A company that runs an open-weight language model on its own server for internal use is a deployer under the EU AI Act, not a provider. For most internal uses its obligations today are AI literacy measures (Article 4), staying clear of the prohibited practices (Article 5) and the transparency duties of Article 50. The heavier high-risk rules for Annex III use cases were postponed to 2 December 2027 by Regulation (EU) 2026/1744, in force since 27 July 2026. Under the GDPR the company is a controller whether it calls a US API or hosts the model itself. Self-hosting removes the processor contract and the international transfer for the model step, and leaves security, impact assessment and lawfulness with the company.
This article is information for IT and compliance leads, read from the legal texts and the Commission and EDPB pages on 16 November 2026. It is not legal advice. Running a model yourself does not make any processing compliant: that depends on your purposes, your legal basis, your safeguards and your documentation, and on advice from your data protection officer or counsel.
The AI Act timeline as amended in 2026
The AI Act is Regulation (EU) 2024/1689. It was amended by Regulation (EU) 2026/1744 of 8 July 2026, often called the Digital Omnibus on AI, published in the Official Journal on 24 July 2026 and in force on 27 July 2026. The omnibus is adopted law. Its main effect for most companies is that the high-risk dates moved.
| Date | What applies | Changed by the omnibus? |
|---|---|---|
| 1 August 2024 | The AI Act entered into force | No |
| 2 February 2025 | Prohibited AI practices and AI literacy obligations apply | The date stands; the wording of Article 4 was replaced |
| 2 August 2025 | Governance rules and the obligations for general-purpose AI (GPAI) models apply | No |
| 2 August 2026 | General date of application, with exceptions. The AI Office and national authorities implement, supervise and enforce | The date stands; the high-risk rules were carved out of it |
| 2 December 2026 | New prohibitions added by the omnibus (systems producing non-consensual intimate content or child sexual abuse material). Providers of generative systems placed on the market before 2 August 2026 must comply with the marking duty in Article 50(2) by this date | New |
| 2 December 2027 | High-risk rules for the Annex III use cases | Postponed from 2 August 2026 |
| 2 August 2028 | High-risk rules for AI embedded in products regulated under Annex I | Postponed from 2 August 2027 |
| 2 August 2030 | Providers and deployers of high-risk systems intended for public authorities must comply | No |
As of 16 November 2026 the prohibitions, the literacy duty, the GPAI model rules and Article 50 apply, and the next date is 2 December 2026.
Provider or deployer: which one are you?
Article 3 defines both roles. A provider is a person or body that "develops an AI system or a general-purpose AI model or that has an AI system or a general-purpose AI model developed and places it on the market or puts the AI system into service under its own name or trademark, whether for payment or free of charge". A deployer is a person or body "using an AI system under its authority except where the AI system is used in the course of a personal non-professional activity".
A company that downloads open weights and runs them behind an internal chat tool is a deployer. The organisation that trained and published the model is the provider of the GPAI model, and most of the Act's documentation and conformity duties sit with providers.
Three ways the line moves:
- You offer the system to others under your name. A chatbot product sold to customers makes you a provider of that AI system.
- Article 25. A deployer becomes the provider of a high-risk system if it puts its name or trademark on one, substantially modifies one, or changes the intended purpose of a system, including a general-purpose one, so that it becomes high-risk. Pointing an internal chat model at CV screening is the textbook example of a changed purpose.
- Fine-tuning. Whether fine-tuning an open model makes a company a GPAI model provider is addressed in the Commission's GPAI guidelines. We did not read them for this article, so we give no answer here: take advice before you fine-tune and distribute a model.
What must a company deploying an open-weight model internally do?
| Article | Applies when | What it asks of a deployer |
|---|---|---|
| Article 4, AI literacy | Always, since 2 February 2025 | As replaced by the omnibus: "take measures to support the development of AI literacy of their staff". The text adds that this "does not require providers or deployers to guarantee any specific level of AI literacy of any individual" |
| Article 5, prohibited practices | Always, for everyone | Do not build or use the banned applications, including the ones added from 2 December 2026 |
| Article 50, transparency | When the use case triggers it | For example Article 50(4): a deployer of a system that generates a deep fake "shall disclose that the content has been artificially generated or manipulated". Notices also apply to emotion recognition and biometric categorisation, and to AI-generated text published to inform the public |
| Article 26, deployers of high-risk systems | Only if the use is high-risk; from 2 December 2027 for Annex III | Use the system according to its instructions, assign human oversight to competent people, keep input data relevant, monitor operation, report incidents and suspend use on risk |
| Article 25 | If you rebrand, substantially modify or repurpose into high-risk | You take on the provider's obligations for that high-risk system |
In practice, an internal assistant for drafting, coding and search needs four things on file: an inventory entry, an acceptable-use policy that rules out Article 5 practices and names the uses that would be high-risk, a record of the literacy measures taken, and a check of whether any published output triggers Article 50.
The open-source exemptions and where they stop
There are two, and neither is a general pass for deployers.
Article 2(12) says the Regulation "does not apply to AI systems released under free and open-source licences, unless they are placed on the market or put into service as high-risk AI systems or as an AI system that falls under Article 5 or 50". So an open-source system used for a high-risk purpose, a prohibited practice or a use covered by the transparency article is back inside the Act.
Article 53(2) concerns the providers of GPAI models. It lifts the duties in Article 53(1)(a) and (b), the technical documentation and the information for downstream providers, for models "released under a free and open-source licence that allows for the access, usage, modification, and distribution of the model, and whose parameters, including the weights, the information on the model architecture, and the information on model usage, are made publicly available". Its last sentence reads: "This exception shall not apply to general-purpose AI models with systemic risks". The copyright policy and the public summary of training content in points (c) and (d) still apply to the model provider. We did not check whether the omnibus changed Article 53, so read the consolidated text.
What this means for a deployer:
- The exemption belongs to the model's provider. It does not reduce your duties under Articles 4, 5, 26 or 50.
- The licence has to allow access, use, modification and distribution. Several popular open-weight models are published under non-commercial or custom licences. Whether a given licence meets the test is a legal question about that licence, and "open weights" is not the same thing as "free and open-source".
- A model with systemic risk does not benefit, however it is licensed.
GDPR when you call a US LLM API
If prompts contain personal data (names in e-mails, customer tickets, HR files), sending them to an API is processing by a third party on your behalf.
- Article 28. The API vendor is your processor. The controller "shall use only processors providing sufficient guarantees", and a written contract is required. Check the data processing agreement for retention, sub-processors, use for training and audit rights.
- Chapter V. If the vendor processes outside the EEA, the transfer needs a basis: an adequacy decision (Article 45), appropriate safeguards such as standard contractual clauses (Article 46) or, narrowly, a derogation (Article 49). Article 48 says third-country court or authority orders are not recognised unless based on an international agreement.
- The EU-US Data Privacy Framework. The adequacy decision, (EU) 2023/1795 of 10 July 2023, is in force as far as the Commission's page shows, and covers transfers to US companies certified under the framework. Its future is not settled: according to press and legal reports we did not verify against the Court's own register, the General Court dismissed the challenge in Latombe v Commission (T-553/23) on 3 September 2025 and an appeal (C-703/25 P) was described as pending in mid-2026. Plan for both outcomes: confirm the vendor's certification and have standard contractual clauses in the contract as a fallback.
- Article 35. A data protection impact assessment is required for processing likely to result in a high risk, "in particular using new technologies".
Where the API is the better choice: low volumes, data that is not personal, and tasks that need the quality of a frontier model. The API versus self-hosted cost comparison covers the money side.
GDPR when you run the model yourself
Self-hosting removes the AI vendor as processor and removes the transfer for the inference step. It leaves you as controller with these duties:
- Article 32, security of processing. Access control, encryption in transit and at rest, logging, patching and backups are now yours. The server arrives with a clean OS and root access.
- Article 35. The DPIA question does not go away because the model is local.
- Article 28 again, for hosting. The company that hosts the server is a processor if personal data is stored on it, so a data processing agreement is still needed. MassiveGRID is a GDPR-compliant company, ISO 27001 and ISO 9001 certified; the GDPR compliance page describes the commitments.
- Chapter V, for location. The transfer analysis now depends on where the server is and who operates it. We do not publish the location of the GPU servers: if residency matters, ask sales which data center a server will be delivered in before ordering and write the answer into your records.
- The model itself. EDPB Opinion 28/2024, adopted on 18 December 2024, makes three points that reach deployers. Whether a model is anonymous is assessed case by case, and it must be "very unlikely" that individuals can be identified from it or that personal data can be extracted through queries. Legitimate interest can be a legal basis for developing or deploying a model, subject to a three-step test. And a model developed with unlawfully processed personal data can affect the lawfulness of its deployment unless it has been duly anonymised. Keep what the model's provider publishes about training data on file.
- The 2026 guidelines. The EDPB lists Guidelines 02/2026 on anonymisation and Guidelines 03/2026 on web scraping in the context of generative AI, both dated 8 July 2026, with their consultations closed. Watch for the final versions.
API or self-hosted: a decision table by data category
The table is a starting point for a discussion with your DPO, not a ruling.
| Data in the prompt | Hosted API | Self-hosted model | Usual lean |
|---|---|---|---|
| Public or non-personal data (marketing copy, public documentation, open-source code) | No GDPR issue for the prompt; contract terms on confidentiality still matter | Works, but the server costs the same whether used or not | API, unless volume is high |
| Ordinary personal data (customer support tickets, CRM notes) | Article 28 contract, transfer basis, DPIA screening, retention settings | Article 32 security and DPIA screening; no AI processor, no transfer for inference | Either; decided by volume and by how much vendor management you want |
| Special categories and employee data (health, HR files, union membership) | All of the above with a full DPIA; the transfer question becomes the hard part | Full DPIA and strict access control; data stays on one machine | Self-hosted |
| Trade secrets and unreleased source code | Not a GDPR matter, but retention and training terms are critical | No third party sees the material | Self-hosted |
| Data under professional secrecy (legal, medical, audit) | Professional rules may require client consent in addition to the GDPR | Removes the AI provider; the hosting contract and location remain | Self-hosted, see the posts on law firms and healthcare |
| Any data, in a use that would be high-risk under Annex III | Article 26 duties from 2 December 2027 | The same Article 26 duties, plus the Article 25 risk if you modify or repurpose | Decided by the use, not the hosting |
On cost, paying per token costs less at low volume. A dedicated server is a fixed bill: the GPU dedicated servers start at $449.99 per month for 24 GB of video memory and $1,656.15 for 96 GB, each a bare-metal machine for one customer. The private AI hosting page maps workloads to servers, and the VRAM guide explains which models fit.
Frequently Asked Questions
Does the EU AI Act apply to a self-hosted open-source LLM used only internally?
Yes, in part. The company is a deployer. The AI literacy duty in Article 4 and the prohibitions in Article 5 apply now, Article 50 applies if a use triggers it, and Article 26 applies from 2 December 2027 if the use falls under the Annex III high-risk list. The open-source exemption in Article 2(12) does not cover high-risk uses or uses under Articles 5 or 50.
Did the AI Act high-risk deadline of August 2026 move?
Yes. Regulation (EU) 2026/1744, in force since 27 July 2026, postponed the high-risk rules for Annex III use cases from 2 August 2026 to 2 December 2027, and those for AI embedded in Annex I regulated products from 2 August 2027 to 2 August 2028.
Is calling a US LLM API with personal data legal under the GDPR?
It can be, with an Article 28 processing contract and a valid transfer basis under Chapter V. The EU-US Data Privacy Framework adequacy decision of 10 July 2023 is in force as far as the Commission's page showed on 16 November 2026, and a court challenge to it was reported as pending on appeal, so keep standard contractual clauses as a fallback. This is not legal advice.
Does self-hosting an LLM make us GDPR compliant?
No. It removes the AI vendor as a processor and the transfer for the inference step. You remain the controller, with duties for security (Article 32), impact assessment (Article 35), a legal basis and a contract with the hosting provider. If the server location matters, confirm it with sales before ordering a GPU server.
Does the Article 53(2) open-source exemption help a company that only uses the model?
Not directly. It relieves providers of open-source GPAI models from the technical documentation and downstream-information duties, and it does not apply to models with systemic risk. A deployer's own obligations under Articles 4, 5, 26 and 50 are unchanged.
Run the model on a server that is yours alone
MassiveGRID GPU dedicated servers are bare metal with one customer per machine and root access: an RTX PRO 4000 with 24 GB for $449.99 per month, or an RTX PRO 6000 with 96 GB from $1,656.15, each with a 1 Gbps port and unlimited traffic, hosted in an ISO 27001 certified data center. If data residency matters, ask sales which data center a server will be delivered in before ordering.
See Private AI HostingFurther Reading
- Regulation (EU) 2024/1689, the AI Act (Articles 2, 3, 4, 5, 25, 26, 50, 53)
- Regulation (EU) 2026/1744 of 8 July 2026 amending the AI Act
- European Commission: AI Act application timeline (updated 19 October 2026)
- European Commission: AI omnibus enters into force (27 July 2026)
- Regulation (EU) 2016/679, the GDPR (Articles 28, 32, 35 and Chapter V)
- EDPB Opinion 28/2024 on AI models and personal data
- European Commission: EU-US data transfers and the Data Privacy Framework