Sovereign AI in Europe: what it means for where your models run
Sovereign AI means control over the data, the hardware and the law that applies to them. What the AI Act, GDPR and the CLOUD Act mean for hosting AI in the EU.
Sovereign AI means keeping control over three things: where your data and models physically sit, which country's law can reach them, and who can switch the hardware off or change the model under you. For most European companies it is less a political slogan than a practical question: which servers run our models, who operates them, and what happens if a foreign court asks for the data.
Below: where the term comes from, what EU law actually says about it in October 2026, what the EU is building, and the questions to ask any AI hosting provider.
Where the term comes from
NVIDIA popularised the phrase. Its own definition is about countries: sovereign AI is "a nation's capabilities to produce artificial intelligence using its own infrastructure, data, workforce and business networks". Governments took it up as a reason to fund national compute.
For a company, the same idea shrinks to a smaller set of questions. A useful way to split it is into three layers:
| Layer | The question | What decides it |
|---|---|---|
| Data | Where are the prompts, documents and outputs stored and processed? | The data centre's location and the provider's subprocessors |
| Jurisdiction | Which country's courts and agencies can demand access? | Where the operating company and its parent are registered, not only where the server is |
| Control | Who can stop the service, change the model or read the weights? | Whether you run open models on hardware you control, or call someone else's API |
The second layer is the one most often missed. A server in Frankfurt run by a company headquartered in the United States is in the EU physically, but not fully outside US law, as the next section shows.
What the law says in 2026
GDPR. Personal data can leave the European Economic Area only under the conditions in Chapter V of the GDPR (Articles 44 to 50): an adequacy decision, standard contractual clauses or another listed safeguard. For the United States, the Commission adopted the EU-US Data Privacy Framework adequacy decision in July 2023. On 3 September 2025 the EU General Court dismissed a challenge to it (Latombe v Commission, case T-553/23). An appeal to the Court of Justice (case C-703/25 P) was filed on 31 October 2025, so the framework is valid today but not settled. The two previous frameworks, Safe Harbor and Privacy Shield, were both struck down by the Court of Justice.
The US CLOUD Act. Since March 2018, US law (18 U.S.C. § 2713) requires US providers to disclose data in their "possession, custody, or control, regardless of whether such communication, record, or other information is located within or outside of the United States." This is why the operator's nationality matters as much as the server's address. Whether a given request would succeed in a specific case is a question for lawyers, not for a hosting page.
The EU AI Act. Regulation (EU) 2024/1689 entered into force on 1 August 2024. Bans on prohibited practices and the AI literacy duty have applied since 2 February 2025, and obligations for general-purpose AI model providers since 2 August 2025. The Digital Omnibus on AI, Regulation (EU) 2026/1744, entered into force on 27 July 2026 and moved the high-risk deadlines: high-risk uses listed in Annex III (such as employment, education and credit) now apply from 2 December 2027, and AI built into regulated products (Annex I) from 2 August 2028. Most other rules, including the transparency duties in Article 50, applied from 2 August 2026.
The AI Act regulates what an AI system does and who provides it. It does not require that a model runs on servers in the EU. Location matters mostly through GDPR, contracts and the risk of foreign access.
The Data Act. Regulation (EU) 2023/2854 has applied since 12 September 2025. It gives cloud customers the right to switch providers, and Article 29 phases out switching charges: until 12 January 2027 providers may charge only their costs for the switch, and from that date they may not charge for switching at all. For sovereignty this matters more than it sounds: being able to leave is part of being in control.
What the EU is building
The EU's answer at infrastructure level comes in two sizes.
- AI factories. Supercomputing centres with AI-optimised capacity, run through the EuroHPC Joint Undertaking. On 12 March 2025 EuroHPC selected six more, in Austria, Bulgaria, France, Germany, Poland and Slovenia, bringing the total to 13 at the time. The Bulgarian one, BRAIN++ (Bulgarian Robotics & AI Nexus), is located at Sofia Tech Park.
- AI gigafactories. In February 2025 the Commission announced InvestAI, aiming to mobilise €200 billion for AI, including a €20 billion fund for AI gigafactories. On 30 July 2026 EuroHPC opened the call: up to seven gigafactories in at least seven Member States, with public money meant to unlock more than €20 billion of private investment. Tenders close on 12 November 2026, selection is expected in early 2027, and operation within 18 months of selection.
In practice this means the large public EU capacity will mostly come online from 2028, and much of it is aimed at research, start-ups and very large training runs. A company that needs a private model in production next quarter will still rent from commercial providers.
Questions to ask an AI hosting provider
These work for any provider, including us. If the answers are vague, the sovereignty claim is too.
- Where exactly is the hardware? The data centre, the city and the company that operates the building.
- Who is the operator, and who owns it? Ask for the legal entity and its parent company. This decides which laws, including foreign ones, can reach it.
- Is the server shared? Many GPU clouds split one machine between customers. If you process personal data or keep model weights you care about, ask whether you get the whole machine.
- Which subprocessors touch the data? Support teams, backups, monitoring tools and their locations.
- Which model, and can it change? A hosted API can be updated or withdrawn by its owner. Open-weight models that you run yourself stay the same until you change them.
- How do you leave? How data and weights are exported, how fast, and whether you pay for it. The Data Act now limits the last part.
Where a dedicated EU server fits
Sovereignty is a scale, not a yes or no. This is how the usual options compare:
| If you... | What usually fits | The trade-off |
|---|---|---|
| need frontier models and accept a US provider | A hyperscaler or model API, EU region, with a GDPR data processing agreement | Easiest, but the operator remains under US law |
| run open models on personal or confidential data, in production | A dedicated GPU server from an EU-registered operator, in an EU data centre | You manage the software stack yourself |
| need thousands of GPUs for training research | EuroHPC AI factories and, from about 2028, gigafactories | Access through calls and applications, not on demand |
| process data that cannot leave your own building | Your own hardware on site | Highest control, highest upfront cost and staff time |
NEOXIS Corp. is a Bulgarian company running one 8 × RTX PRO 6000 server in a data centre near Sofia: 768 GB of GPU memory, root access, and the whole machine for one customer. What fits in that memory, and how teams use it for private models, is on our private LLM hosting page. It is not a hyperscaler: one location, one configuration, and you run your own software. Pricing starts at $1.70 per GPU-hour, about $9,930 a month for the whole machine; details are on the pricing page.
Questions
Does the EU AI Act require AI to be hosted in the EU? No. It sets duties based on what an AI system does and who provides or deploys it. Where it runs matters mainly through GDPR rules on personal data transfers and through your contracts.
Is an EU region of a US cloud "sovereign"? It keeps data physically in the EU, which helps with GDPR. The provider is still a US company subject to the CLOUD Act, so it does not answer the jurisdiction question on its own. Some providers offer separate EU-operated structures; ask how they are owned and staffed.
Is the EU-US Data Privacy Framework still valid? Yes, as of October 2026. The General Court upheld it in September 2025, and an appeal is pending before the Court of Justice.
What is the difference between an AI factory and an AI gigafactory? Both are EU-backed compute centres. Gigafactories are the larger tier, planned for training very large models, with the call for them closing on 12 November 2026.
This article is a technical overview, not legal advice. For a specific case, ask a lawyer who works with GDPR and the AI Act.
Sources
All pages were checked on 18 October 2026.
- NVIDIA: What Is Sovereign AI?, definition.
- European Commission: AI Act, regulatory framework and timeline, including the AI Omnibus dates.
- Regulation (EU) 2024/1689, the AI Act, EUR-Lex.
- AI Act Explorer: Digital Omnibus on AI, Regulation (EU) 2026/1744, Articles 111 and 113 as amended.
- Regulation (EU) 2016/679, the GDPR, Chapter V, Articles 44 to 50.
- IAPP: European General Court dismisses Latombe challenge, upholds EU-US Data Privacy Framework, September 2025.
- WilmerHale: European Court of Justice to review challenge to EU-U.S. Data Privacy Framework, 1 December 2025, appeal C-703/25 P.
- Cornell Law School LII: 18 U.S. Code § 2713, added by the CLOUD Act on 23 March 2018.
- European Commission: Data Act, applicable from 12 September 2025, and Regulation (EU) 2023/2854, Article 29.
- EuroHPC JU: EuroHPC JU selects additional AI Factories, 12 March 2025.
- European Commission: Commission launches InvestAI initiative to mobilise €200 billion of investment in AI, February 2025.
- EuroHPC JU: The EuroHPC Joint Undertaking launches the AI Gigafactories Call, 30 July 2026.
- NEOXIS: pricing and server specifications.
