The Algorithmic Calculation Problem: Why Foundation Models Cannot Solve Socialist Planning

The Algorithmic Calculation Problem: Why Foundation Models Cannot Solve Socialist Planning | Research Square window.SnipcartSettings = { analytics: { enabled: false } }; (function() { var accessVector = localStorage.getItem('access_vector') || ''; window.dataLayer = window.dataLayer || []; if (accessVector) { window.dataLayer.push({ user: { profile: { profileInfo: { snid: accessVector } } } }); } })(); (function(w,d,s,l,i){w[l]=w[l]||[];w[l].push({'gtm.start':new Date().getTime(),event:'gtm.js'});var f=d.getElementsByTagName(s)[0],j=d.createElement(s),dl=l!='dataLayer'?'&l='+l:'';j.async=true;j.src='https://www.googletagmanager.com/gtm.js?id='+i+dl;f.parentNode.insertBefore(j,f);})(window,document,'script','dataLayer','GTM-K279D39R'); Browse Preprints In Review Journals COVID-19 Preprints AJE Video Bytes Research Tools Research Promotion AJE Professional Editing AJE Rubriq About Preprint Platform In Review Editorial Policies Our Team Advisory Board Help Center Sign In Submit a Preprint Cite Share Download PDF Research Article The Algorithmic Calculation Problem: Why Foundation Models Cannot Solve Socialist Planning Craig Wright This is a preprint; it has not been peer reviewed by a journal. https://doi.org/ 10.21203/rs.3.rs-8856269/v1 This work is licensed under a CC BY 4.0 License Status: Posted Version 1 posted You are reading this latest preprint version Abstract Large language models and generative AI systems have reignited debate over the feasibility of non-market economic coordination. A growing literature contends that modern AI renders the Hayekian knowledge problem technologically obsolete. This paper argues that these claims rest on a misidentification of the epistemic status of foundation model capabilities. Drawing on mechanistic interpretability research (Elhage et al., 2021 ; Olah et al., 2020 ; Anthropic, 2023 ) and a novel dataset of 127 AI-native firms (2022–2024), I develop three arguments. First, foundation models suffer from a derivative knowledge problem : they compress statistical regularities from market-coordinated data but cannot generate the forward-looking knowledge that entrepreneurial discovery produces under genuine uncertainty. I formalize this distinction by separating computational optimization within a known possibility space from entrepreneurial discovery that expands the possibility space itself, and I demonstrate that the specific epistemic object missing from non-market telemetry is opportunity-cost structure —the counterfactual valuations that only competitive bidding reveals. Second, AI capital exhibits extreme Lachmannian heterogeneity and plan-dependence, generating Austrian business cycle dynamics visible in the 2024 GPU shortage and subsequent overcapacity. Third, AI simultaneously lowers barriers to competitive entry at the application layer while concentrating complementary assets at the infrastructure layer—a pattern I term “democratized disruption with oligopolistic infrastructure.” Empirical analysis using Cox proportional hazard models reveals that model-provider dependency and pivot frequency are significant predictors of AI firm survival, while infrastructure-layer market structure shapes but does not determine application-layer outcomes. The findings suggest structural limits to algorithmic coordination that are not reducible to computational constraints. economic calculation socialist calculation debate large language models knowledge problem spontaneous order capital heterogeneity foundation models artificial intelligence institutional economics market coordination 1. Introduction The socialist calculation debate, inaugurated by Mises ( 1920 ) and refined through the interwar contributions of Hayek ( 1937 , 1945 ), Lange ( 1936 ), and Lerner ( 1944 ), addressed whether rational economic coordination is possible without market-generated prices. The Austrian position—that prices generated through private property and competitive exchange carry epistemic content that cannot be replicated by a central authority—was widely regarded as vindicated by the empirical record of planned economies (Boettke, 1993 ; Lavoie, 1985 ). Yet the emergence of large language models (LLMs) and generative AI systems has prompted a reconsideration. A growing literature spanning accelerationist Marxism (Bastani, 2019 ; Srnicek, 2017 ), digital socialism (Morozov, 2019 ; Phillips and Rozworski, 2019 ), and techno-utopianism (Buterin, 2023 ) contends that modern AI has rendered the knowledge problem technologically obsolete. This paper argues that these claims rest on a fundamental misidentification of both what foundation models do and why the calculation problem is not merely a computational limitation. I develop three interconnected arguments. The first is epistemological: foundation models suffer from what I call the derivative knowledge problem . These systems compress statistical regularities from data that was itself produced through market coordination. They do not—and structurally cannot—generate the forward-looking knowledge that entrepreneurial discovery produces under genuine uncertainty. I formalize this distinction by separating computational optimization within a known possibility space from entrepreneurial discovery that expands the possibility space itself, and I demonstrate that the specific epistemic object absent from non-market data is opportunity-cost structure—the counterfactual valuations that competitive bidding reveals. The second argument is capital-theoretic: AI investment exhibits the radical heterogeneity and plan-dependence that Lachmann ( 1956 ) identified as characteristic of capital goods in a world of genuine uncertainty. The 2024 GPU shortage and subsequent overcapacity provide a textbook illustration. The third argument concerns market structure: AI simultaneously lowers barriers to entrepreneurial entry while concentrating essential complementary assets—a pattern I term “democratized disruption with oligopolistic infrastructure.” Empirically, the paper draws on a proprietary dataset of 127 AI-native firms founded between 2022 and 2024. Using Cox proportional hazard models with robustness checks for alternative censoring assumptions and stricter dependency definitions, I examine how model-provider dependency, founding capital, team composition, and pivot frequency predict firm survival. The results are consistent with the theoretical framework: AI markets exhibit vigorous application-layer competition whose outcomes are shaped—but not determined—by infrastructure-layer market structure. Three case studies of frontier model providers (OpenAI, Anthropic, Mistral) illustrate the institutional diversity that market processes generate and that centralized planning could not anticipate. The paper proceeds as follows. Section 2 revisits the socialist calculation debate with attention to the specific claims made by proponents of AI-enabled planning. Section 3 develops the derivative knowledge problem, including a formal distinction between optimization and discovery and a demonstration that opportunity-cost structure is irreconstructible from non-market telemetry. Section 4 applies Lachmannian capital theory to AI investment patterns. Section 5 analyzes the paradox of democratized disruption. Section 6 presents empirical evidence, including survival analysis with robustness checks. Section 7 concludes by reflecting on the structural limits of algorithmic coordination. 2. The Calculation Debate Revisited: From Analog to Digital Utopianism The original Misesian challenge was not, as is sometimes supposed, merely that socialist planners would lack sufficient computational power to solve a system of simultaneous equations. As Lavoie ( 1985 ) and Boettke ( 1998 ) have persuasively argued, the knowledge at issue in the calculation debate is not the sort of propositional knowledge that can, even in principle, be codified and processed by a central authority. Hayek’s ( 1945 ) contribution was to clarify that the relevant knowledge is dispersed, tacit, contextual, and continuously changing—“the knowledge of the particular circumstances of time and place” (p. 521). Market prices serve as epistemic signals precisely because they aggregate this otherwise inaccessible knowledge through the competitive process of entrepreneurial bidding. Contemporary advocates of AI planning have revived the computational interpretation of the calculation problem, arguing that what Hayek described as an information-processing challenge can now be overcome by sufficiently powerful machine learning systems. Morozov ( 2019 ) suggests that platform companies like Amazon and Alibaba already engage in forms of non-market coordination at enormous scale. Bastani ( 2019 ) contends that AI-driven automation will make the marginal cost of production approach zero, dissolving the need for price-mediated allocation. Phillips and Rozworski ( 2019 ) point to Walmart’s internal logistics algorithms as evidence that planned economies “already exist” within the capitalist firm. These arguments engage the calculation debate at the wrong level of abstraction. The question is not whether a sufficiently powerful algorithm can process information—it is whether the relevant information can exist in the absence of market institutions. As Kirzner ( 1984 ) emphasized, entrepreneurial discovery is not a matter of optimizing within a known constraint set; it is a process of alertness to hitherto unnoticed opportunities. The knowledge that a particular resource combination would satisfy an as-yet-unarticulated consumer preference does not exist somewhere waiting to be aggregated. It comes into existence only through the entrepreneurial act itself, embedded within an institutional context of private property and residual claimancy. The distinction between processing existing information and generating new information is the conceptual pivot on which the entire debate turns—and it is precisely this distinction that the AI planning literature fails to engage. 3. The Derivative Knowledge Problem The central claim of this paper is that foundation models—large-scale neural networks trained on massive corpora of text, code, and other data—suffer from a derivative knowledge problem that makes them structurally incapable of substituting for market coordination. This is not a claim about computational limitations that might be overcome with larger models or better architectures. It is a claim about the epistemic status of the data on which these models are trained and, more fundamentally, about the logical distinction between two kinds of knowledge generation. 3.1 The Epistemic Provenance of Training Data Consider what a large language model learns. A model like GPT-4 or Claude is trained on hundreds of billions of tokens of text drawn from the internet, books, academic papers, and code repositories. Recent advances in mechanistic interpretability (Elhage et al., 2021 ; Olah et al., 2020 ) have revealed that these models develop internal representations that capture complex semantic and relational structures. Anthropic’s ( 2023 ) work on “toy models of superposition” demonstrates that neural networks can learn to represent more features than they have dimensions, achieving remarkable information compression. But the crucial question, from the standpoint of economic calculation, is: where does this information come from? The training data is not a neutral description of an objective reality. It is the textual residue of billions of human decisions, transactions, evaluations, and communications—the overwhelming majority of which were made within market-coordinated institutional frameworks. When a language model “learns” that certain products are associated with certain price ranges, that particular business strategies tend to succeed or fail, or that specific resource combinations are valued by consumers, it is learning the condensed output of market processes. The model’s knowledge is derivative of the market order in a strong sense: it is not merely informed by market outcomes but constituted by them. Mechanistic interpretability provides enabling evidence for the compression capacity of these models, though the provenance argument rests on a separate foundation. Elhage et al. ( 2021 ) demonstrate that transformer models learn attention patterns that track complex relational structures, and Olah et al. ( 2020 ) show that neural network features decompose into interpretable components corresponding to human-recognizable concepts. These findings establish that foundation models achieve genuine structured compression rather than mere surface-level pattern matching. But the provenance claim does not rest on interpretability alone. It rests on the institutional origin of the corpora: when models process economic text—financial reports, business analyses, product reviews—the relational structures they compress (relative prices, comparative quality assessments, opportunity costs) are themselves outputs of market coordination. The model’s internal representations are not generic information-processing structures that happened to be trained on economic data; they are market-specific epistemic artifacts whose evaluative content derives from the competitive price system that generated the training distribution. Interpretability reveals that the model has learned rich structure; institutional analysis reveals where that structure comes from. 3.2 The Counterfactual Collapse This derivative quality becomes decisive when we consider the counterfactual scenario that proponents of AI planning envision. Suppose a socialist planning authority were to deploy a foundation model trained on the entirety of the internet as of 2025. The model would contain rich knowledge about relative valuations, supply chain logistics, consumer preferences, and production techniques. But this knowledge was generated under market institutions. Remove those institutions and the model’s knowledge begins to decay in a manner that the model itself cannot correct. This is because the model cannot generate the new knowledge that market processes continuously produce. Entrepreneurial discovery reveals previously unknown production possibilities, latent consumer preferences, and resource complementarities. These discoveries are not drawn from a pre-existing stock of information; they are genuinely creative acts that depend on the institutional incentive structures of market competition (Kirzner, 1997 ). A foundation model can interpolate within the space of existing market-generated knowledge, but it cannot extrapolate into the space of knowledge that would have been generated by market processes that no longer exist. The analogy to Neurath’s ship is instructive, though it runs in the opposite direction from how it is typically deployed. Neurath ( 1921 ) argued that we must rebuild the ship of knowledge while sailing on it. The proponent of AI planning proposes something more radical: to sail on a ship built from the blueprints of a demolished vessel, without the institutional infrastructure that generated those blueprints. 3.3 Optimization Versus Discovery: A Formal Distinction The derivative knowledge argument can be sharpened by formalizing the distinction between two epistemic operations that the AI planning literature conflates: computational optimization and entrepreneurial discovery . Computational optimization operates within a defined possibility space . Let Ω denote the set of feasible states of the world, f(·) an objective function mapping states to values, and C a constraint set. Optimization consists of identifying x* ∈ Ω such that f(x*) ≥ f(x) for all x ∈ C. Crucially, Ω, f, and C are given in advance. The computational challenge is combinatorial: the search space may be vast, and the function may be non-convex. But the elements of the problem—what goods exist, what preferences obtain, what production possibilities are available—are specified ex ante. Foundation models excel at this kind of operation. They compress high-dimensional patterns from training data and can efficiently identify solutions within the possibility space that the data defines. This is what underlies their remarkable performance in code generation, text completion, and pattern recognition: they operate within the statistical manifold of their training distribution. Entrepreneurial discovery, by contrast, expands the possibility space itself . The entrepreneur does not select x* from a known Ω; she identifies a previously unrecognized Ω′ ⊃ Ω, or a previously unarticulated f′ that revalues existing elements of Ω, or a novel production possibility that was not in C. Following Knight ( 1921 ), this is not risk (calculable probability over known outcomes) but genuine uncertainty (unforeseeable structural novelty). The introduction of the iPhone did not optimize within the existing mobile phone possibility space; it redefined what a mobile device was. Airbnb did not compute the optimal hotel allocation; it discovered that the category “accommodation” was larger than anyone had recognized. This distinction is not a matter of degree—as though discovery were merely very difficult optimization. It is a categorical difference in epistemic operation. Optimization is a mapping within a representation; discovery is a transformation of the representation. A foundation model, however large, operates within the representation defined by its training data. It can recombine elements of that representation in novel ways—and this recombinatory capacity is genuinely impressive—but it cannot transcend the representation itself. The model has no access to what Shackle ( 1972 ) called “the void”: the space of possibilities that have not yet been imagined, let alone observed. This formal distinction has a direct implication for the calculation debate. A socialist planning authority equipped with foundation models would possess an extraordinarily powerful optimization engine—capable of allocating known resources to known ends with impressive efficiency. But it would lack the discovery engine that market competition provides: the institutional mechanism through which unknown resources, unknown ends, and unknown production possibilities are continuously brought into existence. The planning problem is not that the planner cannot compute; it is that the planner cannot know what to compute about . 3.4 Objections and Replies: Continuous Retraining and Decentralized Input The strongest objection to the derivative knowledge argument takes the following form: could a planning authority not continuously retrain its foundation models on post-market data, thereby generating fresh knowledge without market institutions? Three variants of this objection deserve careful engagement. The first variant suggests that administrative data —production reports, consumption surveys, logistics telemetry—could substitute for market-generated data. The difficulty is that administrative data records what happened rather than what could have happened . Market prices encode opportunity costs: the value of the next-best alternative use of a resource, as assessed by the marginal bidder who did not obtain it. Administrative data records the allocation that was made but contains no information about the allocations that were foregone. Without opportunity cost information, a foundation model retrained on administrative data would learn the surface pattern of economic activity without the evaluative structure that makes coordination possible. The second variant proposes decentralized data input —allowing citizens to report preferences, needs, and local conditions directly to the planning system. This is the most sophisticated version of the objection because it addresses the Hayekian concern about dispersed knowledge. But it confronts two difficulties. First, agents lack incentives to reveal truthful preferences in the absence of the discipline imposed by budget constraints and competitive exchange. The incentive-compatibility problem that Hurwicz ( 1960 , 1972 ) formalized applies with full force: without a mechanism that makes truthful revelation individually rational, preference data will be systematically distorted. Second, and more fundamentally, many of the relevant preferences do not exist until they are evoked by entrepreneurial offerings. Consumers did not report a preference for smartphones, ride-sharing, or social media before these products existed. Entrepreneurial discovery is not the aggregation of pre-existing preferences; it is the co-creation of preferences and products through market interaction. The third variant invokes synthetic data generation —using AI models to simulate market dynamics and generate training data endogenously. This proposal is self-defeating. A model trained on its own outputs converges to the statistical regularities of its initial training distribution rather than discovering new information. The phenomenon of “model collapse” documented by Shumailov et al. ( 2024 ) shows that recursive self-training degrades model quality precisely because it eliminates the exogenous variation that contact with genuine human behavior provides. The market is not merely a source of data; it is the source of the surprise —the unforeseeable novelty—on which productive model training depends. 3.5 The Irreconstructibility of Opportunity-Cost Structure The objections addressed above share a common assumption: that the knowledge deficit facing a post-market planning system is essentially a data quantity problem—that what planners lack is sufficient information about preferences, endowments, and technologies, and that better telemetry or larger datasets could close the gap. This subsection argues that the deficit is instead a data structure problem. The specific epistemic object that non-market telemetry cannot reconstruct is opportunity-cost structure : the pattern of counterfactual valuations that competitive bidding reveals and that no quantity of observational data can replicate. Consider a concrete case. A market price of $ 12 per bushel for wheat does not merely record the transaction price at which wheat changed hands. It encodes: (i) the marginal buyer’s valuation, (ii) the marginal seller’s reservation price, (iii) the next-best alternative use of the land, labor, water, and capital that produced the wheat, as assessed by the bidders who did not obtain those resources, and (iv) the valuation of the goods not produced because resources were allocated to wheat rather than to their next-best use. Items (iii) and (iv) are the opportunity-cost structure. They are not observable in any administrative record because they describe states of the world that did not obtain . They exist as information only because the competitive bidding process forced agents to reveal their valuations by bearing the cost of acting on them. The irreconstructibility claim can be stated formally. Let x denote the observed allocation and v(x) the market price vector. Let Ω denote the set of all feasible allocations, and let v(ω) for ω ∈ Ω denote the valuations that would obtain under alternative allocations. Market prices embed information about the gradient ∇v at x—how valuations change at the margin for small reallocation—because the competitive process is precisely a mechanism for eliciting marginal willingness to pay across alternative uses. Administrative telemetry observes only x and (at best) v(x): the allocation that was made and the prices at which it transacted. It is silent on ∇v because the counterfactual allocations ω ≠ x were never tried, and the agents who would have bid on those allocations were never forced to reveal their valuations. This is not a practical difficulty that better sensors or more comprehensive monitoring could overcome. It is a structural feature of non-market institutional environments. Opportunity costs are revealed by competitive bidding; they are not observed by administrative recording. The distinction maps directly onto the mechanism-design literature: as Hurwicz ( 1972 ) demonstrated, a social choice function achieves informational efficiency only if the institutional mechanism incentivizes agents to reveal private valuations. Without such a mechanism, the information does not enter the system—not because agents choose to withhold it, but because the institutional structure provides no occasion for its articulation. The implication for foundation models is decisive. A model retrained on non-market data would learn the surface texture of economic activity—what was produced, where, when, in what quantities—but would lack the evaluative depth that opportunity-cost structure provides. It would know that 10,000 bushels of wheat were shipped from Region A to Region B, but not whether the land in Region A would have been more productively employed growing soybeans, the labor more valuably engaged in manufacturing, or the transport capacity more efficiently allocated to a different commodity route. Without this evaluative depth, the model cannot assess whether any given allocation is better or worse than its alternatives —which is to say, it cannot perform economic calculation in the sense that Mises ( 1920 ) identified as the core function of market prices. The algorithmic calculation problem is not that the algorithm lacks data; it is that the data it can access is observationally thin on precisely the dimension—counterfactual valuation—that economic coordination requires. 4. Capital Heterogeneity and Plan-Dependence in AI Markets Lachmann’s ( 1956 ) capital theory emphasizes two properties of capital goods that distinguish the Austrian approach from neoclassical production theory: heterogeneity and plan-dependence. Capital goods are not interchangeable units of a homogeneous aggregate but specific, multi-attribute resources whose economic significance depends on the entrepreneurial plans within which they are embedded. A blast furnace is not merely “capital”; it is a highly specific asset whose value depends on complementary inputs, expected future demand, the availability of skilled labor, and the broader production structure within which it operates. AI capital exhibits these Lachmannian properties in extreme form. A GPU cluster configured for training large language models represents an enormous fixed investment whose economic value depends critically on complementary assets: training data (which may be proprietary), model architectures (which embody tacit engineering knowledge), human capital (ML researchers with highly specific expertise), and downstream applications (which may not yet exist). The plan-dependence of AI capital creates vulnerabilities characteristic of Austrian business cycle dynamics. 4.1 The 2024 GPU Shortage as Malinvestment The GPU shortage that emerged in late 2023 and persisted through much of 2024 provides a compelling illustration. Following the release of ChatGPT in November 2022, demand for NVIDIA’s H100 and A100 GPUs surged dramatically. Major technology companies placed orders worth billions of dollars. Cloud computing providers expanded GPU capacity. Startups acquired computing resources at significant premiums. By mid-2024, the secondary market for H100 GPUs had reached prices well above NVIDIA’s list price. From the standpoint of capital theory, this pattern is recognizable as a boom-phase elongation of the production structure. The expectation of future AI revenues—fueled by the capabilities demonstrated by GPT-4 and subsequent models—led to massive investment in higher-order capital goods (GPU manufacturing capacity, data center construction, power infrastructure) predicated on assumptions about future consumer valuations that may or may not materialize. The subsequent emergence of overcapacity concerns in late 2024 and early 2025, as some AI startups failed to find product-market fit and hyperscaler capital expenditure growth rates decelerated, suggests the early stages of the Hayekian “reveal”—the point at which malinvestment becomes apparent as entrepreneurial plans prove inconsistent with underlying consumer time preferences. 4.2 Complementarity and Specificity in AI Production The plan-dependence of AI capital is amplified by complex complementarity relationships. A foundation model’s value depends not only on the compute used to train it but on the quality and composition of its training data, the fine-tuning applied for specific applications, the deployment infrastructure supporting inference, and the business model connecting the model’s capabilities to willingness-to-pay. Each of these complementary assets exhibits its own form of specificity and plan-dependence, creating a web of interdependent entrepreneurial expectations. This web of complementarities means that errors in any component can cascade through the production structure. A training data strategy that proves legally vulnerable (as several AI companies have discovered in copyright litigation), an architectural choice superseded by a rival’s innovation, or a business model that fails to capture sufficient value—any of these can render enormous complementary investments economically obsolete. The capital loss is not merely quantitative but qualitative: the specific configuration of assets loses its economic rationale, and the assets cannot simply be “recombined” without entrepreneurial judgment about alternative plan structures. 5. Democratized Disruption with Oligopolistic Infrastructure Kirzner’s ( 1973 ) theory of entrepreneurial discovery posits that alert entrepreneurs, by noticing and exploiting previously overlooked price discrepancies, drive the market toward coordination. The competitive process is fundamentally equilibrating: entrepreneurial profit opportunities exist because the market is not yet fully coordinated, and entrepreneurial action tends to eliminate these opportunities. This framework has provided an influential theoretical foundation for understanding how markets generate and disseminate knowledge. The AI market complicates this picture in ways that call for refinement rather than abandonment. AI lowers barriers to entrepreneurial entry at the application layer while simultaneously concentrating essential complementary assets at the infrastructure layer. This creates what I term “democratized disruption with oligopolistic infrastructure”—a market structure in which the entrepreneurial discovery process operates vigorously at one level of the production structure while being constrained by market power at another. 5.1 Application-Layer Entrepreneurship The release of foundation model APIs—beginning with OpenAI’s API in 2020 and expanding rapidly through 2023–2025—has dramatically lowered the cost of AI-powered entrepreneurship. An entrepreneur with a domain-specific insight can now build an AI application without training their own model, managing GPU infrastructure, or employing a team of machine learning researchers. The result has been an explosion of AI-native startups: our dataset of 127 firms founded between 2022 and 2024 encompasses applications ranging from legal document analysis to agricultural advisory services to personalized education. The median founding team size is 3.2 people, and the median initial capital requirement is approximately $ 150,000—a fraction of the cost that would have been required five years earlier. This application-layer entrepreneurship is quintessentially Kirznerian. These entrepreneurs are alert to specific, localized opportunities to apply general-purpose AI capabilities in particular market niches. They possess contextual knowledge—about the pain points of legal professionals, the information needs of smallholder farmers, the learning styles of different student populations—that cannot be centrally aggregated. Their competitive entry generates new knowledge about the value of AI capabilities in specific applications, knowledge that did not exist before the entrepreneurial act. 5.2 Infrastructure-Layer Concentration At the infrastructure layer, however, a different dynamic obtains. The training of frontier foundation models requires capital investments on the order of $ 100 million to $ 1 billion, access to proprietary or carefully curated datasets of enormous scale, and teams of researchers with highly specialized expertise. As of early 2025, the number of organizations capable of training genuinely frontier models can be counted on two hands: OpenAI, Anthropic, Google DeepMind, Meta, Mistral, and a small number of others. This concentration does not arise from market failure in any conventional sense. It reflects genuine economies of scale in model training, the accumulated specificity of complementary assets, and the path-dependent nature of organizational knowledge. But it does mean that application-layer entrepreneurs are dependent on infrastructure-layer incumbents for access to foundational capabilities. The relationship between these two layers—competitive and entrepreneurial at the application level, oligopolistic and capital-intensive at the infrastructure level—creates dynamics that existing theoretical frameworks do not fully address. This layered structure has an important implication for the calculation debate. Even if a planning authority could replicate the infrastructure layer—assembling the compute, data, and talent required to train frontier models—it could not replicate the application-layer discovery process. The thousands of entrepreneurs experimenting with AI applications are conducting a massively parallel, decentralized search across the space of possible uses, guided by local knowledge and disciplined by profit and loss. No central authority could design this search process because the search space itself is unknown in advance. 6. Empirical Evidence: AI Firm Dynamics and Institutional Diversity 6.1 Data and Methodology The empirical component of this paper draws on a proprietary dataset of 127 AI-native firms founded between January 2022 and December 2024. Firms were identified through a systematic search of Crunchbase, PitchBook, and Y Combinator’s public portfolio, filtered to include only firms whose primary product or service relies on foundation model capabilities (either via API access or open-weight deployment). The sample excludes firms whose AI use is incidental to a non-AI core business, as well as infrastructure-layer firms (GPU cloud providers, model training companies) to maintain analytical focus on application-layer entrepreneurship. For each firm, I recorded founding date, initial capital raised (seed and pre-seed), founding team size, team composition (proportion of technical versus domain-specialist founders), primary use case (categorized into seven sectors: legal, healthcare, education, financial services, creative tools, enterprise software, and other), model provider dependency (primary API provider or open-weight model), number of significant pivots (defined as a change in primary use case, target market, or model provider), and operational status as of January 2025 (active, ceased operations, or acquired). Data were collected from Crunchbase records, SEC filings, press releases, company websites, and 34 semi-structured interviews with founders and early-stage investors conducted between June and November 2024. Selection bias is a potential concern: firms that failed very early may not appear in Crunchbase or PitchBook records. To mitigate this, I supplemented the database search with Y Combinator batch records (which include failed firms) and cross-referenced with TechCrunch’s shutdown tracker. Survivorship bias in the interview sample is addressed by including 8 interviews with founders of firms that had ceased operations by the interview date. The case study methodology follows Yin ( 2018 ); case studies of OpenAI, Anthropic, and Mistral were constructed from public corporate communications, regulatory filings, published research, and 12 background interviews conducted under Chatham House rules. 6.2 Summary Statistics Table 1 presents summary statistics for the 127 firms in the sample, disaggregated by operational status as of January 2025. Table 1 Summary Statistics by Operational Status (N = 127) Variable All Firms Active Ceased Acquired N 127 93 23 11 Initial capital ($K), median 150 185 90 210 Founding team size, mean 3.2 3.5 2.4 3.1 Technical founders (%), mean 62.4 58.7 78.3 63.6 Pivots, mean 0.81 0.94 0.48 0.55 Single-provider dependent (%) 71.7 64.5 91.3 81.8 OpenAI as primary provider (%) 58.3 53.8 73.9 63.6 Several patterns are notable. Ceased firms had significantly smaller founding teams (mean 2.4 versus 3.5 for active firms, p < 0.01, two-tailed t-test) and lower initial capital (median $ 90K versus $ 185K, p < 0.05, Mann-Whitney U). Ceased firms were substantially more likely to be dependent on a single model provider (91.3% versus 64.5%), suggesting that provider diversification may function as a form of entrepreneurial hedging against infrastructure-layer risk. Interestingly, active firms had a higher mean number of pivots (0.94 versus 0.48 for ceased firms), consistent with the interpretation that pivoting reflects adaptive entrepreneurial learning rather than failure. 6.3 Survival Analysis To examine the determinants of AI firm survival more rigorously, I estimate a Cox proportional hazard model with firm failure (cessation of operations) as the event of interest. The observation window runs from each firm’s founding date to January 2025 or the date of cessation, whichever comes first. Acquired firms are treated as right-censored. The model takes the form: h(t | X) = h₀(t) exp(β₁ ln(Capital) + β₂ TeamSize + β₃ TechShare + β₄ Pivots + β₅ SingleProvider + γ Sector) where h₀(t) is the baseline hazard, Capital is initial funding in thousands of dollars (log-transformed), TeamSize is the number of founders, TechShare is the proportion of founders with technical backgrounds, Pivots is the cumulative count of significant pivots, SingleProvider is a binary indicator for dependence on a single model provider, and Sector is a vector of sector fixed effects. Table 2 reports the results. Table 2 Cox Proportional Hazard Model: Determinants of AI Firm Failure Variable Hazard Ratio Std. Error p-value ln(Initial Capital) 0.71 0.12 0.031** Team Size 0.68 0.14 0.018** Technical Share 1.89 0.41 0.044** Pivot Count 0.54 0.19 0.009*** Single Provider Dep. 2.47 0.63 0.003*** Sector FE — — Joint: 0.127 N = 127; Events = 23; Concordance = 0.74 Notes : *** p < 0.01, ** p < 0.05. Hazard ratios below 1 indicate reduced failure risk. The results reveal several patterns consistent with the theoretical framework. Single-provider dependency is the strongest predictor of failure: firms dependent on a single model provider face a hazard ratio of 2.47 (p < 0.003), indicating roughly 2.5 times the failure risk of firms with provider diversification. This is consistent with the capital-theoretic argument about plan-dependence: firms whose entrepreneurial plans are tightly coupled to a single infrastructure provider face elevated risk from API pricing changes, capability shifts, or provider strategic pivots. Pivot count has a protective effect (HR = 0.54, p < 0.01): each additional pivot reduces failure risk by approximately 46%. This finding challenges the view that pivoting signals distress. In the context of AI entrepreneurship, pivots appear to reflect adaptive learning—the kind of entrepreneurial responsiveness to feedback that Kirzner ( 1997 ) associates with the discovery process. Active firms are those that updated their plans in response to market signals; failed firms disproportionately maintained their original plans. The positive hazard ratio for technical founder share (HR = 1.89, p < 0.044) is counterintuitive but interpretable. Teams with very high proportions of technical founders may lack the domain-specific market knowledge needed to identify viable applications—precisely the kind of contextual knowledge that Hayek ( 1945 ) argued cannot be centrally aggregated. The most resilient firms in the sample combine technical capability with domain expertise, consistent with the Austrian emphasis on entrepreneurship as requiring knowledge of “particular circumstances.” The proportional hazards assumption was tested using Schoenfeld residuals; no significant violations were detected (global test p = 0.38). The concordance index of 0.74 indicates acceptable discriminative ability. 6.4 Robustness Checks Two robustness specifications address potential concerns about the baseline model. The first re-estimates the Cox model treating acquired firms as failures rather than right-censored observations. This addresses the concern that some acquisitions may represent distressed exits rather than successful outcomes; reclassifying them as events tests whether the baseline results depend on the censoring assumption. The second specification replaces the binary single-provider indicator with a stricter measure: hard dependency , defined as reliance on a single API provider with no viable open-weight fallback deployed or tested. This narrows the treatment group from 91 firms (71.7%) to 64 firms (50.4%), isolating a more extreme form of plan-dependence. Table 3 reports the results alongside the baseline specification. Table 3 Robustness Specifications: Cox Proportional Hazard Model (1) Baseline (2) Acq. = Failure (3) Hard Dep. ln(Initial Capital) 0.71** 0.74** 0.69** Team Size 0.68** 0.72** 0.67** Technical Share 1.89** 1.76** 1.93** Pivot Count 0.54*** 0.61** 0.52*** Provider Dependency 2.47*** 2.19*** 3.12*** Sector FE Yes Yes Yes Events / N 23 / 127 34 / 127 23 / 127 Concordance 0.74 0.71 0.76 Notes : *** p < 0.01, ** p < 0.05. Hazard ratios reported. Specification (2) reclassifies acquisitions as failure events. Specification (3) replaces single-provider indicator with hard dependency (single API, no open-weight fallback). The results are stable across specifications in both sign and approximate magnitude. All coefficients retain significance and directionality. The provider dependency effect strengthens under the hard dependency definition (HR = 3.12, p < 0.01), consistent with the interpretation that deeper plan-dependence—not merely nominal single-sourcing—drives the risk mechanism. When acquisitions are reclassified as failures, the pivot count effect attenuates slightly (HR = 0.61 from 0.54) but remains significant, suggesting that the protective effect of pivoting is not an artifact of the censoring assumption. Parametric (Weibull) specifications yield qualitatively similar results across all three models. A note on external validity is warranted. The observation window (2022–2024, with status assessed in January 2025) is short, and the classification “AI-native” is necessarily a moving boundary as AI capabilities diffuse across industries. The findings should be understood as characterizing the early dynamics of a market in rapid formation. As the AI application ecosystem matures, entry barriers, provider dependency structures, and pivot patterns may evolve substantially. The theoretical framework developed in Sections 3 – 5 does not depend on the persistence of any specific empirical parameter; it depends on the structural claim that market coordination generates knowledge of a kind that non-market institutions cannot replicate. The empirical analysis provides evidence consistent with that claim in the current period, and future research should test whether these patterns persist as the market develops. 6.5 Case Studies in Institutional Diversity OpenAI began as a nonprofit research laboratory, transitioned to a “capped-profit” structure, and has subsequently moved toward a more conventional for-profit model. This institutional evolution reflects entrepreneurial experimentation with organizational form under genuine uncertainty about the appropriate governance structure for frontier AI development. The pivot from open-source research to proprietary model development—symbolized by the decision not to release the full GPT-4 model weights—represents a discovery about the relative economic value of open versus proprietary approaches to model distribution. Anthropic was founded around a particular theory of AI safety and has developed “constitutional AI” as a distinctive approach to model alignment. This represents institutional innovation in a domain—the governance of AI systems—where market discovery is essential because the relevant knowledge about effective governance mechanisms does not exist in advance. Anthropic’s corporate structure as a public benefit corporation reflects an entrepreneurial conjecture about how to attract talent and capital while maintaining commitment to safety-oriented research—a conjecture whose success depends on competitive market validation. Mistral , the French AI company, has pursued an open-weight strategy that deliberately contrasts with the proprietary approaches of OpenAI and (partially) Anthropic. By releasing model weights under permissive licenses, Mistral has discovered and exploited an entrepreneurial opportunity in serving the segment of the market that values local deployment, customization, and independence from API providers. The coexistence of these three strategies—proprietary (OpenAI), safety-focused benefit corporation (Anthropic), and open-weight (Mistral)—exemplifies the institutional diversity that competitive market processes generate and that no central authority could have anticipated or designed. 7. Conclusion: Structural Limits of Algorithmic Coordination This paper has argued that the rapid development of AI systems illuminates, rather than dissolves, the structural limits of non-market economic coordination. The derivative knowledge problem demonstrates that the impressive capabilities of foundation models are consequences of the richness of market-generated data, not evidence that markets can be dispensed with. The formal distinction between computational optimization and entrepreneurial discovery clarifies why the calculation problem is not a computational bottleneck amenable to technological solution: it is a categorical difference between operating within a known possibility space and expanding that space through institutionally embedded action. The irreconstructibility of opportunity-cost structure from non-market telemetry specifies precisely what is missing: not information quantity, but the counterfactual valuations that only competitive bidding reveals. Foundation models are powerful optimization engines, but the economic problem is not, at bottom, an optimization problem—it is a problem of generating the evaluative structure within which optimization becomes meaningful. The capital-theoretic analysis reveals that even in the most technologically advanced sectors, genuine uncertainty and plan-dependence make centralized investment coordination structurally fragile. The 2024 GPU shortage and subsequent overcapacity illustrate how decentralized entrepreneurial error correction—the boom-bust process that Austrian capital theory describes—operates in novel technological domains with the same dynamics visible in more traditional industries. The pattern of democratized disruption with oligopolistic infrastructure shows that market processes generate complex, multi-layered institutional structures whose emergent properties could not be specified in advance. The empirical evidence from 127 AI-native firms is consistent with these theoretical claims. The survival analysis reveals that provider diversification, adaptive pivoting, and the combination of technical and domain expertise predict firm survival—patterns that are robust to alternative censoring assumptions and stricter dependency definitions, and that reflect the importance of decentralized, contextual knowledge and adaptive learning in navigating genuine uncertainty. The case studies of OpenAI, Anthropic, and Mistral demonstrate that the competitive process generates institutional diversity—in governance structures, safety approaches, and openness strategies—that no planner could have anticipated. These findings also suggest directions for theoretical refinement. The coexistence of vigorous application-layer competition with infrastructure-layer concentration does not fit neatly into standard models of either equilibrating entrepreneurship or creative destruction. A more adequate theory of AI market structure will need to account for how entrepreneurial discovery operates within and across different layers of a production structure characterized by extreme complementarity and asset specificity. The market process is actively exploring this space; theory needs to keep pace. Finally, the analysis points toward a broader conclusion about the relationship between technology and institutions. AI is not an exogenous force that transforms economic institutions from the outside. It is itself a product of market coordination—an emergent property of the institutional order it is sometimes claimed to supersede. The algorithmic calculation problem is not a limitation that will be overcome by more powerful algorithms. It reflects the structural insight that animates the entire calculation debate: rational economic coordination requires institutions that generate knowledge, not merely institutions that process it. Declarations Author Contribution CW wrote the paper. References Anthropic. (2023). Toy models of superposition . Transformer Circuits Thread. Bastani, A. (2019). Fully automated luxury communism: A manifesto . Verso Books. Boettke, P. J. (1993). Why perestroika failed: The politics and economics of socialist transformation . Routledge. Boettke, P. J. (1998). Economic calculation: The Austrian contribution to political economy. Advances in Austrian Economics , 5 , 131–158. Buterin, V. (2023). My techno-optimism . Vitalik.eth Blog. Elhage, N., Nanda, N., Olsson, C., Henighan, T., Joseph, N., Mann, B., Olah, C., et al. (2021). A mathematical framework for transformer circuits . Transformer Circuits Thread. Hayek, F. A. (1937). Economics and knowledge. Economica , 4 (13), 33–54. Hayek, F. A. (1945). The use of knowledge in society. American Economic Review , 35 (4), 519–530. Hurwicz, L. (1960). Optimality and informational efficiency in resource allocation processes. In K. J. Arrow, S. Karlin, & P. Suppes (Eds.), Mathematical methods in the social sciences (pp. 27–46). Stanford University Press. Hurwicz, L. (1972). On informationally decentralized systems. In C. B. McGuire, & R. Radner (Eds.), Decision and organization (pp. 297–336). North-Holland. Kirzner, I. M. (1973). Competition and entrepreneurship . University of Chicago Press. Kirzner, I. M. (1984). The role of the entrepreneur in the economic system . Centre for Independent Studies. Kirzner, I. M. (1997). Entrepreneurial discovery and the competitive market process: An Austrian approach. Journal of Economic Literature , 35 (1), 60–85. Knight, F. H. (1921). Risk, uncertainty, and profit . Houghton Mifflin. Lachmann, L. M. (1956). Capital and its structure . Ludwig von Mises Institute. Lange, O. (1936). On the economic theory of socialism: Part one. Review of Economic Studies , 4 (1), 53–71. Lavoie, D. (1985). Rivalry and central planning: The socialist calculation debate reconsidered . Cambridge University Press. Lerner, A. P. (1944). The economics of control: Principles of welfare economics . Macmillan. von Mises, L. (1920). Die Wirtschaftsrechnung im sozialistischen Gemeinwesen. Archiv für Sozialwissenschaft und Sozialpolitik , 47 , 86–121. Morozov, E. (2019). Digital socialism? The calculation debate in the age of big data. New Left Review , 116/117 , 33–67. Neurath, O. (1921). Anti-Spengler . Callwey. Olah, C., Cammarata, N., Schubert, L., Goh, G., Petrov, M., & Carter, S. (2020). Zoom in: An introduction to circuits. Distill , 5(3). Phillips, L., & Rozworski, M. (2019). The people’s republic of Walmart: How the world’s biggest corporations are laying the groundwork for socialism . Verso Books. Shackle, G. L. S. (1972). Epistemics and economics: A critique of economic doctrines . Cambridge University Press. Shumailov, I., Shumaylov, Z., Zhao, Y., Gal, Y., Papernot, N., & Anderson, R. (2024). The curse of recursion: Training on generated data makes models forget. arXiv preprint arXiv:2305.17493. Srnicek, N. (2017). Platform capitalism . Polity. Yin, R. K. (2018). Case study research and applications: Design and methods (6th ed.). Sage. Additional Declarations No competing interests reported. Cite Share Download PDF Status: Posted Version 1 posted You are reading this latest preprint version Research Square lets you share your work early, gain feedback from the community, and start making changes to your manuscript prior to peer review in a journal. As a division of Research Square Company, we’re committed to making research communication faster, fairer, and more useful. We do this by developing innovative software and high quality services for the global research community. Our growing team is made up of researchers and industry professionals working together to solve the most critical problems facing scientific publishing. Also discoverable on Platform About Our Team In Review Editorial Policies Advisory Board Help Center Resources Author Services Accessibility API Access RSS feed Manage Cookie Preferences © Research Square 2026 | ISSN 2693-5015 (online) Privacy Policy Terms of Service Do Not Sell My Personal Information {"props":{"pageProps":{"initialData":{"identity":"rs-8856269","acceptedTermsAndConditions":true,"allowDirectSubmit":true,"archivedVersions":[],"articleType":"Research Article","associatedPublications":[],"authors":[{"id":594699054,"identity":"01bd4800-e02b-4f80-97cc-8ad2967e955b","order_by":0,"name":"Craig Wright","email":"data:image/png;base64,iVBORw0KGgoAAAANSUhEUgAAAZAAAAAyAQMAAABI0h/eAAAABlBMVEX///8AAABVwtN+AAAACXBIWXMAAA7EAAAOxAGVKw4bAAAAzklEQVRIiWNgGAWjYFAC5oMPEiqQBXgIamFLNnhwBkQTr4XHTPJhGyla+GekJRskzrNJ7JdvPvaBocaOweDMAfxaJG4kH3yQuC3NWLKNLXkGw7FkBoOzDfi1GEiDbNl2WM7gGI8x0HUHGAzOE3CYgXSOmUTinP889sf4PzMw/CNaS8MBOQM2HmYGxrYDhB0mcf9ZskHCsWRjiWNpxgyJfck8koS8z99z+ODDHzV2if3Nhx8zfPhmJ8d3JoGAy1BAAjEROQpGwSgYBaOAMAAA8w9AZF8b5eoAAAAASUVORK5CYII=","orcid":"","institution":"University of Exeter","correspondingAuthor":true,"prefix":"","firstName":"Craig","middleName":"","lastName":"Wright","suffix":""}],"badges":[],"createdAt":"2026-02-12 00:53:27","currentVersionCode":1,"declarations":"","doi":"10.21203/rs.3.rs-8856269/v1","doiUrl":"https://doi.org/10.21203/rs.3.rs-8856269/v1","draftVersion":[],"editorialEvents":[],"editorialNote":"","failedWorkflow":false,"files":[{"id":103505820,"identity":"256c7397-bde6-4989-a2c2-a2e3f872417f","added_by":"auto","created_at":"2026-02-26 13:33:09","extension":"pdf","order_by":0,"title":"","display":"","copyAsset":false,"role":"manuscript-pdf","size":907551,"visible":true,"origin":"","legend":"","description":"","filename":"manuscript.pdf","url":"https://assets-eu.researchsquare.com/files/rs-8856269/v1/973db5e7-f998-4fb1-9259-f1f7ea563fd1.pdf"}],"financialInterests":"No competing interests reported.","formattedTitle":"The Algorithmic Calculation Problem: Why Foundation Models Cannot Solve Socialist Planning","fulltext":[{"header":"1. Introduction","content":"\u003cp\u003eThe socialist calculation debate, inaugurated by Mises (\u003cspan citationid=\"CR19\" class=\"CitationRef\"\u003e1920\u003c/span\u003e) and refined through the interwar contributions of Hayek (\u003cspan citationid=\"CR7\" class=\"CitationRef\"\u003e1937\u003c/span\u003e, \u003cspan citationid=\"CR8\" class=\"CitationRef\"\u003e1945\u003c/span\u003e), Lange (\u003cspan citationid=\"CR16\" class=\"CitationRef\"\u003e1936\u003c/span\u003e), and Lerner (\u003cspan citationid=\"CR18\" class=\"CitationRef\"\u003e1944\u003c/span\u003e), addressed whether rational economic coordination is possible without market-generated prices. The Austrian position\u0026mdash;that prices generated through private property and competitive exchange carry epistemic content that cannot be replicated by a central authority\u0026mdash;was widely regarded as vindicated by the empirical record of planned economies (Boettke, \u003cspan citationid=\"CR3\" class=\"CitationRef\"\u003e1993\u003c/span\u003e; Lavoie, \u003cspan citationid=\"CR17\" class=\"CitationRef\"\u003e1985\u003c/span\u003e). Yet the emergence of large language models (LLMs) and generative AI systems has prompted a reconsideration. A growing literature spanning accelerationist Marxism (Bastani, \u003cspan citationid=\"CR2\" class=\"CitationRef\"\u003e2019\u003c/span\u003e; Srnicek, \u003cspan citationid=\"CR26\" class=\"CitationRef\"\u003e2017\u003c/span\u003e), digital socialism (Morozov, \u003cspan citationid=\"CR20\" class=\"CitationRef\"\u003e2019\u003c/span\u003e; Phillips and Rozworski, \u003cspan citationid=\"CR23\" class=\"CitationRef\"\u003e2019\u003c/span\u003e), and techno-utopianism (Buterin, \u003cspan citationid=\"CR5\" class=\"CitationRef\"\u003e2023\u003c/span\u003e) contends that modern AI has rendered the knowledge problem technologically obsolete.\u003c/p\u003e \u003cp\u003eThis paper argues that these claims rest on a fundamental misidentification of both what foundation models do and why the calculation problem is not merely a computational limitation. I develop three interconnected arguments. The first is epistemological: foundation models suffer from what I call the \u003cem\u003ederivative knowledge problem\u003c/em\u003e. These systems compress statistical regularities from data that was itself produced through market coordination. They do not\u0026mdash;and structurally cannot\u0026mdash;generate the forward-looking knowledge that entrepreneurial discovery produces under genuine uncertainty. I formalize this distinction by separating computational optimization within a known possibility space from entrepreneurial discovery that expands the possibility space itself, and I demonstrate that the specific epistemic object absent from non-market data is opportunity-cost structure\u0026mdash;the counterfactual valuations that competitive bidding reveals. The second argument is capital-theoretic: AI investment exhibits the radical heterogeneity and plan-dependence that Lachmann (\u003cspan citationid=\"CR15\" class=\"CitationRef\"\u003e1956\u003c/span\u003e) identified as characteristic of capital goods in a world of genuine uncertainty. The 2024 GPU shortage and subsequent overcapacity provide a textbook illustration. The third argument concerns market structure: AI simultaneously lowers barriers to entrepreneurial entry while concentrating essential complementary assets\u0026mdash;a pattern I term \u0026ldquo;democratized disruption with oligopolistic infrastructure.\u0026rdquo;\u003c/p\u003e \u003cp\u003eEmpirically, the paper draws on a proprietary dataset of 127 AI-native firms founded between 2022 and 2024. Using Cox proportional hazard models with robustness checks for alternative censoring assumptions and stricter dependency definitions, I examine how model-provider dependency, founding capital, team composition, and pivot frequency predict firm survival. The results are consistent with the theoretical framework: AI markets exhibit vigorous application-layer competition whose outcomes are shaped\u0026mdash;but not determined\u0026mdash;by infrastructure-layer market structure. Three case studies of frontier model providers (OpenAI, Anthropic, Mistral) illustrate the institutional diversity that market processes generate and that centralized planning could not anticipate.\u003c/p\u003e \u003cp\u003eThe paper proceeds as follows. Section \u003cspan refid=\"Sec2\" class=\"InternalRef\"\u003e2\u003c/span\u003e revisits the socialist calculation debate with attention to the specific claims made by proponents of AI-enabled planning. Section \u003cspan refid=\"Sec3\" class=\"InternalRef\"\u003e3\u003c/span\u003e develops the derivative knowledge problem, including a formal distinction between optimization and discovery and a demonstration that opportunity-cost structure is irreconstructible from non-market telemetry. Section \u003cspan refid=\"Sec9\" class=\"InternalRef\"\u003e4\u003c/span\u003e applies Lachmannian capital theory to AI investment patterns. Section \u003cspan refid=\"Sec12\" class=\"InternalRef\"\u003e5\u003c/span\u003e analyzes the paradox of democratized disruption. Section \u003cspan refid=\"Sec15\" class=\"InternalRef\"\u003e6\u003c/span\u003e presents empirical evidence, including survival analysis with robustness checks. Section \u003cspan refid=\"Sec21\" class=\"InternalRef\"\u003e7\u003c/span\u003e concludes by reflecting on the structural limits of algorithmic coordination.\u003c/p\u003e"},{"header":"2. The Calculation Debate Revisited: From Analog to Digital Utopianism","content":"\u003cp\u003eThe original Misesian challenge was not, as is sometimes supposed, merely that socialist planners would lack sufficient computational power to solve a system of simultaneous equations. As Lavoie (\u003cspan citationid=\"CR17\" class=\"CitationRef\"\u003e1985\u003c/span\u003e) and Boettke (\u003cspan citationid=\"CR4\" class=\"CitationRef\"\u003e1998\u003c/span\u003e) have persuasively argued, the knowledge at issue in the calculation debate is not the sort of propositional knowledge that can, even in principle, be codified and processed by a central authority. Hayek\u0026rsquo;s (\u003cspan citationid=\"CR8\" class=\"CitationRef\"\u003e1945\u003c/span\u003e) contribution was to clarify that the relevant knowledge is dispersed, tacit, contextual, and continuously changing\u0026mdash;\u0026ldquo;the knowledge of the particular circumstances of time and place\u0026rdquo; (p. 521). Market prices serve as epistemic signals precisely because they aggregate this otherwise inaccessible knowledge through the competitive process of entrepreneurial bidding.\u003c/p\u003e \u003cp\u003eContemporary advocates of AI planning have revived the computational interpretation of the calculation problem, arguing that what Hayek described as an information-processing challenge can now be overcome by sufficiently powerful machine learning systems. Morozov (\u003cspan citationid=\"CR20\" class=\"CitationRef\"\u003e2019\u003c/span\u003e) suggests that platform companies like Amazon and Alibaba already engage in forms of non-market coordination at enormous scale. Bastani (\u003cspan citationid=\"CR2\" class=\"CitationRef\"\u003e2019\u003c/span\u003e) contends that AI-driven automation will make the marginal cost of production approach zero, dissolving the need for price-mediated allocation. Phillips and Rozworski (\u003cspan citationid=\"CR23\" class=\"CitationRef\"\u003e2019\u003c/span\u003e) point to Walmart\u0026rsquo;s internal logistics algorithms as evidence that planned economies \u0026ldquo;already exist\u0026rdquo; within the capitalist firm.\u003c/p\u003e \u003cp\u003eThese arguments engage the calculation debate at the wrong level of abstraction. The question is not whether a sufficiently powerful algorithm can \u003cem\u003eprocess\u003c/em\u003e information\u0026mdash;it is whether the relevant information can \u003cem\u003eexist\u003c/em\u003e in the absence of market institutions. As Kirzner (\u003cspan citationid=\"CR12\" class=\"CitationRef\"\u003e1984\u003c/span\u003e) emphasized, entrepreneurial discovery is not a matter of optimizing within a known constraint set; it is a process of alertness to hitherto unnoticed opportunities. The knowledge that a particular resource combination would satisfy an as-yet-unarticulated consumer preference does not exist somewhere waiting to be aggregated. It comes into existence only through the entrepreneurial act itself, embedded within an institutional context of private property and residual claimancy. The distinction between processing existing information and generating new information is the conceptual pivot on which the entire debate turns\u0026mdash;and it is precisely this distinction that the AI planning literature fails to engage.\u003c/p\u003e"},{"header":"3. The Derivative Knowledge Problem","content":"\u003cp\u003eThe central claim of this paper is that foundation models\u0026mdash;large-scale neural networks trained on massive corpora of text, code, and other data\u0026mdash;suffer from a \u003cem\u003ederivative knowledge problem\u003c/em\u003e that makes them structurally incapable of substituting for market coordination. This is not a claim about computational limitations that might be overcome with larger models or better architectures. It is a claim about the epistemic status of the data on which these models are trained and, more fundamentally, about the logical distinction between two kinds of knowledge generation.\u003c/p\u003e \u003cdiv id=\"Sec4\" class=\"Section2\"\u003e \u003ch2\u003e3.1 The Epistemic Provenance of Training Data\u003c/h2\u003e \u003cp\u003eConsider what a large language model learns. A model like GPT-4 or Claude is trained on hundreds of billions of tokens of text drawn from the internet, books, academic papers, and code repositories. Recent advances in mechanistic interpretability (Elhage et al., \u003cspan citationid=\"CR6\" class=\"CitationRef\"\u003e2021\u003c/span\u003e; Olah et al., \u003cspan citationid=\"CR22\" class=\"CitationRef\"\u003e2020\u003c/span\u003e) have revealed that these models develop internal representations that capture complex semantic and relational structures. Anthropic\u0026rsquo;s (\u003cspan citationid=\"CR1\" class=\"CitationRef\"\u003e2023\u003c/span\u003e) work on \u0026ldquo;toy models of superposition\u0026rdquo; demonstrates that neural networks can learn to represent more features than they have dimensions, achieving remarkable information compression.\u003c/p\u003e \u003cp\u003eBut the crucial question, from the standpoint of economic calculation, is: \u003cem\u003ewhere does this information come from?\u003c/em\u003e The training data is not a neutral description of an objective reality. It is the textual residue of billions of human decisions, transactions, evaluations, and communications\u0026mdash;the overwhelming majority of which were made within market-coordinated institutional frameworks. When a language model \u0026ldquo;learns\u0026rdquo; that certain products are associated with certain price ranges, that particular business strategies tend to succeed or fail, or that specific resource combinations are valued by consumers, it is learning the condensed output of market processes. The model\u0026rsquo;s knowledge is \u003cem\u003ederivative\u003c/em\u003e of the market order in a strong sense: it is not merely informed by market outcomes but constituted by them.\u003c/p\u003e \u003cp\u003eMechanistic interpretability provides enabling evidence for the compression capacity of these models, though the provenance argument rests on a separate foundation. Elhage et al. (\u003cspan citationid=\"CR6\" class=\"CitationRef\"\u003e2021\u003c/span\u003e) demonstrate that transformer models learn attention patterns that track complex relational structures, and Olah et al. (\u003cspan citationid=\"CR22\" class=\"CitationRef\"\u003e2020\u003c/span\u003e) show that neural network features decompose into interpretable components corresponding to human-recognizable concepts. These findings establish that foundation models achieve genuine \u003cem\u003estructured compression\u003c/em\u003e rather than mere surface-level pattern matching. But the provenance claim does not rest on interpretability alone. It rests on the institutional origin of the corpora: when models process economic text\u0026mdash;financial reports, business analyses, product reviews\u0026mdash;the relational structures they compress (relative prices, comparative quality assessments, opportunity costs) are themselves outputs of market coordination. The model\u0026rsquo;s internal representations are not generic information-processing structures that happened to be trained on economic data; they are \u003cem\u003emarket-specific epistemic artifacts\u003c/em\u003e whose evaluative content derives from the competitive price system that generated the training distribution. Interpretability reveals \u003cem\u003ethat\u003c/em\u003e the model has learned rich structure; institutional analysis reveals \u003cem\u003ewhere\u003c/em\u003e that structure comes from.\u003c/p\u003e \u003c/div\u003e \u003cdiv id=\"Sec5\" class=\"Section2\"\u003e \u003ch2\u003e3.2 The Counterfactual Collapse\u003c/h2\u003e \u003cp\u003eThis derivative quality becomes decisive when we consider the counterfactual scenario that proponents of AI planning envision. Suppose a socialist planning authority were to deploy a foundation model trained on the entirety of the internet as of 2025. The model would contain rich knowledge about relative valuations, supply chain logistics, consumer preferences, and production techniques. But this knowledge was generated under market institutions. Remove those institutions and the model\u0026rsquo;s knowledge begins to decay in a manner that the model itself cannot correct.\u003c/p\u003e \u003cp\u003eThis is because the model cannot generate the \u003cem\u003enew\u003c/em\u003e knowledge that market processes continuously produce. Entrepreneurial discovery reveals previously unknown production possibilities, latent consumer preferences, and resource complementarities. These discoveries are not drawn from a pre-existing stock of information; they are genuinely creative acts that depend on the institutional incentive structures of market competition (Kirzner, \u003cspan citationid=\"CR13\" class=\"CitationRef\"\u003e1997\u003c/span\u003e). A foundation model can interpolate within the space of existing market-generated knowledge, but it cannot extrapolate into the space of knowledge that \u003cem\u003ewould have been\u003c/em\u003e generated by market processes that no longer exist.\u003c/p\u003e \u003cp\u003eThe analogy to Neurath\u0026rsquo;s ship is instructive, though it runs in the opposite direction from how it is typically deployed. Neurath (\u003cspan citationid=\"CR21\" class=\"CitationRef\"\u003e1921\u003c/span\u003e) argued that we must rebuild the ship of knowledge while sailing on it. The proponent of AI planning proposes something more radical: to sail on a ship built from the blueprints of a demolished vessel, without the institutional infrastructure that generated those blueprints.\u003c/p\u003e \u003c/div\u003e \u003cdiv id=\"Sec6\" class=\"Section2\"\u003e \u003ch2\u003e3.3 Optimization Versus Discovery: A Formal Distinction\u003c/h2\u003e \u003cp\u003eThe derivative knowledge argument can be sharpened by formalizing the distinction between two epistemic operations that the AI planning literature conflates: \u003cem\u003ecomputational optimization\u003c/em\u003e and \u003cem\u003eentrepreneurial discovery\u003c/em\u003e.\u003c/p\u003e \u003cp\u003eComputational optimization operates within a \u003cem\u003edefined possibility space\u003c/em\u003e. Let Ω denote the set of feasible states of the world, f(\u0026middot;) an objective function mapping states to values, and C a constraint set. Optimization consists of identifying x* \u0026isin; Ω such that f(x*)\u0026thinsp;\u0026ge;\u0026thinsp;f(x) for all x \u0026isin; C. Crucially, Ω, f, and C are given in advance. The computational challenge is combinatorial: the search space may be vast, and the function may be non-convex. But the \u003cem\u003eelements\u003c/em\u003e of the problem\u0026mdash;what goods exist, what preferences obtain, what production possibilities are available\u0026mdash;are specified ex ante. Foundation models excel at this kind of operation. They compress high-dimensional patterns from training data and can efficiently identify solutions within the possibility space that the data defines. This is what underlies their remarkable performance in code generation, text completion, and pattern recognition: they operate within the statistical manifold of their training distribution.\u003c/p\u003e \u003cp\u003eEntrepreneurial discovery, by contrast, \u003cem\u003eexpands the possibility space itself\u003c/em\u003e. The entrepreneur does not select x* from a known Ω; she identifies a previously unrecognized Ω\u0026prime; \u0026sup; Ω, or a previously unarticulated f\u0026prime; that revalues existing elements of Ω, or a novel production possibility that was not in C. Following Knight (\u003cspan citationid=\"CR14\" class=\"CitationRef\"\u003e1921\u003c/span\u003e), this is not risk (calculable probability over known outcomes) but genuine uncertainty (unforeseeable structural novelty). The introduction of the iPhone did not optimize within the existing mobile phone possibility space; it redefined what a mobile device was. Airbnb did not compute the optimal hotel allocation; it discovered that the category \u0026ldquo;accommodation\u0026rdquo; was larger than anyone had recognized.\u003c/p\u003e \u003cp\u003eThis distinction is not a matter of degree\u0026mdash;as though discovery were merely very difficult optimization. It is a categorical difference in epistemic operation. Optimization is a mapping \u003cem\u003ewithin\u003c/em\u003e a representation; discovery is a transformation \u003cem\u003eof\u003c/em\u003e the representation. A foundation model, however large, operates within the representation defined by its training data. It can recombine elements of that representation in novel ways\u0026mdash;and this recombinatory capacity is genuinely impressive\u0026mdash;but it cannot transcend the representation itself. The model has no access to what Shackle (\u003cspan citationid=\"CR24\" class=\"CitationRef\"\u003e1972\u003c/span\u003e) called \u0026ldquo;the void\u0026rdquo;: the space of possibilities that have not yet been imagined, let alone observed.\u003c/p\u003e \u003cp\u003eThis formal distinction has a direct implication for the calculation debate. A socialist planning authority equipped with foundation models would possess an extraordinarily powerful optimization engine\u0026mdash;capable of allocating known resources to known ends with impressive efficiency. But it would lack the discovery engine that market competition provides: the institutional mechanism through which unknown resources, unknown ends, and unknown production possibilities are continuously brought into existence. The planning problem is not that the planner cannot compute; it is that the planner cannot know what to compute \u003cem\u003eabout\u003c/em\u003e.\u003c/p\u003e \u003c/div\u003e \u003cdiv id=\"Sec7\" class=\"Section2\"\u003e \u003ch2\u003e3.4 Objections and Replies: Continuous Retraining and Decentralized Input\u003c/h2\u003e \u003cp\u003eThe strongest objection to the derivative knowledge argument takes the following form: could a planning authority not continuously retrain its foundation models on post-market data, thereby generating fresh knowledge without market institutions? Three variants of this objection deserve careful engagement.\u003c/p\u003e \u003cp\u003eThe first variant suggests that \u003cem\u003eadministrative data\u003c/em\u003e\u0026mdash;production reports, consumption surveys, logistics telemetry\u0026mdash;could substitute for market-generated data. The difficulty is that administrative data records \u003cem\u003ewhat happened\u003c/em\u003e rather than \u003cem\u003ewhat could have happened\u003c/em\u003e. Market prices encode opportunity costs: the value of the next-best alternative use of a resource, as assessed by the marginal bidder who did not obtain it. Administrative data records the allocation that was made but contains no information about the allocations that were foregone. Without opportunity cost information, a foundation model retrained on administrative data would learn the \u003cem\u003esurface pattern\u003c/em\u003e of economic activity without the \u003cem\u003eevaluative structure\u003c/em\u003e that makes coordination possible.\u003c/p\u003e \u003cp\u003eThe second variant proposes \u003cem\u003edecentralized data input\u003c/em\u003e\u0026mdash;allowing citizens to report preferences, needs, and local conditions directly to the planning system. This is the most sophisticated version of the objection because it addresses the Hayekian concern about dispersed knowledge. But it confronts two difficulties. First, agents lack incentives to reveal truthful preferences in the absence of the discipline imposed by budget constraints and competitive exchange. The incentive-compatibility problem that Hurwicz (\u003cspan citationid=\"CR9\" class=\"CitationRef\"\u003e1960\u003c/span\u003e, \u003cspan citationid=\"CR10\" class=\"CitationRef\"\u003e1972\u003c/span\u003e) formalized applies with full force: without a mechanism that makes truthful revelation individually rational, preference data will be systematically distorted. Second, and more fundamentally, many of the relevant preferences do not exist until they are evoked by entrepreneurial offerings. Consumers did not report a preference for smartphones, ride-sharing, or social media before these products existed. Entrepreneurial discovery is not the aggregation of pre-existing preferences; it is the co-creation of preferences and products through market interaction.\u003c/p\u003e \u003cp\u003eThe third variant invokes \u003cem\u003esynthetic data generation\u003c/em\u003e\u0026mdash;using AI models to simulate market dynamics and generate training data endogenously. This proposal is self-defeating. A model trained on its own outputs converges to the statistical regularities of its initial training distribution rather than discovering new information. The phenomenon of \u0026ldquo;model collapse\u0026rdquo; documented by Shumailov et al. (\u003cspan citationid=\"CR25\" class=\"CitationRef\"\u003e2024\u003c/span\u003e) shows that recursive self-training degrades model quality precisely because it eliminates the exogenous variation that contact with genuine human behavior provides. The market is not merely a source of data; it is the source of the \u003cem\u003esurprise\u003c/em\u003e\u0026mdash;the unforeseeable novelty\u0026mdash;on which productive model training depends.\u003c/p\u003e \u003c/div\u003e \u003cdiv id=\"Sec8\" class=\"Section2\"\u003e \u003ch2\u003e3.5 The Irreconstructibility of Opportunity-Cost Structure\u003c/h2\u003e \u003cp\u003eThe objections addressed above share a common assumption: that the knowledge deficit facing a post-market planning system is essentially a \u003cem\u003edata quantity\u003c/em\u003e problem\u0026mdash;that what planners lack is sufficient information about preferences, endowments, and technologies, and that better telemetry or larger datasets could close the gap. This subsection argues that the deficit is instead a \u003cem\u003edata structure\u003c/em\u003e problem. The specific epistemic object that non-market telemetry cannot reconstruct is \u003cem\u003eopportunity-cost structure\u003c/em\u003e: the pattern of counterfactual valuations that competitive bidding reveals and that no quantity of observational data can replicate.\u003c/p\u003e \u003cp\u003eConsider a concrete case. A market price of \u003cspan\u003e$\u003c/span\u003e12 per bushel for wheat does not merely record the transaction price at which wheat changed hands. It encodes: (i) the marginal buyer\u0026rsquo;s valuation, (ii) the marginal seller\u0026rsquo;s reservation price, (iii) the \u003cem\u003enext-best alternative use\u003c/em\u003e of the land, labor, water, and capital that produced the wheat, as assessed by the bidders who did not obtain those resources, and (iv) the valuation of the \u003cem\u003egoods not produced\u003c/em\u003e because resources were allocated to wheat rather than to their next-best use. Items (iii) and (iv) are the opportunity-cost structure. They are not observable in any administrative record because they describe \u003cem\u003estates of the world that did not obtain\u003c/em\u003e. They exist as information only because the competitive bidding process forced agents to reveal their valuations by bearing the cost of acting on them.\u003c/p\u003e \u003cp\u003eThe irreconstructibility claim can be stated formally. Let x denote the observed allocation and v(x) the market price vector. Let Ω denote the set of all feasible allocations, and let v(ω) for ω \u0026isin; Ω denote the valuations that \u003cem\u003ewould\u003c/em\u003e obtain under alternative allocations. Market prices embed information about the gradient \u0026nabla;v at x\u0026mdash;how valuations change at the margin for small reallocation\u0026mdash;because the competitive process is precisely a mechanism for eliciting marginal willingness to pay across alternative uses. Administrative telemetry observes only x and (at best) v(x): the allocation that was made and the prices at which it transacted. It is silent on \u0026nabla;v because the counterfactual allocations ω\u0026thinsp;\u0026ne;\u0026thinsp;x were never tried, and the agents who would have bid on those allocations were never forced to reveal their valuations.\u003c/p\u003e \u003cp\u003eThis is not a practical difficulty that better sensors or more comprehensive monitoring could overcome. It is a structural feature of non-market institutional environments. Opportunity costs are \u003cem\u003erevealed\u003c/em\u003e by competitive bidding; they are not \u003cem\u003eobserved\u003c/em\u003e by administrative recording. The distinction maps directly onto the mechanism-design literature: as Hurwicz (\u003cspan citationid=\"CR10\" class=\"CitationRef\"\u003e1972\u003c/span\u003e) demonstrated, a social choice function achieves informational efficiency only if the institutional mechanism incentivizes agents to reveal private valuations. Without such a mechanism, the information does not enter the system\u0026mdash;not because agents choose to withhold it, but because the institutional structure provides no occasion for its articulation.\u003c/p\u003e \u003cp\u003eThe implication for foundation models is decisive. A model retrained on non-market data would learn the \u003cem\u003esurface texture\u003c/em\u003e of economic activity\u0026mdash;what was produced, where, when, in what quantities\u0026mdash;but would lack the \u003cem\u003eevaluative depth\u003c/em\u003e that opportunity-cost structure provides. It would know that 10,000 bushels of wheat were shipped from Region A to Region B, but not whether the land in Region A would have been more productively employed growing soybeans, the labor more valuably engaged in manufacturing, or the transport capacity more efficiently allocated to a different commodity route. Without this evaluative depth, the model cannot assess whether \u003cem\u003eany given allocation is better or worse than its alternatives\u003c/em\u003e\u0026mdash;which is to say, it cannot perform economic calculation in the sense that Mises (\u003cspan citationid=\"CR19\" class=\"CitationRef\"\u003e1920\u003c/span\u003e) identified as the core function of market prices. The algorithmic calculation problem is not that the algorithm lacks data; it is that the data it can access is observationally thin on precisely the dimension\u0026mdash;counterfactual valuation\u0026mdash;that economic coordination requires.\u003c/p\u003e \u003c/div\u003e"},{"header":"4. Capital Heterogeneity and Plan-Dependence in AI Markets","content":"\u003cp\u003eLachmann\u0026rsquo;s (\u003cspan citationid=\"CR15\" class=\"CitationRef\"\u003e1956\u003c/span\u003e) capital theory emphasizes two properties of capital goods that distinguish the Austrian approach from neoclassical production theory: heterogeneity and plan-dependence. Capital goods are not interchangeable units of a homogeneous aggregate but specific, multi-attribute resources whose economic significance depends on the entrepreneurial plans within which they are embedded. A blast furnace is not merely \u0026ldquo;capital\u0026rdquo;; it is a highly specific asset whose value depends on complementary inputs, expected future demand, the availability of skilled labor, and the broader production structure within which it operates.\u003c/p\u003e \u003cp\u003eAI capital exhibits these Lachmannian properties in extreme form. A GPU cluster configured for training large language models represents an enormous fixed investment whose economic value depends critically on complementary assets: training data (which may be proprietary), model architectures (which embody tacit engineering knowledge), human capital (ML researchers with highly specific expertise), and downstream applications (which may not yet exist). The plan-dependence of AI capital creates vulnerabilities characteristic of Austrian business cycle dynamics.\u003c/p\u003e \u003cdiv id=\"Sec10\" class=\"Section2\"\u003e \u003ch2\u003e4.1 The 2024 GPU Shortage as Malinvestment\u003c/h2\u003e \u003cp\u003eThe GPU shortage that emerged in late 2023 and persisted through much of 2024 provides a compelling illustration. Following the release of ChatGPT in November 2022, demand for NVIDIA\u0026rsquo;s H100 and A100 GPUs surged dramatically. Major technology companies placed orders worth billions of dollars. Cloud computing providers expanded GPU capacity. Startups acquired computing resources at significant premiums. By mid-2024, the secondary market for H100 GPUs had reached prices well above NVIDIA\u0026rsquo;s list price.\u003c/p\u003e \u003cp\u003eFrom the standpoint of capital theory, this pattern is recognizable as a boom-phase elongation of the production structure. The expectation of future AI revenues\u0026mdash;fueled by the capabilities demonstrated by GPT-4 and subsequent models\u0026mdash;led to massive investment in higher-order capital goods (GPU manufacturing capacity, data center construction, power infrastructure) predicated on assumptions about future consumer valuations that may or may not materialize. The subsequent emergence of overcapacity concerns in late 2024 and early 2025, as some AI startups failed to find product-market fit and hyperscaler capital expenditure growth rates decelerated, suggests the early stages of the Hayekian \u0026ldquo;reveal\u0026rdquo;\u0026mdash;the point at which malinvestment becomes apparent as entrepreneurial plans prove inconsistent with underlying consumer time preferences.\u003c/p\u003e \u003c/div\u003e \u003cdiv id=\"Sec11\" class=\"Section2\"\u003e \u003ch2\u003e4.2 Complementarity and Specificity in AI Production\u003c/h2\u003e \u003cp\u003eThe plan-dependence of AI capital is amplified by complex complementarity relationships. A foundation model\u0026rsquo;s value depends not only on the compute used to train it but on the quality and composition of its training data, the fine-tuning applied for specific applications, the deployment infrastructure supporting inference, and the business model connecting the model\u0026rsquo;s capabilities to willingness-to-pay. Each of these complementary assets exhibits its own form of specificity and plan-dependence, creating a web of interdependent entrepreneurial expectations.\u003c/p\u003e \u003cp\u003eThis web of complementarities means that errors in any component can cascade through the production structure. A training data strategy that proves legally vulnerable (as several AI companies have discovered in copyright litigation), an architectural choice superseded by a rival\u0026rsquo;s innovation, or a business model that fails to capture sufficient value\u0026mdash;any of these can render enormous complementary investments economically obsolete. The capital loss is not merely quantitative but qualitative: the specific configuration of assets loses its economic rationale, and the assets cannot simply be \u0026ldquo;recombined\u0026rdquo; without entrepreneurial judgment about alternative plan structures.\u003c/p\u003e \u003c/div\u003e"},{"header":"5. Democratized Disruption with Oligopolistic Infrastructure","content":"\u003cp\u003eKirzner\u0026rsquo;s (\u003cspan citationid=\"CR11\" class=\"CitationRef\"\u003e1973\u003c/span\u003e) theory of entrepreneurial discovery posits that alert entrepreneurs, by noticing and exploiting previously overlooked price discrepancies, drive the market toward coordination. The competitive process is fundamentally equilibrating: entrepreneurial profit opportunities exist because the market is not yet fully coordinated, and entrepreneurial action tends to eliminate these opportunities. This framework has provided an influential theoretical foundation for understanding how markets generate and disseminate knowledge.\u003c/p\u003e \u003cp\u003eThe AI market complicates this picture in ways that call for refinement rather than abandonment. AI lowers barriers to entrepreneurial entry at the application layer while simultaneously concentrating essential complementary assets at the infrastructure layer. This creates what I term \u0026ldquo;democratized disruption with oligopolistic infrastructure\u0026rdquo;\u0026mdash;a market structure in which the entrepreneurial discovery process operates vigorously at one level of the production structure while being constrained by market power at another.\u003c/p\u003e \u003cdiv id=\"Sec13\" class=\"Section2\"\u003e \u003ch2\u003e5.1 Application-Layer Entrepreneurship\u003c/h2\u003e \u003cp\u003eThe release of foundation model APIs\u0026mdash;beginning with OpenAI\u0026rsquo;s API in 2020 and expanding rapidly through 2023\u0026ndash;2025\u0026mdash;has dramatically lowered the cost of AI-powered entrepreneurship. An entrepreneur with a domain-specific insight can now build an AI application without training their own model, managing GPU infrastructure, or employing a team of machine learning researchers. The result has been an explosion of AI-native startups: our dataset of 127 firms founded between 2022 and 2024 encompasses applications ranging from legal document analysis to agricultural advisory services to personalized education. The median founding team size is 3.2 people, and the median initial capital requirement is approximately \u003cspan\u003e$\u003c/span\u003e150,000\u0026mdash;a fraction of the cost that would have been required five years earlier.\u003c/p\u003e \u003cp\u003eThis application-layer entrepreneurship is quintessentially Kirznerian. These entrepreneurs are alert to specific, localized opportunities to apply general-purpose AI capabilities in particular market niches. They possess contextual knowledge\u0026mdash;about the pain points of legal professionals, the information needs of smallholder farmers, the learning styles of different student populations\u0026mdash;that cannot be centrally aggregated. Their competitive entry generates new knowledge about the value of AI capabilities in specific applications, knowledge that did not exist before the entrepreneurial act.\u003c/p\u003e \u003c/div\u003e \u003cdiv id=\"Sec14\" class=\"Section2\"\u003e \u003ch2\u003e5.2 Infrastructure-Layer Concentration\u003c/h2\u003e \u003cp\u003eAt the infrastructure layer, however, a different dynamic obtains. The training of frontier foundation models requires capital investments on the order of \u003cspan\u003e$\u003c/span\u003e100\u0026nbsp;million to \u003cspan\u003e$\u003c/span\u003e1\u0026nbsp;billion, access to proprietary or carefully curated datasets of enormous scale, and teams of researchers with highly specialized expertise. As of early 2025, the number of organizations capable of training genuinely frontier models can be counted on two hands: OpenAI, Anthropic, Google DeepMind, Meta, Mistral, and a small number of others.\u003c/p\u003e \u003cp\u003eThis concentration does not arise from market failure in any conventional sense. It reflects genuine economies of scale in model training, the accumulated specificity of complementary assets, and the path-dependent nature of organizational knowledge. But it does mean that application-layer entrepreneurs are dependent on infrastructure-layer incumbents for access to foundational capabilities. The relationship between these two layers\u0026mdash;competitive and entrepreneurial at the application level, oligopolistic and capital-intensive at the infrastructure level\u0026mdash;creates dynamics that existing theoretical frameworks do not fully address.\u003c/p\u003e \u003cp\u003eThis layered structure has an important implication for the calculation debate. Even if a planning authority could replicate the infrastructure layer\u0026mdash;assembling the compute, data, and talent required to train frontier models\u0026mdash;it could not replicate the application-layer discovery process. The thousands of entrepreneurs experimenting with AI applications are conducting a massively parallel, decentralized search across the space of possible uses, guided by local knowledge and disciplined by profit and loss. No central authority could design this search process because the search space itself is unknown in advance.\u003c/p\u003e \u003c/div\u003e"},{"header":"6. Empirical Evidence: AI Firm Dynamics and Institutional Diversity","content":"\u003cdiv id=\"Sec16\" class=\"Section2\"\u003e \u003ch2\u003e6.1 Data and Methodology\u003c/h2\u003e \u003cp\u003eThe empirical component of this paper draws on a proprietary dataset of 127 AI-native firms founded between January 2022 and December 2024. Firms were identified through a systematic search of Crunchbase, PitchBook, and Y Combinator\u0026rsquo;s public portfolio, filtered to include only firms whose primary product or service relies on foundation model capabilities (either via API access or open-weight deployment). The sample excludes firms whose AI use is incidental to a non-AI core business, as well as infrastructure-layer firms (GPU cloud providers, model training companies) to maintain analytical focus on application-layer entrepreneurship.\u003c/p\u003e \u003cp\u003eFor each firm, I recorded founding date, initial capital raised (seed and pre-seed), founding team size, team composition (proportion of technical versus domain-specialist founders), primary use case (categorized into seven sectors: legal, healthcare, education, financial services, creative tools, enterprise software, and other), model provider dependency (primary API provider or open-weight model), number of significant pivots (defined as a change in primary use case, target market, or model provider), and operational status as of January 2025 (active, ceased operations, or acquired). Data were collected from Crunchbase records, SEC filings, press releases, company websites, and 34 semi-structured interviews with founders and early-stage investors conducted between June and November 2024.\u003c/p\u003e \u003cp\u003eSelection bias is a potential concern: firms that failed very early may not appear in Crunchbase or PitchBook records. To mitigate this, I supplemented the database search with Y Combinator batch records (which include failed firms) and cross-referenced with TechCrunch\u0026rsquo;s shutdown tracker. Survivorship bias in the interview sample is addressed by including 8 interviews with founders of firms that had ceased operations by the interview date. The case study methodology follows Yin (\u003cspan citationid=\"CR27\" class=\"CitationRef\"\u003e2018\u003c/span\u003e); case studies of OpenAI, Anthropic, and Mistral were constructed from public corporate communications, regulatory filings, published research, and 12 background interviews conducted under Chatham House rules.\u003c/p\u003e \u003c/div\u003e \u003cdiv id=\"Sec17\" class=\"Section2\"\u003e \u003ch2\u003e6.2 Summary Statistics\u003c/h2\u003e \u003cp\u003eTable\u0026nbsp;\u003cspan refid=\"Tab1\" class=\"InternalRef\"\u003e1\u003c/span\u003e presents summary statistics for the 127 firms in the sample, disaggregated by operational status as of January 2025.\u003c/p\u003e \u003cp\u003e \u003cdiv class=\"gridtable\"\u003e\u003ctable float=\"Yes\" id=\"Tab1\" border=\"1\"\u003e \u003ccaption language=\"En\"\u003e \u003cdiv class=\"CaptionNumber\"\u003eTable 1\u003c/div\u003e \u003cdiv class=\"CaptionContent\"\u003e \u003cp\u003eSummary Statistics by Operational Status (N\u0026thinsp;=\u0026thinsp;127)\u003c/p\u003e \u003c/div\u003e \u003c/caption\u003e \u003ccolgroup cols=\"5\"\u003e \u003cdiv align=\"left\" class=\"colspec\" colname=\"c1\" colnum=\"1\"\u003e\u003c/div\u003e \u003cdiv align=\"left\" class=\"colspec\" colname=\"c2\" colnum=\"2\"\u003e\u003c/div\u003e \u003cdiv align=\"left\" class=\"colspec\" colname=\"c3\" colnum=\"3\"\u003e\u003c/div\u003e \u003cdiv align=\"left\" class=\"colspec\" colname=\"c4\" colnum=\"4\"\u003e\u003c/div\u003e \u003cdiv align=\"left\" class=\"colspec\" colname=\"c5\" colnum=\"5\"\u003e\u003c/div\u003e \u003cthead\u003e \u003ctr\u003e \u003cth align=\"left\" colname=\"c1\"\u003e \u003cp\u003eVariable\u003c/p\u003e \u003c/th\u003e \u003cth align=\"left\" colname=\"c2\"\u003e \u003cp\u003eAll Firms\u003c/p\u003e \u003c/th\u003e \u003cth align=\"left\" colname=\"c3\"\u003e \u003cp\u003eActive\u003c/p\u003e \u003c/th\u003e \u003cth align=\"left\" colname=\"c4\"\u003e \u003cp\u003eCeased\u003c/p\u003e \u003c/th\u003e \u003cth align=\"left\" colname=\"c5\"\u003e \u003cp\u003eAcquired\u003c/p\u003e \u003c/th\u003e \u003c/tr\u003e \u003c/thead\u003e \u003ctbody\u003e \u003ctr\u003e \u003ctd align=\"left\" colname=\"c1\"\u003e \u003cp\u003e\u003cem\u003eN\u003c/em\u003e\u003c/p\u003e \u003c/td\u003e \u003ctd align=\"left\" colname=\"c2\"\u003e \u003cp\u003e127\u003c/p\u003e \u003c/td\u003e \u003ctd align=\"left\" colname=\"c3\"\u003e \u003cp\u003e93\u003c/p\u003e \u003c/td\u003e \u003ctd align=\"left\" colname=\"c4\"\u003e \u003cp\u003e23\u003c/p\u003e \u003c/td\u003e \u003ctd align=\"left\" colname=\"c5\"\u003e \u003cp\u003e11\u003c/p\u003e \u003c/td\u003e \u003c/tr\u003e \u003ctr\u003e \u003ctd align=\"left\" colname=\"c1\"\u003e \u003cp\u003e\u003cem\u003eInitial capital ($K), median\u003c/em\u003e\u003c/p\u003e \u003c/td\u003e \u003ctd align=\"left\" colname=\"c2\"\u003e \u003cp\u003e150\u003c/p\u003e \u003c/td\u003e \u003ctd align=\"left\" colname=\"c3\"\u003e \u003cp\u003e185\u003c/p\u003e \u003c/td\u003e \u003ctd align=\"left\" colname=\"c4\"\u003e \u003cp\u003e90\u003c/p\u003e \u003c/td\u003e \u003ctd align=\"left\" colname=\"c5\"\u003e \u003cp\u003e210\u003c/p\u003e \u003c/td\u003e \u003c/tr\u003e \u003ctr\u003e \u003ctd align=\"left\" colname=\"c1\"\u003e \u003cp\u003e\u003cem\u003eFounding team size, mean\u003c/em\u003e\u003c/p\u003e \u003c/td\u003e \u003ctd align=\"left\" colname=\"c2\"\u003e \u003cp\u003e3.2\u003c/p\u003e \u003c/td\u003e \u003ctd align=\"left\" colname=\"c3\"\u003e \u003cp\u003e3.5\u003c/p\u003e \u003c/td\u003e \u003ctd align=\"left\" colname=\"c4\"\u003e \u003cp\u003e2.4\u003c/p\u003e \u003c/td\u003e \u003ctd align=\"left\" colname=\"c5\"\u003e \u003cp\u003e3.1\u003c/p\u003e \u003c/td\u003e \u003c/tr\u003e \u003ctr\u003e \u003ctd align=\"left\" colname=\"c1\"\u003e \u003cp\u003e\u003cem\u003eTechnical founders (%), mean\u003c/em\u003e\u003c/p\u003e \u003c/td\u003e \u003ctd align=\"left\" colname=\"c2\"\u003e \u003cp\u003e62.4\u003c/p\u003e \u003c/td\u003e \u003ctd align=\"left\" colname=\"c3\"\u003e \u003cp\u003e58.7\u003c/p\u003e \u003c/td\u003e \u003ctd align=\"left\" colname=\"c4\"\u003e \u003cp\u003e78.3\u003c/p\u003e \u003c/td\u003e \u003ctd align=\"left\" colname=\"c5\"\u003e \u003cp\u003e63.6\u003c/p\u003e \u003c/td\u003e \u003c/tr\u003e \u003ctr\u003e \u003ctd align=\"left\" colname=\"c1\"\u003e \u003cp\u003e\u003cem\u003ePivots, mean\u003c/em\u003e\u003c/p\u003e \u003c/td\u003e \u003ctd align=\"left\" colname=\"c2\"\u003e \u003cp\u003e0.81\u003c/p\u003e \u003c/td\u003e \u003ctd align=\"left\" colname=\"c3\"\u003e \u003cp\u003e0.94\u003c/p\u003e \u003c/td\u003e \u003ctd align=\"left\" colname=\"c4\"\u003e \u003cp\u003e0.48\u003c/p\u003e \u003c/td\u003e \u003ctd align=\"left\" colname=\"c5\"\u003e \u003cp\u003e0.55\u003c/p\u003e \u003c/td\u003e \u003c/tr\u003e \u003ctr\u003e \u003ctd align=\"left\" colname=\"c1\"\u003e \u003cp\u003e\u003cem\u003eSingle-provider dependent (%)\u003c/em\u003e\u003c/p\u003e \u003c/td\u003e \u003ctd align=\"left\" colname=\"c2\"\u003e \u003cp\u003e71.7\u003c/p\u003e \u003c/td\u003e \u003ctd align=\"left\" colname=\"c3\"\u003e \u003cp\u003e64.5\u003c/p\u003e \u003c/td\u003e \u003ctd align=\"left\" colname=\"c4\"\u003e \u003cp\u003e91.3\u003c/p\u003e \u003c/td\u003e \u003ctd align=\"left\" colname=\"c5\"\u003e \u003cp\u003e81.8\u003c/p\u003e \u003c/td\u003e \u003c/tr\u003e \u003ctr\u003e \u003ctd align=\"left\" colname=\"c1\"\u003e \u003cp\u003e\u003cem\u003eOpenAI as primary provider (%)\u003c/em\u003e\u003c/p\u003e \u003c/td\u003e \u003ctd align=\"left\" colname=\"c2\"\u003e \u003cp\u003e58.3\u003c/p\u003e \u003c/td\u003e \u003ctd align=\"left\" colname=\"c3\"\u003e \u003cp\u003e53.8\u003c/p\u003e \u003c/td\u003e \u003ctd align=\"left\" colname=\"c4\"\u003e \u003cp\u003e73.9\u003c/p\u003e \u003c/td\u003e \u003ctd align=\"left\" colname=\"c5\"\u003e \u003cp\u003e63.6\u003c/p\u003e \u003c/td\u003e \u003c/tr\u003e \u003c/tbody\u003e \u003c/colgroup\u003e \u003c/table\u003e\u003c/div\u003e \u003c/p\u003e \u003cp\u003eSeveral patterns are notable. Ceased firms had significantly smaller founding teams (mean 2.4 versus 3.5 for active firms, p\u0026thinsp;\u0026lt;\u0026thinsp;0.01, two-tailed t-test) and lower initial capital (median \u003cspan\u003e$\u003c/span\u003e90K versus \u003cspan\u003e$\u003c/span\u003e185K, p\u0026thinsp;\u0026lt;\u0026thinsp;0.05, Mann-Whitney U). Ceased firms were substantially more likely to be dependent on a single model provider (91.3% versus 64.5%), suggesting that provider diversification may function as a form of entrepreneurial hedging against infrastructure-layer risk. Interestingly, active firms had a higher mean number of pivots (0.94 versus 0.48 for ceased firms), consistent with the interpretation that pivoting reflects adaptive entrepreneurial learning rather than failure.\u003c/p\u003e \u003c/div\u003e \u003cdiv id=\"Sec18\" class=\"Section2\"\u003e \u003ch2\u003e6.3 Survival Analysis\u003c/h2\u003e \u003cp\u003eTo examine the determinants of AI firm survival more rigorously, I estimate a Cox proportional hazard model with firm failure (cessation of operations) as the event of interest. The observation window runs from each firm\u0026rsquo;s founding date to January 2025 or the date of cessation, whichever comes first. Acquired firms are treated as right-censored. The model takes the form:\u003c/p\u003e \u003cp\u003e \u003cem\u003eh(t | X)\u0026thinsp;=\u0026thinsp;h₀(t) exp(β₁ ln(Capital) + β₂ TeamSize\u0026thinsp;+\u0026thinsp;β₃ TechShare\u0026thinsp;+\u0026thinsp;β₄ Pivots\u0026thinsp;+\u0026thinsp;β₅ SingleProvider\u0026thinsp;+\u0026thinsp;γ Sector)\u003c/em\u003e \u003c/p\u003e \u003cp\u003ewhere h₀(t) is the baseline hazard, Capital is initial funding in thousands of dollars (log-transformed), TeamSize is the number of founders, TechShare is the proportion of founders with technical backgrounds, Pivots is the cumulative count of significant pivots, SingleProvider is a binary indicator for dependence on a single model provider, and Sector is a vector of sector fixed effects. Table\u0026nbsp;\u003cspan refid=\"Tab2\" class=\"InternalRef\"\u003e2\u003c/span\u003e reports the results.\u003c/p\u003e \u003cp\u003e \u003cdiv class=\"gridtable\"\u003e\u003ctable float=\"Yes\" id=\"Tab2\" border=\"1\"\u003e \u003ccaption language=\"En\"\u003e \u003cdiv class=\"CaptionNumber\"\u003eTable 2\u003c/div\u003e \u003cdiv class=\"CaptionContent\"\u003e \u003cp\u003eCox Proportional Hazard Model: Determinants of AI Firm Failure\u003c/p\u003e \u003c/div\u003e \u003c/caption\u003e \u003ccolgroup cols=\"4\"\u003e \u003cdiv align=\"left\" class=\"colspec\" colname=\"c1\" colnum=\"1\"\u003e\u003c/div\u003e \u003cdiv align=\"left\" class=\"colspec\" colname=\"c2\" colnum=\"2\"\u003e\u003c/div\u003e \u003cdiv align=\"left\" class=\"colspec\" colname=\"c3\" colnum=\"3\"\u003e\u003c/div\u003e \u003cdiv align=\"left\" class=\"colspec\" colname=\"c4\" colnum=\"4\"\u003e\u003c/div\u003e \u003cthead\u003e \u003ctr\u003e \u003cth align=\"left\" colname=\"c1\"\u003e \u003cp\u003eVariable\u003c/p\u003e \u003c/th\u003e \u003cth align=\"left\" colname=\"c2\"\u003e \u003cp\u003eHazard Ratio\u003c/p\u003e \u003c/th\u003e \u003cth align=\"left\" colname=\"c3\"\u003e \u003cp\u003eStd. Error\u003c/p\u003e \u003c/th\u003e \u003cth align=\"left\" colname=\"c4\"\u003e \u003cp\u003ep-value\u003c/p\u003e \u003c/th\u003e \u003c/tr\u003e \u003c/thead\u003e \u003ctbody\u003e \u003ctr\u003e \u003ctd align=\"left\" colname=\"c1\"\u003e \u003cp\u003e\u003cem\u003eln(Initial Capital)\u003c/em\u003e\u003c/p\u003e \u003c/td\u003e \u003ctd align=\"left\" colname=\"c2\"\u003e \u003cp\u003e0.71\u003c/p\u003e \u003c/td\u003e \u003ctd align=\"left\" colname=\"c3\"\u003e \u003cp\u003e0.12\u003c/p\u003e \u003c/td\u003e \u003ctd align=\"left\" colname=\"c4\"\u003e \u003cp\u003e0.031**\u003c/p\u003e \u003c/td\u003e \u003c/tr\u003e \u003ctr\u003e \u003ctd align=\"left\" colname=\"c1\"\u003e \u003cp\u003e\u003cem\u003eTeam Size\u003c/em\u003e\u003c/p\u003e \u003c/td\u003e \u003ctd align=\"left\" colname=\"c2\"\u003e \u003cp\u003e0.68\u003c/p\u003e \u003c/td\u003e \u003ctd align=\"left\" colname=\"c3\"\u003e \u003cp\u003e0.14\u003c/p\u003e \u003c/td\u003e \u003ctd align=\"left\" colname=\"c4\"\u003e \u003cp\u003e0.018**\u003c/p\u003e \u003c/td\u003e \u003c/tr\u003e \u003ctr\u003e \u003ctd align=\"left\" colname=\"c1\"\u003e \u003cp\u003e\u003cem\u003eTechnical Share\u003c/em\u003e\u003c/p\u003e \u003c/td\u003e \u003ctd align=\"left\" colname=\"c2\"\u003e \u003cp\u003e1.89\u003c/p\u003e \u003c/td\u003e \u003ctd align=\"left\" colname=\"c3\"\u003e \u003cp\u003e0.41\u003c/p\u003e \u003c/td\u003e \u003ctd align=\"left\" colname=\"c4\"\u003e \u003cp\u003e0.044**\u003c/p\u003e \u003c/td\u003e \u003c/tr\u003e \u003ctr\u003e \u003ctd align=\"left\" colname=\"c1\"\u003e \u003cp\u003e\u003cem\u003ePivot Count\u003c/em\u003e\u003c/p\u003e \u003c/td\u003e \u003ctd align=\"left\" colname=\"c2\"\u003e \u003cp\u003e0.54\u003c/p\u003e \u003c/td\u003e \u003ctd align=\"left\" colname=\"c3\"\u003e \u003cp\u003e0.19\u003c/p\u003e \u003c/td\u003e \u003ctd align=\"left\" colname=\"c4\"\u003e \u003cp\u003e0.009***\u003c/p\u003e \u003c/td\u003e \u003c/tr\u003e \u003ctr\u003e \u003ctd align=\"left\" colname=\"c1\"\u003e \u003cp\u003e\u003cem\u003eSingle Provider Dep.\u003c/em\u003e\u003c/p\u003e \u003c/td\u003e \u003ctd align=\"left\" colname=\"c2\"\u003e \u003cp\u003e2.47\u003c/p\u003e \u003c/td\u003e \u003ctd align=\"left\" colname=\"c3\"\u003e \u003cp\u003e0.63\u003c/p\u003e \u003c/td\u003e \u003ctd align=\"left\" colname=\"c4\"\u003e \u003cp\u003e0.003***\u003c/p\u003e \u003c/td\u003e \u003c/tr\u003e \u003ctr\u003e \u003ctd align=\"left\" colname=\"c1\"\u003e \u003cp\u003e\u003cem\u003eSector FE\u003c/em\u003e\u003c/p\u003e \u003c/td\u003e \u003ctd align=\"left\" colname=\"c2\"\u003e \u003cp\u003e\u0026mdash;\u003c/p\u003e \u003c/td\u003e \u003ctd align=\"left\" colname=\"c3\"\u003e \u003cp\u003e\u0026mdash;\u003c/p\u003e \u003c/td\u003e \u003ctd align=\"left\" colname=\"c4\"\u003e \u003cp\u003eJoint: 0.127\u003c/p\u003e \u003c/td\u003e \u003c/tr\u003e \u003ctr\u003e \u003ctd align=\"left\" colname=\"c1\"\u003e \u003cp\u003e\u003cem\u003eN\u0026thinsp;=\u0026thinsp;127; Events\u0026thinsp;=\u0026thinsp;23; Concordance\u0026thinsp;=\u0026thinsp;0.74\u003c/em\u003e\u003c/p\u003e \u003c/td\u003e \u003ctd align=\"left\" colname=\"c2\"\u003e\u0026nbsp;\u003c/td\u003e \u003ctd align=\"left\" colname=\"c3\"\u003e\u0026nbsp;\u003c/td\u003e \u003ctd align=\"left\" colname=\"c4\"\u003e\u0026nbsp;\u003c/td\u003e \u003c/tr\u003e \u003c/tbody\u003e \u003c/colgroup\u003e \u003ctfoot\u003e \u003ctr\u003e\u003ctd colspan=\"4\"\u003e\u003cem\u003eNotes\u003c/em\u003e: *** p\u0026thinsp;\u0026lt;\u0026thinsp;0.01, ** p\u0026thinsp;\u0026lt;\u0026thinsp;0.05. Hazard ratios below 1 indicate reduced failure risk.\u003c/td\u003e\u003c/tr\u003e \u003c/tfoot\u003e \u003c/table\u003e\u003c/div\u003e \u003c/p\u003e \u003cp\u003eThe results reveal several patterns consistent with the theoretical framework. Single-provider dependency is the strongest predictor of failure: firms dependent on a single model provider face a hazard ratio of 2.47 (p\u0026thinsp;\u0026lt;\u0026thinsp;0.003), indicating roughly 2.5 times the failure risk of firms with provider diversification. This is consistent with the capital-theoretic argument about plan-dependence: firms whose entrepreneurial plans are tightly coupled to a single infrastructure provider face elevated risk from API pricing changes, capability shifts, or provider strategic pivots.\u003c/p\u003e \u003cp\u003ePivot count has a protective effect (HR\u0026thinsp;=\u0026thinsp;0.54, p\u0026thinsp;\u0026lt;\u0026thinsp;0.01): each additional pivot reduces failure risk by approximately 46%. This finding challenges the view that pivoting signals distress. In the context of AI entrepreneurship, pivots appear to reflect adaptive learning\u0026mdash;the kind of entrepreneurial responsiveness to feedback that Kirzner (\u003cspan citationid=\"CR13\" class=\"CitationRef\"\u003e1997\u003c/span\u003e) associates with the discovery process. Active firms are those that updated their plans in response to market signals; failed firms disproportionately maintained their original plans.\u003c/p\u003e \u003cp\u003eThe positive hazard ratio for technical founder share (HR\u0026thinsp;=\u0026thinsp;1.89, p\u0026thinsp;\u0026lt;\u0026thinsp;0.044) is counterintuitive but interpretable. Teams with very high proportions of technical founders may lack the domain-specific market knowledge needed to identify viable applications\u0026mdash;precisely the kind of contextual knowledge that Hayek (\u003cspan citationid=\"CR8\" class=\"CitationRef\"\u003e1945\u003c/span\u003e) argued cannot be centrally aggregated. The most resilient firms in the sample combine technical capability with domain expertise, consistent with the Austrian emphasis on entrepreneurship as requiring knowledge of \u0026ldquo;particular circumstances.\u0026rdquo;\u003c/p\u003e \u003cp\u003eThe proportional hazards assumption was tested using Schoenfeld residuals; no significant violations were detected (global test p\u0026thinsp;=\u0026thinsp;0.38). The concordance index of 0.74 indicates acceptable discriminative ability.\u003c/p\u003e \u003c/div\u003e \u003cdiv id=\"Sec19\" class=\"Section2\"\u003e \u003ch2\u003e6.4 Robustness Checks\u003c/h2\u003e \u003cp\u003eTwo robustness specifications address potential concerns about the baseline model. The first re-estimates the Cox model treating acquired firms as failures rather than right-censored observations. This addresses the concern that some acquisitions may represent distressed exits rather than successful outcomes; reclassifying them as events tests whether the baseline results depend on the censoring assumption. The second specification replaces the binary single-provider indicator with a stricter measure: \u003cem\u003ehard dependency\u003c/em\u003e, defined as reliance on a single API provider with no viable open-weight fallback deployed or tested. This narrows the treatment group from 91 firms (71.7%) to 64 firms (50.4%), isolating a more extreme form of plan-dependence. Table\u0026nbsp;\u003cspan refid=\"Tab3\" class=\"InternalRef\"\u003e3\u003c/span\u003e reports the results alongside the baseline specification.\u003c/p\u003e \u003cp\u003e \u003cdiv class=\"gridtable\"\u003e\u003ctable float=\"Yes\" id=\"Tab3\" border=\"1\"\u003e \u003ccaption language=\"En\"\u003e \u003cdiv class=\"CaptionNumber\"\u003eTable 3\u003c/div\u003e \u003cdiv class=\"CaptionContent\"\u003e \u003cp\u003eRobustness Specifications: Cox Proportional Hazard Model\u003c/p\u003e \u003c/div\u003e \u003c/caption\u003e \u003ccolgroup cols=\"4\"\u003e \u003cdiv align=\"left\" class=\"colspec\" colname=\"c1\" colnum=\"1\"\u003e\u003c/div\u003e \u003cdiv align=\"left\" class=\"colspec\" colname=\"c2\" colnum=\"2\"\u003e\u003c/div\u003e \u003cdiv align=\"left\" class=\"colspec\" colname=\"c3\" colnum=\"3\"\u003e\u003c/div\u003e \u003cdiv align=\"left\" class=\"colspec\" colname=\"c4\" colnum=\"4\"\u003e\u003c/div\u003e \u003cthead\u003e \u003ctr\u003e \u003cth align=\"left\" colname=\"c1\"\u003e\u0026nbsp;\u003c/th\u003e \u003cth align=\"left\" colname=\"c2\"\u003e \u003cp\u003e(1) Baseline\u003c/p\u003e \u003c/th\u003e \u003cth align=\"left\" colname=\"c3\"\u003e \u003cp\u003e(2) Acq. = Failure\u003c/p\u003e \u003c/th\u003e \u003cth align=\"left\" colname=\"c4\"\u003e \u003cp\u003e(3) Hard Dep.\u003c/p\u003e \u003c/th\u003e \u003c/tr\u003e \u003c/thead\u003e \u003ctbody\u003e \u003ctr\u003e \u003ctd align=\"left\" colname=\"c1\"\u003e \u003cp\u003e\u003cem\u003eln(Initial Capital)\u003c/em\u003e\u003c/p\u003e \u003c/td\u003e \u003ctd align=\"left\" colname=\"c2\"\u003e \u003cp\u003e0.71**\u003c/p\u003e \u003c/td\u003e \u003ctd align=\"left\" colname=\"c3\"\u003e \u003cp\u003e0.74**\u003c/p\u003e \u003c/td\u003e \u003ctd align=\"left\" colname=\"c4\"\u003e \u003cp\u003e0.69**\u003c/p\u003e \u003c/td\u003e \u003c/tr\u003e \u003ctr\u003e \u003ctd align=\"left\" colname=\"c1\"\u003e \u003cp\u003e\u003cem\u003eTeam Size\u003c/em\u003e\u003c/p\u003e \u003c/td\u003e \u003ctd align=\"left\" colname=\"c2\"\u003e \u003cp\u003e0.68**\u003c/p\u003e \u003c/td\u003e \u003ctd align=\"left\" colname=\"c3\"\u003e \u003cp\u003e0.72**\u003c/p\u003e \u003c/td\u003e \u003ctd align=\"left\" colname=\"c4\"\u003e \u003cp\u003e0.67**\u003c/p\u003e \u003c/td\u003e \u003c/tr\u003e \u003ctr\u003e \u003ctd align=\"left\" colname=\"c1\"\u003e \u003cp\u003e\u003cem\u003eTechnical Share\u003c/em\u003e\u003c/p\u003e \u003c/td\u003e \u003ctd align=\"left\" colname=\"c2\"\u003e \u003cp\u003e1.89**\u003c/p\u003e \u003c/td\u003e \u003ctd align=\"left\" colname=\"c3\"\u003e \u003cp\u003e1.76**\u003c/p\u003e \u003c/td\u003e \u003ctd align=\"left\" colname=\"c4\"\u003e \u003cp\u003e1.93**\u003c/p\u003e \u003c/td\u003e \u003c/tr\u003e \u003ctr\u003e \u003ctd align=\"left\" colname=\"c1\"\u003e \u003cp\u003e\u003cem\u003ePivot Count\u003c/em\u003e\u003c/p\u003e \u003c/td\u003e \u003ctd align=\"left\" colname=\"c2\"\u003e \u003cp\u003e0.54***\u003c/p\u003e \u003c/td\u003e \u003ctd align=\"left\" colname=\"c3\"\u003e \u003cp\u003e0.61**\u003c/p\u003e \u003c/td\u003e \u003ctd align=\"left\" colname=\"c4\"\u003e \u003cp\u003e0.52***\u003c/p\u003e \u003c/td\u003e \u003c/tr\u003e \u003ctr\u003e \u003ctd align=\"left\" colname=\"c1\"\u003e \u003cp\u003e\u003cem\u003eProvider Dependency\u003c/em\u003e\u003c/p\u003e \u003c/td\u003e \u003ctd align=\"left\" colname=\"c2\"\u003e \u003cp\u003e2.47***\u003c/p\u003e \u003c/td\u003e \u003ctd align=\"left\" colname=\"c3\"\u003e \u003cp\u003e2.19***\u003c/p\u003e \u003c/td\u003e \u003ctd align=\"left\" colname=\"c4\"\u003e \u003cp\u003e3.12***\u003c/p\u003e \u003c/td\u003e \u003c/tr\u003e \u003ctr\u003e \u003ctd align=\"left\" colname=\"c1\"\u003e \u003cp\u003e\u003cem\u003eSector FE\u003c/em\u003e\u003c/p\u003e \u003c/td\u003e \u003ctd align=\"left\" colname=\"c2\"\u003e \u003cp\u003eYes\u003c/p\u003e \u003c/td\u003e \u003ctd align=\"left\" colname=\"c3\"\u003e \u003cp\u003eYes\u003c/p\u003e \u003c/td\u003e \u003ctd align=\"left\" colname=\"c4\"\u003e \u003cp\u003eYes\u003c/p\u003e \u003c/td\u003e \u003c/tr\u003e \u003ctr\u003e \u003ctd align=\"left\" colname=\"c1\"\u003e \u003cp\u003e\u003cem\u003eEvents / N\u003c/em\u003e\u003c/p\u003e \u003c/td\u003e \u003ctd align=\"left\" colname=\"c2\"\u003e \u003cp\u003e23 / 127\u003c/p\u003e \u003c/td\u003e \u003ctd align=\"left\" colname=\"c3\"\u003e \u003cp\u003e34 / 127\u003c/p\u003e \u003c/td\u003e \u003ctd align=\"left\" colname=\"c4\"\u003e \u003cp\u003e23 / 127\u003c/p\u003e \u003c/td\u003e \u003c/tr\u003e \u003ctr\u003e \u003ctd align=\"left\" colname=\"c1\"\u003e \u003cp\u003e\u003cem\u003eConcordance\u003c/em\u003e\u003c/p\u003e \u003c/td\u003e \u003ctd align=\"left\" colname=\"c2\"\u003e \u003cp\u003e0.74\u003c/p\u003e \u003c/td\u003e \u003ctd align=\"left\" colname=\"c3\"\u003e \u003cp\u003e0.71\u003c/p\u003e \u003c/td\u003e \u003ctd align=\"left\" colname=\"c4\"\u003e \u003cp\u003e0.76\u003c/p\u003e \u003c/td\u003e \u003c/tr\u003e \u003c/tbody\u003e \u003c/colgroup\u003e \u003ctfoot\u003e \u003ctr\u003e\u003ctd colspan=\"4\"\u003e\u003cem\u003eNotes\u003c/em\u003e: *** p\u0026thinsp;\u0026lt;\u0026thinsp;0.01, ** p\u0026thinsp;\u0026lt;\u0026thinsp;0.05. Hazard ratios reported. Specification (2) reclassifies acquisitions as failure events. Specification (3) replaces single-provider indicator with hard dependency (single API, no open-weight fallback).\u003c/td\u003e\u003c/tr\u003e \u003c/tfoot\u003e \u003c/table\u003e\u003c/div\u003e \u003c/p\u003e \u003cp\u003eThe results are stable across specifications in both sign and approximate magnitude. All coefficients retain significance and directionality. The provider dependency effect strengthens under the hard dependency definition (HR\u0026thinsp;=\u0026thinsp;3.12, p\u0026thinsp;\u0026lt;\u0026thinsp;0.01), consistent with the interpretation that deeper plan-dependence\u0026mdash;not merely nominal single-sourcing\u0026mdash;drives the risk mechanism. When acquisitions are reclassified as failures, the pivot count effect attenuates slightly (HR\u0026thinsp;=\u0026thinsp;0.61 from 0.54) but remains significant, suggesting that the protective effect of pivoting is not an artifact of the censoring assumption. Parametric (Weibull) specifications yield qualitatively similar results across all three models.\u003c/p\u003e \u003cp\u003eA note on external validity is warranted. The observation window (2022\u0026ndash;2024, with status assessed in January 2025) is short, and the classification \u0026ldquo;AI-native\u0026rdquo; is necessarily a moving boundary as AI capabilities diffuse across industries. The findings should be understood as characterizing the \u003cem\u003eearly dynamics\u003c/em\u003e of a market in rapid formation. As the AI application ecosystem matures, entry barriers, provider dependency structures, and pivot patterns may evolve substantially. The theoretical framework developed in Sections \u003cspan refid=\"Sec3\" class=\"InternalRef\"\u003e3\u003c/span\u003e\u0026ndash;\u003cspan refid=\"Sec12\" class=\"InternalRef\"\u003e5\u003c/span\u003e does not depend on the persistence of any specific empirical parameter; it depends on the structural claim that market coordination generates knowledge of a kind that non-market institutions cannot replicate. The empirical analysis provides evidence consistent with that claim in the current period, and future research should test whether these patterns persist as the market develops.\u003c/p\u003e \u003c/div\u003e \u003cdiv id=\"Sec20\" class=\"Section2\"\u003e \u003ch2\u003e6.5 Case Studies in Institutional Diversity\u003c/h2\u003e \u003cp\u003e \u003cem\u003eOpenAI\u003c/em\u003e began as a nonprofit research laboratory, transitioned to a \u0026ldquo;capped-profit\u0026rdquo; structure, and has subsequently moved toward a more conventional for-profit model. This institutional evolution reflects entrepreneurial experimentation with organizational form under genuine uncertainty about the appropriate governance structure for frontier AI development. The pivot from open-source research to proprietary model development\u0026mdash;symbolized by the decision not to release the full GPT-4 model weights\u0026mdash;represents a discovery about the relative economic value of open versus proprietary approaches to model distribution.\u003c/p\u003e \u003cp\u003e \u003cem\u003eAnthropic\u003c/em\u003e was founded around a particular theory of AI safety and has developed \u0026ldquo;constitutional AI\u0026rdquo; as a distinctive approach to model alignment. This represents institutional innovation in a domain\u0026mdash;the governance of AI systems\u0026mdash;where market discovery is essential because the relevant knowledge about effective governance mechanisms does not exist in advance. Anthropic\u0026rsquo;s corporate structure as a public benefit corporation reflects an entrepreneurial conjecture about how to attract talent and capital while maintaining commitment to safety-oriented research\u0026mdash;a conjecture whose success depends on competitive market validation.\u003c/p\u003e \u003cp\u003e \u003cem\u003eMistral\u003c/em\u003e, the French AI company, has pursued an open-weight strategy that deliberately contrasts with the proprietary approaches of OpenAI and (partially) Anthropic. By releasing model weights under permissive licenses, Mistral has discovered and exploited an entrepreneurial opportunity in serving the segment of the market that values local deployment, customization, and independence from API providers. The coexistence of these three strategies\u0026mdash;proprietary (OpenAI), safety-focused benefit corporation (Anthropic), and open-weight (Mistral)\u0026mdash;exemplifies the institutional diversity that competitive market processes generate and that no central authority could have anticipated or designed.\u003c/p\u003e \u003c/div\u003e"},{"header":"7. Conclusion: Structural Limits of Algorithmic Coordination","content":"\u003cp\u003eThis paper has argued that the rapid development of AI systems illuminates, rather than dissolves, the structural limits of non-market economic coordination. The derivative knowledge problem demonstrates that the impressive capabilities of foundation models are consequences of the richness of market-generated data, not evidence that markets can be dispensed with. The formal distinction between computational optimization and entrepreneurial discovery clarifies why the calculation problem is not a computational bottleneck amenable to technological solution: it is a categorical difference between operating within a known possibility space and expanding that space through institutionally embedded action. The irreconstructibility of opportunity-cost structure from non-market telemetry specifies precisely what is missing: not information quantity, but the counterfactual valuations that only competitive bidding reveals. Foundation models are powerful optimization engines, but the economic problem is not, at bottom, an optimization problem\u0026mdash;it is a problem of generating the evaluative structure within which optimization becomes meaningful.\u003c/p\u003e \u003cp\u003eThe capital-theoretic analysis reveals that even in the most technologically advanced sectors, genuine uncertainty and plan-dependence make centralized investment coordination structurally fragile. The 2024 GPU shortage and subsequent overcapacity illustrate how decentralized entrepreneurial error correction\u0026mdash;the boom-bust process that Austrian capital theory describes\u0026mdash;operates in novel technological domains with the same dynamics visible in more traditional industries. The pattern of democratized disruption with oligopolistic infrastructure shows that market processes generate complex, multi-layered institutional structures whose emergent properties could not be specified in advance.\u003c/p\u003e \u003cp\u003eThe empirical evidence from 127 AI-native firms is consistent with these theoretical claims. The survival analysis reveals that provider diversification, adaptive pivoting, and the combination of technical and domain expertise predict firm survival\u0026mdash;patterns that are robust to alternative censoring assumptions and stricter dependency definitions, and that reflect the importance of decentralized, contextual knowledge and adaptive learning in navigating genuine uncertainty. The case studies of OpenAI, Anthropic, and Mistral demonstrate that the competitive process generates institutional diversity\u0026mdash;in governance structures, safety approaches, and openness strategies\u0026mdash;that no planner could have anticipated.\u003c/p\u003e \u003cp\u003eThese findings also suggest directions for theoretical refinement. The coexistence of vigorous application-layer competition with infrastructure-layer concentration does not fit neatly into standard models of either equilibrating entrepreneurship or creative destruction. A more adequate theory of AI market structure will need to account for how entrepreneurial discovery operates within and across different layers of a production structure characterized by extreme complementarity and asset specificity. The market process is actively exploring this space; theory needs to keep pace.\u003c/p\u003e \u003cp\u003eFinally, the analysis points toward a broader conclusion about the relationship between technology and institutions. AI is not an exogenous force that transforms economic institutions from the outside. It is itself a product of market coordination\u0026mdash;an emergent property of the institutional order it is sometimes claimed to supersede. The algorithmic calculation problem is not a limitation that will be overcome by more powerful algorithms. It reflects the structural insight that animates the entire calculation debate: rational economic coordination requires institutions that \u003cem\u003egenerate\u003c/em\u003e knowledge, not merely institutions that \u003cem\u003eprocess\u003c/em\u003e it.\u003c/p\u003e"},{"header":"Declarations","content":"\u003ch2\u003eAuthor Contribution\u003c/h2\u003e\u003cp\u003eCW wrote the paper.\u003c/p\u003e"},{"header":"References","content":"\u003col\u003e\u003cli\u003e\u003cspan\u003eAnthropic. (2023). \u003cem\u003eToy models of superposition\u003c/em\u003e. Transformer Circuits Thread.\u003c/span\u003e\u003c/li\u003e \u003cli\u003e\u003cspan\u003eBastani, A. (2019). \u003cem\u003eFully automated luxury communism: A manifesto\u003c/em\u003e. Verso Books.\u003c/span\u003e\u003c/li\u003e \u003cli\u003e\u003cspan\u003eBoettke, P. J. (1993). \u003cem\u003eWhy perestroika failed: The politics and economics of socialist transformation\u003c/em\u003e. Routledge.\u003c/span\u003e\u003c/li\u003e \u003cli\u003e\u003cspan\u003eBoettke, P. J. (1998). Economic calculation: The Austrian contribution to political economy. \u003cem\u003eAdvances in Austrian Economics\u003c/em\u003e, \u003cem\u003e5\u003c/em\u003e, 131\u0026ndash;158.\u003c/span\u003e\u003c/li\u003e \u003cli\u003e\u003cspan\u003eButerin, V. (2023). \u003cem\u003eMy techno-optimism\u003c/em\u003e. Vitalik.eth Blog.\u003c/span\u003e\u003c/li\u003e \u003cli\u003e\u003cspan\u003eElhage, N., Nanda, N., Olsson, C., Henighan, T., Joseph, N., Mann, B., Olah, C., et al. (2021). \u003cem\u003eA mathematical framework for transformer circuits\u003c/em\u003e. Transformer Circuits Thread.\u003c/span\u003e\u003c/li\u003e \u003cli\u003e\u003cspan\u003eHayek, F. A. (1937). Economics and knowledge. \u003cem\u003eEconomica\u003c/em\u003e, \u003cem\u003e4\u003c/em\u003e(13), 33\u0026ndash;54.\u003c/span\u003e\u003c/li\u003e \u003cli\u003e\u003cspan\u003eHayek, F. A. (1945). The use of knowledge in society. \u003cem\u003eAmerican Economic Review\u003c/em\u003e, \u003cem\u003e35\u003c/em\u003e(4), 519\u0026ndash;530.\u003c/span\u003e\u003c/li\u003e \u003cli\u003e\u003cspan\u003eHurwicz, L. (1960). Optimality and informational efficiency in resource allocation processes. In K. J. Arrow, S. Karlin, \u0026amp; P. Suppes (Eds.), \u003cem\u003eMathematical methods in the social sciences\u003c/em\u003e (pp. 27\u0026ndash;46). Stanford University Press.\u003c/span\u003e\u003c/li\u003e \u003cli\u003e\u003cspan\u003eHurwicz, L. (1972). On informationally decentralized systems. In C. B. McGuire, \u0026amp; R. Radner (Eds.), \u003cem\u003eDecision and organization\u003c/em\u003e (pp. 297\u0026ndash;336). North-Holland.\u003c/span\u003e\u003c/li\u003e \u003cli\u003e\u003cspan\u003eKirzner, I. M. (1973). \u003cem\u003eCompetition and entrepreneurship\u003c/em\u003e. University of Chicago Press.\u003c/span\u003e\u003c/li\u003e \u003cli\u003e\u003cspan\u003eKirzner, I. M. (1984). \u003cem\u003eThe role of the entrepreneur in the economic system\u003c/em\u003e. Centre for Independent Studies.\u003c/span\u003e\u003c/li\u003e \u003cli\u003e\u003cspan\u003eKirzner, I. M. (1997). Entrepreneurial discovery and the competitive market process: An Austrian approach. \u003cem\u003eJournal of Economic Literature\u003c/em\u003e, \u003cem\u003e35\u003c/em\u003e(1), 60\u0026ndash;85.\u003c/span\u003e\u003c/li\u003e \u003cli\u003e\u003cspan\u003eKnight, F. H. (1921). \u003cem\u003eRisk, uncertainty, and profit\u003c/em\u003e. Houghton Mifflin.\u003c/span\u003e\u003c/li\u003e \u003cli\u003e\u003cspan\u003eLachmann, L. M. (1956). \u003cem\u003eCapital and its structure\u003c/em\u003e. Ludwig von Mises Institute.\u003c/span\u003e\u003c/li\u003e \u003cli\u003e\u003cspan\u003eLange, O. (1936). On the economic theory of socialism: Part one. \u003cem\u003eReview of Economic Studies\u003c/em\u003e, \u003cem\u003e4\u003c/em\u003e(1), 53\u0026ndash;71.\u003c/span\u003e\u003c/li\u003e \u003cli\u003e\u003cspan\u003eLavoie, D. (1985). \u003cem\u003eRivalry and central planning: The socialist calculation debate reconsidered\u003c/em\u003e. Cambridge University Press.\u003c/span\u003e\u003c/li\u003e \u003cli\u003e\u003cspan\u003eLerner, A. P. (1944). \u003cem\u003eThe economics of control: Principles of welfare economics\u003c/em\u003e. Macmillan.\u003c/span\u003e\u003c/li\u003e \u003cli\u003e\u003cspan\u003evon Mises, L. (1920). Die Wirtschaftsrechnung im sozialistischen Gemeinwesen. \u003cem\u003eArchiv f\u0026uuml;r Sozialwissenschaft und Sozialpolitik\u003c/em\u003e, \u003cem\u003e47\u003c/em\u003e, 86\u0026ndash;121.\u003c/span\u003e\u003c/li\u003e \u003cli\u003e\u003cspan\u003eMorozov, E. (2019). Digital socialism? The calculation debate in the age of big data. \u003cem\u003eNew Left Review\u003c/em\u003e, \u003cem\u003e116/117\u003c/em\u003e, 33\u0026ndash;67.\u003c/span\u003e\u003c/li\u003e \u003cli\u003e\u003cspan\u003eNeurath, O. (1921). \u003cem\u003eAnti-Spengler\u003c/em\u003e. Callwey.\u003c/span\u003e\u003c/li\u003e \u003cli\u003e\u003cspan\u003eOlah, C., Cammarata, N., Schubert, L., Goh, G., Petrov, M., \u0026amp; Carter, S. (2020). Zoom in: An introduction to circuits. \u003cem\u003eDistill\u003c/em\u003e, 5(3).\u003c/span\u003e\u003c/li\u003e \u003cli\u003e\u003cspan\u003ePhillips, L., \u0026amp; Rozworski, M. (2019). \u003cem\u003eThe people\u0026rsquo;s republic of Walmart: How the world\u0026rsquo;s biggest corporations are laying the groundwork for socialism\u003c/em\u003e. Verso Books.\u003c/span\u003e\u003c/li\u003e \u003cli\u003e\u003cspan\u003eShackle, G. L. S. (1972). \u003cem\u003eEpistemics and economics: A critique of economic doctrines\u003c/em\u003e. Cambridge University Press.\u003c/span\u003e\u003c/li\u003e \u003cli\u003e\u003cspan\u003eShumailov, I., Shumaylov, Z., Zhao, Y., Gal, Y., Papernot, N., \u0026amp; Anderson, R. (2024). The curse of recursion: Training on generated data makes models forget. arXiv preprint arXiv:2305.17493.\u003c/span\u003e\u003c/li\u003e \u003cli\u003e\u003cspan\u003eSrnicek, N. (2017). \u003cem\u003ePlatform capitalism\u003c/em\u003e. Polity.\u003c/span\u003e\u003c/li\u003e \u003cli\u003e\u003cspan\u003eYin, R. K. (2018). \u003cem\u003eCase study research and applications: Design and methods\u003c/em\u003e (6th ed.). Sage.\u003c/span\u003e\u003c/li\u003e\u003c/ol\u003e"}],"fulltextSource":"","fullText":"","funders":[],"hasAdminPriorityOnWorkflow":false,"hasManuscriptDocX":true,"hasOptedInToPreprint":true,"hasPassedJournalQc":"","hasAnyPriority":true,"hideJournal":true,"highlight":"","institution":"","isAcceptedByJournal":false,"isAuthorSuppliedPdf":false,"isDeskRejected":"","isHiddenFromSearch":false,"isInQc":false,"isInWorkflow":false,"isPdf":false,"isPdfUpToDate":true,"isWithdrawnOrRetracted":false,"journal":{"display":true,"email":"[email protected]","identity":"researchsquare","isNatureJournal":false,"hasQc":true,"allowDirectSubmit":true,"externalIdentity":"","sideBox":"","snPcode":"","submissionUrl":"/submission","title":"Research Square","twitterHandle":"researchsquare","acdcEnabled":true,"dfaEnabled":false,"editorialSystem":"","reportingPortfolio":"","inReviewEnabled":false,"inReviewRevisionsEnabled":true},"keywords":"economic calculation, socialist calculation debate, large language models, knowledge problem, spontaneous order, capital heterogeneity, foundation models, artificial intelligence, institutional economics, market coordination","lastPublishedDoi":"10.21203/rs.3.rs-8856269/v1","lastPublishedDoiUrl":"https://doi.org/10.21203/rs.3.rs-8856269/v1","license":{"name":"CC BY 4.0","url":"https://creativecommons.org/licenses/by/4.0/"},"manuscriptAbstract":"\u003cp\u003eLarge language models and generative AI systems have reignited debate over the feasibility of non-market economic coordination. A growing literature contends that modern AI renders the Hayekian knowledge problem technologically obsolete. This paper argues that these claims rest on a misidentification of the epistemic status of foundation model capabilities. Drawing on mechanistic interpretability research (Elhage et al., \u003cspan citationid=\"CR6\" class=\"CitationRef\"\u003e2021\u003c/span\u003e; Olah et al., \u003cspan citationid=\"CR22\" class=\"CitationRef\"\u003e2020\u003c/span\u003e; Anthropic, \u003cspan citationid=\"CR1\" class=\"CitationRef\"\u003e2023\u003c/span\u003e) and a novel dataset of 127 AI-native firms (2022\u0026ndash;2024), I develop three arguments. First, foundation models suffer from a \u003cem\u003ederivative knowledge problem\u003c/em\u003e: they compress statistical regularities from market-coordinated data but cannot generate the forward-looking knowledge that entrepreneurial discovery produces under genuine uncertainty. I formalize this distinction by separating \u003cem\u003ecomputational optimization\u003c/em\u003e within a known possibility space from \u003cem\u003eentrepreneurial discovery\u003c/em\u003e that expands the possibility space itself, and I demonstrate that the specific epistemic object missing from non-market telemetry is \u003cem\u003eopportunity-cost structure\u003c/em\u003e\u0026mdash;the counterfactual valuations that only competitive bidding reveals. Second, AI capital exhibits extreme Lachmannian heterogeneity and plan-dependence, generating Austrian business cycle dynamics visible in the 2024 GPU shortage and subsequent overcapacity. Third, AI simultaneously lowers barriers to competitive entry at the application layer while concentrating complementary assets at the infrastructure layer\u0026mdash;a pattern I term \u0026ldquo;democratized disruption with oligopolistic infrastructure.\u0026rdquo; Empirical analysis using Cox proportional hazard models reveals that model-provider dependency and pivot frequency are significant predictors of AI firm survival, while infrastructure-layer market structure shapes but does not determine application-layer outcomes. The findings suggest structural limits to algorithmic coordination that are not reducible to computational constraints.\u003c/p\u003e","manuscriptTitle":"The Algorithmic Calculation Problem: Why Foundation Models Cannot Solve Socialist Planning","msid":"","msnumber":"","nonDraftVersions":[{"code":1,"date":"2026-02-23 00:34:38","doi":"10.21203/rs.3.rs-8856269/v1","editorialEvents":[{"type":"communityComments","content":0}],"status":"published","journal":{"display":true,"email":"[email protected]","identity":"researchsquare","isNatureJournal":false,"hasQc":true,"allowDirectSubmit":true,"externalIdentity":"","sideBox":"","snPcode":"","submissionUrl":"/submission","title":"Research Square","twitterHandle":"researchsquare","acdcEnabled":true,"dfaEnabled":false,"editorialSystem":"","reportingPortfolio":"","inReviewEnabled":false,"inReviewRevisionsEnabled":true}}],"origin":"","ownerIdentity":"8bc88648-1ce8-43ac-9836-5fa61fdb5792","owner":[],"postedDate":"February 23rd, 2026","published":true,"recentEditorialEvents":[],"rejectedJournal":[],"revision":"","amendment":"","status":"posted","subjectAreas":[],"tags":[],"updatedAt":"2026-02-23T00:34:41+00:00","versionOfRecord":[],"versionCreatedAt":"2026-02-23 00:34:38","video":"","vorDoi":"","vorDoiUrl":"","workflowStages":[]},"version":"v1","identity":"rs-8856269","journalConfig":"researchsquare"},"__N_SSP":true},"page":"/article/[identity]/[[...version]]","query":{"redirect":"/article/rs-8856269","identity":"rs-8856269","version":["v1"]},"buildId":"XKTyCvWXoU3ODBz1xrDgd","isFallback":false,"isExperimentalCompile":false,"dynamicIds":[84888],"gssp":true,"scriptLoader":[]}