Scherwiller, near Sélestat — the model runs where you are
For many organisations the question is not which model is cleverest, but where the sentences you hand it end up. A litigation file, an operative report, an industrial production plan, a council resolution — those texts have no business sitting with a third party. We install an open language model on a machine dedicated to you, in our technical room in Centre Alsace, and we run it.
Let us start with candour
We would sell sovereign AI badly by lying about our own. The assistant you can try a little further down this page currently goes through a model provider's programming interface: your questions are passed to it. The gateway is already written to query our own machine first and fall back to the API only when it is unavailable — all that is missing is the machine, due when hosting goes live.
Try it, right now
It answers from the information published on this site, makes no commitment on behalf of RB2S, and tells you when it does not know — including about the machine running it right now. No sign-up, nothing is kept.
What “Alsatian” actually means
Plenty of offerings call themselves sovereign because the company is French, while inference runs on an American hyperscaler. The three questions are separate, and we answer them separately.
The model is loaded into memory and executed on physical hardware at 6 rue du Sommerberg. Your requests cross no border, because they never leave the building.
We rack, power, monitor and restart it ourselves. Nothing is subcontracted to an operator whose name you would not know, and your contact is less than an hour's drive away.
A GDPR-compliant processing agreement, French law, French jurisdiction. No extraterritorial legislation such as the Cloud Act layers itself on top.
Who this is for
Local AI is pointless for drafting a job advert. It matters a great deal once the material is confidential by nature, or once handing it to a third party raises a professional, contractual or regulatory problem.
Summarising a file, comparing contract versions, building a chronology: immediately useful work, on documents a solicitor or an accountant cannot pass to a third-party service without thinking twice.
Reports, letters, transcriptions. As soon as health data is involved, hosting and transfers obey strict rules: a dedicated machine on site makes the demonstration considerably simpler.
Manufacturing procedures, test reports, technical documentation, internal code. These are precisely the texts that constitute the competitive advantage — and the ones people hesitate most to paste into a public chat window.
Resolutions, correspondence from residents, public procurement. Public buyers look closely at where processing takes place, and regional infrastructure is easier to justify before a council than a distant framework agreement.
At high volume, per-token billing becomes unpredictable. A dedicated machine has a fixed monthly price, whether you query it ten times or ten thousand times a day.
If you need the strongest models on the market, very long reasoning, or a massive and irregular load, a public API remains the right tool. We would rather lose a rental than install a machine that will disappoint you.
How it is set up
What you want to do, on which documents, for how many people at once. That conversation determines the size of the model, and therefore the machine — never the other way round.
First conversation, no commitmentWe run the candidate model on your own documents, and you judge it on your texts rather than on a demonstration we picked. This is also where you find out whether a smaller model would do.
Over a few daysA dedicated Mac, a GPU server or a Jetson module, depending on what the trial showed. It is ordered, installed in the technical room, and you get root access and a fixed IP address.
Within seven to fifteen daysThe model is exposed to your applications through an interface compatible with the market standard: anything that already talks to a public API talks to your machine by changing one address. Monitoring and backups included.
Then monthly, 30 days' noticeWhat it runs on
The size of model you can run depends first on available memory, and its speed on that memory's bandwidth. These are the orders of magnitude we use when advising — and we redo the arithmetic with your figures.
Prices for dedicated Macs are on the hosting page. Prices for GPU servers and Jetson modules are quoted within 48 hours: they depend on the hardware price at the time of order and on actual electricity consumption.
What we will not tell you
At low volume, a machine left switched on costs more in electricity than moderate API use costs in tokens. The benefit lies elsewhere: confidentiality, and a cost that stops moving as volume grows.
The open models that fit on a company machine are good, sometimes excellent at bounded tasks — summarising, extracting, rewriting, classifying. On long and difficult reasoning, the large proprietary models keep an edge.
The photovoltaic roof covers part of daytime consumption. A machine running day and night consumes well beyond that. We give the actual coverage rate to anyone who asks.
Frequently asked
First conversation
We will tell you which model is enough for it, which machine it needs, and whether a public API would serve you better. Reply within twenty-four working hours, first meeting with no commitment.