AI servers and GPU workstations in Corsica
We install AI servers and GPU workstations on your premises so you can run artificial intelligence models locally, in Bastia and across Corsica. Your documents are never sent to an outside service: the assistant, document search or writing aid runs on your own hardware, under your control.
Run LLMs on premises and keep your data in-house
An LLM, or large language model, is the kind of program that understands and writes text, like the online chat assistants many people now use. Most of these services run on remote servers: every question you ask and every document you share leaves your business.
There are now open models that can be downloaded and run on your own hardware. Installed on a server in your office, they can summarise documents, draft letters, answer questions about your internal procedures or sort emails, without the data ever leaving your network. This is the foundation of the private, local AI we offer through our artificial intelligence service.
AI server or GPU workstation: the right hardware
Running an AI model calls above all for a powerful graphics card, or GPU: a specialised processor able to carry out in parallel the huge number of calculations these models need. Depending on your situation, we offer several options.
- 01
An AI workstation
A desktop computer fitted with one or more GPUs, for a small number of users or to trial local AI before going further.
- 02
A dedicated AI server
A machine installed in a rack or server room, designed to run around the clock and serve the whole team over the network.
- 03
Software set-up
Operating system, drivers, the software that runs the models, a chat interface accessible from the browser, and user accounts.
- 04
Connection to your documents
Through our AI service, we can connect the model to your documents using RAG (retrieval-augmented generation), a technique that lets it answer by drawing on your own sources.
Private AI for sectors where confidentiality matters
- 01
Professional practices
Solicitors, accountants, notaries: professional confidentiality does not sit well with sending client files to an online service.
- 02
Healthcare
Health data is among the most sensitive there is. Processing it on hardware you control means you know exactly where it is.
- 03
Local authorities
Administrative documents and residents' data: local AI lets you explore useful applications without relying on an outside service.
- 04
Engineering consultancies and construction
Standards, specifications and technical reports: an assistant that finds the information in your own documentation.
Choosing a GPU for AI: the criteria that matter
- 01
GPU memory
This is the main criterion. The graphics card's memory (VRAM) determines how large a model can be loaded and how long the documents it can handle. Quantisation, a technique for compressing models, lets larger ones fit, at the cost of a slight loss of accuracy.
- 02
Number of simultaneous users
One person querying the model now and then and a whole team using it all day do not need the same power.
- 03
Intended uses
Drafting, summarising, searching your documents, transcribing meetings: each use points to a particular type of model, and therefore of hardware.
- 04
Environment
A GPU machine draws a lot of power, gives off heat and can be noisy. It needs a ventilated location, a suitable power supply and ideally a UPS (uninterruptible power supply).
- 05
Room to grow
Allowing space and power to add a GPU later saves replacing everything when your needs grow.
From assessing your needs to your AI server going live
- 01
Needs assessment
We identify with you the tasks where local AI brings a real benefit, the documents involved and the number of users, then check the intended location on site, in Haute-Corse or anywhere else on the island.
- 02
Recommendation and quote
A reasoned configuration and a clear, jargon-free quote, with an honest explanation of what the chosen hardware will and will not be able to do.
- 03
Supply, installation and configuration
Installation of the machine, operating system and chosen models, user accounts, network access settings and performance checks with your first users.
- 04
Handover and maintenance
A briefing for your team on good practice and the limits of AI, documentation, then ongoing support. Our training service can help your staff get up to speed.
Security, GDPR and the budget for an AI server
Keeping processing local does not remove the need for good practice. The AI server sits on your internal network, with no direct exposure to the internet; access is through named accounts; the documents the model can consult respect each person's permissions; and useful data is backed up like the rest of your system. GDPR still applies to any personal data processed, but you know exactly where it is and who can access it.
The budget depends mainly on the amount of GPU memory required, and therefore the size of the models you want to run, the number of simultaneous users, the choice between a workstation and a dedicated server, storage for your documents, any electrical or ventilation work and the level of software integration you want. We also weigh this option up with you against a well-governed online AI service, because local is not always the best answer.
Keeping your local AI up to date over time
Open models are improving fast: a more capable or lighter model may appear a few months after installation. We keep track of these developments, test new versions before rolling them out and adjust the configuration to how you actually use it. Maintenance also covers security updates for the operating system and drivers, monitoring of the machine and backups of its settings.
Technologies and methods
- GPU
- VRAM
- Open-source LLMs
- RAG
- Linux
- Docker
- Quantisation
Frequently asked questions
Is local AI as good as an online service?
The largest online models remain more powerful for some complex tasks. For many everyday business uses (summarising, rewording, drafting, searching documents), open models running locally give very good results. We show you what you can reasonably expect before you buy anything.
Is a workstation enough, or do we need a server?
For a handful of users or a trial phase, a workstation with a good GPU is often enough. A dedicated server makes sense when the whole team uses AI continuously, or when several applications run side by side.
Is our data really protected?
Documents and questions stay on your machine, within your network: nothing needs to be sent to an outside AI provider for it to work. On top of that, we add named accounts, a segmented network and backups.
Can the AI work from our own documents?
Yes. With document search (RAG), the model consults your procedures, contracts or archives to answer, and shows which documents it used. That work falls under our artificial intelligence service, which we pair with the hardware.
Is an AI server noisy or power-hungry?
A GPU machine uses more power and runs hotter than an office computer, especially under load. We take this into account from the assessment stage: location, ventilation, power supply and hardware sized to your actual needs.