements incorporate provisions intended to protect existing customers, while also committing to rate stability through the end of the decade [26].
“our vertically integrated, state-regulated model supports our ability to provide reliable power with speed”
— Chris Womack, Chairman and CEO, Southern Company [26]
4.4 The Incremental-Megawatt Ledger
Every Reactor Yield project can be evaluated through what I call an Incremental-Megawatt Ledger, a structured accounting of the costs, benefits and obligations created by each increment of output. The ledger’s purpose is not to prescribe a single allocation but to make the allocation visible, so that regulators, customers and investors can compare projects on a consistent basis and so that disputes can focus on substantive disagreements rather than on missing information. Table 4 sets out its principal entries.
Table 4. The Incremental-Megawatt Ledger
| Ledger entry | Key question | Who typically bears or captures it |
| Incremental capital cost | What engineering, equipment and outage costs does the uprate require? | Utility or generator, often recovered through the financing customer’s tariff or PPA |
| Incremental nuclear output | How many MW and MWh over the remaining licensed life? | Grid (physical); financing customer (contractual attribution) |
| Incremental transmission requirement | Does output exceed existing export capability? | Interconnection customer or rate base, depending on market rules [35] |
| Incremental cooling requirement | Does higher thermal power strain water or discharge permits? | Plant owner; environmental regulators |
| Incremental operating and maintenance cost | Do fuel, maintenance or staffing costs rise? | Plant owner, passed through under contract terms |
| Incremental customer benefit | What system value accrues to non-participants? | General customers (e.g., Georgia Power’s ~$900 million estimate) [1] |
| Incremental hyperscaler payment | What does the financing customer pay, and for which attributes? | Hyperscaler, in exchange for credits, energy or capacity |
| Incremental reliability benefit | How much does firm capacity improve resource adequacy? | System operator and all customers |
| Incremental regulatory obligation | What NRC and state approvals are needed, and on what timeline? | Licensee (NRC); utility (state commission) [11] |
The ledger is a proposed evaluation instrument. Assignments in the third column describe common patterns rather than legal requirements.
4.5 Interconnection Rights as a Hidden Allocation Variable
One ledger entry deserves special emphasis because it is frequently overlooked: the right to inject power into the grid. An uprate that raises a plant’s output above the capacity interconnection rights it already holds may require new studies and upgrades, and in congested regions those studies can impose delays measured in years, eroding the very time-to-megawatt advantage that makes Reactor Yield attractive. The Crane case shows both the problem and a potential remedy, since FERC’s June 2026 approval allowed interconnection rights to be transferred from a fossil plant to a nuclear restart over the objections of PJM’s independent market monitor [35]. Whether similar transfers, surplus-interconnection provisions, or pre-cleared uprate headroom become routine will shape how much of the NRC’s expected pipeline actually reaches customers on schedule, and it suggests that interconnection policy is as much a part of Reactor Yield as reactor engineering.
4.6 Public Utility Regulation Becomes Part of AI Architecture
The Five-Layer AI Economy therefore cannot stop at semiconductors, datacenters and models. State utility commissions increasingly sit between AI demand and physical infrastructure, and a reactor uprate associated with a hyperscaler may require decisions on tariffs, cost allocation, integrated resource planning and consumer protection, while nuclear-safety licensing remains within the federal NRC framework. These institutions are not peripheral to AI development; they increasingly influence the rate at which Layer 1 can expand, and through Layer 1 the rate at which every higher layer can grow. The central policy question should remain neutral and measurable: can arrangements be structured so that large-load customers contribute sufficiently to incremental generation while system benefits and project risks are transparently allocated? Reactor Yield does not predetermine the answer, but it provides the unit of analysis through which different state approaches can be compared on common terms.
Section 5: The Policy and Regulatory Machinery of Reactor Yield
Reactor Yield would remain a niche engineering practice without a policy environment that rewards it, and the most striking development of 2025 and 2026 is the degree to which federal policy has been reorganized around extracting more output from existing reactors. This section traces that machinery through three layers: the executive orders of May 2025 that set explicit uprate targets, the Department of Energy’s UPRISE initiative launched in March 2026 with its financing and match-making tools, and the NRC’s shift toward faster, more predictable review timelines for both uprates and license renewals. It closes with the institutional risks that could slow implementation, because ambitious targets are only as credible as the agencies charged with meeting them.
5.1 The 2025 Executive Orders and the Five-Gigawatt Target
On May 23, 2025, the President signed four executive orders intended to expand American nuclear capacity from roughly 100 gigawatts to 400 gigawatts by 2050, among them Executive Order 14302 on reinvigorating the nuclear industrial base [49]. Section 4 of that order directs the Department of Energy to prioritize work with industry to facilitate five gigawatts of power uprates at existing reactors and to have ten new large reactors with complete designs under construction by 2030, and it instructs the department’s lending programs to prioritize restarts, uprates, and the completion of suspended construction [50].
“To maximize the speed and scale of new nuclear capacity”
— Executive Order 14302, The White House [50]
Legal analysts at Hogan Lovells observed that the five-gigawatt near-term goal is realistically achievable only through uprates and restarts of shut-down units rather than through new construction, and that the Loan Programs Office’s willingness to support such efforts was a positive signal [51]. That interpretation places Reactor Yield at the center of the administration’s near-term nuclear agenda, since the new large reactors called for in the same order cannot, on any realistic schedule, deliver power before the 2030s. It also highlights a structural feature of the targets, which is that the uprate goal is defined in gigawatts delivered by 2030 while the large-reactor goal is defined in projects under construction, an asymmetry that implicitly acknowledges the different time problems these two strategies solve.
5.2 UPRISE: Turning a Target Into a Program
The Department of Energy’s Office of Nuclear Energy launched the Utility Power Reactor Incremental Scaling Effort, or UPRISE, on March 12, 2026, with management by Idaho National Laboratory and an objective of expanding capacity by increasing the output of existing reactors, returning dormant facilities to service, and completing stalled projects [3]. The initiative set targets of 2.5 gigawatts of additional nuclear capacity by 2027 and five gigawatts by 2029, a year ahead of the executive order’s deadline, and it identified license renewal, power uprates, restarts, and efficiency improvements including advanced fuels as its principal levers [3].
“This will be a resurgence for America’s nuclear fleet”
— Rian Bahran, Deputy Assistant Secretary for Nuclear Energy, U.S. Department of Energy [3]
UPRISE’s design aligns closely with the Reactor Yield framework developed in this paper. Its near-term work is organized around establishing the business case through supply-chain readiness reviews, assessments of plant equipment for increased output, and validation of economic models to support investment decisions, and it pairs that analysis with Energy Dominance Financing’s lending capacity and a series of match-making workshops between plant owners and end users [20]. The first such workshop was held at Louisiana State University in May 2026, where department officials emphasized pairing nuclear expansion in the Gulf Coast with growing regional industrial demand [17]. By May 2026, the department reported that UPRISE had already seen planned uprates at six reactors across Hatch, Vogtle and Farley producing a combined 345 megawatts of additional baseload capacity [16].
5.3 The NRC as Accelerant Rather Than Bottleneck
For most of the past two decades, the NRC was widely perceived as a source of delay in nuclear projects, and that perception shaped the economics of uprates by adding regulatory schedule risk to engineering and procurement risk. The agency’s recent conduct suggests a meaningful change in that dynamic, at least for work at existing plants. Its May 2024 commitment to review extended uprates in twelve months, stretch uprates in nine, and measurement-recapture uprates in six, supported by a dedicated project plan and standing cadres of reviewers who carry lessons from one application to the next, directly addresses the schedule uncertainty that historically discouraged uprate investment [11]. The June 2026 approval of the Hatch subsequent license renewal in under twelve months, against a historical average of roughly two and a half years, provided concrete evidence that such targets can be met [14].
The regulatory acceleration matters for Reactor Yield in a specific way. Because the value of an uprate depends on the remaining operating life over which it produces power, faster license renewals raise the value of uprates, and faster uprate reviews shorten time-to-megawatt, so improvements on both fronts compound. At the same time, the NRC’s approval remains a genuine safety determination rather than a formality, and the agency continues to require licensees to demonstrate that plants can operate safely at higher power, as its public guidance on uprate reviews makes clear [6]. The appropriate policy goal is predictability without erosion of rigor, and the evidence of 2026 suggests that the two are compatible.
5.4 Institutional Risks to Implementation
Ambitious targets generate their own risks, and three deserve attention. The first is agency capacity: when the executive orders were issued, the Nuclear Innovation Alliance cautioned that staffing reductions and proposed budget cuts at the Department of Energy could impede the very programs the orders created [52]. The second is supply chain, since extended uprates require large turbine components, generators and main transformers that compete for manufacturing capacity with the broader grid buildout and with new gas turbines. The third is regulatory legitimacy, because public confidence in the NRC’s independence is an asset that, once depleted, would raise the cost of every future nuclear project, including uprates.
“make it harder to implement these executive orders”
— Judi Greenwald, President and CEO, Nuclear Innovation Alliance, on DOE staffing reductions [52]
Independent forecasters have tempered their expectations accordingly. BloombergNEF has estimated net U.S. nuclear growth of roughly seven percent, or about seven gigawatts, by the end of the decade, driven primarily by restarts, uprates and license extensions for reactors whose licenses would otherwise expire before 2035 [53]. That figure is consistent with the Reactor Yield thesis, since it identifies the existing fleet as the dominant source of near-term nuclear growth, while also indicating that the realistic scale of that growth is measured in single-digit gigawatts rather than the tens of gigawatts sometimes implied in public discussion.
Section 6: Existing-Fleet First — The Reactor Yield Pipeline From 2027 to 2030
The most compelling reason to study Reactor Yield now is that the concept is no longer theoretical: a measurable pipeline of uprate applications, a federal program with explicit gigawatt targets, and a set of corporate financing structures all exist simultaneously for the first time. This section examines that pipeline in detail, argues that its roughly two gigawatts should be judged against the right counterfactual, proposes a sequencing principle and a four-track nuclear stack for the 2027–2030 period, and introduces yield density as a way of mapping the hidden geography of AI energy. It closes by tracing how an incremental megawatt at Layer 1 propagates through the entire Five-Layer AI Economy.
6.1 The NRC Pipeline Is Already Forming
The NRC’s expected-application schedule, updated on September 8, 2026, lists 31 uprates expected through 2032, comprising ten measurement-uncertainty-recapture uprates, two stretch uprates and nineteen extended uprates, together representing about 7,336 megawatts thermal and roughly 2,421 megawatts electric [2]. The distribution over time is heavily front-loaded: two extended uprates totaling about 335 megawatts are expected in 2026, seventeen uprates totaling about 875 megawatts in 2027, and six extended uprates totaling about 635 megawatts in 2028, after which the annual volume declines sharply [2]. Figure 3 and Table 5 present the schedule.
Figure 3. NRC expected power-uprate applications by year of filing, 2026–2032 (approximate megawatts electric).
Table 5. NRC expected uprate applications by calendar year (as of September 8, 2026)
| Year | Total uprates | MUR | Stretch | Extended | MW thermal | Approx. MW electric |
| 2026 | 2 | 0 | 0 | 2 | 1,014 | 335 |
| 2027 | 17 | 8 | 2 | 7 | 2,650 | 875 |
| 2028 | 6 | 0 | 0 | 6 | 1,923 | 635 |
| 2029 | 1 | 0 | 0 | 1 | 598 | 197 |
| 2030 | 3 | 2 | 0 | 1 | 754 | 249 |
| 2031 | 1 | 0 | 0 | 1 | 200 | 66 |
| 2032 | 1 | 0 | 0 | 1 | 197 | 65 |
| Total | 31 | 10 | 2 | 19 | 7,336 | 2,421 |
Source: NRC Expected Applications for Power Uprates [2]. Filing dates are expectations; approval and in-service dates follow later, typically aligned with refueling outages.
The unit-level detail reveals the geography and ownership of the pipeline. Named filings in 2027 include measurement-recapture uprates at Duke Energy’s Brunswick Units 1 and 2 and at Wolf Creek, extended uprates at Duke’s McGuire Units 1 and 2 and at Southern Nuclear’s Hatch Units 1 and 2, and stretch uprates at PSEG’s Salem Units 1 and 2; 2028 brings extended uprates at Energy Northwest’s Columbia, Duke’s Catawba Unit 1 and Southern Nuclear’s Vogtle Units 1 and 2; and Vistra’s Perry, Beaver Valley and Davis-Besse follow between 2029 and 2032, consistent with the Meta agreement’s delivery schedule [2][42]. Constellation accounts for a large block of filings whose sites remain proprietary, which is itself informative about the competitive sensitivity of extractable capacity in merchant markets.
It is important to read these numbers precisely. Filing dates are not in-service dates; an extended uprate filed in the second quarter of 2027, reviewed within the NRC’s twelve-month target, and installed during a subsequent refueling outage might reach full output in 2029 or 2030, while measurement-recapture uprates can move considerably faster. The schedule is therefore best understood as a leading indicator of the megawatts likely to arrive between roughly 2028 and 2033, which aligns closely with the period in which AI load growth is expected to be steepest and new large reactors are least likely to be available.
6.2 Two Gigawatts Is Not Trivial
Two to two and a half gigawatts will not solve America’s AI electricity problem, and this paper has been explicit on that point, but neither is the figure insignificant. At the high capacity factors typical of the U.S. fleet, an additional 2.4 gigawatts of nuclear capacity produces round-the-clock electricity comparable to roughly two new large reactors, and it does so without waiting for an entirely new fleet, at sites where much of the supporting infrastructure already exists. The relevant comparison is therefore not 2 gigawatts against total future AI demand but 2 gigawatts available comparatively soon against 2 gigawatts that would otherwise have to be supplied by some other resource during the same window, whether new gas turbines facing multi-year equipment backlogs, renewable and storage projects awaiting interconnection, or delayed data-center energization.
The comparison becomes more favorable still when the pipeline is combined with the other existing-fleet levers. The UPRISE target of five gigawatts by 2029 explicitly counts restarts and efficiency improvements alongside uprates [3], and the restarts under way, including the 835-megawatt Crane unit targeted for 2027 [35], add to the uprate total. Taken together, the existing fleet is plausibly the largest source of new firm, carbon-free capacity that can be delivered in the United States before 2030, which is a remarkable conclusion for assets that, in several cases, were on a path to retirement only six years ago.
6.3 Existing-Fleet First Does Not Mean Existing-Fleet Only
An existing-fleet-first strategy does not reject new large reactors, small modular reactors, microreactors, geothermal, natural gas, renewables, storage, transmission, fuel cells, or demand flexibility. It introduces a sequencing principle instead: first determine what additional safe and economic output can be extracted from installed infrastructure; second, extend economically valuable generating assets where safety requirements are satisfied; and third, build the new generating fleet required for long-term growth. This sequencing avoids treating existing and future nuclear capacity as competitors, recognizing them instead as successive layers of the same supply strategy, and it mirrors the logic of the administration’s executive orders, which pair a near-term uprate target with a longer-term construction target [50].
6.4 The 2027–2030 Nuclear Stack
A coherent nuclear strategy for AI-era electricity demand can therefore be expressed as four parallel tracks, each with its own time horizon, instruments and 2026 evidence base. Reactor Yield occupies the third track but interacts strongly with the first two, because without preservation there is no reactor to uprate, without sufficient remaining life the economics of major uprates weaken, and without new construction uprating eventually encounters physical limits.
Table 6. The four-track nuclear stack for the AI era
| Track | Objective | Principal instruments | 2025–2026 evidence |
| A — Preserve | Keep safe, economic plants operating | Long-term PPAs, state support, capacity markets | Meta–Vistra PPAs for plants once headed to retirement [4] |
| B — Extend | Secure subsequent license renewals to 80 years | NRC SLR under streamlined 12-month reviews | Hatch licensed to 2054 and 2058 [18][14] |
| C — Uprate (Reactor Yield) | Extract additional safe output from existing units | MUR, stretch and extended uprates; hyperscaler and federal financing | ~2,421 MWe NRC pipeline; Google, Meta and customer-enabled uprates [2][1] |
| D — Build | Add new large, small modular and advanced reactors | EO 14302 targets, federal loans, corporate offtake | Ten large reactors under construction targeted by 2030 [50] |
Restarts of previously retired units, such as Crane and Palisades, sit between Tracks A and C and draw on the same financing and interconnection tools.
6.5 Mapping America’s Reactor Yield
A national Reactor Yield analysis would examine plants across Georgia, the Carolinas, Virginia, Pennsylvania, New Jersey, Ohio, Michigan, Illinois, Minnesota, Arizona, Washington, Texas and every other state with an operating fleet, not on the presumption that every reactor should be uprated but in order to give each reactor a comparable technical-economic profile. The analysis would ask, for each unit, how much incremental output is technically plausible; what equipment must be replaced; what the expected capital cost is; how long engineering, procurement and installation would require; what NRC licensing action would be necessary; whether additional transmission is needed; whether adequate cooling capacity is available; how much operating life remains; whether subsequent license renewal is under way or feasible; and whether a large-load customer capable of supporting financing is present in the region. The result would be an entirely different map of American nuclear infrastructure, one that displays not only existing capacity but potential incremental reactor yield, and the DOE’s plant-equipment assessments under UPRISE are the logical starting point for building it [20].
6.6 From Reactor Count to Yield Density
By 2030, states competing for AI infrastructure may increasingly compare their nuclear systems using a metric I call yield density, defined as the incremental nuclear megawatts realistically available per existing nuclear site. States with established fleets may discover that their infrastructure advantage is larger than nominal installed capacity suggests, because a site capable of safely producing another 100 megawatts through uprating possesses a very different economic profile from one already operating at its practical maximum. This creates a hidden geography of AI energy in which the Southeast, with Southern Company’s Hatch, Vogtle and Farley programs and Duke Energy’s McGuire, Catawba and Brunswick filings, and the Ohio–Pennsylvania corridor, with Vistra’s three uprated plants, emerge as regions of unusually high near-term yield density [2][16].
6.7 Reactor Yield and the Five-Layer AI Economy
The Five-Layer framework makes the wider significance of Reactor Yield visible, because it shows how an engineering change at a single plant propagates upward through the entire AI stack. The IEA’s formulation of the underlying dependency is concise, and it is the premise on which the entire framework rests.
“There is no AI without energy”
— International Energy Agency, Energy and AI [7]
Table 7. How an incremental nuclear megawatt propagates through the Five-Layer AI Economy
| Layer | Effect of additional Reactor Yield |
| Layer 1 — Energy | Uprates create additional firm, carbon-free electricity at existing sites, often with shorter time-to-megawatt than new generation. |
| Layer 2 — Chips | Additional firm power allows more accelerators to be energized rather than stranded awaiting interconnection. |
| Layer 3 — Datacenters | More firm capacity expands feasible campus size and utilization in constrained regions such as the Southeast and PJM. |
| Layer 4 — Models | Training and inference schedules become less constrained by local electricity availability. |
| Layer 5 — Applications and Agents | Expanded compute supports broader deployment of continuous AI services, autonomous agents, robotics and machine-driven economic activity. |
The propagation is conceptual; the magnitude of effects at each layer depends on regional grid conditions and on complementary investments in transmission and flexibility.
One turbine replacement at Layer 1 can therefore propagate through the entire stack, and that is the deeper meaning of Reactor Yield: it treats the physical capability of existing reactors as a strategic reserve for the digital economy, one that can be drawn upon through engineering and capital rather than waiting for the construction cycle of an entirely new fleet.
Section 7: What Have We Learned? Eight Pillars of Reactor Yield
The preceding sections examined Reactor Yield from the perspectives of engineering, demand, finance, allocation, policy and geography. This section distills those analyses into eight pillars, extending the five with which this research began by adding three that emerged from the 2026 evidence: the transformation of the regulator from bottleneck to accelerant, the binding role of grid and supply-chain constraints, and the obligation to evaluate Reactor Yield honestly against its alternatives and its critics. Each pillar is intended to be both a finding and a guide to action for utilities, hyperscalers, regulators and investors.
Pillar 1 — The Existing Nuclear Fleet Is Not a Fixed Asset
Installed nuclear capacity should not be treated as a permanent ceiling. Many reactors contain additional generating potential that can be unlocked through improved measurement, engineering changes, equipment replacement and licensing, and the NRC’s record of 171 approved uprates adding about eight gigawatts electric demonstrates that this is established practice rather than a novel technology [6]. What artificial intelligence changes is the economic value of that practice, because the megawatts are familiar while the demand environment is not, and as electricity scarcity intensifies around major datacenter regions, incremental firm megawatts at existing sites become more strategically valuable than at any time in the fleet’s history.
Pillar 2 — Time-to-Megawatt Is Becoming as Important as Cost-per-Megawatt
The AI infrastructure cycle places a premium on speed that conventional levelized-cost analysis does not capture. New reactors may be essential to America’s longer-term electricity future, yet even optimistic learning-curve estimates imply construction schedules of 80 to 96 months for a next AP1000 [33], which means they cannot satisfy requirements emerging between 2027 and 2030. Uprates can occupy part of that temporal gap, and the correct question is therefore not whether uprates are better than new reactors but which electricity resource can arrive when the load needs it, a reframing that turns nuclear economics from a static cost comparison into a scheduling problem.
Pillar 3 — Hyperscalers Can Become Producers of Incremental Capacity, Not Merely Consumers
The Google–Georgia Power, Meta–Vistra and Constellation transactions show that AI companies now possess the financial capacity and long-duration electricity requirements to finance investments that expand generating capability rather than merely reallocating it. This creates a potentially historic transition from electricity buyer to infrastructure counterparty to capacity creator, and if the model spreads, AI companies may help finance uprates, life extensions, transmission reinforcement and eventually entirely new generation, altering the relationship between the technology and utility sectors in ways that would have seemed implausible a decade ago [4][1].
Pillar 4 — Every Incremental Megawatt Requires an Allocation Framework
If a hyperscaler helps create additional electricity from an existing nuclear asset, regulators, utilities and customers need clarity about who finances the project, who assumes construction risk, who receives the electricity economically, who benefits from additional grid capacity, who pays for associated transmission, and how benefits are distributed between participating and non-participating customers. Reactor Yield therefore belongs not only to nuclear engineering but to utility economics, corporate finance, infrastructure planning and public regulation, and instruments such as the Incremental-Megawatt Ledger, together with greater transparency about the economic substance of redacted agreements [13], will determine whether the model retains public legitimacy.
Pillar 5 — The Near-Term Opportunity Is Measured in Megawatts per Reactor, Not Reactors per Decade
The AI-era nuclear conversation gravitates naturally toward dramatic numbers: how many reactors, how many small modular reactors, how many gigawatts, how many trillions of dollars. Reactor Yield offers a smaller but more actionable question about how many additional megawatts each existing reactor can safely produce, and the NRC currently expects uprate applications totaling roughly 2.4 gigawatts electric through 2032 [2]. That number may rise, fall or shift as operators make investment decisions and reviews proceed, but the conceptual shift remains: before counting reactors that have not yet been built, the United States can count megawatts that may still be hidden inside reactors that already exist.
Pillar 6 — Regulation Has Become an Accelerant for the Existing Fleet
The historical assumption that the NRC is primarily a source of delay is increasingly inaccurate for work at existing plants. Published review targets of six to twelve months for uprates [11], and the approval of Hatch’s subsequent license renewal in under a year against a prior average of roughly two and a half years [14], show a regulator capable of predictable, timely decisions without abandoning safety review. Because faster renewals raise the lifetime value of uprates while faster uprate reviews shorten time-to-megawatt, the combined effect compounds, and preserving both the speed and the credibility of that regulatory performance is itself a Reactor Yield strategy.
Pillar 7 — Reactor Yield Is Bounded by the Grid, the Cooling System and the Supply Chain
The reactor is rarely the binding constraint on extractable capacity. The Crane restart demonstrated that interconnection limits can threaten to delay even an existing plant by years until regulators approve creative remedies such as transferring capacity rights [21][35], while cooling-water availability, thermal-discharge permits, and the manufacturing backlog for large turbines, generators and transformers can all cap the economically rational portion of a plant’s yield curve. A Reactor Yield strategy that ignores transmission export capability or equipment lead times will overstate what the fleet can deliver, and the most valuable planning work may therefore lie outside the reactor building, in switchyards, substations and supplier factories.
Pillar 8 — Reactor Yield Must Be Judged Honestly Against Its Alternatives and Its Critics
Reactor Yield is strongest when it is presented as one instrument within a portfolio rather than as a panacea. Duke University research shows that modest load flexibility could let existing systems absorb tens of gigawatts of new demand [38], the IEA expects renewables, storage and natural gas to supply the largest shares of data-center growth [7], and critics at MIT and Stanford question both the sustainability of the datacenter buildout and the reliability of nuclear cost promises [37][34]. These perspectives do not refute Reactor Yield; they discipline it. Uprates are most defensible where they are cheapest per megawatt, fastest to deliver, paired with long remaining licensed life, and combined with flexibility and grid investment, and the policy objective should be to identify those opportunities rigorously rather than to maximize uprating for its own sake.
“There is unprecedented interest in powering data centers with nuclear reactors”
— Prof. Aditi Verma, University of Michigan, Nuclear Engineering and Radiological Sciences [54]
Professor Verma’s research program, which is developing an open-source decision-support tool integrating power-system data, geospatial constraints, legal considerations and safety parameters for siting decisions [54], exemplifies the kind of balanced, transparent analysis that the eighth pillar calls for, and a comparable public tool for mapping extractable capacity across the existing fleet would be a natural complement.
Conclusion: Why Reactor Yield Matters
The artificial-intelligence infrastructure race is usually described through gigantic numbers: hundreds of billions of dollars of annual capital expenditure, gigawatts of contracted datacenter load, millions of accelerators, new transmission corridors, new gas plants, new reactors, small modular reactors and multigigawatt campuses bound by long-term power contracts. Against that backdrop, Georgia Power and Google’s approximately 96 megawatts can appear almost insignificant, and a reasonable observer might dismiss the announcement as a footnote to the larger story. This paper has argued that such a dismissal would be a mistake, because those megawatts represent a different way of thinking about infrastructure, one in which America’s energy problem is understood not exclusively as a construction problem but partly as an extraction problem.
The logic of extraction is disciplined and concrete. Before constructing the next generating system, determine whether the existing system contains unused productive capability. Before waiting for the next reactor vessel, examine the turbine, the pumps, the generator, the transformer, the cooling system, the instrumentation and the licensed operating margin, and then ask what combination of engineering, capital and regulatory approval could convert those latent capabilities into electricity. That sequence of questions is what this paper calls Reactor Yield, and the term fits because yield describes precisely what the analysis seeks to uncover: additional productive output from capital that has already been built, financed, licensed and connected.
The distinction is likely to become more important between 2027 and 2030. America’s new-nuclear ambitions remain essential, and nothing in this analysis argues against them, but the country also possesses a large fleet of operating reactors with decades of institutional knowledge, operating history, grid interconnections and, increasingly, licensed lives extending into the 2050s, as Hatch’s June 2026 renewal demonstrated [18]. Some of those reactors will have little practical additional yield, others will offer modest measurement gains, some will justify substantial equipment upgrades, and a smaller number will combine long remaining lives, favorable engineering conditions, strong transmission and nearby hyperscale demand in ways that make their incremental megawatts unusually valuable. Finding those reactors, with the rigor of a national inventory rather than the happenstance of individual deals, is the next analytical task.
The evidence assembled here produces a simple but potentially powerful infrastructure sequence for the remainder of the decade: preserve what already operates, extend what can safely operate longer, extract additional output where engineering and economics justify it, and build the new fleet required for the decades beyond. The sequence does not replace the nuclear renaissance; it may well accelerate it, by generating near-term revenue, sustaining the workforce and supply chain, demonstrating new financing structures, and buying time for first-of-a-kind advanced reactors to mature. Within the Five-Layer AI Economy, that acceleration matters because every incremental nuclear megawatt begins at Layer 1 but can propagate upward through the entire stack, from electricity to chips, from chips to datacenters, from datacenters to models, and from models to applications, agents, robots and machine-driven economic activity.
The future of AI energy will ultimately be measured by how many new generating resources America can build, and new reactors will be part of that measure. For the next several years, however, one of the most revealing measures may be much simpler, and it is the question with which this paper began and with which it ends.
How much more can the reactors we already built safely produce?
That is why this paper is called Reactor Yield.
Footnotes and Endnotes:
[1] Georgia Power (statements by Aaron Mitchell and Lucia Tian, Google). “Georgia Power agreement with Google to provide approximately $900 million in projected benefits for customers and advance nuclear energy in Georgia.” Georgia Power News Hub, September 21, 2026. https://www.georgiapower.com/news-hub/press-releases/agreement-with-google-provides-$900-million-customer-benefits-advances-nuclear-energy.html
[2] U.S. Nuclear Regulatory Commission. “Expected Applications for Power Uprates (by unit).” NRC, page updated September 8, 2026. https://www.nrc.gov/reactors/operating/licensing/power-uprates/status-power-apps/expected-applications
[3] U.S. Department of Energy, Office of Nuclear Energy (statement by Rian Bahran). “The Nation’s Nuclear Reactor Fleet Is on the Rise: Introducing the UPRISE Initiative.” DOE, March 12, 2026. https://www.energy.gov/node/4856801
[4] Vistra Corp. (statements by Jim Burke and Stacey Doré). “Vistra and Meta Announce Agreements to Support Nuclear Plants in PJM and Add New Nuclear Generation to the Grid.” Vistra Investor Relations, January 9, 2026. https://investor.vistracorp.com/2026-01-09-Vistra-and-Meta-Announce-Agreements-to-Support-Nuclear-Plants-in-PJM-and-Add-New-Nuclear-Generation-to-the-Grid
[5] Insider Monkey / Yahoo Finance. “Google Is Helping Upgrade Southern’s Nuclear Plants. Can 96 MW Ease Its AI Power Bottleneck?.” Yahoo Finance, September 2026. https://finance.yahoo.com/energy/articles/google-helping-upgrade-southern-nuclear-184432017.html
[6] U.S. Nuclear Regulatory Commission. “Backgrounder on Power Uprates for Nuclear Plants.” NRC Office of Public Affairs. https://www.nrc.gov/reading-rm/doc-collections/fact-sheets/power-uprates
[7] International Energy Agency. “Energy and AI: Executive Summary.” IEA, 2025. https://www.iea.org/reports/energy-and-ai/executive-summary
[8] International Energy Agency (statement by Fatih Birol). “Data centre electricity use surged in 2025, even with tightening bottlenecks driving a scramble for solutions.” IEA News, April 2026. https://www.iea.org/news/data-centre-electricity-use-surged-in-2025-even-with-tightening-bottlenecks-driving-a-scramble-for-solutions
[9] International Energy Agency (statement by Fatih Birol). “AI is set to drive surging electricity demand from data centres while offering the potential to transform how the energy sector works.” IEA News, April 10, 2025. https://www.iea.org/news/ai-is-set-to-drive-surging-electricity-demand-from-data-centres-while-offering-the-potential-to-transform-how-the-energy-sector-works
[10] U.S. Energy Information Administration. “Plant Vogtle Unit 4 begins commercial operation.” EIA Today in Energy, May 1, 2024. https://www.eia.gov/todayinenergy/detail.php?id=61963
[11] U.S. Nuclear Regulatory Commission. “Power Uprate Review Readiness.” NRC. https://www.nrc.gov/reactors/operating/licensing/power-uprates/optimizepoweruprates.html
[12] U.S. Nuclear Regulatory Commission. “Approved Applications for Power Uprates.” NRC. https://www.nrc.gov/reactors/operating/licensing/power-uprates/status-power-apps/approved-applications
[13] World Nuclear News. “Google agreement to support Georgia plant uprates.” World Nuclear News, September 2026. https://www.world-nuclear-news.org/articles/google-agreement-to-support-Georgia-plant-uprates
[14] American Nuclear Society, Nuclear Newswire. “Hatch SLR approved by NRC in under 12 months.” ANS, June 12, 2026. https://www.ans.org/news/2026-06-12/article-8118/hatch-slr-approved-by-nrc-in-under-12-months/
[15] Power Engineering. “Google to pay for nuclear uprates at Georgia Power’s Hatch, Vogtle plants.” Power Engineering, September 2026. https://www.power-eng.com/nuclear/google-to-pay-for-nuclear-uprates-at-georgia-powers-hatch-vogtle-plants/
[16] U.S. Department of Energy, Office of Nuclear Energy. “One Year After Executive Orders, U.S. Nuclear Energy Renaissance Is in Full Swing.” DOE, May 23, 2026. https://www.energy.gov/ne/articles/one-year-after-executive-orders-us-nuclear-energy-renaissance-full-swing
[17] Louisiana State University. “U.S. Department of Energy Workshops at LSU Accelerate Future of Nuclear Energy.” LSU, May 2026. https://lsu.edu/blog/2026/05/uprise-meeting.php
[18] Georgia Power / Southern Company. “Georgia’s Plant Hatch receives 20-year license renewal from the Nuclear Regulatory Commission.” Southern Company Newsroom, June 15, 2026. https://southerncompany.mediaroom.com/2026-06-15-Georgias-Plant-Hatch-receives-20-year-license-renewal-from-the-Nuclear-Regulatory-Commission
[19] Levin, Danny (Industrial Info Resources). “One More U.S. Nuclear License Renewed Another 20 Years.” Industrial Info Resources, June 17, 2026. https://www.industrialinfo.com/news/article/one-more-us-nuclear-license-renewed-another-20-years–359150
[20] American Public Power Association. “DOE initiative aims to increase nuclear power output through uprates and restarts.” Public Power, March 2026. https://www.publicpower.org/periodical/article/doe-initiative-aims-increase-nuclear-power-output-through-uprates-and-restarts
[21] Power Engineering. “Constellation files at FERC to keep Crane nuclear restart on 2027 timeline.” Power Engineering, 2026. https://www.power-eng.com/business/constellation-files-at-ferc-to-keep-crane-nuclear-restart-on-2027-timeline/
[22] Shehabi, A., Smith, S. J., Hubbard, A., Newkirk, A., Lei, N., Siddik, M. A. B., Holecek, B., Koomey, J., Masanet, E., and Sartor, D. (Lawrence Berkeley National Laboratory). “2024 United States Data Center Energy Usage Report (LBNL-2001637).” LBNL, December 2024. https://energyanalysis.lbl.gov/publications/2024-lbnl-data-center-energy-usage-report
[23] Smith, S. J., Hubbard, A., Newkirk, A., Ganeshalingam, M., Holecek, B., Sartor, D., Mills, M., and Shehabi, A. (LBNL). “United States Data Center Energy Usage Report: 2025 Update.” LBNL, 2025. https://eta-publications.lbl.gov/bibcite/export/bibtex/bibcite_reference/36976
[24] Enlit World. “AI and data centre electricity use continues to surge, IEA finds.” Enlit World, April 17, 2026. https://www.enlit.world/library/ai-and-data-centre-electricity-use-continue-to-surge-iea-finds
[25] Investing.com (Southern Company Q2-2026 earnings call; remarks by CFO David Poroch). “Earnings call transcript: Southern Company tops Q2 2026 EPS estimate but misses revenue.” Investing.com, July 30, 2026. https://www.investing.com/news/transcripts/earnings-call-transcript-southern-company-tops-q2-2026-eps-estimate-but-misses-revenue-93CH-4825848
[26] BigGo Finance (remarks by Southern Company CEO Chris Womack). “Southern Company EPS Soars 23% to $1.13 as Data Center Load Surges 55% and OpenAI Deal Highlights 17 GW Contracted Pipeline.” BigGo Finance, July 30, 2026. https://finance.biggo.com/news/US_SO_2026-07-30
[27] Data Center Frontier. “Southern’s 17 GW Pipeline Puts AI Power Demand Into Utility Math.” Data Center Frontier, August 18, 2026. https://www.datacenterfrontier.com/energy/article/55398810/southerns-17-gw-pipeline-puts-ai-power-demand-into-utility-math
[28] U.S. Department of Energy, Office of Nuclear Energy. “Advantages and Challenges of Nuclear-Powered Data Centers.” DOE, February 12, 2026. https://www.energy.gov/ne/articles/advantages-and-challenges-nuclear-powered-data-centers
[29] Crownhart, Casey (with comments by Patrick White, Nuclear Innovation Alliance, and Urvi Parekh, Meta). “Can nuclear power really fuel the rise of AI?.” MIT Technology Review, May 20, 2025. https://www.technologyreview.com/2025/05/20/1116339/ai-nuclear-power-energy-reactors/
[30] Powerstack (interview with Prof. Jesse Jenkins, Princeton University ZERO Lab). “Jesse Jenkins on the fastest clean firm(a) power.” Powerstack, June 22, 2026. https://powerstack.currence.ai/p/powerstack-jesse-jenkins-on-the-fastest-clean-firm-a-power-7c92
[31] Center for Strategic and International Studies. “White House Executive Orders Target Ambitious Nuclear Deployment in the United States and Abroad.” CSIS, June 10, 2025. https://www.csis.org/analysis/white-house-executive-orders-target-ambitious-nuclear-deployment-united-states-and-abroad
[32] Spangler, R., Qin, S., Larsen, L. M., Bolisetti, C., Abou-Jaoude, A., Shirvan, K., et al. (INL, MIT, ORNL). “Potential Cost Reduction in New Nuclear Deployments Based on Recent AP1000 Experience.” OSTI.GOV, 2025. https://www.osti.gov/biblio/2571083
[33] Shirvan, Koroush (Massachusetts Institute of Technology, CANES). “2024 Total Cost Projection of Next AP1000 (MIT-ANP-TR-201).” MIT Center for Advanced Nuclear Energy Systems, July 2024. https://web.mit.edu/kshirvan/www/research/ANP201%20TR%20CANES.pdf
[34] Lovins, Amory B. (Stanford University). “Nuclear power is failing, and AI can’t rescue it.” Utility Dive, September 5, 2025. https://www.utilitydive.com/news/nuclear-power-smr-ai-amory-lovins/758660/
[35] Mugglehead Investment Magazine. “Constellation wins waiver to accelerate Crane nuclear plant return.” Mugglehead, June 2026. https://mugglehead.com/constellation-wins-waiver-to-accelerate-crane-nuclear-plant-return/
[36] World Nuclear News (remarks by Joseph Dominguez, Constellation). “Constellation seeks regulator’s help for 2027 plant restart.” World Nuclear News, 2026. https://world-nuclear-news.org/articles/constellation-seeks-regulators-help-for-2027-plant-restart
[37] Zewe, Adam (MIT News; comments by Noman Bashir, MIT CSAIL / MIT Climate and Sustainability Consortium). “Explained: Generative AI’s environmental impact.” MIT News, January 17, 2025. https://news.mit.edu/2025/explained-generative-ai-environmental-impact-0117
[38] Norris, T. H., Profeta, T., Patiño-Echeverri, D., and Cowie-Haskell, A. (Duke University Nicholas Institute). “Rethinking Load Growth: Assessing the Potential for Integration of Large Flexible Loads in US Power Systems (NI R 25-01).” Nicholas Institute for Energy, Environment & Sustainability, February 2025. https://nicholasinstitute.duke.edu/publications/rethinking-load-growth
[39] Enlit World (statement by Tyler Norris, Duke University). “Significant flexible load potential for US grid finds Duke study.” Enlit World, 2025. https://www.enlit.world/library/significant-flexible-load-potential-for-us-grid-finds-duke-study
[40] The Current (Georgia). “Google to help fund upgrades for Georgia Power nuclear reactors.” The Current, September 24, 2026. https://thecurrentga.org/2026/09/24/google-to-help-fund-upgrades-for-georgia-power-nuclear-reactors/
[41] DatacenterDynamics. “Google inks deal to support nuclear uprates at Georgia Power’s Vogtle and Hatch plants.” DCD, September 2026. https://www.datacenterdynamics.com/en/news/google-inks-deal-to-support-nuclear-uprates-at-georgia-powers-vogtle-and-hatch-plants/
[42] World Nuclear News. “Amazon and Meta agreements boost Vistra nuclear plants.” World Nuclear News, February 27, 2026. https://world-nuclear-news.org/articles/ppas-support-extended-operations-at-vistra-plants
[43] Meta Platforms. “Meta Announces Nuclear Energy Projects, Unlocking Up to 6.6 GW to Power American Leadership in AI Innovation.” Meta Newsroom, January 2026. https://about.fb.com/news/2026/01/meta-nuclear-energy-projects-power-american-ai-leadership/
[44] Constellation Energy Corporation. “Form 8-K, Exhibit 99.1: Second Quarter 2026 Results.” U.S. SEC EDGAR, August 6, 2026. https://www.sec.gov/Archives/edgar/data/0001868275/000186827526000097/ceg-20260806991.htm
[45] 24/7 Wall St.. “Constellation Energy Corp (CEG) Q2 2026 Earnings.” 24/7 Wall St., August 6, 2026. https://247wallst.com/cards/constellation-energy-corp-q2-2026-earnings-ceg-01kzbbgcrtg4chyhg2r07hf2a4
[46] Investing.com. “Constellation Q2 2026 slides: nuclear contracts fuel guidance raise.” Investing.com, August 6, 2026. https://www.investing.com/news/company-news/constellation-q2-2026-slides-nuclear-contracts-fuel-guidance-raise-93CH-4843602
[47] American Nuclear Society, Nuclear Newswire. “Constellation news archive (Dresden 30-MWe uprate PPA with Walmart; FERC Crane waiver).” ANS, 2026. https://www.ans.org/news/tag-constellation/
[48] Silverstein, Ken. “The AI Boom Is Making Nuclear Power Bankable Again.” Forbes, July 26, 2026. https://www.forbes.com/sites/kensilverstein/2026/07/26/the-ai-boom-is-making-nuclear-power-bankable-again/
[49] U.S. Department of Energy, Office of Nuclear Energy. “9 Key Takeaways from President Trump’s Executive Orders on Nuclear Energy.” DOE, 2026. https://www.energy.gov/ne/articles/9-key-takeaways-president-trumps-executive-orders-nuclear-energy
[50] The White House. “Executive Order 14302: Reinvigorating the Nuclear Industrial Base.” White House Presidential Actions, May 23, 2025. https://www.whitehouse.gov/presidential-actions/2025/05/reinvigorating-the-nuclear-industrial-base/
[51] Hogan Lovells. “Understanding President Trump’s Executive Orders on promoting nuclear energy and removing regulatory barriers to licensing.” Hogan Lovells, June 4, 2025. https://www.hoganlovells.com/en/publications/understanding-president-trumps-executive-orders-on-promoting-nuclear-energy
[52] Utility Dive (statement by Judi Greenwald, Nuclear Innovation Alliance). “Trump aims for 400 GW of nuclear by 2050, 10 large reactors under construction by 2030.” Utility Dive, May 28, 2025. https://www.utilitydive.com/news/trump-aims-for-400-gw-of-nuclear-by-2050-10-large-reactors-under-construct/749107/
[53] CarbonCredits.com (citing BloombergNEF). “2026: The Year Nuclear Power Reclaims Relevance With 15 Reactors, AI Demand, and China’s Expansion.” CarbonCredits.com, May 7, 2026. https://carboncredits.com/2026-the-year-nuclear-power-reclaims-relevance-with-15-reactors-ai-demand-and-chinas-expansion/
[54] University of Michigan Nuclear Engineering & Radiological Sciences (Prof. Aditi Verma). “Investigating the future of nuclear-powered AI data centers.” UMich NERS, December 4, 2025. https://ners.engin.umich.edu/2025/12/04/investigating-the-future-of-nuclear-powered-ai-data-centers/
