The Sovereign Stack Needs a Sovereign Model Layer

Europe’s sovereignty debate has concentrated on where data sits and which law governs it. The model that reads that data belongs in the same argument, and it is often the layer an organization controls least.

Earlier this year we asked a plain question: under what law does your AI live? The answer rested on three tests: where the model runs, whether your data trains someone else’s system, and whether everything can be audited.

Those tests still stand. This series has since covered ownership, cost and capability. This piece adds a question the original argument left open: who decides whether your model keeps running, and in what form?

What happened in June

On 12 June 2026, the US government directed a leading American AI provider to suspend access to its two newest models for all foreign nationals, citing national security. To comply, the provider switched both models off for every customer. Its other models stayed available, and within three weeks the controls were lifted and the provider began restoring access.

The reasons are a matter for others. For a European enterprise, the mechanism is what counts. The data had not moved and nobody’s compliance position had changed, yet a decision taken in one jurisdiction reached customers worldwide at once. No data residency clause could have prevented it. Reuters reported soon afterwards that large European companies were moving faster to spread their AI across several providers.

A stack is as sovereign as its weakest layer

Think of enterprise AI as a stack: infrastructure, then data, then the model, then the applications people use. Most of the sovereignty debate has been about the first two.

The European Commission’s Cloud Sovereignty Framework already looks further up. For its own cloud procurement, the Commission asks how far AI models are developed, trained, hosted and governed under EU control, and whether a service can keep operating if vendor support is withdrawn. A bid that misses the required level on any one of eight objectives is rejected.

The principle holds beyond public tenders. An organization can keep its data in Europe, on European infrastructure, and still rely on a model that someone else can change or withdraw.

What open models change

An open model is one whose trained files are published, so you can download it and run it on infrastructure you control. That shifts three things.

Continuity. A model on your own infrastructure keeps running. No provider can switch it off, retire it or reprice it, and a new version arrives when you choose to adopt it.

Jurisdiction. When you run the model yourself, no model provider handles your requests. That removes one party a foreign authority could compel.

Accountability. The version you tested stays in production until you change it, which makes a regulator’s questions easier to answer.

Three checks before you call it sovereign

Open is a starting point. A model still has to pass three checks.

Where it runs. An open model on a US-controlled cloud stays within the legal exposure we set out in that earlier article. Sovereignty comes from the model and the infrastructure together.

What the license allows. “Open” covers a wide range of terms, and some limit how a model can be used. Review them as you would any vendor contract.

Where it came from. Many of the strongest open models come from outside Europe, several from China. Run on your own infrastructure, such a model sends no data back to its maker, but fair questions remain about its training and future procurement rules. Test and document each one for its intended use.

What this looks like by industry

In financial services and insurance, DORA requires exit strategies for technology services supporting critical or important functions. An open model already tested on your own infrastructure gives that strategy a working fallback.

In healthcare, logistics and the public sector, continuity carries as much weight as confidentiality. A clinical documentation process, a customs workflow or a citizen service cannot be left waiting on a decision made elsewhere.

Decide by what is at stake

Sovereignty need not mean local models everywhere, or European technology only. Open-ended, low-sensitivity work can stay with a hosted frontier model. For work you cannot afford to lose or cannot send outside, keep the model under your control, and give anything hosted a tested alternative.

Running models yourself brings testing, monitoring and access control in-house, and this is where the platform matters. Sirma.AI Enterprise gives access to more than 500 models, including open models through its integration with Hugging Face, and routes each workload by sensitivity, cost and performance, on-premise, in a private cloud or on a European sovereign cloud. If a model becomes unavailable, the workload can move to another without rebuilding what sits on top of it.

Where this leaves you

Under what law does your AI live? The answer now has to cover the model as well as the data. For every model your operations depend on, three things are worth knowing: who can change it, who can withdraw it, and what happens to your business on the day they do.

If you want to know which of your workloads depend on a model you do not control, we can map them with you and show what a fallback would look like.

Learn more about our platform: Sirma.AI

Sources: