Digital sovereignty
Replacing the dependency
on American companies
This is not a political opinion, it is a technical property: does your system keep working if a supplier decides otherwise? We build so that the answer is yes.
The problem, straight
What a dependency costs
A dependency only hurts at the moment it reminds you it is there. Here are the five moments, all of which have already happened to somebody:
- The terms change. The supplier changes its prices, its limits or its usage rules, and you sign nothing: it is already done.
- The service stops. A closed API, a suspended account, a country no longer served — and your work chain stops with it.
- The data lives elsewhere. It sits with them, under their backups, and a foreign law can compel them to hand it over without asking you.
- The roadmap is not yours. What you need is not on the next release plan. You wait.
- The cost varies. Usage pricing cannot be budgeted: it grows with volume, and your volume decides.
The replacement
What we replace, and with what
Nothing exotic: free software, installed at your site, which we know how to run and keep up to date.
The last line is not a promise: open your browser console on this page, there is no request to a foreign domain.
Two examples
This is not theory, it is work already done
Droidian, on an abandoned phone
We installed a complete Linux system — Debian — on a 2018 Huawei P20 Lite that its manufacturer no longer supports: same device, another system, and nobody decides in your place when updates stop. The phone stays a phone: it calls, it takes the network, it fits in a pocket.
This is not a demonstration project: it is the system we use.
Odysseus, the assistant that stays home
A free agent harness (AGPL-3.0) turned into a native desktop application, with no embedded browser: the server computes, the client displays. Models running on your hardware plug straight into it — which is what makes it possible to compare several models on the same question, blind, and to choose on evidence rather than reputation.
The licence and the provenance stay in the repository, with the code.
The hardware
A 2018 card and 30 GB of memory are enough
It is the question that always comes back: “surely you need a $20,000 server to run a large model?” No. And the reasoning is simple.
A large model does not have to fit inside the graphics card's memory. What fits there is only the part that is active while it works. The rest — which is nearly all of it — can live in system memory and on disk, and be called in as needed. That is exactly what a mixture-of-experts (MoE) model is: at each step, a small part of the parameters is engaged, not the whole set.
Our reference workstation: a GeForce RTX 2080 Ti — 11 GB of video memory, a 2018 card — backed by 30 GB of DDR4 system memory. Nothing exotic, nothing new.
What matters
The card's memory is not the limit of the model. It is the limit of what you keep within immediate reach.
What really helps
System memory in quantity — 30 GB changes everything — and a disk fast enough to feed the next step.
What comes out
An open model, served at your site, with no usage bill and no data leaving the room.
And performance?
On the tasks a process actually contains — reading a document, filling a field, writing a query, structuring an answer, classifying an image — an open model served on modest hardware does the work usually asked of a paid service. Open models in 2026 are no longer toys: what cost a fortune two years ago now runs on an 11 GB card — a GeForce RTX 2080 Ti, released in 2018 — and 30 GB of DDR4 system memory.
What that changes for a company, more than the price: the data stays, the cost is fixed, and no supplier can change the rules or cut the service. Three things a cloud service cannot promise, by construction.
What it does not replace, and we say so: a frontier model on every subject at once and on the first attempt at an unseen problem. On a targeted task, the comparison holds; elsewhere, measure before asserting.
What it is not
Sovereignty is not autarky
We are not asking anyone to rebuild everything and throw everything away. A company is entitled to use American software that works well. What sovereignty adds is the exit: being able to leave with your data, your rules and your process, the day the supplier changes its mind.
Buying from a supplier who can close the door on you is not a problem as long as the door stays open. We build so that it stays open.