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

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.