How To Build Real AI Sovereignty
Antoni Kozelski, CEO/founder of Vstorm, an agentic AI engineering firm helping companies implement agentic AI to transform their businesses.
gettyGartner forecasts that worldwide sovereign cloud spending will reach $80 billion in 2026, growing more than 35% from 2025. And yet the European Commission reports that the EU still depends on non-EU providers for over 80% of its key digital products, services and infrastructure. Money is pouring into the word “sovereign.” The dependence it is meant to fix has barely moved.
In my own work, I’ve found the reason is rarely the data center. It is ownership: which layers of the stack a company actually holds, which it rents and who can revoke them. Nearly every client who asks us for a “sovereign” system means something different by the word.
The instinct to treat sovereignty as a location (e.g., “Our data sits in-country, and therefore, we are sovereign.”) is the first and most expensive mistake leaders can make. Location is not control. The CLOUD Act lets U.S. authorities compel a U.S.-incorporated provider to hand over data held anywhere in the world, including in an EU data center. The postcode on the server changes nothing. What governs the system is the law that governs the company operating it.
Jurisdiction follows the company, not the data center. Data residency asks where information sits. AI sovereignty asks who holds legal and operational authority over it and whether a foreign entity could alter, disable or disclose it without consent.
No one has finished writing the rules yet. The EU’s Cloud and AI Development Act is still moving through negotiation, its sovereignty framework unsettled. Open standards for agentic systems are still being drafted; the Agentic AI Foundation, hosted by the Linux Foundation and launched in late 2025, exists to coordinate those standards, and my company contributes to that work.
When the standards are still forming, every architecture decision is a sovereignty decision, whether a team recognizes it or not.
I have worked with many organizations that believed they had bought sovereignty, but in reality, they had not. The failure modes I’ve observed are strikingly similar each time:
1. They treat data residency as sovereignty. A local region and a national label are often mistaken for control, but if the control plane, the billing or the support team answers to a foreign parent, the data is still reachable. This is what the industry now calls sovereignty washing.
2. They rent the intelligence layer without noticing. A company can own its prompts, its data and even its process logic and still not own the codebase the system runs on, paying indefinitely for the one layer that decides what the system is allowed to do.
3. They confuse model access with ownership. The models are available to everyone. Calling a closed model through an API is not the same as holding open weights you can run, move or swap on infrastructure you control.
4. They leave the exit untested. The real proof of ownership is whether you can leave. If taking your data and your logic to another provider means a rebuild, you did not own them; you were renting them all along.
Sovereignty is a spectrum, not a switch, and it’s a mistake to chase all of it at once. No single company controls the whole stack—the chips and the critical minerals beneath them are transnational and out of anyone’s sole reach. Real returns come from sequencing, not scale: Own the layers that would hurt most to lose, prove you can exit and then decide what to rent deliberately.
The order is simple: Start with the layers whose loss would cost the most: your data and the process logic that encodes how you run. Next, prove you can move before you scale. Export the data, swap the model and change the provider without re-architecting. Finally, own the deepest layers you reasonably can and rent the rest as a choice, not a default.
1. They are portable. Providers, models and infrastructure can change without a rebuild, but portability is the working test of agentic AI without vendor lock-in.
2. Jurisdiction sits with an entity the owner controls, not a foreign parent that can be compelled from outside.
3. The code and the process logic are held, not licensed indefinitely.
The cost of getting this wrong is now measurable. The Commission’s own target is to cut “EU reliance on non-EU technology providers from more than 80 percent today to 40 percent by 2030,” an admission of how deep the exposure runs. Under the EU AI Act, data-governance failures on high-risk systems carry fines up to €35 million or up to 7% of annual worldwide turnover. And there is a physical floor: The IEA projects data centers will draw around 1,000 TWh by 2030, roughly Japan’s entire electricity consumption. A system you cannot power is sovereign in name only.
The “sovereign” label is a commodity. What decides true sovereignty is which layers you actually own. Again, sovereignty is not the postcode on the server. It is the deed to the house.
Forbes Technology Council is an invitation-only community for world-class CIOs, CTOs and technology executives. Do I qualify?
