Guide
Self-hosted LLM and GDPR: what your own AI server solves
With your own AI server in Germany, the biggest GDPR worries go away: no model provider reads along, the model needs no transfer to the US, and you decide what is stored. What that means for contracts, logs, erasure and the DPIA, and what stays on your list.
Most GDPR worries about AI come down to one question: who else sees the data? With ChatGPT and other cloud AI services, every prompt goes to the provider. With a self-hosted LLM on your own GPU server, it stays with you. That one change settles the hardest questions.
Self-hosted LLM vs API: what changes
The DSK, the joint body of Germany’s data protection authorities, calls technically closed AI systems the better choice for data protection. Closed means: the AI runs in a sealed-off environment, a small group of people uses it, they control what goes in and out, and the provider doesn’t train on it. Your own server provides the sealed-off environment. Who gets an account is your call.
| Model API (for example OpenAI) | Your own GPU server | |
|---|---|---|
| Where prompts are processed | On the provider’s servers | On your server in Germany |
| What gets stored | What the provider’s terms say | What you decide |
| Transfer to the US | Often, with US providers | None for the model |
| Data processing agreement | With the API provider | With the data centre and whoever maintains the server |
With your own server, you depend on fewer promises from others.
No transfer to the US for the model
Personal data may only leave the EU under strict conditions; for the US they rest on the EU-US Data Privacy Framework of 2023. Even an EU region of a US provider doesn’t fully settle it: under a US law, the CLOUD Act, the provider can be required to hand over data it controls, wherever it is stored.
With a dedicated server from a European host, that question doesn’t come up for the model. Hetzner states that it processes and stores data for its dedicated servers only in the EU, and your server stands in Germany.
Who needs a data processing agreement, and with whom
If you run an AI only for your own purposes, you usually remain the sole controller, the one who decides what happens to the data. A data processing agreement (DPA) is needed with everyone who handles the data on your behalf.
With a rented server you need two. One with the data centre, because hosting counts as processing on your behalf even if the data centre never looks at your data; at Hetzner you conclude it in the customer account. The other with whoever maintains the server. With Private AI, that’s us, and we sign a DPA with every client.
Logs and chat history: you decide what is kept
With an API, the provider decides what it logs. On your own server, you do. On a Private AI server, prompts and answers are not logged unless you decide otherwise. We maintain the machine, meaning updates, load, disk and health, not your conversations.
For chat histories in employee accounts, the DSK suggests a privacy-friendly default: no history beyond the session, unless the person wants to keep it. Backups are copies too and belong in your deletion plan.
Erasure and access requests
People can ask what you store about them and have it deleted. On your own server that gets much easier: chat histories, uploaded files and search indexes sit on disks you control. You find a person’s data, delete it and can show that you did.
The model itself only learns from your data if you fine-tune it. Fine-tuning with LoRA produces a small adapter file next to the base model. If someone asks for erasure, you remove their record, train the adapter again and delete the old one.
Whether a model refuses to answer is not a data protection measure. Protection comes from who has access, what is stored and how you delete. That holds for uncensored models just the same.
When you need a DPIA for AI
A data protection impact assessment (DPIA) is a written risk check. You need one when processing is likely to pose a high risk to people. For AI, the DSK expects that often, and its list of cases that always need a DPIA includes AI that talks to customers or evaluates people.
So a customer-facing chat usually needs a DPIA, while an internal tool that summarises technical manuals usually doesn’t. What counts is what the AI does. Your own server makes the assessment easier to write: you know where the data goes, what is stored and who has access, without relying on a provider’s documentation.
What stays your job
Your own server takes the provider out of the picture. The everyday data protection work stays with you, as with any software that handles personal data:
- Decide what the AI is for and on what legal basis you use personal data.
- Mention the AI in your privacy notice.
- Answer access and erasure requests.
- Add the AI to your record of processing activities. The entry is short: fewer recipients, no third country for the model.
- Check what the AI says about people before you rely on it.
- Don’t let the AI decide about people on its own, for example who gets invited to a job interview.
- Report a data breach to the authority within 72 hours.
Private AI: your own AI server in Germany
This is what I build with Private AI. Your own AI runs on a dedicated GPU server at Hetzner in Germany, with open-weight models, uncensored variants included, adapted to your data and tested against your own questions before going live. No model provider in between. Prompts and answers are processed on your server and not logged unless you decide otherwise. You get a private chat interface and an API for your own tools and agents.
We sign a data processing agreement with every client and keep one admin login for maintenance until you want full handover. Then root access and every key are yours alone. The first call is free: Private AI.
Common questions
Is a self-hosted LLM GDPR-compliant?
It solves the biggest questions: no model provider sees your prompts, and with the server in the EU the model needs no transfer to the US. Legal basis, privacy notice and deletion stay with you, as with any software that handles personal data.
Do I need a DPA if I self-host an LLM?
With hardware in your own building that nobody outside can reach, usually not. With a rented server you need one with the data centre (Hetzner offers it in the customer account) and one with whoever maintains the server with admin access. With Private AI, we sign one with every client.
Is ChatGPT GDPR compliant?
It depends on the product, the contract and how you use it. With any AI API, your prompts go to the provider and are handled under its terms, and with a US provider the transfer question comes on top. With your own server in the EU, neither applies to the model.
Do I need a DPIA for AI?
Often, yes. Germany's data protection authorities expect a high risk in many AI uses, for example a chat that advises customers. Whether you need one depends on what the AI does. With your own server the assessment is easier to write, because you know exactly where the data goes.
On-premise LLM or rented GPU server: which is better for data protection?
Both work. In your own building you need no DPA with a data centre, but power, cooling and physical security are yours. A dedicated server in a German data centre needs that DPA and can still run as a closed system, the setup Germany's data protection authorities prefer.
Sources
- General Data Protection Regulation (GDPR), official text, EUR-Lex eur-lex.europa.eu
- DSK: Guidance on AI and data protection (in German), May 2024 datenschutzkonferenz-online.de
- DSK: Short paper on processing on behalf (in German) datenschutzkonferenz-online.de
- DSK: List of processing that requires a DPIA (in German) datenschutzkonferenz-online.de
- Hetzner Docs: Data Protection at Hetzner docs.hetzner.com
Hetzner is a trademark of its owner. Naming it does not imply endorsement by, or a business relationship with, OCTOO AI.