AI Infrastructure as Public Capacity: Dependency, Pluralism, and Institutional Choice

Abstract

Artificial intelligence depends on combinations of compute, energy, data, models, software, interfaces, standards, skills, organisations, and allocation rules. Governing each component separately can obscure how some combinations become reusable foundations on which institutions and communities depend. This article asks when a dated AI resource arrangement has infrastructural significance for specified actors and functions, and when access to that arrangement supports genuine public capacity. A continuing hypothetical university-compute case translates the diagnostic without replacing its stacked technical and institutional structure. Infrastructural significance is relational rather than intrinsic. Cross-context reuse is necessary but not sufficient; it must be connected to material dependency, difficult substitution or exit, bottleneck control, or serious public consequences. These conditions form a diagnostic, not a score. Public capacity means the real and durable opportunity of plural public, academic, civic, community, and smaller organisational actors to develop, evaluate, contest, adapt, and redirect AI-related resources. Three figures distinguish diagnostic conditions, conversion from access to capacity, and polycentric governance of stacked dependencies. The paper reports no original empirical study and makes no inference from intelligence or infrastructure control to authority, consciousness, legitimacy, or moral status.