Private AI, on a DGX Station dedicated to you
An assistant that answers from your own mail, files and documents, built for the work you actually do, running on an AI supercomputer the size of a tower PC (a deskside supercomputer) that serves you and nobody else. It stands in my facility, beside the machine that runs your estate. Nothing is sent to OpenAI, Google or Microsoft.
It is the most I offer, and it is added to the managed service rather than sold on its own: it answers from data that already lives on the machine dedicated to you.
What it does for you
It answers from your own papers
Ask it about a contract, a thread of mail or ten years of minutes. It answers from your own documents and names the ones it used, so you can check.
Each person, only what they may open
Across a family’s or a firm’s whole estate, it answers each person only from what that person is already allowed to open. Whoever asks about salaries without access to that folder gets nothing from it.
It does the work you repeat
Drafts in the shape of your own documents, the summary of a folder, the comparison you make every month. Once I know which tasks you repeat, I build the tools for them.
The machine
- What it is
An NVIDIA DGX Station: the GB300 Grace Blackwell Ultra chip, 748 GB of memory shared by the processor and the graphics unit, 252 GB of it fast graphics memory (HBM3e), and up to 20 petaflops by NVIDIA’s own figure. NVIDIA says it runs models of up to a trillion parameters, the measure of how large a language model is.
- Who builds it
NVIDIA designs it, and its partners build and sell it under their own names: Dell Pro Max with GB300, HP ZGX Fury, ASUS ExpertCenter Pro ET900N G3, MSI XpertStation WS300, Supermicro Super AI Station and Gigabyte W775-V10. I choose among them for delivery and support, and order it when you sign.
- Why it is possible now
A model of this size used to need racks of servers in a datacentre. Now it fits in one machine, and that machine can stand next to the one that already runs your estate.
- Where it stands, and whose it is
In my facility in Busto Arsizio, beside the machine dedicated to you, connected to it and to nothing that would carry your data out. I buy it out of the price and it stays mine, like that machine; your documents, your questions and every answer are yours.
What I do
— Size it
Which models, for how many people, how fast it has to answer: decided with you before anything is ordered.
— Build it
I connect it to your mail, files and documents, set who may ask about what, and build the tools for the tasks you repeat. Where adapting a model to your own material (fine-tuning) makes it better at your work, I do that too.
— Bring it up to date
Better open models arrive every few months. Every quarter I review yours and move it to a newer one when that is better for your work, not because it is new.
— Say what it cannot do
Open models run privately are good and getting better, and the largest models the big companies run are still ahead on some tasks. You hear what yours can and cannot do before you pay for it.
What it costs
€400,000 a year, on top
The DGX Station, the build, the models and the quarterly review, in one figure. There is no hardware quote: I buy the machine out of it and it stays mine. Paid at signature, then each year in advance.
€50,000 a year, the service underneath
Private AI is added to the managed service, which moves your mail, files and documents onto a machine dedicated to you: €100,000 the first year, €50,000 after. Together, the first year comes to €500,000 and each year after to €450,000.
Every figure here is before VAT, which is added where it applies. The terms are those of the service, as if I stop sets them out.
Questions
Is it as good as ChatGPT?
On your own documents it can do better, because it answers from them and ChatGPT does not have them. On general questions, the largest models the big companies run are still ahead on some tasks. This machine runs open models of hundreds of billions of parameters, and before you pay I tell you plainly where yours falls short.
Does anything reach OpenAI, Google, Microsoft or NVIDIA?
No. The models run on the machine, and your documents and your questions stay on it. Software updates come in; nothing of yours goes out.
Can I have it without moving my mail and files?
No. It answers from data that lives on the machine dedicated to you, so it is added to the managed service, never in place of it.
Who can read my questions?
You, and the people you allow. As its administrator I can technically reach that machine, as I can reach everything I run for you; who can reach your data says what that means and how it is limited.
When does it start answering?
Once your estate is on the machine dedicated to you and the DGX Station has arrived. Delivery depends on the maker, and I give you its date when I order. It answers from your documents first; the tools for the work you repeat follow as we find it.
Tell me what you would ask it
A sentence is enough. I reply within two working days with a time to talk.