Procurement dependence asks whether public agencies retain audit rights, portability, exit clauses, data access, and continuity guarantees.

asymmetric interdependence, not as the opposite of sovereignty.

Franco and coauthors provide a concrete digital-age bridge through the case of Mercado Libre (Franco et al. 2024). Their analysis complicates any simple local/foreign distinction. A regional platform can be technologically sophisticated, data-intensive, and commercially powerful while remaining structurally dependent on US cloud providers and proprietary infrastructures. At the same time, that regional platform can exercise extractive power over local users, sellers, and data subjects. This layered position is exactly what an AI sovereignty profile must capture. Dependence does not only run from Latin American states to Global North firms. It can also be reproduced through regional intermediaries that are subordinate upward and dominant downward.

This is where the bridge to digital and AI dependence becomes clear. Valente and Grohmann argue that Latin American critical data studies should go beyond the general language of data colonialism by drawing on regional traditions such as dependency theory, labor overexploitation, and liberation thought (Valente and Grohmann 2024). Their contribution is important because it prevents the paper from treating dependence as a neutral technical term. In the region, dependence carries the memory of center-periphery relations, unequal exchange, epistemic asymmetries, and constrained development. At the same time, V3 should not remain inside classical dependency theory. The AI stack has its own mechanisms: cloud concentration, proprietary APIs, accelerator supply chains, model ecosystems, standards bodies, data pipelines, procurement contracts, and platform governance.

ependency theory argued that underdevelopment was not simply an earlier stage on the path to development, but a structural position within an international system that linked development and underdevelopment through unequal functions, external constraints, and internal alliances (Cardoso and Faletto 1977). That tradition is not a ready-made measurement model for AI, but it offers a powerful starting point: dependence is relational, historically produced, and mediated by domestic political and economic structures. Cardoso and Faletto are especially useful because they avoid a simplistic externalism. Dependence is not only something imposed from outside; it is internalized through domestic class relations, elite alliances, productive structures, and institutional choices (Cardoso and Faletto 1977).

Dependence is the third concept, and it should be structured with the same theoretical seriousness as state capacity and political authority.

Roberts argues that digital sovereignty should not be reduced to descriptive control over infrastructure; it must be evaluated normatively, by asking whether control serves autonomy, rights, and democratic self-determination (Roberts 2024). Santaniello similarly shows that digital sovereignty claims are often contradictory and politically instrumental; they can be used by states, corporations, and other actors in ways that conceal dependencies or justify centralization (Santaniello 2025). Mügge asks the question directly in the AI context: sovereignty for whom, to what end, and to whose benefit (Mügge 2024)?

Hawkins, Lehdonvirta, and Wu

State capacity is the first concept because AI sovereignty cannot be reduced to declarations, strategies, or legal claims. A state may announce a national AI plan, publish ethical principles, and sign international agreements while still lacking the administrative, technical, fiscal, infrastructural, and absorptive capacity to shape how AI is actually built, procured, deployed, audited, and contested. In this sense, state capacity is not the same as state ambition. It is the organized ability to act.

ILIA 2026 is better treated, for now, as a measurement-design source that documents continuity, changes, and possible extensions.

1 Introduction

The synthesis process also incorporates collaborator feedback. Andres Carvallo’s comments emphasized that the introduction should make the “why” explicit and that the method should read as a polished academic and institutional process rather than as a repository log. Mitchell’s comments proposed a three-pillar structure organized around capacity, governance, and dependence.

The analysis builds on an 88-source methodological corpus covering AI sovereignty, digital sovereignty, data sovereignty, compute sovereignty, AI readiness, digital dependency, regional development, public-sector governance, and critical data studies

Fourth, it identifies the conceptual ground on which a later methodological annex should map indicators, discard weak subindicators, propose weights, and test the profile against external indices such as R&D intensity, ICT development, GDP, ILIA’s official score, and AI preparedness.

The scope of this version is deliberately conceptual. It prepares the argument and the measurement logic, but it does not yet compute revised weights, rerun country rankings, or implement cross-index diagnostics.

This paper makes four contributions to ILIA’s next stage. First, it defines AI sovereignty as a relational problem of agency rather than as a fixed status of technological independence. Second, it proposes a three-concept framework for the region: state capacity, political authority, and dependence. Third, it clarifies why the current ILIA data can support a preliminary state-capacity and governance/authority baseline, but cannot yet support a completed dependence-aware AI sovereignty index. Fourth, it identifies the conceptual ground on which a later methodological annex should map indicators, discard weak subindicators, propose weights, and test the profile against external indices such as R&D intensity, ICT development, GDP, ILIA’s official score, and AI preparedness.