AI without a US cloud: GDPR and EU AI Act checklist for self-hosted LLMs (as of December 2026) | ComputeBox
KI & GPU
AI without a US cloud: GDPR and EU AI Act checklist for self-hosted LLMs (as of December 2026)
The AI Act has not been postponed, only the high-risk obligations have. This checklist shows what you need to do for a self-hosted LLM by December 2, 2026: deadlines, GDPR building blocks, Art. 50 and a sovereign architecture.
The AI Act has not been postponed. Only the high-risk obligations have. The transparency rules have applied since August, and the next deadline expires on December 2, 2026. This checklist shows you what you need to do for a self-hosted LLM now.
In short
Art. 50 of the AI Act has applied since August 2, 2026: chatbots must identify themselves as AI, and deepfakes must be labeled.
On December 2, 2026, the transition period for machine-readable marking of existing generative systems ends (Art. 50(2)). Two new prohibitions take effect on the same day.
The high-risk obligations only start on December 2, 2027 (Annex III) and August 2, 2028 (Annex I).
For the GDPR you need three things: a data processing agreement with your host, a data location in the EU and a deletion policy for chats and logs.
An open-weight model on your own GPU in a German data center spares you the third-country transfer assessment for prompts and responses.
Note: This article is not legal advice. Details at the end.
German companies want AI from their own country. According to a Bitkom survey from November 2, 2026, 96% would prefer to use AI from German providers. 87% pay attention to the country of origin when choosing. US providers are now an option for only 41%, down from 51% a year earlier.
The same survey also shows the limit: 64% would not accept compromises on price or features for the sake of origin. Sovereign AI therefore has to pay off and work in everyday use.
Then there is the calendar. Since the Digital Omnibus, you often read that the AI Act has been "postponed". That is only true for part of it. Anyone who has put the transparency obligations on hold because of this has been in default since August.
The Digital Omnibus on AI is Regulation (EU) 2026/1744. It was published in the Official Journal on July 24, 2026, and has applied since July 27, 2026. It postpones Chapter III, meaning the high-risk rules. Everything else keeps its date (ITMR, Cooley).
Date
What applies
Status
since August 2, 2026
Transparency obligations under Art. 50: AI notice for chatbots, disclosure of deepfakes and of AI-generated texts on matters of public interest. New generative systems must mark their outputs in a machine-readable way from day one.
Applies. Not postponed.
December 2, 2026
End of the four-month transition period for Art. 50(2) for generative systems placed on the market before August 2, 2026. Plus two new prohibitions in Art. 5: AI systems for non-consensual intimate imagery and for child sexual abuse material.
New through the Omnibus.
December 2, 2027
High-risk obligations for stand-alone systems under Annex III, for example recruitment, education, creditworthiness, biometrics.
Postponed from August 2, 2026.
August 2, 2028
High-risk obligations for AI in regulated products under Annex I, for example medical devices and machinery.
Postponed from August 2, 2027.
Already in force for some time are the original prohibitions and AI literacy under Art. 4 (since February 2025), as well as the rules for general-purpose AI models (since August 2025). The Omnibus has softened the AI literacy requirement: you have to promote your staff's AI literacy, but you no longer have to guarantee a specific level.
An internal chat assistant is usually not a high-risk system. The postponement to 2027 and 2028 then does not affect you at all. Your obligations come from Art. 50 and the GDPR, and both apply today.
Pay attention to the use case. If your LLM pre-screens job applications or evaluates employees, you end up in Annex III. Then your deadline runs until December 2027.
The December deadline hits existing systems. If your generative system has been running since before August 2, 2026, machine-readable marking must be in place by December 2, 2026. Systems launched later got no grace period.
The new prohibitions also apply to existing systems. If you run image models, check the safety filters before December. With open-weight models, you are responsible for making sure they are active.
Self-hosting solves the transfer problem, but not the GDPR. As soon as names, customer data or employee data end up in prompts, you are processing personal data. You need four building blocks.
Building block 1: Data processing agreement with your host#
Your server sits with a service provider. The provider processes the data on your behalf, even if it never looks at your chats. Art. 28 GDPR requires a data processing agreement (DPA) for this.
DPA with your host signed and filed
Sub-processors checked, especially the data center operator
Your host's technical and organisational measures (TOMs) are on file
Processing entered in the record of processing activities under Art. 30 GDPR
With a US provider, you need a legal basis for the third-country transfer. Currently, that is usually the EU-US Data Privacy Framework. It applies, but it is currently under review by the Court of Justice of the EU (CJEU): the appeal in the Latombe case is pending under case number C-703/25 P (HÄRTING). The CJEU has already struck down two predecessor agreements.
If your model runs on a server in Germany with a German provider, this question does not arise for prompts and responses.
Server location is stated in the contract, not just on the website
Backups and snapshots are also stored in the EU
No external APIs active in your stack that pass prompts on (web search, cloud models, external embeddings)
Building block 3: Chats and prompt storage in Open WebUI#
Open WebUI stores every chat on the server in its database. By default, that is a SQLite file in the data volume; for production use, the documentation recommends PostgreSQL. Uploaded files and the vector database for RAG sit alongside it.
Important for your data protection concept: in the default configuration, admins can read all users' chats. The Open WebUI documentation is clear about this: assume that every admin can read every chat.
This is how you restrict it:
The first two switches only lock the interface. Anyone with access to the database, the host or the backups can still get to the content. Keep this group small and document it.
For sensitive groups, you can enforce temporary chats. The server then stores no history:
Admin access to chats turned off, or justified and documented
Retention period for chats defined and technically enforced. Do not rely on automatic deletion.
Access and erasure per user tested (Art. 15 and 17 GDPR)
Data volume encrypted, for example with LUKS
Works council involved, if you have one. A system that makes employee chats analyzable is usually subject to co-determination.
Art. 50 requires two things: people must know that they are talking to an AI, and generated content must be recognizable as such. Who has to do what depends on the role.
Paragraph
Who
Obligation
Exception
Para. 1
Provider
Users are informed that they are interacting with an AI system.
The AI is obvious to a reasonably attentive person.
Para. 2
Provider
Generated text, images, audio and video are marked in a machine-readable way.
Purely assistive functions that do not substantially alter the content.
Para. 4
Deployer
Disclose deepfakes. Disclose AI-generated texts on matters of public interest.
Texts that have undergone human review and for which someone holds editorial responsibility.
Para. 5
Both
The notice is clear, accessible and given at the latest at the first interaction.
With self-hosting, you are often both. If you combine an open-weight model with an interface into your own assistant and put it into service under your own name, you may count as the provider of that AI system. Then paragraphs 1 and 2 apply to you as well. Clarify this classification early, because the December deadline depends on it.
For a customer chatbot on your website or a phone assistant, the case is clear: the notice comes before the first message. Not in the footer and not only in the privacy policy.
One sentence is enough:
With an internal assistant, your staff usually know they are using an AI. The exception for obvious cases then seems likely to apply. A banner still only takes you two minutes and saves you the discussion.
The assistant's name makes clear it is an AI, for example via WEBUI_NAME
Banner set in Open WebUI (admin settings, Interface section)
System prompt instructs the model to identify itself as an AI when asked
External chatbots and voice assistants show or say the notice before the first interaction
This is the part with the December 2, 2026 deadline. The regulation says what must be achieved, but not how. The gap is filled by the European Commission's voluntary code of practice of June 10, 2026. It calls for at least two layers, typically signed metadata plus an invisible watermark. For plain text, one layer is enough according to the code, because text does not carry metadata (Freshfields).
What this means for your stack:
Images, audio, video: Write provenance data into the metadata, for example following the C2PA standard, and add a watermark. Check whether your export strips the metadata again.
Text: This is the hardest part in an open-source stack. Ollama and vLLM do not mark text on their own. Decide which technique you will use and keep a generation log.
Exceptions: If your assistant only corrects spelling or rephrases lightly, the assistive exception applies. According to it-blue, the Commission's guidelines of July 20, 2026, also exclude source code, very short character strings and purely machine-to-machine outputs.
List of all generative systems created, with the date they went live (before or after August 2, 2026)
Decided for each system: provider role yes or no
Marking technique defined and tested for each content type
Decision documented, even if you rely on an exception
This point mainly affects agencies and editorial teams. If you publish deepfakes, meaning deceptively real images, voices or videos, you must label them visibly. The code provides an EU icon for this. An equivalent label with the abbreviation "AI" also works (Bird & Bird).
For texts on matters of public interest, the labeling requirement does not apply if a human has reviewed the text and someone holds editorial responsibility. Put this approval process in writing.
Approval process for AI texts documented: who reviews, who is responsible?
Violations of Art. 50 can cost up to 15 million euros or 3% of worldwide annual turnover. For SMEs, the lower of the two amounts applies. In Germany, the Federal Network Agency (Bundesnetzagentur) is responsible. The German implementing act KI-MIG has applied since July 29, 2026 (IHK).
A "sovereign" architecture: model, GPU and data in one hand#
Your setup is sovereign when three things come together: a model you run yourself, a GPU that belongs only to you, and data that does not leave the server.
You download the weights once, and from then on the model runs locally. Prompts and responses do not go to any model provider. Only the download itself contacts the model registry.
License checked: does it allow commercial use? Apache 2.0 and MIT are unproblematic, other licenses come with conditions.
Model version and source documented
For image models: safety filters active, see the new prohibitions from December 2, 2026
With shared GPUs, your requests take turns with those of other customers on the same card. A dedicated GPU with no time-slicing belongs only to you for the rental period. You share VRAM and compute with no one. That makes response times predictable and tenant isolation easy to explain when your data protection officer asks.
The server alone is not enough. What matters is what it sends out.
Outbound connections limited to what is necessary by firewall (updates, model download)
Web search, external model APIs and cloud embeddings disabled in Open WebUI
Embeddings for RAG run locally on the same GPU
Access only via HTTPS and login, ideally also via VPN or IP allowlist
Monitoring, backups and error tracking also run in the EU
To set up Ollama and Open WebUI, see our tutorial on self-hosting an LLM with Ollama. For many concurrent users, switching to vLLM pays off: see our tutorial on vLLM for production inference.
For a team, the RTX 6000 Ada with 48 GB is the right choice; for a pilot or a single department, the L4 with 24 GB is enough. VRAM determines which model you can load.
NVIDIA L4
RTX 6000 Ada (team instance)
VRAM
24 GB
48 GB
Model size
13B to 14B at 8-bit, about 30B with 4-bit quantization
up to 70B quantized (approx. 42 GB)
Good for
Pilot project, single department, RAG with embeddings, agency with a few concurrent users
Team assistant for the whole company, large context windows, several models in parallel, fine-tuning
Both instances come with what the checklist above requires:
German data center. Your server is located in Germany. That settles the data location.
Dedicated GPU with no time-slicing. The card belongs only to you for the rental period.
Cancel monthly. You start with a pilot and don't commit for years. No setup fee.
Flat monthly price. No billing per token or per second. That matters when AI from Germany must not cost more than the US alternative.
Ready in minutes. Root access via SSH, then you install Ollama or vLLM.
Our tip: Start with the L4 if you first want to check whether a 14B model is enough for your tasks. Switch to the RTX 6000 Ada as soon as several teams work at the same time or you need a 70B model.
No. Only the high-risk obligations have been postponed: to December 2, 2027, for Annex III and to August 2, 2028, for Annex I. The transparency obligations under Art. 50 have applied since August 2, 2026.
Does the AI Act also apply to a purely internal LLM?
Yes. However, an internal assistant is usually not a high-risk system. AI literacy, the prohibitions and Art. 50 remain relevant, insofar as you generate or publish content.
Do I need a DPA even though the host never looks at my data?
Yes. Hosting counts as processing on behalf of a controller as soon as personal data is stored on the server.
Is a model from the US running on a German server a third-country transfer?
No. The model's origin is not the data flow. An open-weight model runs locally and does not send prompts to its maker.
Do I have to visibly label every AI-generated text?
No. Visible labeling concerns deepfakes and texts on matters of public interest without editorial review. Machine-readable marking is a separate obligation and falls on the provider of the system.
This article is technical guidance and not legal advice. It reflects the situation as of December 2026. Whether your system is a high-risk system, whether you are a provider or a deployer, and which exceptions apply to you depends on the individual case.
So talk to your data protection officer before you put an LLM with personal data into production. If you don't have one, bring in external data protection advisers or a specialized law firm. Only the regulation texts in the Official Journal of the EU and the information provided by the competent authorities are binding.