Thoughts on Digital Sovereignty

By · Updated

We talk about AI sovereignty as if it's a binary choice - build everything yourself, or hand over control. The real question is more concrete: if your AI provider were switched off tomorrow, what in your organization would stop working? For most companies, and most countries, the honest answer is "more than we'd like, and we don't fully know what." Sovereignty, done well, isn't about independence from the market. It's about knowing exactly where you depend, and designing on purpose so that no single decision made by someone else can stop you.

Key takeaways

  • Digital sovereignty is not autarky: it means choosing deliberately where to depend, and on what terms.
  • The AI stack has five layers: infrastructure, data, models, applications, and people and governance.
  • For each layer and use case, choose a posture: Own, Steer or Rent.
  • Only what is both critical and hard to replace justifies the cost of owning outright.
  • Everything you rent needs an exit path that has actually been tested.

What is digital sovereignty?

Sovereignty is not autarky. Nobody needs to train their own foundation model, run their own data centers or fabricate their own chips to be sovereign - and almost nobody could, even if they wanted to. What sovereignty actually means, for a country or a company, is deciding, layer by layer, what you own outright, what you steer while still depending on someone else's technology, and what you simply rent because it isn't worth the cost of controlling - as long as you keep a tested way out of everything you rent. Seen this way, nobody is choosing between dependency and independence. Everyone is choosing where to depend, and on what terms.

The AI Stack: five layers of sovereignty

We can distinguish five distinct layers in the AI stack. Each of those has their own sovereignty questions that need to be considered.

Diagram of the five layers of the AI stack: infrastructure, data, models, applications, and people and governance


The first layer usually gets the most attention. The Infrastructure layer has a physical side (energy, buildings, chips) and a logical/virtual component (operating a digital cloud). In this layer the question to ask is which jurisdictions can legally compel access to your systems, or switch them off.
Above that sits your data: where it lives, in what format, and whether you could actually take all of it with you in a form you could use elsewhere.
Then come the models themselves - foundation models and whatever you've fine-tuned or built on top of them - and whether you could swap one out without rebuilding every application that depends on it.
Applications are where your business logic actually lives: is it encoded in your own systems, or hard-wired into a vendor's platform?
Finally, people and governance: the in-house skills and decision rights to make informed choices about all of the above. This last layer quietly decides the rest. An organization without its own expertise cannot be sovereign at any other layer, no matter how carefully it architects its infrastructure.

Digital sovereignty for nations

Nations may have different concerns in relation to sovereignty. Westphalian or legal sovereignty is the state's exclusive authority over its territory, extended to data and networks (data localisation, lawful access, "cyber sovereignty"). The UN Charter, Art. 2(1), agreed by all 193 UN member states, talks about sovereign equality of states. Strategic autonomy is the capacity to act alone where necessary and with partners where possible; reducing dependency on non-EU hyperscalers, chips and platforms. This has been the EU's framing since 2019. Developmental or decolonial sovereignty is often the framing in the Global South, where nations want to avoid "data colonialism", keeping value from local data local, building local-language AI, and having a voice in global rule-making.
The UN frames sovereignty as sovereign equality plus capacity: every state has the right to govern technology on its territory, but that right is hollow without compute, data, skills and a seat at the table.

Digital sovereignty for organizations

Digital sovereignty as enterprise risk: where an organization depends on AI providers For a company, sovereignty becomes practical the moment you stop asking "build or buy" for everything at once, and instead set a deliberate posture, layer by layer and use case by use case: own what you build and fully control, because it defines your competitive edge; steer what you depend on, by keeping real leverage through architecture, contracts and in-house expertise even while using someone else's technology; and rent the rest freely, as long as the exit is real and has actually been tested. Only what is both critical to you and hard to replace is worth the cost of owning outright. Everything critical but replaceable gets steered. Everything else gets rented. In practice, this comes down to a few habits worth building into any AI decision: put an abstraction layer between your applications and any model, so that switching providers is a configuration change rather than a rebuild; keep your data in open formats, in a known location, contractually excluded from training someone else's model, with an export you have actually tested; never sign an AI contract without an exit plan and a switching-cost estimate written down beforehand; invest in capability before control, since ownership without expertise is just a more expensive dependency; and keep your own register of AI systems and risk ratings, rather than inheriting whatever the vendor hands you. None of this requires building your own AI. It requires deciding, on purpose, what you can't afford to lose control of - and making sure everything else has a door marked exit.

Determining the right sovereignty strategy

Here is a simple framework to help you determine your sovereignty strategy. For each layer in the AI stack, and for each important use case, set one of three postures:
Own: you build or host it and control it fully. Expensive; reserve it for what defines you.
Steer: you use others' technology but keep control through architecture, contracts and in-house expertise (open-weight models you host, EU-hosted cloud, strong exit clauses).
Rent: you use the market's best option freely, with a tested exit path.
The posture follows from two questions: how critical is this to your business or your risk, and how easy is it to replace?

AI dependency matrix: criticality versus replaceability, mapping to the Own, Steer and Rent postures


Only the top-left quadrant justifies the cost of owning; everything critical but replaceable is steered, and the rest is rented with an exit.

Design Principles

As you start designing the AI approach for your organization there are a few principles that hold true no matter what the use case. You might want to adopt them as general principles in your strategy document.

  1. Portability by default
    Put an abstraction layer between your applications and any model, so switching is a configuration change, not a project.
  2. Your data stays yours
    Open formats, known location, contractual ban on training with it, and a full export you have actually tested.
  3. Exit before entry
    No AI contract without an exit plan and an estimated switching cost, written down before signing.
  4. Capability before control
    Invest in people first. Ownership without expertise is just a more expensive dependency.
  5. Govern it yourself
    Your own register of AI systems and risk ratings, aligned with the EU AI Act, rather than inheriting the vendor's.

Sovereignty-by-Design Scorecard

Screenshot of the AI Sovereignty Scorecard, a sovereignty-by-design assessment tool
You can do a quick assessment of your own AI Sovereignty readiness using my AI Sovereignty Scorecard™. Pick a specific usecase, an AI application in your organization. Answer a few questions in the online scorecard and your score will be calculated automatically. This will give you a quick assessment of whether you are exposed to risks for that application, whether the risks are managed, or whether you are fully sovereign.

Frequently asked questions

What is digital sovereignty in AI?

It is not independence from the market. It means deciding, layer by layer, what you own, what you steer and what you rent, while keeping a tested way out of everything you rent.

What are the layers of the AI stack?

Infrastructure (energy, buildings, chips and cloud), data, models, applications, and people and governance.

What is the Own, Steer, Rent framework?

A way to set a posture for each layer and use case. Own what defines you, steer what you depend on through architecture, contracts and expertise, and rent the rest with a tested exit. The choice follows from how critical the capability is and how easy it is to replace.

How can an organization reduce AI vendor lock-in?

Follow five design principles. Portability by default: put an abstraction layer between your applications and any model, so switching is a configuration change, not a project. Your data stays yours: open formats, a known location, a contractual ban on training with it, and a full export you have actually tested. Exit before entry: no AI contract without an exit plan and an estimated switching cost, written down before signing. Capability before control: invest in people first, because ownership without expertise is just a more expensive dependency. Govern it yourself: keep your own register of AI systems and risk ratings, aligned with the EU AI Act, rather than inheriting the vendor's.