What EU data residency means for AI and analytics workloads
EU data residency applies to every tool in an AI or analytics pipeline that touches the data. A customer contract or a data protection officer usually asks for it in one line. So which parts of an AI or analytics stack does that line cover, and how do you check a provider?
Take one example, which is hypothetical. A Berlin SaaS team summarises support tickets with a large language model (LLM) and charts the results in an analytics dashboard. One ticket makes four stops:
- At rest: it is stored in the team's data warehouse.
- In processing: it is sent to an LLM API, which reads it and writes a summary.
- Support access: an engineer at one of the vendors opens it to debug a failed job.
- Subprocessors: each tool hands parts of the work to its own vendors, such as hosting or payment providers.
Residency has to hold at every stop. A warehouse in Frankfurt does not help if the LLM call runs in Virginia. The example is simplified: it does not show how likely each gap is, or what closing it costs. We built Lyceum as a sovereign European AI inference provider for open-source models, and our serverless models run in Paris and Finland. The last section shows how we answer each stop, and AI data residency, explained properly covers the storage and processing split in depth.
The five checks for EU data residency
The five checks below turn "EU data residency" into questions a provider can answer in writing. Run them for every tool the ticket touches.
| Check | What to ask the provider | Where the answer is | The example ticket |
|---|---|---|---|
| Data at rest | Where are data, backups and logs stored? | Location statement, DPA | Warehouse in an EU region |
| Data in processing | Where does compute run, including the LLM call? | Region list, model or endpoint page | LLM API location decides it |
| Support access | Who can open customer data, and from which countries? | DPA, access policy | Remote access from outside the EEA is a transfer |
| Subprocessors | Which vendors get which data, and where are they? | Subprocessor list, notice period | Each vendor needs its own answer |
| Non-EU law exposure | Is the provider, or its parent, subject to laws such as the US CLOUD Act? | Contracting entity, terms, transfer assessment | A US parent can be ordered to disclose |
A data processing agreement (DPA) is the contract under GDPR Article 28 that binds a provider to process your data only on your instructions. Most answers in the table live there. Our guide to subprocessors in AI inference covers the fourth row in depth.
Why an EU region alone does not settle residency
An EU region settles where the bytes sit. It does not settle who can reach them, and the GDPR cares about both.
The GDPR has no data-localisation rule. Article 44 allows transfers outside the EEA only under Chapter V, the GDPR's rules for international transfers. The European Data Protection Board (EDPB), the body of EU data protection authorities, adds the access point. Its Guidelines 05/2021 treat remote access from a third country as a transfer. That holds "even if it takes place only by means of displaying personal data on a screen", for example in support situations.
The same guidelines name a second case: an EU processor that is subject to third-country law. Such a processor may receive government access requests. GDPR Article 28(1) lets you use only processors that give sufficient guarantees, and for this processor, the EDPB writes, reliability "may be in doubt".
The US CLOUD Act is the usual example. The law dates from March 2018. It obliges US cloud and communication providers to disclose data in their "possession, custody, or control", wherever it is located. The US Department of Justice describes it as a tool to reach evidence held by US-based global providers. GDPR Article 48 answers from the EU side. A third-country order is recognised only when it rests on an international agreement, such as a legal assistance treaty. Without one, the order alone gives the provider no GDPR basis to hand the data over.
Article 83(5) sets the stakes. Breaking the transfer rules can cost up to €20 million or 4% of worldwide annual turnover, whichever is higher. We explain the hyperscaler case in the CLOUD Act problem, and the wider argument in sovereignty beyond hosting.
EU-US Data Privacy Framework status in October 2026
The EU-US Data Privacy Framework (DPF) is still in force in October 2026. It is an adequacy decision, adopted on 10 July 2023. That is the Commission's finding that participating US companies protect personal data well enough for it to flow to them freely. The Commission's adequacy list still names the United States for those companies, next to Switzerland, the United Kingdom and Japan.
The decision is under challenge. The General Court dismissed Philippe Latombe's action on 3 September 2025. He appealed to the Court of Justice on 31 October 2025, asking it to annul the decision (Case C-703/25 P). Until that court rules, the DPF applies. Its safeguards limit US intelligence access to what is necessary and proportionate, answering the Court of Justice's Schrems II judgment of July 2020.
We recommend the example team does two things. First, check that each US analytics or LLM vendor appears on the Data Privacy Framework list. Second, make the DPA name a fallback in case the court annuls the decision. That fallback is usually standard contractual clauses (SCCs), the Commission's model transfer contract.
EU AI Act and Data Act dates to plan around
The EU AI Act, Regulation (EU) 2024/1689, applies in stages. The AI Omnibus, an amending law (Regulation (EU) 2026/1744), entered into force on 27 July 2026 and moved the high-risk dates.
| Date | What applies | Source |
|---|---|---|
| 2 August 2025 | Obligations for providers of general-purpose AI models, which are trained for many tasks, such as large language models | AI Act, Article 113 |
| 2 August 2026 | Most of the Act, including the Article 50 duty to tell people they are dealing with an AI system | AI Act, Article 113 |
| 2 December 2026 | Generative AI systems already on the market before 2 August 2026 must mark their output as AI-generated (Article 50(2)) | Omnibus, Article 111(4) |
| 12 January 2027 | Switching charges, the fees for moving to another cloud provider, end, data transfer-out fees included | Data Act, Article 29 |
| 2 December 2027 | Stand-alone high-risk systems (Annex III), such as hiring or credit scoring | Omnibus, Article 113 |
| 2 August 2028 | High-risk AI inside regulated products (Annex I) | Omnibus, Article 113 |
The Data Act row matters for residency. Until 12 January 2027, providers may charge switching fees, but no more than their direct costs. From that date they may charge none. That makes moving data to an EU provider cheaper. Our EU AI Act timeline covers each obligation, and our guide to Article 50 marking covers the December 2026 date.
How our serverless inference answers the five checks
Lyceum Serverless Inference answers the five checks with facts from our legal and data-handling page. Here is the example ticket's second stop, the LLM call, run on our API.
| Check | Our answer |
|---|---|
| Data at rest | Prompts and outputs are not stored on disk or in a database. The opening part of recent prompts stays in GPU memory until newer requests replace it, which is what makes cached input cheaper. |
| Data in processing | Paris and Finland for serverless models. Locations for dedicated endpoints (GPUs reserved for you) and GPU virtual machines are agreed per contract. |
| Support access | Our team cannot see prompts or responses. We never log them or write them to a database, so there is nothing to open. |
| Subprocessors | Mainly in the EU and EEA. Anything outside is covered by GDPR safeguards such as SCCs. |
| Non-EU law exposure | You contract with Lyceum Technology Germany GmbH or Lyceum Switzerland GmbH. Not every subprocessor has its headquarters in the EU, so ask for the list and run this check on each entry. |
The one record a request leaves is a usage record: your organisation and user, the model, token counts and the time. It never contains the prompt or the response, which is what zero-retention LLM APIs means in practice. Our team can see account details, usage records and system logs, and works from Berlin and Zurich.
Our DPA is Annex 1 of our terms. You are the controller, who decides why data is processed, and we are the processor, who processes it for you. We tell you before we add or replace a subprocessor, and you can object within 10 business days. The subprocessor list is available on request. Our ISO 27001:2022 audit is in progress. Service incidents appear on our status page.
Choosing a European host does not discharge your own duties as controller. This table covers only the LLM call, so run the same checks on your warehouse, analytics tool and support desk. Our press facts and model list show what we serve today.
Summary
EU data residency holds only when five checks pass for every tool that touches your data: where it is stored, where it is processed, who can open it, which subprocessors receive it and which non-EU laws reach the provider. An EU region settles the first two, while the other three depend on what each provider commits to in writing. For LLM calls on our API, our DPA and legal page make that commitment for all five.
If a customer contract requires EU processing for your LLM calls, send our DPA to your reviewer first. You can read it on our legal page, and request the subprocessor list and our TOMs (technical and organisational measures) there.
Sources
- EUR-Lex: Regulation (EU) 2016/679 General Data Protection Regulation
- edpb.europa.eu
- law.cornell.edu
- justice.gov
- commission.europa.eu
- European Commission: Adequacy decisions
- eur-lex.europa.eu
- dataprivacyframework.gov
- EUR-Lex: Regulation (EU) 2024/1689, the EU AI Act
- eur-lex.europa.eu
- EUR-Lex: Regulation (EU) 2023/2854, the EU Data Act
- National Law Review: EU Digital Omnibus on AI enters into force (31 July 2026)