An AI installed on your premises, without the cloud: for whom, how and at what cost?
Updated on 23 October 2026: Macs added, plus a link to our hardware comparison.
The cloud is not the only option
When people talk about generative AI, they think of ChatGPT or Copilot: online services, powerful, but which mean sending your texts and documents to external servers. There is an increasingly credible alternative: "open" models (Mistral, Llama, Qwen, Gemma…) that you download and run on your own computer or server.
At ExsIT, we use this kind of setup every day — for translation, writing and document processing. It is no longer an enthusiast's hobby project: it is a working tool.
Who is it relevant for?
- Regulated professions: medical, paramedical, legal, accounting, notarial;
- HR departments: payslips, appraisals, personnel files;
- Businesses with trade secrets: design offices, R&D, industrial subcontractors;
- Organisations that want a fixed cost rather than a per-user subscription.
What can you expect?
Local models do not always match the largest online models on complex reasoning. But for a large share of office tasks, they do the job very well:
- summarising and rephrasing documents;
- drafting emails or letters;
- translating (FR / NL / EN in particular);
- querying a base of internal documents ("what does our procedure say about…?");
- extracting structured information from a text.
What hardware?
The decisive factor is the graphics card memory (VRAM), which determines the size of model you can use and how fast it runs.
| Configuration | Typical use | Order of magnitude |
|---|---|---|
| Recent PC without a dedicated graphics card | Small models, tests, occasional use | Existing hardware |
| Workstation with a 12–16 GB GPU | Comfortable individual use (models of ~8–14 billion parameters) | Mid-range workstation |
| Server with a GPU of 24 GB or more | Shared use by a small team | SME server-type investment |
| Mac with unified memory (Mac mini, Mac Studio) | Larger models, silence, low power consumption | From individual workstation to small team server |
Specific machines, prices and measured speeds: our comparison Local AI: which hardware for which budget? reviews six configurations from €1,069 to €6,649 (October 2026 prices), with the AI model suited to each.
On the software side, tools such as Ollama or LM Studio make installation far more accessible than two years ago, with a chat interface similar to ChatGPT.
The limits to know
- A local model has no internet access: it does not know the news.
- Quality depends heavily on the choice of model and its configuration.
- The installation has to be maintained (updates, backups, secure access) — like any server.
A hybrid solution, often the best
In practice, many SMEs combine the two: a managed online assistant for routine, low-stakes tasks, and a local AI for sensitive documents. This is the approach we recommend most often (see also AI and GDPR).
Shall we talk?
We size the hardware, install and configure your local AI, and train your users. The hardware is chosen for your real use — without selling you what you do not need.