Foundational models are commonly framed as advanced technological tools or commercial products. This article advances a different perspective: foundational models should be understood as strategic assets embedded in complex infrastructures of capital, compute, data, and institutional control.

Drawing from political economy, the article argues that the scale, capital intensity, and dependency structures associated with foundational models produce durable power asymmetries that cannot be addressed through application-level regulation alone. By reframing foundational models as infrastructural assets, the article highlights the limits of current governance approaches and outlines key implications for state capacity, sovereignty, and democratic oversight.

1. Introduction: Beyond the Tool Narrative

Public and regulatory debates around artificial intelligence continue to frame foundational models primarily as technological artifacts: systems that can be assessed, regulated, or constrained according to their observable outputs and applications.

This framing implicitly treats models as neutral instruments whose risks emerge mainly at the point of use.

Such a perspective obscures a core transformation. Foundational models are not merely tools embedded in products; they are increasingly infrastructural assets whose development, ownership, and operation reorganize economic power, institutional capacity, and governance authority.

Their significance lies less in any single application than in the structural dependencies they generate across sectors and jurisdictions.

From a political economy standpoint, the central question is not whether foundational models are powerful, but how that power is produced, accumulated, and governed.

2. Strategic Assets in Political Economy

In classical political economy and security studies, an asset is considered strategic when it exhibits a combination of the following characteristics:

  • High fixed and sunk costs
  • Strong economies of scale
  • Limited substitutability
  • Network and coordination effects
  • Capacity to generate dependency
  • Relevance to state interests and public functions

Foundational models meet all these criteria.

The costs associated with training frontier models (compute, energy, engineering labor, data acquisition, and infrastructure) are increasingly prohibitive.

As scale becomes a primary determinant of model performance, competitive entry narrows structurally, favoring a small number of actors capable of mobilizing extraordinary capital and infrastructure.

This shifts foundational models from the domain of competitive software markets into the realm of strategic economic assets, comparable not to applications, but to financial clearing systems, energy grids, or telecommunications backbones.

3. The Foundational Model Stack as a Governance Structure

Understanding foundational models as strategic assets requires moving beyond the model itself to examine the stack of infrastructures on which it depends. This stack can be analytically decomposed into four interdependent layers:

  • Physical Layer:

Semiconductors, energy systems, data centers, and global supply chains.

  • Infrastructural Layer:

Cloud platforms, distributed computing environments, and orchestration systems.

  • Model Layer:

Training processes, architectures, weights, and fine-tuning.

  • Access Layer:

APIs, products, licensing terms, deployment constraints, and usage policies.

Each layer constitutes a governance choke point. Control over the stack enables actors not only to innovate, but to shape who can participate, under what conditions, and at what cost.

Governance, in this sense, is exercised not primarily through law, but through architectural and contractual decisions embedded in sociotechnical systems.

4. Scale, Concentration, and Structural Power

The economics of foundational models exhibit strong increasing returns to scale. Performance improvements are closely tied to computing power, data volume and quality, and optimization capacity. This dynamic produces three structural outcomes:

  • Market Concentration:

A small number of firms dominate model training and deployment.

  • Vertical Integration Control:

Over compute, cloud, and model layers consolidates within the same organizations.

  • Dependency Creation Downstream:

Actors—startups, governments, and institutions—become reliant on upstream providers for access to foundational capabilities.

From a governance perspective, this matters because competition alone cannot correct these asymmetries.

Even where multiple firms exist, their shared reliance on similar infrastructural constraints reproduces concentration rather than mitigating it.

Recent cross-firm agreements between nominal competitors, such as the AI partnership between Apple and Google, illustrate how infrastructural constraints can align strategic interests, blurring the boundary between competition and coordination at the foundational layer.

5. Governance Beyond Regulation

Most current AI governance initiatives focus on application-level regulation: risk classification, safety assessments, transparency requirements, and post-deployment monitoring. While necessary, these approaches leave the core political economy of foundational models largely untouched.

Governance at the infrastructural level involves different questions:

  • Who is able to train foundational models at scale?
  • Who controls access to weights, APIs, and deployment rights?
  • Who can audit or contest model behavior?
  • Who bears systemic risks arising from widespread deployment?

Increasingly, answers to these questions are determined inside corporate governance structures, not public institutions.

6. State Capacity and Model Dependency

The strategic nature of foundational models has direct implications for state capacity. Governments face a growing mismatch between their regulatory ambitions and their operational dependence on privately controlled AI infrastructure.

This dependency is particularly pronounced outside the core economies that host large-scale compute and cloud infrastructure. For many states, participation in the AI economy takes the form of consumption, fine-tuning, or data provision, rather than foundational development.

From a political economy perspective, this dynamic resembles earlier forms of technological dependency, now mediated through models rather than machinery or energy. The result is an emerging asymmetry between states that govern AI infrastructures and those that merely use them.

7. Rethinking Policy: Infrastructure, Not Applications

If foundational models are strategic assets, governance approaches must shift accordingly. This does not imply immediate nationalization or centralized control, but it does require infrastructure-level policy thinking.

Key governance directions include:

  • Transparency obligations tied to training scale and compute usage
  • Institutional oversight mechanisms at the model and infrastructure level
  • Public or semi-public investment in foundational capabilities
  • Separation between model ownership and downstream deployment power
  • International coordination on compute access and standards

These are not technical fixes, but institutional design challenges that reflect the strategic role of AI infrastructures in contemporary economies.

8. Conclusion: From Innovation to Institutional Design

Managing innovation or mitigating isolated risks no longer addresses all the challenges posed by foundational models. Today, the real challenge is achieving institutional alignment in the face of the new strategic assets that concentrate power, generate dependency, and reshape state-market relations.