EU AI Act and DSGVO for Self-Hosted LLMs: What Actually Applies

If your company runs an LLM in the EU, you face two regulatory layers that are frequently confused: the AI Act (Regulation (EU) 2024/1689), which regulates AI systems and models, and the DSGVO/GDPR (Regulation (EU) 2016/679), which regulates processing of personal data — which includes prompts and outputs.¹² Self-hosting an open-source model does not exempt you from either. But it changes which obligations apply and who carries them, in ways that often work in your favour compared to calling a foreign API.

This guide covers the actual legal state as of 24 September 2026, including the timeline amendments of the Digital Omnibus on AI, which is adopted law — Regulation (EU) 2026/1744, published in the Official Journal on 24 July 2026 and in force since 27 July 2026 — not a pending proposal.³ Every article number and date below is cited to the primary sources in the footnotes.

1. The AI Act risk pyramid — and where a self-hosted LLM lands

The key insight: the AI Act classifies by use, not by technology. A general-purpose LLM serving an internal knowledge base or code completion is not high-risk by virtue of being an LLM. High-risk status attaches only through Article 6 — either as a safety component of a regulated product (Annex I) or through a listed use case (Annex III): employment and worker management, education, credit scoring, essential services, law enforcement, migration, administration of justice, and biometrics.¹

A self-hosted chatbot for employees or customers is therefore, by itself, a transparency-tier system, not a high-risk one. Reaching for the model weights yourself does not push you up the pyramid.

When your deployment does become Annex III high-risk

Two role traps matter here. First: if you take a non-high-risk system — including a general-purpose model you host yourself — and modify its intended purpose so it becomes high-risk, you become the provider of that high-risk system under Article 25(1)(c), with the full Article 16 obligation stack (risk management, technical documentation, conformity assessment, registration).¹ The same applies if you put your own name on an existing high-risk system (Art. 25(1)(a)) or substantially modify one (Art. 25(1)(b); "substantial modification" is defined in Art. 3(23)). Internal use is no defence: Article 43(4) requires a fresh conformity assessment for substantial modifications even if the modified system stays with the current deployer.

Second: fines for high-rule violations reach €15 million or 3% of worldwide turnover, and Art. 5 breaches up to €35 million or 7% — but only from the dates those rules apply to you. Which brings us to the timeline.

2. The GPAI / timeline question — current legal state (September 2026)

Much of what you will find online still describes the Digital Omnibus as a "proposal". It is not. The AI strand was adopted as Regulation (EU) 2026/1744 (signed 8 July 2026, OJ 24 July 2026, in force 27 July 2026) after political agreement in May 2026 and the Parliament and Council votes in June 2026.³ The high-risk dates it moved are now fixed calendar dates, not conditional on standards. What applies to you as a deployer or self-hosting company today:

Supervision of your AI-literacy measures starts 3 August 2026 — enforcement is live now.⁴

The practically relevant point: the deferral is a reprieve, not a repeal. If your assistant touches an Annex III use case, the full Chapter III regime lands on 2 December 2027 — sooner than most migration and documentation projects finish. And open-source self-hosting does not create a documentation holiday.

Open-source models: Art. 53(2) — an exemption for the provider, not for you

Article 53(2) exempts GPAI providers whose models are released under a free and open-source licence (weights, architecture and usage information publicly accessible) from the technical-documentation and downstream-information duties of Art. 53(1)(a) and (b).⁵ That exemption (a) never applies to systemic-risk models — for those, the training-compute presumption of Art. 3(63), 10²⁵ FLOP, is the relevant line — (b) does not remove the copyright policy and training-data summary duties under Art. 53(1)(c)/(d), and (c) is a provider obligation. As the company deploying the model you neither inherit these duties nor lose anything by them. If anything, an open-source release under a qualifying licence means the transparency information you need for your own compliance file usually already exists publicly.

If you fine-tune the model and offer the resulting model commercially, you can leave deployer territory and become a GPAI provider yourself — at that point Art. 53(1) applies to you (with the Art. 53(2) exemption available if you release the fine-tune under a qualifying open licence and your training compute stays below the Art. 3(63) presumption).

3. DSGVO/GDPR: what self-hosting actually changes

The columns tell the compliance story: self-hosting converts an outsider-trust problem into a configuration problem. You no longer need contractual comfort that a US inference provider is not building profiles from your employees' prompts. Your prompt data is personal data the moment it relates to an identifiable person — a developer pasting a customer schema into a coding assistant is the canonical example — and all GDPR principles apply to it: lawfulness (Art. 6 — typically contract or legitimate interest, each requiring the documented three-step assessment the EDPB set out in Opinion 28/2024), purpose limitation (Art. 5(1)(b)), data minimisation (Art. 5(1)(c)), storage limitation, integrity, and accountability.⁶²

Two details deserve precision:

• Your GPU hoster may be a processor. If infrastructure staff could access data, Art. 28(3) Auftragsverarbeitung applies. Where the infrastructure sits in the EU, transfer analysis shifts to a supported-provider comfort check.
• Retention of prompts for fine-tuning is a new purpose. Re-using chat logs to fine-tune your model is not covered by the original Art. 6 basis for the chat service absent a documented compatibility analysis under Art. 6(4) or a fresh legitimate-interest assessment; both approaches work if documented.

4. Transparency (Art. 50) and AI literacy (Art. 4) — live now

Article 50 has applied since 2 August 2026 and was not deferred by the Omnibus.⁷³ For a typical self-hosted assistant:

• Art. 50(1) — if the system interacts directly with natural persons (customer chatbot, AI agent writing emails), users must be told they are interacting with AI at the first interaction, unless obvious. This binds you as provider of your in-house system.
• Art. 50(2) — machine-readable marking of synthetic output, applicable to systems placed on the market from 2 Aug 2026; systems already on the market before that date have until 2 Dec 2026 (new Art. 111(4)).³ Internal-only assistants whose output is never published arguably generate no synthetic content "published" to third parties, but state this assessment in writing rather than assuming it.
• Art. 50(3)/(4) — deployer duties for emotion recognition, biometric categorisation and deepfakes; usually out of scope for a coding assistant, in scope the moment you expose generated media.
• Commission guidelines (adopted 20 July 2026) and the voluntary AI Office Code of Practice on Transparency of AI-Generated Content provide the practical implementation path; signing the Code creates a presumption-of-conformity comfort for the marking duties.⁷

AI literacy (Art. 4) has applied since 2 February 2025 — verified; enforcement of supervision measures runs since 3 August 2026. Two corrections to popular summaries are worth noting: the Omnibus deferred neither the date nor the existence of the duty; the adopted text rewrote Art. 4 to require measures supporting AI literacy of staff operating AI systems (a lower bar than the original "ensure sufficient level"), but the duty itself is unchanged in existence and date.³ For a self-hosting company this is one training deck and a role-based curriculum — cheap, and the first thing an authority asks for.

5. Checklist: a defensible self-hosted stack

6. Bottom line

Self-hosting an open-source LLM in the EU is, under the AI Act, usually the lowest-friction deployment: transparency-tier rather than high-risk, no dependence on a foreign provider's GPAI compliance, and a clean controller position under the DSGVO. It is not regulation-free — Art. 50, Art. 4, and the full GDPR stack apply today — but every remaining obligation is one you control in your own infrastructure rather than one you must police in someone else's contract.

If your company operates in Germany and needs this evaluated for a concrete stack, the following articles may be useful: How LLMs work, Quantization explained, and KV cache explained.

This article is general information about the legal situation as of 24 September 2026, not legal advice. Binding classification of your specific system requires an individual assessment.