---
title: "Sovereign AI: definition and example – Carl-Uwe Würfel"
description: "Sovereign AI explained: what digital sovereignty means for AI, why data location isn’t enough, and how to tell whether a company can switch providers."
url: https://carl-wuerfel.de/en/glossary/souveraene-ki
author: Carl-Uwe Würfel
---

1. [Home](https://carl-wuerfel.de/en)
  2. [Knowledge](https://carl-wuerfel.de/en/knowledge)
  3. [Glossary](https://carl-wuerfel.de/en/glossary)
  4. Sovereign AI

# Sovereign AI

Sovereign AI means a company can make its own decisions about its AI models, data, and providers: where processing happens, which law applies, and whether the model or provider can be replaced.

## On this page

  1. Public administration turned sovereignty into testable goals
  2. Data location doesn’t settle which law applies
  3. The ability to switch is built into the architecture
  4. Contracts define the way out
  5. Open models and local operation are tools with their own dependencies
  6. What sovereign implementation looks like
  7. Where the term stops working
  8. Frequently asked questions
  9. Sources

Sovereignty means being able to switch when it counts.

_Figure: A check, using the example of an insurance broker that has a US provider’s language model summarize claim reports. Before the model goes into a core process, the broker clarifies which law the provider is subject to, whether data, prompts, and test cases can be exported, whether the model can be switched through the company’s own layer without rebuilding, whether the contract covers notice periods and exit, and whether the in-house team can evaluate a replacement model. Only then is the dependency known and a switch possible when it counts._

Digital sovereignty is the ability of people and institutions to perform their roles in the digital world autonomously, on their own terms, and securely. Germany’s IT Planning Council based its 2021 public-sector IT strategy on this definition from the Competence Center for Public IT (ÖFIT). Sovereign AI applies the idea to AI models and tools. Governments use the term too; here, the focus is companies.

Sovereignty doesn’t mean independence from every vendor. Even the European Commission writes that the top level of its Cloud Sovereignty Framework, full control with no critical dependency outside the EU, is not currently relevant because of supply chains for chips and hardware. Sovereignty means being able to switch when it counts.

For executives, this becomes concrete once AI runs core processes. A provider raises prices, retires a model, changes its terms, or is compelled to hand data to an authority outside the EU. Whether the company can respond is decided beforehand, in its architecture and its contracts.

Example

An insurance broker uses a language model from a US provider to summarize claim reports. All applications call the model through a dedicated layer; prompts and test cases live in the broker’s own system. When the provider retires the model version in use, the team tests a successor and an open-weight model at a European host on the same test cases and switches without changing any application. Reports with health data now go to the European host.

Chapter 1 of 7

## Public administration turned sovereignty into testable goals.

The IT Planning Council cites a heavy, sometimes critical dependence of public-sector IT on individual vendors, which puts control of that IT, and of data protection, at risk. It therefore sets three goals: the ability to switch, design capability, and influence on vendors. The strategy explicitly includes the skills needed to run alternatives.

In October 2025, the European Commission presented a Cloud Sovereignty Framework for its own procurement. It scores eight objectives, among them legal sovereignty, data and AI, operations, supply chain, and technology. Its data and AI criteria include customer-only control of encryption keys, traceable access and AI model usage, and processing confined to Europe with no fallback to third countries.

Neither binds companies, but both work as a checklist for AI services.

Chapter 2 of 7

## Data location doesn’t settle which law applies.

Since the 2018 CLOUD Act, 18 U.S.C. § 2713 has required providers of communication and remote computing services to disclose data in their possession, custody, or control when legally compelled, wherever the data is stored. A 2019 Justice Department white paper says this applies to providers under US jurisdiction, which isn’t limited to US companies; whether a parent controls a foreign subsidiary’s data depends on the facts.

EU law pushes back. Under Article 48 GDPR, third-country rulings and administrative decisions demanding personal data may be recognized or enforced only if based on an international agreement such as a mutual legal assistance treaty; other GDPR transfer grounds are unaffected. For non-personal data, Article 32 of the Data Act requires cloud providers to guard against unlawful government access from third countries. A provider can get caught between two legal systems.

A data center in Frankfurt only answers where. What that means for a specific service is for data protection and IT lawyers to assess.

Chapter 3 of 7

## The ability to switch is built into the architecture.

Model providers retire older models. Anthropic, for example, gives customers at least 60 days’ notice; after that, requests to the old model fail. With the model name hard-coded, every retirement means changing every application.

A dedicated layer between applications and models, such as an [LLM gateway](https://carl-wuerfel.de/en/glossary/llm-gateway), puts that decision in one place. Applications call one interface, and the model behind it is a setting. A switch is tested against fixed cases and made without rebuilding applications.

Only what the company holds can move: prompts, test cases, and evaluation criteria belong in its own systems, not just the provider’s console, and knowledge sources and [vector databases](https://carl-wuerfel.de/en/glossary/vector-database) must be exportable. A proprietary model fine-tuned at a provider can’t move, because the weights stay there. The deeper an application relies on one provider’s special features, the harder the switch.

Chapter 4 of 7

## Contracts define the way out.

The Data Act, Regulation (EU) 2023/2854, has applied since September 12, 2025; its Chapter VI covers switching between data processing services such as cloud services. Article 23 requires providers to remove obstacles to moving to another provider or to the customer’s own infrastructure. Article 25 requires switching clauses in the contract, including a transition period of no more than 30 calendar days to transfer exportable data. From January 12, 2027, Article 29 bans switching charges. Whether a given AI service is covered needs case-by-case review.

Either way, a good AI service contract covers the data export format, use of inputs for training, notice periods for model retirements and price changes, processing location, and subcontractors.

Where a service has no real alternative, an exit plan with a tested export and a named replacement helps.

Chapter 5 of 7

## Open models and local operation are tools with their own dependencies.

Open-weight models can run in a company’s own data center or at a European hosting provider. Downloaded weights stay available even after the developer stops offering the model.

Running a model in-house, the company needs hardware, people for operations and updates, and a checked license, because many open models come under vendor licenses with their own conditions. Without in-house skills, the dependency just shifts from the model provider to whoever runs the open model.

Sovereignty therefore doesn’t require giving up US providers. A company can deliberately use a powerful cloud model when the data is cleared for it and switching remains technically and contractually possible. The decision is made per task and per type of data.

Chapter 6 of 7

## What sovereign implementation looks like.

It starts with an inventory: which AI services the company uses, with which data, from which provider, and under which law. That includes tools employees use without approval. Data is classified by confidentiality, with the permitted services defined for each level.

Applications call models through a dedicated layer. Switching models is a tested configuration change, and at least one backup model is measured regularly against the same test cases.

The export of the most important data has been run at least once. Contracts include exit terms and notice periods. The review repeats, because models, prices, and laws keep changing.

Chapter 7 of 7

## Where the term stops working

Sovereign AI describes the ability to decide for oneself, not a location or a product; an offering labeled “sovereign” doesn’t prove it. Local AI is one operating model that can contribute. Data sovereignty is a subset that covers only control over data. Data protection is a legal obligation; sovereignty is a strategic goal.

## Frequently asked questions

What is digital sovereignty?

Digital sovereignty is the ability of people and institutions to perform their roles in the digital world autonomously, on their own terms, and securely. For companies, it mainly means being able to choose technology and providers, and to switch.

Why does digital sovereignty matter for companies?

A company that can’t switch providers loses leverage over prices, terms, and how its data is handled. With AI the risk is higher, because providers regularly retire and replace models.

Is an EU data center enough for sovereign AI?

No, location only settles where processing happens. If the provider is subject to a third country’s law, such as the US CLOUD Act, it can be required to disclose data there even when that data is stored in Europe.

What does the CLOUD Act mean for AI services?

The CLOUD Act makes clear that providers under US jurisdiction must disclose data in their control even when it’s stored outside the US. How that squares with the GDPR for a specific service is a question for lawyers.

Is open source automatically sovereign?

No. Open models make switching and in-house operation easier, but they require hardware, expertise, and a checked license. Without in-house skills, the dependency just moves to the service provider running the model.

How can a company become less dependent on AI providers?

Technically, a dedicated layer between applications and models helps most, since the model can then be switched through a setting. It also takes in-house test cases, exportable data, contracts with exit terms, and a team that can evaluate a replacement model.

## Further reading

  * [**AI glossary**](https://carl-wuerfel.de/en/glossary)
  * [**AI governance: permissions, access, logs**](https://carl-wuerfel.de/en/ai-governance)
  * [**LLM gateway**](https://carl-wuerfel.de/en/glossary/llm-gateway)
  * [**Local and on-premise AI**](https://carl-wuerfel.de/en/glossary/lokale-ki)
  * [**Data governance**](https://carl-wuerfel.de/en/glossary/data-governance)
  * [ Term**Shadow AI**](https://carl-wuerfel.de/en/glossary/schatten-ki)
  * [ Term**Open-source and open-weight models**](https://carl-wuerfel.de/en/glossary/open-source-llm)

## Sources

  1. [IT Planning Council (IT-Planungsrat): Strategy for strengthening the digital sovereignty of public administration IT, version 1.0, January 2021](https://www.it-planungsrat.de/fileadmin/it-planungsrat/foederale-zusammenarbeit/Gremien/AG_Cloud/20210104_Strategy_for_strengthening_the_digital_sovereignty_of_public_administration_IT_1.0.pdf)
  2. [European Commission: Cloud Sovereignty Framework, implementation guidance, 2026](https://commission.europa.eu/document/download/2ad80a48-166f-4c77-a513-80c53ca2a128_en)
  3. [Cornell Law School, Legal Information Institute: 18 U.S.C. § 2713, Required preservation and disclosure of communications and records, added by the CLOUD Act, 2018](https://www.law.cornell.edu/uscode/text/18/2713)
  4. [U.S. Department of Justice: Promoting Public Safety, Privacy, and the Rule of Law Around the World. The Purpose and Impact of the CLOUD Act, white paper, April 2019](https://www.justice.gov/criminal/media/999601/dl?inline)
  5. [European Union: Regulation (EU) 2016/679 (General Data Protection Regulation), Article 48, 2016](https://eur-lex.europa.eu/legal-content/EN/TXT/?uri=CELEX:32016R0679)
  6. [European Union: Regulation (EU) 2023/2854 (Data Act), Articles 23, 25, 29, 32, and 50, 2023](https://eur-lex.europa.eu/legal-content/EN/TXT/?uri=CELEX:32023R2854)
  7. [Anthropic: Model deprecations, documentation, as of November 2026](https://platform.claude.com/docs/en/about-claude/model-deprecations)
