Sovereignty Should Be a Moat, Not a Wall

Europe debates where AI data sits. It should debate who owns the knowledge.

During my recent business trip to Europe, I listened to Satya Nadella’s conversation with Larry Fink at Davos. As someone managing an AI platform team in European enterprise software, the tension is familiar. On one side: the company-wide push for every employee to adopt AI. On the other: compliance standards still catching up with the technology — thorough by design, but not always easy to translate into day-to-day practice.

These two pressures don’t cancel each other out. They shape each other.

Sovereignty is a moat, not a wall

In Europe, sovereignty usually means defense. Keep data within our borders. Control where it flows. The instinct is protective. But there’s a difference between data sovereignty and knowledge sovereignty — and most conversations confuse the two.

Data sovereignty asks: where does our data sit? Knowledge sovereignty asks: who controls the intelligence built from that data?

Nadella frames it directly:

“If your firm is not able to embed the tacit knowledge of the firm in a set of weights in a model that you control — by definition, you have no sovereignty. That means you’re leaking enterprise value through some model company somewhere.”

When an AI feature is just a prompt passed to an API and a response displayed back, the intelligence lives in the model — not in the product. The provider captures the value. The product is a shell.

European companies sit on deep domain expertise — German industrial precision, French luxury craft, Nordic fintech. The question isn’t whether to use AI. It’s whether the AI reasons on your terms, using your intellectual property.

And the defensive posture carries its own risk. As Nadella put it:

“Europe actually should be much more concerned about access to their industrial companies, their financial services companies, of data from the US and the rest of the world — as opposed to just thinking that somehow by protecting Europe you’re going to be competitive.”

Most internal AI solutions today — including many of ours — are still thin wrappers around model APIs. The prompt goes in, the response comes out. That’s a reasonable starting point. But it means the intelligence lives in the model, not in the product. The next step is embedding domain knowledge into the reasoning process — retrieval-augmented generation and structured prompts that encode institutional expertise. That’s where sovereignty becomes real.

When concerns arise about data leaving company infrastructure, the natural response is caution — sometimes it’s easier to restrict than to find a solution. The challenge is that restrictions also pause learning. One path worth exploring: architectures where sensitive data stays within controlled boundaries while still benefiting from model capabilities. Private deployments, on-premise inference, careful data classification. The goal is a setup teams can trust and build on, not one they have to work around.

Multi-model over mono-model

The most common question I hear from teams: “Which model should we use for this?”

Nadella’s answer from the talk is clear: it’s a multi-model world. The right question is not which model to pick. It’s how to orchestrate them.

He breaks it into three capabilities:

• Orchestration — routing tasks to the right model for the job
• Context engineering — feeding your domain knowledge into the reasoning process
• Distillation — training your own smaller models from the reasoning traces of larger ones

The combination is where your IP lives. Not in any single model, but in how you wire them together with your data.

”Can I bring in all the models — closed source, open source, build my own model — orchestrate them and feed it my data to change the trajectory of some outcome I care about? That’s it. That’s the entire picture.”

Our team manages a platform with access to many different LLMs, but real usage scales to one or two models. That’s common. The value a platform should consider bringing to users isn’t more model choices — it’s helping them focus on context engineering and domain knowledge. Which data feeds the prompt. How the output maps back to a business process. That’s where the leverage is.

Flattening the hierarchy

Nadella made a point about how AI changes information flow inside organizations. In traditional structures — especially common in European corporations — information flows up, decisions flow down. Senior people hold context. That’s how authority works.

AI redistributes access to information and domain knowledge. When anyone in the organization can query a reasoning engine trained on company data, the bottleneck shifts. The scarce resource is no longer access to information. It’s knowing which decisions should be automated, which should be delegated to AI, and which need collective judgment from both humans and machines.

This changes what management looks like. Less gatekeeping information, more curating context. Less reviewing outputs, more defining the constraints and intent that shape them.

Compliance reviews used to follow a sequential approval chain. A senior supervisor reviewed every AI integration for data handling, model behavior, and regulatory alignment. The queue grew. Teams waited.

We explored a different approach: codifying compliance criteria into checklists and automated guardrails so teams could self-assess. Senior supervisors shifted to reviewing edge cases rather than routine deployments. The expertise didn’t disappear — it got encoded into the process and became accessible to more people.

Conclusion

Compared to Singapore or Asia, Europe has always felt tranquil and sophisticated to me — a place where history is distilled into architecture, art, and philosophy. Modern civilization built on deep foundations. As AI reshapes how companies think, build, and compete, the question for Europe is the same one it has faced before: follow fast, or lead differently?