A series of high-profile incidents involving frontier AI providers has left businesses asking a question that, until recently, many hadn’t seriously considered. What happens when the technology you’ve built your operations around is no longer entirely under your control?
First came Anthropic’s shutdown, ordered by the US Government with just 90 minutes’ notice. Then came Grok Build’s privacy scandal. More recently, OpenAI admitted to an agent autonomously breaching Hugging Face, raising fresh questions about governance as AI systems become more capable.
Individually, each incident is significant. Taken together, they point to something bigger. They remind us that AI sits within political systems, commercial relationships and regulatory environments that organisations often have very little influence over.
The conversation around AI risk has understandably centred on privacy, compliance and data residency. Those issues still matter, but they’re no longer the whole story. The bigger question now is dependency.
Understanding What Sovereignty Is — and What It Isn’t
I’ve noticed the conversation around sovereignty changing over the past year. Customers would typically ask where their data would be stored. Today, they’re far more likely to ask who actually controls the systems their AI runs on, and what happens if they suddenly lose access to them. That’s a more difficult question to answer, because sovereignty extends well beyond geography.
At its core, sovereign AI is about ownership: being able to deploy AI without introducing dependencies that could threaten resilience, continuity or long-term flexibility, regardless of where a model was originally developed. One way I explain it is by comparing AI agents to employees. They’re increasingly becoming digital members of the workforce, taking responsibility for tasks that were previously carried out by people. If an organisation rents that workforce from a single AI provider, then it’s also outsourcing a critical part of its operational capability.
That doesn’t mean every organisation needs to own every part of the AI stack. Power and physical infrastructure sit at the bottom, followed by compute, then the models themselves and finally the applications running on top. Every layer can be owned by a different organisation. At one end of the spectrum, data remains in-country while everything else is controlled elsewhere. Further along, models run locally, but on infrastructure somebody else operates. Further still, organisations run the models themselves. At the deepest end, they control the infrastructure too. Where an organisation chooses to sit on that spectrum depends on what it cannot afford to lose.
A government running classified workloads will naturally need more control than most commercial businesses. A hospital doesn’t necessarily need to own a data centre, but it does need confidence that the systems supporting patient care can’t suddenly become unavailable because of decisions made by a third party thousands of miles away.
Sovereignty Has to Be Verifiable
The more organisations talk about sovereign AI, the more important it becomes to distinguish genuine sovereignty from clever marketing. Not every provider is transparent about which layers of the stack it actually controls, and if enough vendors begin describing their products as sovereign, the word could quickly lose its value in much the same way that sustainability claims have in other industries.
That’s where the idea of “sovereignty washing” comes in. A solution is marketed as locally controlled because the contract is domestic, the support team is based in-country and the data never leaves national borders, while the infrastructure underneath it is still operated by a foreign provider. A UK organisation could buy a sovereign AI service from a British supplier and still find that the systems it depends on ultimately sit on US infrastructure.
That doesn’t automatically make it the wrong choice. For many organisations, it may be the right balance between cost, capability and control. But it should be a conscious decision.
The Challenges of Renting Versus the Benefits of Owning
It’s easy to see why US frontier AI providers became the default. They offered the most capable models, invested heavily in infrastructure and consistently moved faster than the rest of the market. Combined with growing geopolitical concerns around Chinese AI, they became the natural choice for organisations looking to deploy AI at scale.
That momentum has spread well beyond early adopters. Even in highly regulated industries, where new technology is typically scrutinised more closely, adoption has accelerated. 75% of financial services firms actively use AI. So do 79% of legal professionals, while 65% of clinicians say their use of organisation-provided AI tools has increased over the past year. For many of these organisations, choosing a US provider simply made commercial sense.
The conversation is becoming more nuanced as AI moves deeper into day-to-day operations. Cost is part of the picture. Token prices continue to fall, but enterprise AI spending keeps rising as usage outpaces price reductions, with 93% of organisations already exceeding their AI budgets. Resilience is another. If an external provider suffers an outage, changes its commercial terms or faces government intervention, organisations running critical workloads have little control over the outcome.
Owning more of the AI stack changes both equations. Costs become more predictable as usage grows, while organisations gain greater visibility into how their systems operate and far more control over what happens when circumstances change.
Why a Good Exit Strategy Requires the Right Vendor Partnership
Sovereignty isn’t something organisations can switch on overnight. For regulated businesses, moving AI workloads typically takes three to four years, and very little of that time is spent moving technology. Most of it goes into governance, procurement and the operational work needed to move critical processes without disrupting the organisation.
That’s why the planning has to happen well before anyone needs to leave a provider. While 72% of businesses are factoring sovereign AI into their 2026 roadmaps, only around three in ten have a detailed strategy, budget or migration plan in place. There’s clearly an appetite to reduce dependency, but far fewer organisations have worked through what that would actually involve if they had to make the move.
That starts with understanding which workloads matter most, what an exit would actually cost and whether migration plans have been tested before they’re ever needed. That process also changes the way they evaluate vendors. Three questions usually tell you most of what you need to know. Can sovereignty claims be verified independently, whenever the customer chooses? Does that verification extend across every layer of the stack, or only the one named in the contract? And if the provider decided to, could it access the customer’s data? If the answer to that final question is yes, then what’s being offered is a policy rather than a guarantee.
