What does sovereign AI mean in Europe?

Sovereign AI in Europe is AI that stays under EU law: its data, its model, its operators and the company behind them. What is sovereign AI? covers the idea in general, with five layers of control and a test for each. This article is about the European rules that give those layers their content.

Three things set the European version apart. Data protection law limits where personal data may go. The EU’s top court has twice struck down the arrangement that let it go to the US. Location is not jurisdiction: a provider under US control can be ordered to hand over data it stores in Frankfurt. And the EU now regulates the exit, with a right to switch cloud providers.

This article explains the rules. It is not legal advice.

The GDPR lets personal data leave the EU only under the conditions of its Chapter V. The Court of Justice set the bar in its 2020 Schrems II judgment. Data sent abroad under standard contractual clauses must enjoy “a level of protection essentially equivalent to that guaranteed within the EU by the GDPR”. The assessment has to include what the other country’s public authorities can access. On that ground the Court declared the EU-US Privacy Shield invalid, as it had declared Safe Harbour invalid in 2015.

Its successor is the EU-US Data Privacy Framework. Since the Commission’s adequacy decision of 10 July 2023, personal data can flow freely to US companies that take part in it. The Commission adopted it after the US added safeguards on intelligence access and a redress mechanism for Europeans.

For an AI platform, this means a prompt that carries an employee’s name to a US model provider can be lawful today. It rests on a decision whose two predecessors fell. A platform that can move that work to a model in the EU, or on your own servers, does not depend on the next ruling.

The US CLOUD Act of 2018 reaches any provider under US jurisdiction. US authorities can require it to disclose data in its possession, custody or control, wherever that data is stored. The European Data Protection Board and the European Data Protection Supervisor assessed it in a joint response to the European Parliament in July 2019. Their conclusion: providers handling data under the GDPR can face a conflict of laws. Under Article 48 of the GDPR, a foreign court order alone does not make a transfer lawful. It needs an international agreement, such as a mutual legal assistance treaty, or another ground in Chapter V.

The two bodies also left a question open: how the concept of control applies to EU companies affiliated with US ones. That is why the legal entity matters as much as the region. An EU data center run by a US company’s subsidiary settles where the data is. It leaves open who can be ordered to hand it over.

The Data Act adds a duty to say so. Its Article 28 requires providers of data processing services to state on their websites which jurisdiction their infrastructure is subject to. They must also describe how they prevent government access from outside the EU to non-personal data, where that access would conflict with EU law. Ask any vendor for that page.

The AI Act, Regulation (EU) 2024/1689, has applied in general since 2 August 2026. Its bans on certain practices came first, on 2 February 2025, and its rules for general purpose AI models on 2 August 2025. The Digital Omnibus on AI, Regulation (EU) 2026/1744, moved the rules for high risk systems. They now apply from 2 December 2027 for the uses listed in Annex III. For AI in the products of Annex I, they apply from 2 August 2028.

Running a system on your own servers does not take it out of the Act’s scope. Sovereignty helps with what the Act asks of you as a deployer. A deployer of a high risk system assigns people to oversee it. It keeps the system’s automatic logs for at least six months, as far as they are under its control. What is AI governance for internal tools? covers those duties. Logs that sit with a vendor abroad are harder to account for.

The Commission’s AI Continent Action Plan, launched on 9 April 2025, starts with infrastructure. It plans a network of AI Factories around Europe’s supercomputers, 13 of them already being deployed at launch. It adds AI Gigafactories with about 100,000 advanced AI chips each, to train large models in the EU. A proposed Cloud and AI Development Act aims to at least triple the EU’s data center capacity in five to seven years.

For a buyer, the more useful piece is the Commission’s Cloud Sovereignty Framework. The Commission built it to buy sovereign cloud for the EU institutions, and it turns sovereignty into criteria a tender can score. It measures eight objectives, from legal and operational to supply chain, technology and security. Its assurance levels run from SEAL-0 to SEAL-4. To qualify, a provider had to reach SEAL-2. At that level, in the Commission’s words, providers “abide by the EU laws and regulations without requiring additional technical measures by the customer to protect its data.” The Commission welcomes any organization that wants to use it.

The Data Act, Regulation (EU) 2023/2854, has applied since 12 September 2025. Its Chapter VI makes providers of data processing services remove the obstacles to switching. Article 23 covers a switch to another provider’s service “or to on-premises ICT infrastructure”. The contract has to set out the notice period, the transition and the retrieval window. It also has to list every category of data and digital assets you can take with you.

Read that list for more than the files. An AI platform also holds the links between records, the access rules and the tools built on them. What is AI vendor lock-in? covers what leaving costs at each layer.

Ask where five things sit, and ask for the answers in writing:

1. The data. Every copy, including indexes, logs and backups, and the country of each.
2. The model. Who runs it, under which law, and which records a prompt carries to it. A hosted model outside the EU turns some prompts into transfers.
3. The operations. Who holds administrator access, who gives support, and from which country they work.
4. The legal entity. Which company signs the contract, who owns it, and which jurisdiction its infrastructure is subject to. Under the Data Act, a cloud provider has to publish the last answer.
5. The exit. The notice period, the transition, and what you can take with you in which format. Ask whether you can move to your own servers.

The Cloud Sovereignty Framework asks the same questions with a score attached. Its eight objectives make a checklist that a buyer and a security team can share, in public procurement or not.

Curiosity GmbH is a bootstrapped company in Munich and a full member of the KI Bundesverband. Curiosity Studio runs on premises or in your private cloud, with EU data residency. A cloud deployment we host runs in European data centers, on Hetzner. Air gapped deployments are supported, with an offline license, local models and usage reporting switched off. What is air gapped AI? covers what that takes.

The model is a setting. 19 providers are built in, and open weights run on your own servers through Ollama or any OpenAI compatible endpoint. The models page shows which models run in the US, in the EU or on premises. In the cloud deployments we host, OpenAI is the optional default, and you can choose another. What is a model agnostic AI platform? explains how the switch works.

Some things leave a connected install by default. Studio sends usage and error reports to Curiosity, and you can switch them off. Its front end reports product usage to Mixpanel, on EU servers, with each user’s name and email, and has no setting to stop that. The subprocessors page lists both. Our paper How Curiosity processes your data follows a record from the connector to your users.