Sovereign Agency under Physical Constraint: A Critical Review and Quantitative Augmentation of Game-Theoretic and Infrastructure-Centric Models of National AI Compute Ecosystems

Sovereign artificial intelligence has moved from a legal debate about data residency to a practical question of agency across the whole computational value chain: energy, connectivity, data centres, compute, data, models, applications, skills and governance. This paper reviews an AI-generated report that models national compute ecosystems through infrastructure economics, low-precision arithmetic, game theory and Monte Carlo risk analysis, and augments it into a verified academic treatment. Each empirical claim in the source was checked against institutional sources. Several were confirmed: the scale of the CNAS Sovereign AI Index, the UK Council for Science and Technology's definition of sovereignty as agency, the European AI Gigafactories call and its deadline changes, Bangladesh's draft National AI Policy 2026–2030, the OECD's 2026 due diligence guidance and the specifications of current rack-scale systems. Others were corrected, notably the infrastructure share of projects and the temporary nature of a mid-2026 suspension of foreign access to frontier models; unverifiable figures were removed. The augmentation adds three computed analyses. A synthetic quantisation experiment shows why block-scaled 4-bit formats are central to doing more with scarce national compute. A corrected analysis of the source's openness model shows that its optimum is a single planner's choice, not a Nash equilibrium, and derives its comparative statics. A Monte Carlo schedule model shows that headline probabilities of infrastructure delay depend entirely on assumed distributions and should be reported with them. The paper closes with implications for capital-constrained states, with Bangladesh as a reference case, framed at the level of strategy and policy.