For businesses

AI on your own hardware, with your data at home

We install open-weight models on a machine in your business. They answer from inside your network, without sending your customers’ text to an AI provider and without a bill per query.

Book a 15-minute call How we work

In short

The models run on a machine you own, inside your network. Prompts and documents do not leave it.
You choose the hardware, with our measured recommendation. We do not sell machines.
We start with one concrete task, a written scope and a fixed quote before we start.

What it is used for

Three tasks that usually fit on a single machine. Yours may be another: we look at it on the call.

Search your own documents

Plain-language questions over contracts, procedures or spec sheets, with the citation of the document the answer comes from.

Meeting notes

Transcription and summaries of recordings without uploading the audio to any outside service.

Sort and draft replies

Incoming email, tickets or orders, sorted by topic, with a draft reply a person reviews.

Where everything lives

Your computers and your software talk to a machine in your office. The machine holds the models and the GPU. What you ask does not cross the edge of your network.

What changes compared with a cloud service

Your customers’ text does not reach a third party, so there is no extra AI provider in your record of processing.
The cost is the machine and the work to set it up, not a fee per user or per query.
You decide when a model is updated. Nothing changes without you knowing.
In exchange, the machine is yours: someone in your business has to own it.

What the law says

Two laws, with the article and a link to the official text (checked 2026-10-07). Running locally helps with GDPR, but it does not replace your obligations.

GDPR

Regulation (EU) 2016/679

If the model runs on your machine, your customers’ text does not travel to an AI provider. There is no international transfer to justify (Chapter V, Arts. 44 to 49).
You remain the controller. The security measures in Art. 32 (access, encryption, backups) are yours, now over one more machine.
If we access the system to install or maintain it, we sign a processor agreement (Art. 28).

EU AI Act

Regulation (EU) 2024/1689

Running the model locally does not change your role: if you use the system in your business, you are a "deployer".
The people who use it need AI literacy (Art. 4). We cover it with a training session at handover.
If the system talks to your customers, it must make clear it is an AI (Art. 50). If it were high-risk, your obligations are in Art. 26.

The application date of each AI Act obligation is in our AI Act guide.

Which machine you need

It depends mostly on memory: the model has to fit whole. You can check it yourself with the VRAM explorer and compare machines in the hardware catalogue.

What we measured on our machine

NVIDIA DGX Spark (GB10, 128 GB unified memory), measured 2026-10-04. Each figure links to the post with the method.

llama3.1:8b 41.3 tokens/s 9.2 GB loaded
qwen2.5-coder:7b 41 tokens/s 6.6 GB loaded
gemma4:26b 60.7 tokens/s 17 GB loaded
faster-whisper small 0.25× real time 30 min of audio in about 7.5 min

A smaller machine will give other numbers. That is why we measure on your tasks before recommending one.

How we work

Four steps. You know what happens in each one before it does.

A 15-minute call

You tell us the task and the data. We tell you whether a machine of your own makes sense, or not.
Scope in writing

What is delivered, with which data, on which machine and against which success criterion. A fixed quote before we start.
Deployment on your hardware

We install the models on your machine, measure them on your tasks and connect them to your systems.
Handover and training

Documentation, a training session for your team and the machine running in your name. Support afterwards, if you want it.

What we deliver

The models installed on your machine and connected to the agreed task.
A table of what we measured on your machine: model, latency, memory and date.
The technical description for your record of processing and your inventory of AI uses.
A training session for the people who will use it (Art. 4 of the AI Act).

When not to do this

If the task needs the largest model on the market. An open model on one machine gets a lot right, not everything.
If you will use it now and then and the data is not sensitive. A cloud subscription may be simpler.
If nobody in your business will look after the machine after handover.

Frequently asked questions

Does it need an internet connection?: Not to answer. The model runs on the machine. Model and software updates happen when you authorise them, and the machine can stay with no route to the internet if your case calls for it.
Which models do you install?: Open-weight models (for example from the Llama, Qwen or Gemma families), chosen for the task and for your machine’s memory. Before we recommend one, we measure it on examples of your work.
What if the machine breaks?: The hardware is yours and carries its maker’s warranty. We leave documented how to reinstall the models and settings, plus a copy of the configuration, so a replacement does not start from scratch.
How is it quoted?: By scope, in writing, after the call. It depends on the task, the integrations and whether you already have the machine. We do not publish rates.

More answers on data, training and the AI Act in the FAQ.

Let’s talk for 15 minutes

Tell us the task and the data. We tell you whether a machine of your own makes sense for your business, or not.

Book a 15-minute call

A public body? See sovereign AI for the public sector.

Sources (checked 2026-10-07)

This page explains the frame; it is not legal advice. Written by Jacobo González Jaspe, J4SGON S.L., Valencia.