AI & Technology

It’s time to refocus the sovereign AI debate

By Alwin Magimay, Global Head of AI at PA Consulting and Michael Brown is Senior Vice President of Enterprise AI at Jacobs.  

Ask your business operations lead how much of the business would stop if access to your primary AI model is withdrawn tomorrow. If the answer is uncomfortable, you may have a sovereignty problem.  

Organisations now have access to extraordinary AI capabilities. But much of that capability depends on infrastructure, platforms, and foundation models that sit outside their control. As AI becomes further embedded into critical services, leaders need to think about what happens if access, control, governance or commercial terms change in ways they can’t influence. 

Governments are increasingly recognising this and are investing heavily in domestic AI capabilities as a result. More sovereign AI projects were launched in the first half of 2026 than in all of 2024, with 184 government-backed initiatives across more than 60 countries.  

Yet much of the debate around this investment is stuck on the wrong questions. Sovereign AI is often framed as a race between governments, a question of national capability or a simple question of who owns the model. But this doesn’t actually help a CEO decide what to do on a Monday morning.  

Before leaders can make sensible decisions about sovereign AI, they need to get clear on three common misconceptions. Only then can they determine where sovereignty genuinely matters, where dependency becomes a risk, and what sort of technology infrastructure they’ll need to support it. 

Misconception one: Sovereignty means owning the model 

First, sovereignty isn’t just about owning a model. A sovereign model that can’t be trusted, governed, integrated into operations, or deployed at scale has limited value. True sovereignty is about being able to rely on the technology, rather than simply owning it.  

For leaders running nuclear facilities, energy networks, defence systems, or other national infrastructure, that means having confidence that a model can operate safely, consistently, and accountably within highly regulated environments. It also means measuring how adoption and output is genuinely driving productivity and efficiency across value chains.  

Not every workload requires sovereignty and the sectors that matter most are fewer than people expect. For example, there’s little point in having a sovereign capability if you’re just using it to summarise emails. Leaders need to identify where trust, control, and long-term resilience genuinely matter and start concentrating effort there. 

Some countries are already taking this more focused approach. Rather than attempting to outspend the frontier providers, AI Singapore’s SEA-LION family of open-source Large Language Models has been developed around Southeast Asian languages and cultural contexts. The initiative focuses on underserved use cases in Southeast Asia, where sovereignty adds the most value instead of trying to win a race for the largest model. It shows how sovereign AI can be targeted at areas where local capability adds distinctive value, rather than competing on scale alone. 

Misconception two: Sovereignty means picking one model 

A second misconception is that sovereign AI means organisations must commit themselves to a single model provider. In fact, organisations should look at designing their systems to remain model-agnostic, giving them the ability to switch providers as capabilities evolve rather than tying their future to a single provider.  

This flexibility matters, because AI is continuing to develop at an extraordinary pace. The model that offers the best performance, pricing or governance today may not be the strongest option in two years’ time. Organisations that build their operating models around a single provider risk finding themselves locked into decisions made at the very beginning of the AI era. 

Dependence on one single provider can therefore become a strategic risk, rather than a simple technology decision. The question for leaders is therefore not, ‘Which model should I choose?’, but ‘What level of dependency on a single provider am I prepared to accept?’ What’s even more problematic than choosing the wrong model is losing the freedom to choose at all. 

Misconception three: Sovereignty is just a resilience issue 

Third, the conversation around sovereign AI is often framed as a resilience or a security issue. In reality, economics and infrastructure can be just as important.  

AI is fundamentally a consumption-based technology. The more organisations use it, the more they spend, unlike software costs that tend to reduce as user licences are added. As AI agents become embedded in workflows and decision-making processes, boards will need to check whether their AI operating model is economically sustainable and decide where greater control, resilience and independence justify the investment.  

Sovereignty must therefore also be looked at through a commercial lens. For many use cases, paying for frontier models will remain the right choice. Frontier model consumption has no fixed cost, and marginal costs rising with increased usage. But as AI usage scales, organisations may find that owning more of their AI stack becomes both cheaper and more resilient. This is because running a smaller domain-specific model on infrastructure you control has a high fixed cost, but a lower marginal cost.  

There is a tipping point, and organisations need to work out where that lies. Below it, sovereignty is an insurance premium, but above it, sovereignty starts to pay for itself.  

But sovereignty is about more than the model itself, and the foundations needed to deliver it have to be built at a national level. The UK can’t build sovereign AI capability without sovereign infrastructure, so models are only one part of the equation. Real sovereignty requires secure hosting, access to compute power, development environments, responsible AI expertise, and industry adoption. Without domestic compute capability, for example, organisations may find themselves training supposedly sovereign models elsewhere, undermining the very objective sovereignty is intended to achieve.  

What leaders should do next 

The challenge for leaders is not to wait for policy certainty around sovereign AI. Instead, organisations should start by identifying where trust, resilience, and control are genuinely mission-critical to their operations, and therefore where sovereign AI has the greatest value.  

This can involve starting with your use case inventory, and applying a simple three-question test to identify which workloads genuinely require sovereign AI. In most organisations, the answer will be fewer than ten. 

Next, for every workload that remains, organisations need to define the specific goal that they need sovereign AI to achieve. That could be protecting the continuity of a regulated service, reducing exposure, delivering a lower cost at scale, or providing a capability that isn’t offered by frontier models. If the workload does none of these things, it doesn’t belong in scope for sovereign AI. 

It’s also important to separate the different layers of a technology stack. Preserving portability at the model layer and maintaining control over data and infrastructure can help avoid unnecessary dependencies. Make it easy to switch between AI model providers now, before doing so becomes difficult or expensive. 

Finally, test the economics under pressure. This could look like modelling consumption at three times the current volume and checking whether your decisions still make sense. Token economics that look marginal today often look different as AI moves from experimentation into core business processes. 

Most importantly, organisations need to design for change. The AI market 18 months from now will be different from the market today as leading models, platforms, and providers will inevitably shift, new economics will emerge, and regulation will continue to evolve.  

Ultimately, the sovereign AI debate needs to move beyond the three misconceptions that often dominate it. Sovereignty is not simply about owning a model, committing to a single provider, or treating resilience as the only objective. Organisations need to be deliberate decisions about where control matters, where dependency creates risk, and where investment in infrastructure and capability will create long-term value.  

Related Articles

Back to top button