Skip to content

Sovereign infrastructure

Digital sovereignty in business software

What digital sovereignty means in business software, how to vet a vendor, and what we can say about ourselves, gaps included. For management, IT leads and procurement in mid-sized companies.

11/2026
Legal position as of
8
Questions for any vendor

In short

Sovereign infrastructure describes how independent a company stays from its software vendor. Five layers decide it: location, operator, technology, model, and licence and operations. This page walks through them and shows openly where Beyonetix stands.

Updated: November 2026

Digital sovereignty means a company can trace and help decide where its data sits, who operates the infrastructure and which AI model processes it. The location of a data centre answers only one of five layers.

The five layers

Digital sovereignty in five layers

By 2026 the term appears on almost every vendor page, from hyperscaler subsidiaries to one-person consultancies. Anyone who wants to check it breaks it into its parts. That includes who trained the model and what happens on an acquisition or a new US regulation. A checklist for any vendor, us included, follows at the end.

01 Location Where do the data physically sit, and which law applies there?
02 Operator Who technically runs the servers: the vendor, a parent company in the background or a hyperscaler?
03 Technology What software does the system run on, and which jurisdiction does it come from?
04 Model Which AI model processes the request, who trained it, and does it change unnoticed?
05 Licence and operations Under which licence do the components run, and can every processing step be traced?

The distinction

Location and technology are two different layers

A server in Frankfurt settles the location. How independent the system is depends on the software, the licence terms and the AI model running on it. If the software comes from the US, runs on a US hyperscaler cloud and calls a US language model, the dependency on the US remains, even with the server in Germany. The German location determines which data-protection law governs the processing.

Claims such as “data protection under the GDPR” and “Made in Germany” can both be verified, and both concern location. Data protection under the GDPR does not answer whose model reads the data or which jurisdiction the technology comes from. A law such as the US CLOUD Act can allow access even when the server sits in the EU. “Made in Germany” often names the company seat. Which cloud and which language model are in use usually is not stated there.

So ask beyond the server location: what runs on that server, who owns it, and what happens if that changes? That is exactly what the checklist asks.

Three patterns

Where promises break

Three patterns show where sovereignty promises break.

Location and technology German location, foreign technology A vendor advertises its system with hosting in Germany. The technology underneath comes from a foreign manufacturer and runs in the cloud of a large operator. The server location is German, but the system still depends on the licence terms, update cycles and legal position of that foreign vendor.
Ownership Owners can change A vendor advertises trust, responsibility and sovereignty. Owners and stakes can change, for example through a merger, a sale or new investors. A promise that rests solely on today's ownership structure holds only as long as that structure does.
Verifiability Sovereignty promised without independent verification An operating model promises to process data exclusively in German data centres and without access by the manufacturer. That can only be checked reliably if an independent body monitors the operation and you can see the result.

True partial statements, open questions

Each of these patterns can hold true on its own and concerns one of the five layers. It is often communicated as if it answered all five. A checklist therefore asks about each layer individually.

Further reading

For cloud AI at work, read our article on data protection with cloud AI.

Deadlines and legal position at a glance

Several rules with fixed dates affect the choice of vendor, location and AI model in 2026. Which of them apply to your company is a question for legal advice.

2018
US CLOUD Act
US providers can be compelled to disclose data, regardless of server location.
2020
Schrems II (CJEU)
EU-US data transfers ruled legally fragile.
08/2025
EU AI Act: GPAI obligations
Obligations for providers of general-purpose AI models apply since 2 August 2025.
09/2025
EU Data Act
Most obligations apply since 12 September 2025 and ease provider switching.
12/2025
NIS2 transposition act
In force since December 2025, widening the scope from about 4,500 to about 30,000 affected companies.
06/2026
US Supreme Court / Data Privacy Framework
The 29 June 2026 ruling on FTC independence is part of the debate on the EU-US Data Privacy Framework. An appeal against its confirmation by the General Court is pending before the CJEU.
08/2026
EU AI Act: labelling duty
Since 2 August 2026 AI chatbots must identify themselves. Technical labelling of generated content has a transition period until 2 December 2026.
12/2027
EU AI Act: high-risk obligations
The Digital Omnibus on AI (in force since 27 July 2026) moves them from August 2026 to 2 December 2027.

Frameworks

BSI C5:2026, C3A and CADA at a glance

For context, with no immediate obligation yet: the revised BSI audit standard C5:2026 has been final since April 2026. The new, voluntary framework C3A (“Criteria enabling Cloud Computing Autonomy”) is the first to specifically assess autonomy, with less weight on security. On 3 June 2026 the European Commission presented the draft Cloud and AI Development Act (CADA), a framework with four Union Assurance Levels that makes the sovereignty of cloud services assessable. The draft is still in the legislative process.

As of November 2026, neither framework is binding on mid-sized companies.

The checklist

The eight questions for any vendor

The questions follow from the five layers and the three patterns. They apply to any vendor, Beyonetix included. Each question can be answered or evidenced concretely.

The full version is on the checklist page.
TopicWhat it is about
1. Location Country and data centre of the data
2. Operator Who technically runs the servers
3. Model changes Which model runs and whether a change would be noticed
4. Training Origin and training of the model
5. Ownership Who owns the model
6. Licences Terms of the building blocks in the background
7. Logging Traceability of every AI call
8. Exit What applies on an acquisition or a legal change

Checklist in detail

Warning signs, good answers and our own answers

For each question the dedicated page shows why you ask it, how to recognise a good answer and what we answer ourselves. Read questions 3 to 5 in particular, because that is where a location statement parts ways with an answer about the technology.

Go to the checklist

Technical details

How we operate Beyonetix OS

This is our self-assessment on the eight questions in short form. The answers in detail are on the checklist page.

As of the date at the top of the page. This does not replace legal advice.
PropertyStatement
Hosting A hosting environment in Germany that a partner operates. Where the system runs in a project, we settle with you.
Models Open models that we select and operate. We do not train base models ourselves.
Model layer Many ERP modules call the language models through a central layer. Model choice sits there per task.
Roles per task Text, image, embedding and speech recognition have separate model roles.
Log For each language-model call: model, usage, response time, success or failure.
Model switch Switching the model provider is a setting.
External services Image, video and speech output use declared cloud services. What happens when the main path fails is set per company.
Certification Currently none under ISO 27001 or BSI C5.

Plainly put

Where the gaps are

We name three gaps explicitly, so you do not have to look for them yourself.

  • Not everything runs in the hosting environment. For images, video and speech output we use declared cloud services where our own measurements show them to be the better choice. We do not claim to be “100 percent sovereign”.
  • We currently hold no ISO 27001 or BSI C5 certification. We say so because a claimed certification an auditor cannot find costs more trust than a named gap.
  • We do not develop base models of our own. We select and operate open models.
Go to Beyonetix OS

Point of contact

What Raoul Mayer advises on as a partner

Raoul Mayer

Senior Executive Partner Cybersecurity and Infrastructure

Raoul Mayer advises on infrastructure and security: servers, networks, domains and containers.

Frequently asked

Frequently asked questions about digital sovereignty

What is the difference between data residency and digital sovereignty?

Data residency describes only the physical or legal location where data is stored. Digital sovereignty additionally includes who operates the infrastructure, whose technology and whose AI model processes the data, and who retains control if ownership, contracts or the legal situation change. Data residency is one of five layers of digital sovereignty.

Is “hosting in Germany” enough proof of sovereignty?

A German server location shows which data-protection law governs storage. Which jurisdiction the software in use comes from, who operates it and which AI model handles the processing has to be clarified in addition.

Is a European cloud provider automatically sovereign?

It depends on the case. A European provider can build on US base technology, integrate a US language model, or change through a merger the ownership structure a sovereignty promise originally rested on. The company's headquarters is an indication; contracts and the subcontractor list provide the evidence.

What does the EU AI Act mean concretely for mid-sized companies in 2026?

Since 2 August 2026, AI chatbots must identify themselves as such. For the machine-readable labelling of generated content a transition period runs until 2 December 2026. Obligations for high-risk systems were pushed to 2 December 2027. That is often read as a reprieve, but mainly it gives lead time to prepare. The requirement itself remains.

Is it enough to comply with the GDPR?

The GDPR sets a legal floor for handling personal data, whatever technology or AI model is used. A system can meet that floor and still depend entirely on a non-European technology and model provider. This does not replace legal advice.

Can a vendor be 100 percent sovereign?

In practice rarely completely, once image, video or speech processing with high quality or speed requirements enters the picture. Proven, specialised cloud models exist there, and replacing them fully would today cost quality or speed. A credible answer states where processing is self-hosted, where an external cloud is deliberately used, and why.

Who sees my requests to the language model?

The text models run in the hosting environment in Germany that our partner operates, and every call is logged. Image, video and speech output can run through declared cloud services.

What happens if the main path fails?

For each company it is set whether tasks wait or an external provider steps in, and every fallback is logged.

Can I do without your AI?

Yes. The ERP also runs without AI features. Whatever you add later runs on the same platform.

Discuss the architecture in detail

If you want to put one of these eight questions to us concretely, with your own use case, let's talk about it.