Host Requirements
Choose a plain Linux GPU host with real kernel controls.
nFOX is designed for customer-controlled GPU hosts where the runtime can inspect the system, prepare model files, enforce the Blockhouse sandbox, and write operating evidence.
Baseline Requirements
| Area | Requirement |
|---|---|
| Operating system | Ubuntu 24.04 LTS is recommended. Ubuntu 22.04 LTS may work for some deployments. |
| Architecture | x86_64 Linux host with a real kernel, not an emulated-kernel sandbox. |
| GPU | NVIDIA GPU sized for the selected model, context length, quantization, and expected concurrency. |
| Access | SSH access with permission to install packages, configure runtime services, and review system status. |
| Storage | Enough local disk for model weights, converted artifacts, logs, audit records, and temporary hydration files. |
| Network | Outbound access for setup and approved control-plane communication. Avoid exposing inbound model ports directly. |
Cloud Compatibility
The short version: use a normal GPU VM from a cloud or GPU provider, install Ubuntu 24.04, and keep the host under your administrative control. Bare-metal GPU servers, dedicated GPU VMs, and rented GPU boxes are the expected deployment shape.
This is where sovereign AI becomes practical. A managed AI endpoint, shared GPU pool, dedicated GPU VM, rented GPU box, and owned server do not create the same control boundary. To make a sovereignty claim, the runtime path, prompts, vectors, tokens, tool calls, and audit evidence need to stay inside the enterprise control framework.
Avoid managed notebook products, serverless inference platforms, and hosted container-only environments when you need full Blockhouse enforcement. Those offerings can hide or virtualize the kernel features that containment depends on.
60-Second Preflight
- Confirm you can SSH into the host and run administrative setup commands.
- Confirm the host reports the expected NVIDIA GPU and driver state.
- Confirm the selected model fits the available VRAM and disk budget.
- Confirm the provider is giving you a real Linux VM or bare-metal host, not only a managed notebook.
- Confirm your security team approves outbound-only setup and monitored runtime communication.
When in doubt, choose a simpler GPU VM. Fancy managed AI products often remove the host controls that make private AI evidence useful.