As of 22 November 2026. Step by step to your own data sovereignty: after server, monitoring and backup, AI is the next building block that can stand on your own premises. You get a finished device from us, set up for your task. You run it yourself, and we monitor it on request.
The offer is in preparation. This page shows what runs and has been measured with us today, and says openly what is still missing. The technology behind it is covered in two reports of their own: AI box and AI rig.

The AI box: one board in its own case. Rendering based on photos. Case: Chris Leib.
Summary
| What | How it is meant |
|---|---|
| The device | AI box with one board or AI rig with twelve, on your premises |
| Setup | delivered ready: chat interface in the browser, accounts for your team, the right models loaded |
| Operation | you yourself. On request we monitor it with MrsNanny |
| Your data | The AI runs on your device. Questions and documents go to no AI provider in the cloud |
| Access from outside | only what you explicitly share, via secured share links |
| Cost | buy once, no subscription per employee. Prices come with the binding offer |
What the devices are not: a replacement for the big cloud models. They are small models for clear tasks. In return they are yours, and your data stays in-house.
The two devices
| AI box | AI rig | |
|---|---|---|
| Boards | 1 | 12 |
| Memory for models | about 13 GiB | twelve times about 13 GiB, each board on its own |
| Models | one at a time, up to the class with 14 billion parameters | several at once, with us one per pair of boards |
| Measured with us | 36.1 tokens per second (qwen3:14b, one request) | 16,590 answers in 10 minutes without an error (two simultaneous requests per board) |
| Power | its own power supply. We have not yet measured the consumption ourselves | 200–240 V, two power supplies. 1,432–1,442 W as the sum of the board sensors under full load |
| Placement | small 3D-printed case | 4 rack units in the server room, loud |
| Suits | one team, fixed tasks, getting started | many users and tasks at once |
A rig is not one big computer but twelve separate ones. A model that fits on one board runs twelve times in parallel. We have not yet measured spreading a larger model across several boards.

The live view of our own rig under load. A rig comes with it.
What it is good for
The examples come from our own operation. Whether the quality is enough for your task is something we test together beforehand, with your examples.
| Task | What we can show for it |
|---|---|
| Classifying and summarising alerts | Our daily security report runs this way. 80 alerts in 91 s on six boards; 78 of 80 classified the same in substance as a reference on an RTX 4090 |
| Routing requests to the right model | 39 of 39 test questions classified correctly, about 0.6 s per request |
| Writing reports | in use with us with a model of its own. We have not measured the quality in figures |
| Helping with code | in use with us with a model of its own. We have not measured the quality in figures |
| Describing images | in use with us with a model of its own. We have not measured the quality in figures |
| Chat in the team | Chat interface in the browser. With us ten simultaneous requests were spread across eight boards without an error |
An example from the AI rig
This replay shows how fast and how well a model on your own hardware answers open questions. The six answers on security, privacy, self-responsibility, freedom of speech and data sovereignty were given by our AI rig on 22 November 2026, unabridged and unedited, and arrive here at exactly the speed we measured. “Next question” takes you on. They are model answers, not statements of ours.
How it works
- Register interest. You tell us what the AI should do for you.
- Test together. With your examples we check which model fits and whether a box is enough.
- Binding offer. Device, setup and everything you want with it.
- Build and setup. We deliver the device set up and do the initial setup together with you.
- Operation. You run the device yourself. On request we monitor temperature, fans and load and install updates.
Your data
- The AI runs on your premises. Questions and documents do not leave your network towards an AI provider.
- If we monitor the device, that runs through an encrypted tunnel, as with our other building blocks.
- Only what you want is public. We set up shares one by one, secured.
No device of your own?
If you use MrsNanny and want the AI analysis with it, but do not want to set up a device, you can rent MrRowie SI with it. It then computes on our rig in a data centre in Austria. Your data travels there through an encrypted tunnel, not to a cloud corporation.
Best as a package
Each building block is a step towards your own data sovereignty:
| Step | Building block | Task |
|---|---|---|
| 1 | MrRowie | your own server |
| 2 | MrsNanny | monitoring and updates |
| 3 | MrMiyagi | backup |
| 4 | MrRowie SI | the AI that reads everything and reports little |
| 5 | AI box or AI rig | your own AI for your tasks |
What is finished today and what is still missing
| Finished and in operation with us | Still in progress |
|---|---|
| Our own rig has been computing in the data centre since 19 November 2026 | Prices and the binding offer |
| Fixed models per pair of boards, dispatcher, live view, fan controller | The first delivery to a customer is still to come |
| Daily security report on the rig | The tunnel connection from customers to our rig is not built yet |
| Acceptance per board with load test and arithmetic check | Power draw at the socket, for box and rig |
| Chat interface in the browser in front of the boards | We are currently setting up accounts for several users |
More technology
- AI box BC-250: one board, build and measurements
- AI rig 12× BC-250: twelve boards, operation and build