Digital sovereignty for AI: how to make it verifiable
Digital sovereignty for AI only exists once you can prove where your models run, which data leaves your premises, who used what and how you can leave the provider again. It becomes verifiable through four criteria: location of processing, data flow, complete logging and the replaceability of the model.
For decision makers in brief
- Situation: 71 % of companies use US cloud providers, yet only 8 % would prefer them (Bitkom Cloud Report 2026, from the German digital industry association). No other gap in our industry report is larger.
- Effort: The starting point is not a technology decision, but a data classification. Which data may leave your premises determines architecture and costs more than any choice of model.
- Cost: Whether self hosting or billing per request is cheaper depends on usage volume. You can calculate this with your own values instead of estimating it.
- Risk: If you leave the question open, you lose control not through a decision, but through tolerance. In 39 % of manufacturing companies, AI is merely tolerated (ZEW 2026, the Leibniz Centre for European Economic Research).
The sovereignty gap in figures
The Bitkom Cloud Report 2026 surveyed 603 companies with 20 or more employees, between calendar weeks 14 and 20 of 2026. The result describes a clear tension between what companies operate and what they want to operate.
| Metric | Value | Source |
|---|---|---|
| Use US cloud providers | 71 % | Bitkom Cloud Report 2026 |
| Would prefer US providers | 8 % | Bitkom Cloud Report 2026 |
| Use German providers | 53 % | Bitkom Cloud Report 2026 |
| Would prefer German providers | 91 % | Bitkom Cloud Report 2026 |
| Consider Germany too dependent on US clouds | 85 % (2025: 78 %) | Bitkom 2026 |
| Are rethinking their cloud strategy | 64 % (2025: 50 %) | Bitkom Cloud Report 2026 |
| Accept disadvantages for processing exclusively in Germany | 37 % (previous year: 27 %) | Bitkom Cloud Report 2026 |
| See no equivalent European alternatives | 43 % | Bitkom Cloud Report 2026 |
| See themselves as dependent on non-European digital providers | 68 % | DIHK 2026 (Association of German Chambers of Commerce and Industry) |
The figure that counts is not 91 %, but 37 %. Preferences are cheap, willingness to compromise is expensive. 37 % would use a cloud with data processing exclusively in Germany and accept disadvantages for it, an increase of ten percentage points in twelve months. The counterforce is the 43 % who see no equivalent European alternatives. The market for private AI and AI operated in the EU is therefore limited less by demand than by supply.
At the same time, demand is growing. 42 % of companies use cloud-based AI services today, and 69 % plan to do so within five years. Dependence on US clouds will therefore increase if nothing changes in the architecture. The topic has reached politics: on 18 March 2026, the German federal government adopted a data centre strategy. By 2030, data centre capacity is to at least double and AI and HPC capacity to quadruple, and one field of action explicitly addresses sovereign cloud solutions.
Why AI sharpens the sovereignty question
With conventional office software, the issue is mostly documents and emails. With AI in manufacturing, it is about what makes up the core of a company. A knowledge assistant that gives maintenance technicians answers with source references accesses design data, cost calculations, customer prices, formulations and incident history. No other use case touches so much sensitive material. This is why digital sovereignty for AI weighs more heavily than for standard software.
Accordingly, the topic ranks high among the obstacles. 77 % of companies name data protection requirements as the number one obstacle, even ahead of the shortage of skilled staff (Bitkom 2026, n=604). Our article on scaling AI pilots explains why data access and integration so often become the bottleneck in practice.
The ZEW survey from spring 2026 shows how companies are responding today. In manufacturing, AI is explicitly prohibited in 8 % of companies, only 30 % actively provide AI applications to their employees, and 39 % merely tolerate its use. Tolerance is not governance. Where employees use AI without rules, an approved tool or data classification, exactly the data that companies fear losing ends up in public services. A ban does not solve the problem, it pushes it into shadow IT.
Verifying digital sovereignty for AI: four criteria
Sovereignty is often claimed in offers and rarely proven. The industry report describes what it means operationally. From this, we derive four criteria for verifiable sovereignty that you can apply to any provider and to your own architecture.
| Criterion | Requirement according to the industry report | Question for the provider |
|---|---|---|
| Location of processing | Models run on your own hardware or with processing exclusively in the EU | Where exactly does the model run, and who has physical and administrative access? |
| Data flow | No transfer of design data, cost calculations and customer prices to third parties | Which data leaves the system, to whom, and is this excluded by contract? |
| Logging | Every use is logged so that audits remain possible | Who made which request and when, and can we analyse this ourselves? |
| Replaceability | The model is replaceable so that the dependency does not simply shift | Can we take model, data and configuration with us, and in which format? |
The audit trail
The third criterion decides whether sovereignty is provable or merely promised. A reliable audit trail answers, for any point in time, who used the system, which data was processed and which answer the system gave. This turns sovereignty into a property that an auditor, a data protection officer or a customer can trace. Our page on GDPR compliant AI shows how we document logging, locations and evidence.
The exit plan
The fourth criterion is the one most often overlooked. If you switch from a US provider to a European one, yet are again tied to a single model and a proprietary interface, you have only shifted the dependency. An exit plan sets out before the contract begins in which format models, data and configuration will be handed over and within what period. This is why our process for switching AI provider is published on our website.
A documented case from industry
A documented case from the chemical industry, in which we were not involved, shows that AI on your own infrastructure works at group scale. According to the company, Covestro runs its company-wide assistant COVA on its own infrastructure. The company reports 4,500 weekly users and operating costs of one tenth of a standard solution (Covestro 2025, company statement). The figures have not been independently verified. They do show, however, that self hosting need not mean giving up convenience or cost efficiency.
For the German Mittelstand, mid-sized companies with roughly 50 to 1,000 employees, the question is different. Not every company needs its own hardware. The first criterion explicitly allows both: your own servers or processing exclusively in the EU. What we describe on our page on AI in your own data centre therefore applies equally to both options: what matters is that you know the location and can prove it.
The blind spot in practice
In our analysis of 52 published practitioner conversations (September 2025 to August 2026), data sovereignty runs through as a recurring theme. The initiatives Manufacturing X, Factory X and Process X come up in at least four conversations, always with the same undertone: the initiative exists, yet adoption in the Mittelstand is missing.
What is almost never discussed: where the model runs, who operates it and how operations are secured. Sovereignty is discussed as a principle, hardly as a question of infrastructure. This is precisely where digital sovereignty for AI is decided.
An honest data gap belongs to the picture. There is no public representative figure on how many industrial companies want to run AI on their own premises or in the EU. The industry report therefore argues on the basis of the preference gap in the Bitkom Cloud Report, not on the basis of vendor figures.
Frequently asked questions
Is a data centre in the EU automatically sovereign?
No. The location only meets the first of four criteria. Without logging of every use and without the replaceability of the model, the dependency remains, it simply has a different address.
Do we need our own hardware for sovereign AI?
Not necessarily. The requirement is: models run on your own hardware or with processing exclusively in the EU. Both routes are possible, what matters is that you can prove the location and the data flow.
Why do so many companies use US clouds even though they do not want to?
A major reason lies in the supply. According to the Bitkom Cloud Report 2026, 43 % of companies see no equivalent European alternatives. In addition, existing contracts and tools are rarely actively reassessed.
Is it enough to ban public AI services?
According to the findings of the industry report, no. Bans push usage into shadow IT. A short policy and one approved tool whose data flow you know are more effective.
