Blog · Published · Updated
Private LLMs: data, EU hosting and on-premise
Short answer
A private LLM lets you choose where the answer is computed, but the app has more parts: text extraction, the index, logs and backups. Before you choose EU hosting or an on-premise server, draw a data map from the document to the answer. At Syntalith an app on a local model costs from €3,000 net plus GPU hardware. At 400 questions to documents a month of 12 minutes each, it pays back in 1.6 months.
How we workDetails ↓
- 01Call30 minutes with an engineer, free
- 02PrototypeA working system on your process and data, free
- 03ImplementationTarget and fixed price in the contract; 50% at signing
- 04Acceptance50% once the agreed target is met
- model on your server or in the cloud in the EU
- quoted after the review
- same work in application with EU cloud model
- from €3,000
- task, data and equipment analysis
- €0
We belong to the OpenAI and Anthropic partner programmes and learn through their official courses. What this means ↓
Author: Syntalith · 3 min read
Statistics
Private LLM and data: statistics 2026
1.6 mo.
payback of a €3,000 local app at 400 questions to documents a month: €2,792 of work today, €912 with checking and the GPU server. Scenario.
48.8%
According to Eurostat (2025), 48.8% of EU enterprises that considered AI and did not use it named concerns about data protection and privacy.
49%
According to the European Commission (2026), 49% of EU SMEs and mid-caps name legal uncertainty as a barrier to AI.
Source: European Commission, Flash Eurobarometer 578 (2026) ↗
84%
According to the European Commission (2025), 84% of Europeans think AI at work needs careful management to protect privacy and ensure transparency.
Source: European Commission, Special Eurobarometer 554 (2025) ↗
52.5%
According to Eurostat (2025), 52.5% of EU enterprises that considered AI and did not use it named a lack of clarity about the legal consequences.
Figures checked: 22 October 2026
A server in the office, logs in the cloud
„The data stay with us” can mean the company's building, its cloud account, an EU region or a ban on sharing content with the model provider. Each answer leads to a different architecture and cost. On-premise describes a location and an air gap describes isolation from other networks, and the two are not interchangeable: a server in the office can send logs to a cloud service if monitoring is set up that way.
A document passes through text extraction, search, the model and the app, and each stage may leave temporary files and logs. The model that creates the document's representation for search can also run locally or as an external service. An agent's tools, such as a call to the CRM or an internet search, are separate data flows.
- 01Documenttext extraction locally or in a service
- 02Indexpassages with the user's permissions
- 03Modelwhere the answer is computed
- 04Answerthe chat record and how long it is kept
- 05Backups and logswhere they are and who reads them
What it costs and when it pays back
- Today
- 400 questions to documents × 12 min = 80.0 h (€2,792)
- With the app, an employee checks the output in 25% of the time
- €698
- Hetzner GEX45 GPU server
- €214
- Saving
- €1,880 a month
Payback of €3,000
1.6 months
The volume, the time and the checking share are the scenario's assumptions, which we measure in the trial. €34.90 an hour is the EU average hourly labour cost for 2025 according to Eurostat. Price of the GEX45 server with 24 GB of GPU memory from Hetzner's page on 22 October 2026, without the one-off €209 fee.
The count covers the time spent searching, reading and compiling material. What remains is the time to check the result against the source, 25% of today's work in the scenario. We measure both times in a trial on 20 to 100 cases the team has already handled.
Permissions must work before the answer
A salesperson and a lawyer may use the same model but see different documents, so the app limits search by the user's permissions. An instruction in the prompt does not replace access control. We also test a change of role, the removal of permissions and the deletion of a document, because the index, the cache and the chat record may each need a separate update. A model fine-tuned on confidential examples needs protection too: we limit access to the training set, the adapters and the weights.
- Where the app takes documents from and where it saves results
- What stays in logs, backups and the history of questions
- Who has access to questions, answers and outputs
- Who updates the model, the app and the document set
When EU hosting is enough
If the requirement is an EU region and a data processing agreement, a service in a European cloud region often costs less than your own server. When the requirement forbids internet connections during operation, you need a server on site or an isolated network. The GDPR requires a legal basis for processing, data minimisation, limited retention and security measures, and processing outside the EEA is subject to the transfer rules. The person responsible for data protection assesses these questions, and Syntalith provides the technical description, the data flow and the configuration. Signing off an environment with no connection is described in a local model in manufacturing.
An EU region and a processing agreement are required
a service in an EU region
no hardware
A model without sending content to the provider
a GPU server, for example the GEX45
€214 a month
No internet connections
a server on site and an isolated network
priced after an analysis
Where to start
The most useful inputs are specific requirements: the allowed region, no connections during operation, user roles, the retention period and recovery after a failure. The phrase „the highest security” sets none of these parameters. We deploy local models as part of private LLMs, and we build the view for working with sources as an AI app. For the free analysis, a description of one case and of where the materials sit is enough. Confidential files stay in the company, and you get a written plan within 2 working days.
Worked example · Scenario
How we count the payback: a local model
We run the same count for your company in the free analysis.
€5,817
a month for the work on 1,000 questions
EU cloud version payback: 0.6 mo.
Our assumptions are 10 minutes to find the right instruction, and checking in 15% of the old time. €34.90 an hour is the EU average hourly labour cost for 2025 according to Eurostat, for companies with 10 or more employees. Model cost: Claude Sonnet 5.5, about 6,000 + 500 tokens per answer in the EU cloud version, at the ECB rate of 1.1205 USD per euro on 22 October 2026. We count the payback for an app with a model in the EU cloud (from €3,000 net). We price the model on your server after a call, and the cost of people's work stays the same.
See the full calculation →