By Laas Nury"Sovereign AI" has become a sales term. Cloud providers use it for data centres in Frankfurt, model makers for open-weight releases, and consultants for almost anything European. For a company deciding where its AI should run, the label helps less than the questions behind it.
This article breaks sovereignty into four concrete questions, explains what the law does and does not require, and describes three set-ups that work in practice, from simple to strict. It ends with the questions to put to any AI supplier.
Four questions behind "sovereign"
When people say they want sovereign AI, they usually mean one or more of these:
- Where is the data? Stored and processed in the EU, or somewhere else, including when the AI model reads it.
- Under which law? A US company storing data in Frankfurt remains subject to US law. Under the CLOUD Act, US authorities can require US providers to hand over data they control, wherever it is stored.
- Which model do you depend on? A model you can only reach through one foreign provider's service, or one you could run elsewhere if you had to.
- Who can switch it off? If the provider changes its terms, its prices or its availability in Europe, what happens to the processes that run on it?
Different data and different processes need different answers. That is why "all of it, fully sovereign" is rarely the right goal.
What the law requires, and what it does not
The GDPR does not require EU hosting. Personal data may leave the EU under conditions: to a country or company with an adequacy decision, or with safeguards such as standard contractual clauses. US companies certified under the EU-US Data Privacy Framework count as adequate.
That framework is valid, but contested. The EU General Court upheld it in September 2025. The ruling was appealed to the Court of Justice in October 2025, and that appeal was still pending at the time of writing. The two frameworks before it, Safe Harbour and Privacy Shield, were both struck down by that same court. Companies that depend heavily on US transfers therefore usually keep a plan B.
The EU AI Act does not require EU hosting either. It regulates how AI is used, not where it runs.
So sovereignty is mostly a risk decision, not a legal obligation. It matters most for sensitive personal data, trade secrets, regulated information and processes you cannot afford to have interrupted.
Three workable set-ups
| 1. Mainstream, with EU storage | 2. EU-owned infrastructure | 3. Your own environment | |
|---|---|---|---|
| What it is | Large US AI and cloud services, with data stored in their EU regions and no training on your data | AI running at a European-owned cloud provider, with an EU model or an open-weight model | AI running on your own servers or private cloud, with an open-weight model |
| Where is the data | In the EU, with exceptions to check (support, overflow capacity) | In the EU | Where you put it |
| Under which law | US law applies to the provider | EU law | Your own control |
| Model capability | The most capable models available | Good; strongest on well-defined tasks | Good; depends on the hardware you run |
| Effort for you | Lowest | Moderate | Highest, unless a partner runs it |
| Suits | Most everyday AI use, without sensitive data | Sensitive personal data, confidential business data | Regulated, classified or highly sensitive work, or no internet access |
Most companies end up using two of these side by side: the mainstream set-up for everyday work, and a stricter one for the processes that touch sensitive data.
Set-up 1: mainstream, with EU storage
Large providers now offer EU data storage and contractual commitments not to train on business data. Microsoft, for example, has an EU Data Boundary for Microsoft 365 Copilot and states that prompts and business data are not used to train its models. Read the details, though: its own documentation notes that, under high demand, AI requests can be processed in other regions.
For most AI use without sensitive data, this is a reasonable and practical choice. The capability is high and the effort low.
Set-up 2: EU-owned infrastructure
Here the AI runs at a cloud provider owned and operated in Europe, with a European model or an open-weight model. The data stays in the EU, and a European provider without US parent companies is not subject to US disclosure laws such as the CLOUD Act. Check the ownership, not just the address.
The trade-off is capability and convenience. The most advanced proprietary models are not always available this way. For well-defined business tasks, such as reading documents, classifying requests or preparing bookings, a good open-weight model is often more than enough.
Set-up 3: your own environment
For the strictest cases, the model runs inside your own environment: on your servers, in your private cloud, or on equipment without an internet connection. Nothing leaves your control.
This costs more effort: hardware, updates, monitoring. It is the right answer for regulated or classified information, for sites without connectivity, and for companies whose customers demand it. It is also where the choice of an open-weight model matters most, because you will be running it yourself.
Questions to ask any AI supplier
- Where is our data stored, and where is it processed when the model reads it? Are there exceptions, for support or peak demand?
- Is the company subject to non-EU law that could give authorities access to our data?
- Do you train on our data, or keep it after processing? For how long?
- Which sub-processors are involved, and where are they?
- Which model runs underneath, and could the solution switch to another model if needed?
- What happens when we leave? Can we take our data, our rules and our configuration with us?
Put the answers in your AI register; AI governance on one page shows how.
Where to start
Sort your AI uses by the data they touch. Everyday work without sensitive data can run on mainstream services with EU storage. Processes with sensitive personal data or trade secrets belong on EU-owned infrastructure or in your own environment. Our platform Nightfall exists for exactly that second group: it runs AI where your rules require, on your own servers, in a government cloud or fully offline.
The goal is not maximum sovereignty everywhere. It is the right level for each process, chosen on purpose and written down.
