Our infrastructure
The infrastructure on which your knowledge becomes a model.
Today the value is not in the software but in the model trained for your purpose. Training it, specialising it and running it every day takes adequate compute. We have it, and we know how to build and manage it at your site.
What we have
The servers and workstations we work on every day.
On these machines we run, train and measure the models of our projects. They are also the reference for what we can build for you.
GB300
An NVIDIA GB300 server
The core of our compute. The large, open-weight model runs here.
2 × 96 GB
A workstation with two NVIDIA RTX PRO 6000 Blackwell Max-Q
96 GB of memory per card. This is where we train and specialise models; it can run the large model too.
180 billion
The large model
Open-weight, about 180 billion parameters in total. The same model runs on the server and on the workstation.
40
Requests at once, on the server
The usable capacity, measured by us: several jobs in parallel on the same model.
What it is for
Train, specialise, run.
On this infrastructure an open-weight model learns your purpose and then works for you. Three activities, depending on what is needed.
Train
A new model, when needed
When no existing model is good enough, we train networks of our own from scratch on your data: for example to generate images of materials at true scale.
Specialise
An open model, fine-tuned on you
We take an existing open-weight model and fine-tune it on your data, or we put it inside a procedure built around your case.
Run
The model at work, every day
We put it into service on our infrastructure or yours and measure it on your cases. When a better open model comes out, we verify it on your cases before adopting it.
At your site
Infrastructure like ours, at your site.
If you want the model and the data in the company, we build the infrastructure at your site, with everything it needs: from the first analysis of the workloads to day-to-day management.
-
Design
We start from your purpose and your workloads: which models, how many cards and how much memory, how much power and how much cooling are needed.
-
Build
We choose and assemble the machines, install the models and verify them on your cases before delivery.
-
Install
At your site, together with your IT department: network, power, cooling, access and security.
-
Manage
Updates, checks, new models verified before they go into service. We look after the infrastructure: you use the model.
You can also start on our infrastructure, with the trial on your documents, and move to yours later.
What stays with you
The knowledge, the data, the choices, the decision.
- The knowledge. What the model learns from your data stays yours: it does not become training material for someone else’s AI, and you do not pay for it per use.
- The data. With infrastructure of your own it stays in the company. On ours, we always tell you which model reads it and where it runs.
- The choices, in writing. Which model, under which licence, why, and the measurements on your cases. Open weights do not mean a free licence: we check every licence before using the model.
- The decision. The AI prepares, a person decides: yours.