The only sovereign AI is the one you can walk away from
Pursuing AI sovereignty is complex, but the goal is clear: Control, portability and provider diversity help enterprises avoid dependency and adapt if -- or when -- terms change.
By
Chris Gilmour, CTO at Axians UKGuest Contributor
Published: 19 Oct 2026
Most businesses had their first wake-up call to AI sovereignty needs when Anthropic pulled access to Fable and Mythos in June this year. In compliance with U.S. Department of Commerce export controls, Anthropic cut off all non-U.S. users at a stroke, and access only returned on July 1, once the controls were lifted.
For three weeks, organizations that had begun building on those models lost them. A government had simply changed the terms of use, leaving users sharply reminded that this could happen at any time.
The effect of that keeps rippling outward. This month, the White House's Office of the National Cyber Director asked OpenAI and Anthropic to keep new frontier models away from the U.K.'s AI Security Institute until the U.S. government finishes its own review. That means Britain's ability to independently test the most capable systems now depends, at least in part, on Washington's timetable. At the UN General Assembly, India backed national control over AI, and Germany's foreign minister pointed out that most global AI patents are held by companies in only two major countries.
It's hard not to see the growing demand for sovereign AI as justified. Even if the definition is a mess.
Ask 10 CIOs, get 10 definitions
Ask 10 CIOs what sovereign AI means, and you'll hear data residency, in-country compute, open-weight models, protection from foreign legal reach or simply which country built the model in question. Governments are no clearer. In some capitals, sovereign is shorthand for anything but American; in others, anything but Chinese.
In the U.S., defense contractor L3Harris argues trusted AI must be built domestically, framing domestic control over data, training and updates as a form of deterrence. In Europe, the Tech Sovereignty Package brings together the Chips Act 2.0 and the Cloud and AI Development Act alongside an EU Open Source Strategy. The starting points differ, but the instinct is the same: control.
Sovereignty is a stack, not a postcode
The question of where the data sits is now table stakes for the sovereignty conversation. The more sophisticated organization will weigh data location, the model's country of origin and local regulations. But a maximalist "true" definition of data sovereignty relies on a mesh of the above factors, along with who manufactures and controls the GPUs, accelerators and networking underneath; the legal obligations a provider carries from its country of origin; whether sensitive data stays under appropriate control; and what entities decide how models are trained, governed and deployed.
Sovereignty means ownership, accountability and confidentiality. Each organization has to translate those into what they mean for its own business.
Full sovereignty would require a locally owned stack at every one of those layers: data centers, compute, storage, switching, data center interconnect and WAN, middleware and the AI platform itself. That stack is unlikely ever to be competitive with the global alternatives. That makes compromise almost inevitable, as it has been across IT for decades. The real question is where you make it.
Don't swap one dependency for another
Pursuing sovereignty for its own sake carries a risk. You can replace dependence on a handful of global providers with dependence on a single national one -- as we see in France's push toward Mistral. But that isn't resilience so much as a single point of failure with a flag wrapped around it.
The U.K.'s program appears to recognize this, with the chair of the U.K.'s Sovereign AI Unit having warned that Britain should not shun U.S. tech help in its AI drive. Ministers have even underlined that the much-vaunted sovereign AI fund is not designed to replace foreign cloud providers or achieve total U.K. self-sufficiency but to reduce strategic dependence. The bigger question is ecosystem resilience: avoiding excessive dependence on any one model, provider, cloud platform or AI supply chain.
The real test: Can you walk away?
That brings me to a simpler test. Could your business keep running its AI if any one model, provider or government changed the terms tomorrow? Sovereignty should mean keeping strategic choice and the ability to adapt. In practice, that comes down to two disciplines: model portability and rightsizing.
Treat the ability to switch models or providers as a resilience capability, just as you treat disaster recovery. If swapping a model means re-architecting half your applications, you aren't sovereign because you aren't in control. Then it's a case of accepting that not every process needs the biggest frontier model. Smaller, specialized models are often a better fit for a given process and risk profile, as industry-specific LLMs are already showing. They also widen your choice of where, and on whose infrastructure, they run.
This is urgent because governments increasingly set the terms of access, often out of public view. Under the new U.S. framework, developers can grant the government access to certain frontier models for up to 30 days before making them available to other trusted partners. If a review process you can't see can shape when a model reaches you, portability matters and every AI investment should be judged on "swap-out-ability."
The pressure will come through contracts, not ideology
With all the complexities in achieving AI sovereignty laid out, it's important to remember that the demand is being driven by customers. Public sector contracts will, in all likelihood, mandate a level of AI sovereignty, just as they already set data sovereignty requirements, and that requirement then cascades down supply chains. If you trade with the government or critical infrastructure, expect to prove it.
Europe is moving first, as it often does on regulation. The Cloud and AI Development Act's four-level sovereignty framework sets criteria that cloud providers must meet to serve the EU public sector and, potentially, beyond. The first tier applies from February 2028, with the highest tier mandatory by August 2029. The logic is already reaching AI inference, because sending data into a foreign-hosted model's inference pipeline can recreate the very extraterritorial exposure that sovereignty rules are meant to prevent.
The U.K. is heading the same way. The first competitions under the £100 million Sovereign AI R&D Procurement Scheme opened Aug. 31, covering the NHS, defense, computing efficiency and AI agent security. Organizations that can evidence sovereignty across the stack will find doors opening. Those who can't may find them closing.
Sovereignty is a trust problem
Organizations succeeding with AI tend to have a clear vision of what AI means for the business: the outcomes they want and a roadmap for how AI will change employee experience, customer engagement, efficiency and growth -- backed up by a recognition that AI success is built on trust.
On the ground, most organizations lack a clear vision, don't trust AI, or both. Sovereignty sits at the heart of that trust question: knowing who owns your AI, who controls it and who can see into it. As models grow more capable and organizations move toward greater AI, integrating those with strong governance, identity and security foundations will win out.
Appropriate sovereignty, not absolute sovereignty
Sovereignty means ownership, accountability and confidentiality. Each organization has to translate those into what they mean for its own business. The key question isn't whether you need absolute sovereignty. You won't get it. At least not without an eye-watering price tag.
So, the question becomes what level fits your risk profile, regulatory obligations and strategic objectives. Which layers of my AI stack are exposed? Which of my contracts will soon demand proof of sovereignty? Could I switch models or providers in weeks rather than years? Am I sovereign … enough? Sovereign AI isn't a destination. It's the ability to walk away.
Chris Gilmour is a CTO with more than 20 years of experience working in the networking industry at international ISPs and award-winning systems integrators. He has an extensive network engineering background and has been instrumental in design and management of complex international network and security infrastructure solutions for service providers, research and education institutions, and enterprises.
