Papers

We develop a behavioural model of bank run exposure in a paycheck-to-paycheck economy with loss averse depositors. Income is received through demand deposits, and consumption ratcheting embeds reference dependence in a parsimonious asset-pricing framework. We show that sufficiently high subjective bad-state probabilities endogenously increase liquidity demand and generate equilibrium stress states supporting bank runs. These states define a Bank Run Exposure State Space and yield a martingale representation for exposure dynamics. A proof-of-concept empirical implementation using Call Report data constructs bank-level exposure proxies from funding and lending composition. A regression-weighted composite measure modestly improves fit relative to a retail-share benchmark, with stronger amplification among small banks and during the post-Silicon Valley Bank (SVB) collapse period. The framework highlights how behavioural liquidity demand alters equilibrium reserve holdings and can crowd out productive lending.

This article considers the problem of testing sign agreement among a finite number of parameters. This problem arises in empirical settings such as detecting treatment effects with opposite signs across subgroups, outcomes, or time periods, and testing instrument validity for local average treatment effects. For the null hypothesis that the parameters are either all non-negative or all non-positive, I propose two novel tests: a least favorable test and a conditional test. The least favorable test uses a worst-case null critical value, while the conditional test first screens components with large positive or negative estimates and then tests the remaining sign-unresolved components conditional on the screening event. Unlike existing sign agreement tests, both procedures accommodate arbitrary dependence among estimators; in the special case of independent estimators, the critical values depend only on the dimension and testing levels. We show that both tests control asymptotic size uniformly over a large class of nonparametric distributions. Local asymptotic power analysis reveals a tradeoff: the least favorable test is more powerful near boundary configurations where sign restrictions bind, whereas the conditional test is more powerful when some components are well separated from zero. Simulation evidence supports these theoretical predictions in finite samples.

We study a decision-maker who explores — dynamically choosing what to learn — before stopping to act. We first reduce this dynamic control problem to a static one: any exploration-and-stopping strategy is equivalent to a choice of the joint distribution of the stopped state and the stopping time, subject to one information-budget constraint at each date, and we characterize exactly which distributions are attainable. The reduced problem is a convex program with a linear objective; its dual prices information over time, and the optimal policy concavifies the stopping payoff net of these shadow prices. The curvature of the decision-maker's time preference then governs the shape of optimal exploration: convex time preference induces Poisson exploration, concave time preference confines stopping to a window whose length is controlled by the dispersion of the marginal cost of delay — forcing an initial phase of pure exploration when the window is short — and the linear case lies at the boundary between them. We apply the framework to real options, to the speed–accuracy tradeoff in information acquisition, and to a continuous-time exploration contest.

This paper considers design-based inference on the average treatment effect in finely stratified experiments, where uncertainty arises only from the randomized treatment assignment. We focus on settings in which units are first stratified into groups of fixed size according to baseline covariates and, then within each group, exactly one unit is assigned to treatment. In this setting, we introduce a class of graph-Laplacian variance estimators in which strata form the vertices of a weighted graph and edge weights determine how between-stratum comparisons are aggregated. The canonical estimator of Imai (2008) corresponds to a complete graph with edge weights normalized so that each stratum has weighted degree one, while a paired-stratum estimator arises from a perfect matching graph. For the subclass of degree-calibrated graphs, in which each vertex has weighted degree one, we derive an exact bias identity showing that the corresponding estimators are upward-biased, with bias governed by squared differences in the true stratum-level treatment effects across adjacent strata. As a result, any such estimator may be used for valid inference. The identity further suggests that paired-stratum estimators constructed from a covariate-based perfect matching can induce small biases when treatment effects vary smoothly with the covariates. Without such smoothness, however, we show that paired-stratum estimators can exhibit large worst-case bias, and that, within the class of degree-calibrated estimators, the complete-graph estimator is minimax optimal for normalized bias under a weak bound on treatment-effect heterogeneity. Motivated by this contrast, we propose a regularized graph estimator that controls worst-case normalized bias while preserving much of the locality of the paired-stratum estimator. Simulations illustrate the resulting tradeoff between locality and worst-case protection.

Inspired by possible future markets of autonomous routing and driving (ARAD), we introduce competitive mediator games and their equilibria which generalize the (coarse) correlated equilibria, which have become a popular research area recently as they not only can be more socially efficient than Nash equilibria but also are limits of algorithmic no-regret multi-agent learning dynamics. We discuss the basic properties of competitive mediator games and prove that in the generic setting of anonymous congestion(routing) games with market-share maximizing mediators all competitive mediator equilibria are monopolies whenever one of the mediators is weakly preferred to other mediators by all users. We apply and interpret these results in the context of new markets of competing ARAD service providers. We also provide a comprehensive overview of these markets and discuss the future mechanism design thereof.

In this study, we use electricity demand growth, cooling requirements, and backup system operation to evaluate the environmental and economic implications of artificial intelligence data centers in the United States. Our results indicate that impacts are not determined solely by facility design, but by the broader electricity, water, and land-use systems in which these facilities operate. Emissions are primarily driven by electricity consumption and therefore depend on marginal generation mixes, transmission constraints, and the spatial and temporal distribution of demand. Analysis further shows that local effects include pressures on water resources, increased noise exposure, and land-use changes, with outcomes varying across regions and infrastructure conditions. The assessment of technological and operational measures shows that improvements in energy efficiency, cooling configurations, and operational strategies can reduce these impacts, although their effectiveness depends on system-level conditions. Evaluation of regulatory and market structures suggests that existing frameworks may not fully account for location- and time-specific externalities. These findings support the need for integrated policy approaches that align data center deployment and operation with electricity system characteristics, water availability, and land-use planning to improve overall environmental and economic performance.

We develop a bias-robust causal inference method for observational panel data settings. Such methods typically impute untreated outcomes, so counterfactual error passes straight into the estimated treatment effect while conventional standard errors ignore it. We adapt bias-aware minimax methods, developed for estimating regression coefficients in factor-model panels, to a causal target: the average effect on the treated, which has to be imputed and may vary across units and periods. The estimator corrects the imputed counterfactual with weighted untreated residuals and reports intervals with an explicit allowance for the error that remains. In simulations the proposed method holds nominal coverage where alternatives such as the generalized synthetic control have almost none, especially when the factor rank is underfitted, at the cost of wider intervals. By applying the developed methodology to real data the estimated effect remains significant for counterfactual errors nearly twice the size that the design's placebos typically exhibit.

I study the optimal design and analysis of randomized experiments for estimating finite-population average treatment effects when potential outcomes are known to be bounded, as with binary outcomes. Among all assignment mechanisms and a broad class of affine estimators, worst-case mean-squared error (MSE) is minimized by independent random assignment and an unconventional regression of the support-midpoint-centered outcome on the recentered treatment, with no intercept. This contrasts with the usual prescription of balanced complete randomization and difference-in-means estimation: when outcomes are bounded, randomness in the realized treatment share is informative. The worst-case gain over full-sample complete randomization is asymptotically small, but gains can be first-order relative to other designs: complete within-pair randomization and pair-fixed-effect regression have twice the worst-case MSE. I extend the result to allow for arbitrary estimators. Independent random assignment remains optimal, and the generally-nonlinear optimal estimator can meaningfully reduce worst-case MSE.

We present a model of associative networks that captures how decision makers expand their consideration set through mental associations between alternatives. Our model provides a tractable approach to study how associations shape choice when some alternatives are available and others are merely observable but unavailable. We characterize the model within a random attention framework and demonstrate unique identification of all parameters. This framework delivers a unified account of several prominent choice anomalies, including classic menu effects and their “phantom” counterparts. We illustrate how associative links serve as a strategic variable in applications such as branding, imitation, and platform design.

How much wage dispersion is attributed to workers, firms, and their sorting depends on how wages are adjusted for observed characteristics. Standard AKM decompositions impose a known linear adjustment. We develop Generalized AKM, a framework that permits an unknown smooth covariate function and group-specific nonlinear interactions while preserving the original variance components. We prove consistency and asymptotic normality with heteroskedastic errors and many fixed effects, and characterize the stronger smoothness required for quadratic forms. In Portuguese employer-employee data, adding worker and firm-input controls lowers the bias-corrected worker-effect variance from $0.551$ to $0.474$ of total wage variance, firm-effect variance from $0.144$ to $0.121$, and sorting from $0.080$ to $0.047$. Across three group-specific nonlinear bases, firm-effect variance remains between $0.114$ and $0.117$ and sorting between $0.041$ and $0.042$, while worker-effect variance ranges from $0.474$to $0.491$. Which controls enter matters more for firm variance and sorting than how flexibly they enter; worker variance remains more sensitive to the basis.

National planning counts population, human capital, and artificial-intelligence preparedness in separate ledgers. Demographic accounting has advanced from headcount to skills-adjusted stocks and still debates how much age structure retains once skills are modeled, yet no existing unit carries the conditions under which preparedness becomes productive capacity. This study introduces the Effective Cognitive Population (ECP), a decomposable unit that weights population by capability and by the conditions under which capability is deployed, anchored to the World Bank Human Capital Index Plus (HCI+) and the non-overlapping dimensions of the IMF AI Preparedness Index. The architecture is portable in principle; the case tested here is artificial intelligence, which has a published preparedness index. For 144 countries, HCI+ becomes a productivity level, AI opportunity uses digital infrastructure and innovation integration, conversion governance uses regulation and ethics, and the benchmark is ECP = N H(1 + AC). Against 2024 total output on identical population bases, ECP raises criterion R-squared from 0.849 for the HCI+-adjusted stock to 0.882 and lowers leave-one-country-out RMSE from 0.723 to 0.641, with the working-age comparison identical and bootstrap intervals excluding zero. Eighty-nine of 144 countries move at least ten rank positions from headcount, mostly through the human-capital adjustment itself. Results are stable across denominators, vintages, aggregation forms, and a 27-rule multiverse. The direct A by C interaction is not statistically supported, so the conjunction is a planning rule rather than causal complementarity. ECP is a diagnostic ledger whose scope excludes forecasts of population decline and estimates of AI's causal productivity effect.

Standard CATE estimators become inadequate under strong treatment-effect heterogeneity: confidence intervals for conditional means need not cover individual counterfactual effects. We propose an Individualized Causal Prediction (ICP) framework that constructs finite-sample valid conformal prediction intervals for the individual causal effect of a specific query unit. The method localizes calibration to a causally relevant neighborhood using cosine similarity weighted by Causal Forest variable importance, augments small local samples synthetically, and calibrates intervals with doubly robust AIPW conformity scores satisfying Neyman orthogonality. Under standard identifying assumptions (SUTVA and strong ignorability) and an outcome-independent calibration-set selection condition, the resulting intervals attain marginal coverage at the nominal level. The local design also supports approximately conditional coverage by making calibration scores more representative of the query unit. Experiments on a high-heterogeneity synthetic dataset and the IHDP benchmark demonstrate that local strategies improve point accuracy over global baselines while maintaining nominal or above-nominal coverage.

Operational risk models can estimate event frequencies precisely while leaving the underlying mechanisms weakly resolved. We study this resolution mismatch using motor insurance and road safety. A revisable DAG represents trip-level crash generation; a separate predictive layer links annual rating information to latent driving states; and an observation process maps crashes into recorded liability claims. Compatibility sets collect the structural and crash-to-claim laws that reproduce an observed annual contrast under stated restrictions. External studies enter only through explicit bridge assumptions and sensitivity bounds. The framework therefore distinguishes sampling uncertainty from uncertainty about structure, observation, and study-to-target correspondence. Two limited examples illustrate the gap. A sublinear mileage relation constrains an aggregate accident rate per unit distance but not its mechanism. In French motor-liability data, the 18-20 versus 40-49 claim-frequency relativity is 3.388 in a model including vehicle and geographic variables and 1.235 when the same model also conditions on a medium-resolution bonus-malus score. These are different predictive functionals, not successive causal adjustments. A Spanish culpability estimate is used only to show how a strong cross-study bridge would restrict a toy bookkeeping region. The framework does not estimate the full DAG; it makes explicit which assumptions are needed before an annual predictive contrast can support a mechanism-specific risk statement.

A principal must decide whether to implement a project. She privately knows the cost, an agent privately knows the benefit. Monetary transfers are not available, and compared to the principal, the agent does not fully internalize the cost. We show that the principal-optimal mechanism does not require the agent to report. Instead, it either ignores the agent or endows the agent with free information and full decision authority.

We consider the problem of bargaining when transfers between agents are possible. Such situations are typically modelled as coalitional games with transferable utilities. However this model makes a strong implicit assumption: the outcome can only depend on the total surplus that each coalition of agents can achieve, not on which agents within the coalition generate the surplus. Is this assumption justified? We define a richer model in which solutions may depend on who generates the surplus. In this model, the classical axiomatisation of the Shapley value fails: a broad family of solutions satisfy efficiency, anonymity, the dummy property, and additivity. Nevertheless, we obtain an axiomatisation of the Shapley value in the richer model by adding continuity and individual rationality to the original axioms.

Individuals are often influenced by their peers because deviating from prevailing behavior entails social costs. However, existing peer effects models typically assume that individuals respond similarly to peers who perform better or worse than they do. This paper introduces a novel structural model of asymmetric peer effects in which conformity incentives depend on whether individuals perform below or above each of their peers. We establish that the model admits a unique equilibrium and show that its parameters can be identified and estimated through simple moment conditions. Applying our method to several student outcomes, we uncover strong evidence of asymmetries in peer effects. We then demonstrate that these asymmetries are highly policy-relevant by studying targeted interventions under budget constraints. Ignoring asymmetries leads to inefficient treatment allocation and substantial welfare losses, reducing welfare to levels comparable to those in a benchmark without social interactions.

The paper considers the problem of variable selection for forecasting electricity spot prices. High-dimensional methods such as LASSO and Elastic Net are widely used for this purpose, and while they exhibit strong predictive performance, their tendency to select over-parameterized models raises questions about interpretability. We evaluate the performance of six variable selection procedures, includingthe recently proposed Boosting Multiple Testing (BMT) method, using an extensive dataset from six regional electricity markets. We assess their performance in terms of both out-of-sample forecasting ac-curacy and model parsimony. We find that, although LASSO and Elastic Net achieve similar accuracy and outperform most screening alternatives, BMT matches their forecasting performance while using less than one-tenth as many variables. Our results reveal that BMT offers researchers and practitioners a substantially more interpretable and computationally efficient alternative to shrinkage methods, without any loss of forecasting accuracy. These findings suggest that the over-parameterization typically associated with regularization methods is not a necessary price for predictive accuracy in electricity price forecasting.

We provide a new asymptotic framework to derive approximately optimal treatment assignments when sampling noise from data is compounded by fundamental uncertainty due to partial identification. We recenter the reduced-form parameter around its least-favorable configuration and consider drifting parameter sequences that yield both diminishing levels of sampling uncertainty and of partial identification. We characterize the limiting decision problem as a normal location shift model with a suitable limiting identified set. We apply our results to treatment choice problems with contaminated outcomes, to robust welfare analyses with partially identified consumer surplus, and to the problem of aggregating experimental estimates for policy adoption.

When generative AI makes the deliverable uninformative, a professional-services provider can still certify the preserved human capability to catch the machine's errors, through a liability pledge whose expected cost falls in that capability. The pledge is credible only up to what can be collected, and that ceiling is set by an underwriter, which bears part of the pledge and audits the insured. The range of client stakes over which one certificate separates has a width bounded by the provider's own capital plus the audited share of the underwriter's capacity, so unmonitored capacity adds nothing to it. Blind capital instead relocates that range upward, through a cross-subsidy that exists only under class rating. Monitoring converts capital into width, removes the cross-subsidy, raises the return to preserving skill, and, because audit information leaks, makes the certificate redundant beyond an interior precision.

We examine the effect of political power-sharing on local economic activity. This effect depends on the relative importance of the risks associated with unchecked power and the potential efficiency gains or losses arising from checks and balances. Our research design exploits a geographic discontinuity design due to the haphazard overlap of electoral and administrative boundaries that generates quasi-random variation in the number of politicians governing adjacent regions. We supplement this design using an episode of electoral delimitation that allows us to exploit within-region variation in the number of politicians. We find increasing the number of politicians governing an area can lead to new firm creation, lower unemployment, and greater real economic activity. Our results suggest that non-aligned multiple politicians enhance state efficiency by imposing checks and balances on each other, leading to lower regulatory obstacles, less cronyism, and improved provision of public infrastructure, creating an economically favorable environment for firm creation.

A common problem in social choice is to determine whether there is a social choice procedure, such as a voting method, satisfying some desired criteria. Computer-aided methods such as SAT solving can sometimes answer these questions. However, under typical encodings, a SAT solver may only synthesize a voting method on a finite domain, while we may want one on an infinite domain, such as the domain of all preference profiles for a fixed number of candidates but any finite number of voters. In this paper, we use an approach based on reasoning with constrained Horn clauses and computation with polyhedra to synthesize a voting method on an infinite domain. We then use SMT and Lean to verify its properties. Our main result is a possibility theorem about four well-known criteria from voting theory: the Condorcet winner and loser criteria, positive involvement, and resolvability. Previous work has shown that for five or more candidates, there is no voting method satisfying these axioms, and that for four candidates, there is no method satisfying these core axioms plus one more invariance axiom. Here we show that for four candidates, there does exist a method satisfying the core axioms and more.

The near-term growth of the commercial space economy and sustainability of the low-Earth orbit (LEO) environment depends on the commercial prospects of large LEO satellite telecommunications constellations (“mega-constellations”). If successful, mega-constellations can spur competition between launch providers, which may lead to lower launch prices. They may also demand orbital sustainability services and/or generate additional collision risk, which may in turn affect space sustainability interests. Consumer demand for mega-constellations' internet services is critical to their commercial success. This analysis presents a method to estimate consumer demand for mega-constellations' internet services, using SpaceX's Starlink over December 2021–November 2023 as an example. This analysis finds that Starlink users are highly concentrated in the wealthiest countries. It also finds that consumer demand for Starlink is growing more slowly than orbital capacity is being added and more slowly than demand for non-Starlink internet overall. Continuation of these trends may indicate that the market for mega-constellation internet is less lucrative than previous forecasts have indicated.

This paper studies a linear panel model with an unrestricted individual effect and a time- stationary idiosyncratic disturbance. We first show that stationarity is a strong restriction in a quantile model. In a linear conditional quantile specification with quantile-dependent slopes, equality of the conditional residual distributions across periods generically forces the slope coefficient to be constant over the quantile index. Thus, a stationary-error model identifies a common location coefficient rather than a collection of quantile-specific slope effects. We then develop a fixed-T estimator of this common coefficient. For each period, we run a cross- sectional quantile regression of the outcome on the full history of regressors. Stationarity makes the quantile projection of the composite individual effect and disturbance common across the period-specific regressions. Differences between diagonal and off-diagonal blocks of the resulting projection coefficients therefore identify the common slope whenever T>=2. We combine all such restrictions by a two-step minimum-distance estimator. The estimator is root-n-consistent and asymptotically normal with fixed T, permits unrestricted dependence across periods within an individual, and does not estimate the individual effects. We provide a consistent analytic covariance estimator, a cluster bootstrap, and an overidentification test of the projection restrictions implied by stationarity. Extensive Monte Carlo experiments show adequate performance under various designs.

Section 1502 of the Dodd–Frank Act, enacted in 2010, requires U.S.-listed companies using tin, tantalum, tungsten, and gold (3TG) from the Democratic Republic of the Congo and adjoining countries to disclose information on the minerals' origins. Concerns have been raised that the regulation may have induced a de facto embargo through avoidance of sourcing from the covered region. However, how the price responsiveness of mineral exports evolved under changing institutional and market conditions remains insufficiently understood. Since tungsten production in the covered region is concentrated almost entirely in Rwanda, this study examines the price responsiveness of Rwandan tungsten exports from January 2009 to December 2023. Because missing export quantity data prevent continuous observation of export unit values, we apply the identification approach of Nakano and Nishimura (2025), combining monthly mirror trade data from UN Comtrade with exchange rates and a world average price. An importer fixed-effects model is estimated using export value as the dependent variable, with the sample divided into four periods according to changes in the institutional and market environment. The results reveal substantial temporal variation in price responsiveness. A statistically significant negative price response is observed in Period 1 ($\eta=-20.814$, $p<0.01$), disappears in Period 2 ($\eta=1.814$, $p>0.10$), reappears in Period 3 ($\eta=-5.277$, $p<0.01$), and disappears again in Period 4 ($\eta=0.440$, $p>0.10$). Coefficient-difference tests confirm significant changes between Periods 1 and 2 ($p=0.0021$) and between Periods 3 and 4 ($p=0.0006$). These findings suggest that the price responsiveness of Rwandan tungsten exports varied substantially over time rather than following a uniform trajectory after the regulation.

Artificial intelligence is associated with larger research teams, yet in mathematics, among the most codifiable fields, individual researchers working with AI now produce research-grade results. A span-of-control model reconciles these observations. AI lowers execution cost, which expands laboratory scale, and automates codifiable tasks, which lowers the member share of each unit. Team size is therefore quasi-concave in AI capability, with at most one peak. The model predicts that a fully codifiable team peaks when effective automation coverage reaches a closed-form threshold, typically near complete coverage, and, among fields with shared primitives that possess an interior peak, those with less irreducibly human task content peak first. Under explicit priors, the 90 percent forecast intervals for the fully codifiable peak span 2026 to 2030.

Generative AI is shifting digital commerce from browsing toward agentic search, in which consumers delegate product discovery to AI agents. We compare manual search, which accurately evaluates a limited product set, with agentic search, which screens a broad catalog through noisy representations of preferences and products. Preference complexity is the number of satisfaction-relevant dimensions that are difficult to articulate before search but readily evaluated upon inspection. Consumers have finite attention and choose search intensity: products inspected manually or preference-refinement depth with an agent. We obtain three findings. First, manual search collapses beyond a finite complexity threshold: inspection ceases, mismatch reaches the no-search benchmark, and platform revenue falls to zero. Agentic search avoids this collapse. Once refinement becomes worthwhile, it remains worthwhile as complexity rises; mismatch stays below the no-search benchmark and revenue remains positive, although articulation effort and mismatch may increase. Second, platforms rank the regimes by conversion revenue, whereas consumers also bear search expenditure. When manual inspection is sufficiently inexpensive, agentic search becomes revenue-superior before consumers voluntarily adopt it, creating an adoption lag in which consumers rationally continue manual search. Third, conditional on agentic participation, platforms may assign lower fidelity to consumers with larger attention budgets because they can offset noisier representations through additional refinement, yielding an inverted fidelity allocation. Agentic commerce thus shifts scarcity from product inspection to preference articulation, making consumers' willingness and ability to interact central to voluntary use and platform fidelity design.

We define a category of strategic games in which a morphism is a pair of a map between sets of players and a map between sets of strategy profiles with specific properties and show that this category is well-defined. Three cases are considered for the maps between sets of strategy profiles: they may be order-preserving, order-reflecting or order-embedding with respect to each player's preference relation. We define a presheaf on the category of games valued in a category of sets that sends each strategic game to a set of strategy profiles and present conditions for this presheaf to be well-defined. Two cases are considered for the morphisms in the category of sets: they may be relations or maps. We define a presheaf that sends each strategic game to the set of Nash equilibria (resp. Pareto efficient strategy profiles) and show that this presheaf is well-defined if and only if the maps between sets of strategy profiles are order-reflecting or order-embedding (resp. order-embedding), and the morphisms in the category of sets are relations.

As firms increasingly deploy machine learning for strategic decision-making, understanding algorithmic interactions has become central to operations research and economics. This paper studies learning in infinite-horizon, nonzero-sum linear-quadratic stochastic games under a radically uncoupled information structure, where players are either unaware of opponents or strategically oblivious, observing only a common state and their own action history. Under this minimal information, we analyze an asynchronous decentralized learning process in which each player independently runs a single-agent $\epsilon$-greedy iterated least-squares algorithm. We prove that, despite being unable to identify the system parameters, players' learning dynamics converge almost surely to the complete-information Nash equilibrium and characterize the convergence rate. We then apply the framework to a dynamic Cournot competition with sticky prices. Numerical experiments validate the theoretical results and show that learning under limited information reduces firm profits under both low and high price stickiness, while total surplus declines and market concentration increases when price stickiness is high. Publicly revealing aggregate market output substantially accelerates convergence and mitigates these welfare losses.

This paper explores the idea of promoting well-being and safety in human-AI interactions by forcing AI agents explicitly to empower humans and to manage the power balance between humans and AI agents in a desirable way. Using a principled, partially axiomatic approach based on desirable properties, we design a parametrizable and decomposable objective function for AI systems that represents an inequality- and risk-averse long-term aggregate of human power. It can take into account models of human bounded rationality and social norms, and crucially, considers a wide variety of possible human goals. We prove how certain desiderata enforce particular functional forms and restrict parameter ranges. We exemplify the consequences of softly maximizing this metric in several paradigmatic situations and describe what instrumental sub-goals it will likely imply.

Interconnection queues, not electricity prices, now govern where data centers can be built, and the standard levelized-cost comparison answers a question no developer faces: it assumes a load profile, freezes the grid price while modeling the demand that moves it, and quotes busbar costs a facility cannot buy. This paper evaluates nine on-site supply technologies against a delivered grid whose price is endogenous to projected data-center demand, on a complete-site basis that retains standby charges, with measured GPU training load, delivered fuel prices, production-pathway carbon, and statutory 45V and 48E incentive mechanics. Nothing beats the wire: gas combined cycle produces at 47 USD/MWh but costs about 114 USD per megawatt-hour of complete site energy against a 92 USD grid; four-hour storage is physically capped near 18 percent of annual energy and, charged at the margin, dirtier than the grid; hydrogen from grid-priced power fails on cost and carbon together. An investment inversion converts these findings into capital terms: conversion-hardware learning buys nothing, because free hardware still exceeds the grid for every low-carbon arm, while global electrolyser deployment on sited sub-20 USD/MWh power brings PEM hydrogen power to about 2.2 times the grid at 300 billion USD and 1.9 times at 1 trillion USD (2.7 and 2.3 for the hydrogen engine), with a carbon reduction of roughly 85 percent (6.8-fold) against grid-power production. Grid parity is not purchasable at any budget. On-site supply is an access and depth product; most current investment targets the wrong term.