Local AI and GDPR: what running AI on your own server solves and what stays with you

Is local AI GDPR compliant?

Local AI is not GDPR-compliant by itself, and a cloud assistant is not unlawful by itself. Running the model on your own server takes one party out of the picture, the model provider, together with its contract, its retention of your prompts and any transfer of data to it outside the EU. The duties of the controller stay with your company: a lawful basis, a clear purpose, short retention, access control, answers to people who ask about their data and, where the risk is high, a data protection impact assessment.

This guide is for the owner, managing director or office manager who has heard "AI without the cloud is GDPR-safe" and wants to know what that covers. It separates the GDPR topics that own hosting settles from those it leaves as they are, and ends with the server and the building it needs. It is general information as of October 2026 and no legal advice; your data protection officer or legal counsel confirms the assessment for your company.

Four terms in plain words

The European Commission's pages for businesses define the roles in one sentence each, and the rest of this article uses them.

• Personal data is "any information that relates to an identified or identifiable living individual". A customer's name in an email, an employee's leave request and a voice recording are personal data; a recipe or a machine drawing usually is not, which is why our article on trade secrets and AI covers those separately.
• Controller: "The data controller determines the purposes for which and the means by which personal data is processed." When your staff use an AI assistant for their work, that is normally your company.
• Processor: "The data processor processes personal data only on behalf of the controller." A cloud AI service, a hosting provider or an IT partner with administrator access is typically a processor, and the GDPR requires a binding contract with each of them.
• Transfer: personal data that goes to, or can be reached from, a country outside the European Economic Area (the EU plus Iceland, Liechtenstein and Norway). Chapter V of the GDPR allows that only under its conditions, such as an adequacy decision or standard contractual clauses.

On-premise LLM and data protection: what changes on your own server

With a cloud assistant, every prompt and every uploaded file travels to the provider, which processes it on your behalf. The provider's terms then decide part of what happens to it: how long it is kept, which of its staff and contractors may see it, and in which country it is stored. OpenAI, for example, says on its Enterprise privacy page (updated 8 January 2026) that by default it does not use business data for training, that deleted ChatGPT Business conversations are removed from its systems within 30 days unless longer retention is required by law or needed to protect its services, and that it can sign a data processing agreement. Those are the points your data protection officer reviews before the company approves such a service.

When an open-weight model runs on a server you own, none of that applies to the model. The weights are a file on your disk, and answering a question does not change them or send the question anywhere. There is no model provider to sign a contract with, no provider retention period, and no transfer for the model if the server stands in the EU. The table sorts the main GDPR topics.

Regulation (EU) 2016/679 as summarised on the European Commission's data protection pages for businesses, read 25 October 2026; article numbers as quoted from EUR-Lex in our guides to a DPIA and to GDPR cloud hosting; OpenAI Enterprise privacy page, updated 8 January 2026. The table is our reading, not a legal classification.

The upper rows are what people mean when they say local AI solves GDPR. They are real gains: the provider contract, the transfer question and the provider's own use of inputs are often the points on which an approval stalls. The lower rows are the same for both options.

GDPR-compliant AI for a business: what stays with your company

Five duties do not depend on where the model runs.

A lawful basis and a purpose. The Commission's page lists six grounds, consent, a contract, a legal obligation, vital interests, a task in the public interest and legitimate interests, and says a company "can only process personal data" in those circumstances. An assistant that summarises customer emails uses personal data for a purpose that has to be named. Purpose limitation also means that data collected for one task does not go into the assistant for a different one without a check.

Minimisation and retention. Only the data needed for the purpose, kept for no longer than necessary. On your own server you set the retention of each store yourself, which is an advantage only if someone sets it. A chat history that is never deleted on a local server is the same problem as one kept by a provider.

Information and requests. People have to be told what is done with their data and for how long it is stored. Requests for access, correction or erasure have to be answered "without undue delay, and in principle within 1 month", the Commission writes.

Security. "Appropriate technical and organisational measures" apply to your server as to any other system: access rights, encryption, updates, logged administrator access and a tested restore. A server in your own building that anyone on the network can reach is weaker than a well-run cloud service.

Impact assessment. Where processing "is likely to result in a high risk to the rights and freedoms of individuals", a data protection impact assessment is required before it starts. An assistant with access to HR files or health data will often meet that bar wherever it runs. Our guide to a DPIA for an internal LLM walks through the procedure and the inputs IT supplies for it.

Employee data needs extra care. The query log of an assistant shows who asked what, so it is personal data about your staff, and national rules on employee data can apply in addition to the GDPR. Agree with your data protection officer who may read the log and for what purpose, and involve the works council where your company has one, as our article on the works council and an internal AI assistant describes.

Where personal data sits in an AI assistant

An assistant on company documents is more than a model. It has a chat front end that keeps conversations, a gateway that logs prompts and answers, a search index with passages from your documents, the uploaded files themselves and the backups of all of these. The model's working memory on the GPU holds a conversation only while it is being answered.

Our summary of the data flow in our DPIA guide; the EDPB's Opinion 28/2024, quoted there, considers that models trained on personal data "cannot, in all cases, be considered anonymous".

When a customer asks for erasure, the index has to drop the passages from that customer's documents, the log has to drop or anonymise the entries, and the backups follow when they expire. Plan these functions before go-live. Our RAG on company data guide explains how the search index works, and our guide to removing personal data before it reaches an LLM covers masking names and numbers where the task does not need them.

Our Private AI/ML service builds the assistant on your server with a query log and data and permissions management; tell us which documents and which teams the assistant should serve, and our team proposes the setup and the server for it.

Suppliers that still count

Own hosting removes the model provider, and every other supplier with access to the data stays in the picture.

• Hosting. If the server stands in a data centre and the operator can reach the data, the operator is a processor.
• Engineering and support. An IT partner that installs and maintains the platform with administrator rights is a processor and needs a data processing agreement. Our Private AI/ML page states how we work: engineers work in your environment under your access controls, we take no copies of production data out of it, an NDA comes before technical detail, a data processing agreement is available on request and subprocessors are named in the contract.
• Remote access. For the EDPB, remote access from a third country, for example for support, is a transfer just as storage there is, as our guide to GDPR cloud hosting and the Article 28 contract quotes. Ask every supplier where its support staff work.
• Public services for some tasks. Many companies keep a cloud model for tasks without personal data. Each such connection brings that provider's terms back for the data that goes there. On the platforms we build, public APIs are used only when you enable them for a specific task, and what goes there is visible in the query log.

Where the cloud provider's home jurisdiction is the concern, our article on the CLOUD Act and EU data residency explains what location, operator and keys each change.

Which server, how many users, what it needs in the building

The server size follows from how many conversations run at the same moment, which is a small share of the head count. Our users-per-card rule takes 90 per cent of the 95.6 GiB that the driver reports on a 96 GB RTX PRO 6000, less 3 GiB per card and less the model. gpt-oss-120b, an open-weight model from OpenAI under the Apache 2.0 licence, takes 60.8 GiB, which leaves 22.2 GiB on one card for the memory of the running conversations: 19 conversations of up to 32,000 tokens at the same moment (a token is a word or a part of one).

Example values of our company-size guide (40 per cent of staff in the busiest hour, six requests an hour each, 30 seconds per request, peak factor 2); conversations at 32K by our users-per-card rule for gpt-oss-120b with a 16-bit cache, four cards as two copies split across pairs; power by our estimate (cards at 600 W, processors, plus 15 per cent for the rest of the server).

Two cards are the entry server on our AI servers page: two RTX PRO 6000 Blackwell Server Edition, two AMD EPYC processors, 384 GB of DDR5 ECC memory and two NVMe drives. One card runs the language model; the second is split by MIG, NVIDIA's way of dividing one card into isolated parts, for the search models and a test instance. Two servers let one fail or be updated while the other answers. Four cards per server carry a large company's peak in one machine each. Our company-size guide shows the arithmetic, and a pilot replaces our example values with your own.

For the building: the two-card server draws about 1.8 kW by our estimate and fits one 230 V 16 A circuit, which carries 2.94 kW at 80 per cent; the four-card server draws about 3.4 kW and needs 32 A or three-phase, as our guide to GPU server power on 16 A, 32 A and three-phase feeds works out. Both are rack servers and belong in a server room or a rented rack. Without a room of your own, our Private AI/ML service runs the platform on dedicated hardware in a Tier-3 data centre in Lithuania, inside the EU.

A first checklist for the management

1. Name the first two or three tasks and the kinds of personal data in each, for example customer emails, contracts or HR questions.
2. Ask your data protection officer whether one of them needs an impact assessment, and give them the data flow of the assistant.
3. Set a retention period for each store in the table above and a person who owns deletion.
4. Let the assistant's access rights follow the rights on the file shares, so nobody finds documents through the assistant that they could not open before.
5. List every supplier with access to the server, its location and its data processing agreement.
6. Tell staff which tool to use for which data, and train them before access; our shadow AI guide covers the rules for public chatbots, and our guide to AI literacy training the AI Act duty that applies alongside the GDPR.

Our guide to hybrid AI, confidential data local and the rest in the cloud shows how a company splits tasks between the two, and our comparison of ChatGPT Enterprise and an own AI server sets out what the provider's business plans document.

What we supply

You can buy this system from us: our team sizes the configuration to your workload and budget and replies within one business day. We build the servers to order with two, four or eight RTX PRO 6000 Server Edition cards, assembled and tested before delivery, delivered anywhere in the EU on one EU contract and invoice, with manufacturer warranty on every component handled through us; quantities and lead time are confirmed in writing with the quote. Our Private AI/ML service, delivered by our engineering partner Vixen.UNO, sets up the model, the document search, the query log and the access rights. The legal assessment stays with your data protection officer; we deliver the technical part. Send us the tasks and the number of staff, and we propose a server and a pilot.

FAQ