Beyond AI sovereignty: why the West relies on US frontier models and what comes next

For several years, “AI sovereignty” has been a theme in policy discussions across Europe, Asia, and parts of the Global South. And I have been participating in these conversations amongst the AI practitioners and AI policy wonks for years. The idea was straightforward: nations should retain control over their data, infrastructure, and increasingly, their AI capabilities. But in practice, much of the West, particularly countries aligned with the United States, have quietly moved in a different direction. Rather than building sovereign capability, they have defaulted to reliance on US frontier models. The result is that, for many, the sovereignty debate was never truly resolved, it was bypassed.

The quiet consolidation of AI power

While policymakers debated digital sovereignty, the market consolidated. A small handful of US-based firms now dominate frontier AI:

• They control the most advanced models
• They operate the hyperscale infrastructure required to run them
• They set the pace of capability development

For most Western countries, the path of least resistance has been adoption rather than autonomy. Enterprise systems, government pilots, and even critical services are increasingly built on top of these models. This is not sovereignty. It is dependency.

And in a world where AI sovereignty has effectively given way to dependence, frontier models have become a form of critical infrastructure risk, concentrated, externally controlled, and difficult to substitute when access is constrained.

Why sovereignty lost

There are pragmatic reasons why the sovereignty agenda stalled.

• Cost: Building frontier models and compute infrastructure is prohibitively expensive
• Talent: Expertise is globally concentrated and highly mobile
• Speed: Domestic alternatives struggle to keep pace with frontier development
• Network effects: The best models attract the most users, data, and developers

In this context, insisting on sovereign capability can look economically irrational and strategically slow. So most state and commercial actors chose integration over independence.

The illusion of control

Despite this shift, the language of sovereignty persists. Governments talk about:

• Data localisation
• Regulatory oversight
• Trusted AI frameworks

But these mechanisms operate at the margins, while the core cognitive infrastructure, the models themselves, remains external. However, it merely creates an illusion of control:

• You can regulate inputs and outputs
• You can set compliance requirements
• But you do not control the system’s underlying behaviour or evolution

This is a fundamentally different kind of dependency than previous technology waves.

From sovereignty to strategic dependence

If sovereignty no longer reflects reality, we need a more honest way to frame the situation. If what we are seeing is strategic dependence on US AI providers, and the key questions then shift:

• What level of dependence is acceptable?
• Where are the critical vulnerabilities?
• How do you maintain leverage in asymmetric relationships?

This is closer to energy security or semiconductor supply chains than traditional IT procurement. It requires a shift in thinking from ownership (that is, we no longer control the means of production) to risk management.

Interdependence but not symmetry

It is tempting to describe this as interdependence. But the relationship is not symmetrical, especially if one party can unilaterally deny other parties access to those resources without notice or consultation. The big US firms:

• Set technical standards
• Control update cycles
• Influence global developer ecosystems

While other countries:

• Consume and adapt these systems
• Layer governance and applications on top

This asymmetry matters, as it shapes everything from economic value capture to strategic autonomy.

What comes next

If the sovereignty debate has effectively been skipped, the next phase must deal with the consequences. And that means focusing on:

• Resilience: What happens when access is restricted, degraded, or politicised?
• Substitutability: How easily can systems switch between providers or models?
• Observability: How do you understand and audit systems you do not control?
• Assurance: How can you ascertain that the AI system is and remains safe, compliant, ethical and fit for purpose?
• Leverage: Where can governments and organisations exert influence in the stack?

These are harder, less comfortable, and more realistic questions than sovereignty, and they accept dependency as a starting point.

Rebuilding agency without illusions

The sovereignty window has largely closed for frontier models. Moving forward, if we accept that the goal is not to retroactively achieve full sovereignty, then the task before the rest of us is to rebuild agency within constraint. This might include:

• Investing in niche or specialised models where differentiation is possible
• Strengthening open-source ecosystems as a counterbalance
• Developing interoperability standards to reduce lock-in
• Building institutional capability to govern external AI systems effectively

In other words, working with the system as it exists, rather than the one we might have preferred.

A more honest conversation

The AI sovereignty debate was important. But in much of the West, it has already been overtaken by events. We are now operating in a world where:

• Frontier AI is concentrated in a handful of US firms
• Adoption has outpaced governance
• Dependency is structural, not incidental

Perhaps the question now is no longer whether we should have sovereignty - it is how we navigate a landscape where we do not have sovereignty. And that requires a more honest, and more strategically mature, conversation than we have had so far.