The Quiet Gift to the Cloud
A mid-sized European company uses ChatGPT Enterprise for internal research, Microsoft Copilot in Outlook and Word, Salesforce Einstein for sales analytics, and an AI-powered recruiting tool. Each of these tools processes data – contracts, customer communications, sales history, candidate profiles. Most of it lands on US servers, some flows into vendor training pipelines, and the rest is at minimum used to improve models that the next competitor will also subscribe to.
Nobody actively decided this. It happened because the tools are useful and the path of least resistance runs through default settings. The result: data that represents the company's actual capital becomes input to other people's systems – without anyone ever putting a price on that input.
Why Data Is Suddenly Different
Data was always valuable, but it was passive. It sat in databases, got compressed into reports, and informed decisions. With AI, it has become active – it trains models, it calibrates forecasts, it defines how software behaves.
That's exactly why its strategic weight is shifting. A model is only as good as the data it was trained on. Whoever owns unique data – data no competitor has – can build models no competitor can replicate. Conversely: whoever shares their data gives up that uniqueness.
This isn't theoretical. Every time a prompt is sent to a major vendor without crystal-clear contractual control over data flows, the fine print decides whether the data is only processed or also used for model improvement. "Anonymised" and "aggregated" are elastic terms in this context.
The Three Layers of Data Sovereignty
Data sovereignty splits into three layers that often get confused in practice:
Location. Where is the data physically stored? EU, US, Asia? This question dominates GDPR debates but is only part of the problem. Even data sitting in a Frankfurt data centre can be accessed by a US vendor under the Cloud Act.
Usage rights. What is the vendor allowed to do with the data? Processing for the requested purpose is standard. But: training their own models? Benchmarking? Aggregated insights to third parties? Contract clauses are decisive here, and they rarely get read in standard subscriptions.
Model control. Who controls the model that runs on the data? If the model sits with the vendor, the vendor controls what it can do, when it's updated, whether it's discontinued. Model control is the layer most often overlooked – and strategically the most important.
Real data sovereignty means controlling all three layers. Not for every piece of data, but for the ones that decide competition.
What's Actually at Stake
Three concrete risks rarely discussed openly:
Training contribution to competitors. When a vendor builds a generic model that all customers use, every customer shapes that model with their data. Best case, all customers benefit together. In reality, vendors and the competitors who later subscribe to the same service benefit most. Whoever fed data in early helped train the competition.
Loss of company-specific patterns. Your data contains patterns that are specific – niche customer behaviour, regional quirks, process idiosyncrasies. Those patterns are precisely what makes a competitive advantage. When they flow into a generic model, they get smoothed out. The advantage becomes an average that everyone can use.
Geopolitical dependency. US vendors dominate the AI market. That concentration becomes a risk the moment political decisions restrict access, dictate prices, or tighten export controls. Building a business model on a single foreign platform means accepting a concentration risk that has nothing to do with software quality.
What Data Sovereignty Does Not Mean
Data sovereignty doesn't mean self-hosting everything, training your own LLMs, or avoiding the cloud. That would be expensive activism. It means: making deliberate decisions about which data takes which journey.
An email to a supplier can run through any reasonable AI tool. A roadmap discussion or a contract draft with strategic content should not. The question isn't "cloud yes or no", but "which data belongs in which environment".
In practice, that means tiered architectures: public models for low-stakes work, EU-hosted models for sensitive content, on-premise or edge models for the data that carries the business. That tiering is technically feasible and economically viable today – it wasn't two years ago.
What Changed Technically
Three developments make data sovereignty practical now:
Open models became usable. Models like Llama, Mistral, or DeepSeek now reach quality that two years ago was the preserve of frontier models. They can be self-hosted – in your own data centre, in a sovereign cloud, or at the edge.
Hardware became affordable. Inference hardware that runs a 70-billion-parameter model locally now costs five figures, not seven. For companies whose data carries enough value, that investment is easy to justify.
RAG and fine-tuning matured. Instead of training a model from scratch, you can today fine-tune an open base model on your own data or wire it into your own knowledge base via retrieval. The effort is measured in weeks, not years.
What Companies Should Do Now
Four concrete steps:
Build a data inventory. What data does the company process? Which is strategic, which is operational, which is replaceable? Without that clarity, every further discussion about sovereignty is shadow-boxing.
Map data flows. Which of this data leaves the company through which tools? For each AI tool: what does the contract say, what do the terms say, what actually happens? The answers are often surprising.
Isolate strategic data. For the data that decides competition: build tiered infrastructure. No generic AI tools, no uncontrolled cloud upload, clear contracts with explicit non-training clauses.
Build internal capability. Data sovereignty requires technical skill many companies lack. That needs building – internally or through partners who actually understand the difference between inference, fine-tuning, and retrieval augmentation.
At nh labs, we've seen this awareness grow sharply over recent months. Requests for EU-hosted architectures, on-premise inference, and clearly documented data flows are no longer the exception – they're standard in first conversations.
Bottom Line
Data sovereignty isn't a legal detail or a compliance exercise. It's the question of whether a company controls its competitive base in the AI era or hands it to third parties. Vendors will keep arguing that everything is secure, anonymised, and regulated. Some of that is true – but sovereignty over your own data isn't a leap of faith, it's an architectural decision. Companies making that decision now lock in the lead their data enables. Those postponing it hand that lead away, quarter by quarter.