Your own AI server, operated securely
Local AI for Business
Language models on your own hardware, for data that should not leave your organisation. We assess what you need, support the hardware selection, set everything up, harden it and look after operation.
Local or cloud: three ways to use AI with your own data
Local AI is not the right route for every company. Before you buy hardware, compare it with the other two. We advise on all three and recommend the one that fits your data.
Cloud service under contract
A business account with a model provider, with a data processing agreement and no training on your data. Enough for a lot of data, quick to set up, ongoing costs per user or per use.
Your own tenant
Models in your own cloud subscription, for example in Azure, with region and access rights under your control. Data sits in your tenant, operation stays with the cloud provider.
Local AI in-house
Open language models run on a server or device in your offices or your data centre. No model provider sees the content. In return you cover hardware and power, and we handle setup and support.
Local AI fits when
- you work with data covered by professional secrecy or confidentiality obligations
- internal rules or your customers require certain data to stay in-house
- as a regulated company you do not want to assess and monitor another cloud provider
- many employees use AI every day and per-user costs add up
Which AI server fits
What matters is the number of concurrent users and the response times you expect, more than the size of the company. Three classes cover most cases:
One device in the office
up to about 10 concurrent users
A workstation with a powerful graphics card or a compact computer with a large unified memory. Normal power socket, no server room.
Server for batch processing
overnight analyses and reports
Lots of memory and many processor cores, no graphics card. Suitable for tasks without time pressure. Processor power alone is not enough for fast answers to many users at once.
GPU servers for many users
larger organisations, around 100 employees
Several servers with graphics cards for short response times under many concurrent requests. Needs data centre space, suitable power and cooling.
You buy the hardware directly from the manufacturer or reseller. We do not sell hardware and do not run a data centre for customers.
What we take care of
Needs assessment and selection
We clarify the use cases (search in your own documents, drafting, document processing), the number of concurrent users and the requirements for your data, and derive hardware and models from them.
Procurement support
Specification, comparison of quotes, coordination of delivery. You sign the purchase contract directly with the manufacturer or reseller. We take no referral commission.
Setup
Model server and user interface on your hardware, connection to approved document stores and to your sign-on (single sign-on), user accounts with permissions.
Hardening
Network segmentation, access control and no unintended internet access for the model endpoints. Infrastructure security is our core field.
Operation and maintenance
Updates of models and software, monitoring, capacity planning as usage grows, remotely on your hardware.
Security testing on request
Testing the setup for prompt injection, jailbreaks and data leakage, following the Alliance for Cyber Security guide to penetration testing of language models.
Start small
One task first, then the platform
With the AI starter packages you first solve a single manual task, such as moving receipts into a spreadsheet. The package can run with us and move to your own platform later (operating options 3 and 4).
To the AI starter packages →Practices and law firms
Local AI under professional secrecy
For medical practices, law firms and tax advisers there is a dedicated offer: one device in the office, set up and supported, for patient and client data covered by Section 203 of the German Criminal Code.
Local AI for practices and law firms →Building and securing from one source
Setting up local AI is one thing, operating it securely is another. We advise on information security and run local language models in our own lab. Managing director Christian Gebhardt is a founding and permanent member of the AI Security expert group of the Alliance for Cyber Security and a co-author of its guide to penetration testing of language models (BSI/ACS, 2025).
Security testing for language models →Frequently asked questions
Is using ChatGPT at work always a data protection problem?
Not always. With a business account, a data processing agreement and no training on your data, a cloud service can be used for a lot of data. It gets difficult with data under professional secrecy, with special categories of personal data and wherever customers or supervisors set requirements. Whether a cloud service is enough for your data is a question for your data protection officer; we provide the technical basis for that decision.
What is the difference between local AI and AI from an EU cloud?
With an EU cloud, a provider processes your data in its data centre in the EU, governed by a contract. With local AI, the model runs on your own hardware and no model provider receives the content. Either can be right; it depends on your data and your obligations.
What hardware do we need for our own AI server?
Mainly that depends on the number of concurrent users and the response times you expect. For an office with up to about ten concurrent users a single device is often enough. For overnight analyses a server without a graphics card will do. Many concurrent users need servers with graphics cards. The exact choice follows from the needs assessment.
Do we buy the hardware from you?
No. You buy directly from the manufacturer or reseller and remain the owner. We write the specification, compare quotes technically and coordinate delivery, without commission from suppliers.
Are local models as good as the big cloud models?
For many office tasks such as drafting, summarising and searching your own documents, open models are sufficient. For particularly demanding tasks the largest cloud models are often stronger. That is why we test on your own tasks which model is enough before anything is bought.
This page is a technical overview and does not constitute legal advice.
Check whether local AI fits your organisation
In an initial call we go through use cases, user numbers and your data, and tell you whether local AI is the right route.