Hardware
Sized to your archive, not to a catalogue.
Private AI does not begin with the largest system. It begins with the smallest system that serves your archive, your users, and your latency expectations — with a defined upgrade path above it.
The ladder
Node — one office, one principal
A quiet workstation-class system. Serves compact language models comfortably for private document search, drafting, and correspondence support. Runs on a normal office circuit; heat and noise stay office-compatible.
Vault — the multi-user archive system
A rack server with one or two professional GPUs, protected storage, backups, and role separation. Serves mid-sized models with strong quality for teams — family offices, fiduciaries, law firms. This is the typical first production system.
Cell — the regulated team system
Multi-GPU servers from certified enterprise vendors, with redundancy, SIEM export, change control, and separate development and production environments. Serves large models and higher concurrency for banks, wealth managers, and clinic groups. Direct liquid cooling is recommended at this density.
Rack — the facility programme
Rack-scale systems for institutions with sustained, high-concurrency AI demand or very large models. This tier is a facility engagement, not an appliance: electrical feed, floor loading, liquid cooling, fire safety, and heat-recovery design belong to the programme from day one.
What runs on each tier — honestly
| Tier | Model classes served well | Power class | Cooling |
|---|---|---|---|
| Node | compact models (roughly 7–32 B parameters, quantised) | < 1 kW | air, office-quiet |
| Vault | mid-size models (roughly 30–70 B, quantised where sensible) | 2–8 kW | air; liquid-ready optional |
| Cell | large models (70 B production class and above, memory permitting) | 8–40 kW | precision air or direct liquid |
| Rack | frontier-class open-weight models at scale | 50–200+ kW | liquid, engineered with the facility |
Model-size numbers describe hardware capability classes, not promises. Which model actually serves your workflows is decided by evaluation on your material — and stated in the review with its memory and latency profile.
How we source
Production systems are warrantied OEM or specialist-integrator builds, procured through Swiss channels wherever possible, with onsite support agreements appropriate to the tier. ANULUM designs the architecture, integrates the software and governance layer, runs the evaluation, and supports the system — we are the engineering partner, not a hardware reseller.
Rack-tier honesty: a rack-scale private AI system consumes as much electricity as a small industrial facility and requires professional power and cooling engineering. We will tell you when your use case does not need one — most do not.