Every company weighing AI over its own documents eventually asks the same question: where does all of this actually run? For most tools the answer is a shared cloud platform, and the question ends there. This post is about the other answer, the self-hosted one, what it concretely changes, and how Naxis Assistant, a private knowledge engine built to run either way, treats it.
Single-tenant first, hosting second
The hosting debate hides a more important property: tenancy. A pooled platform holds many companies' documents behind one control plane, which means one authorization bug is everyone's breach and one subpoena, outage or misconfiguration has everyone's name on it.
A single-tenant deployment inverts that. Each client runs a complete instance, its own application, its own database, its own document store, serving nobody else. Isolation stops being a configuration you hope is set correctly and becomes the deployment model itself. That property holds wherever the instance runs, and every Naxis deployment has it by construction; self-hosting then decides who owns the machine underneath it.
What self-hosting changes
With the deployment on your own Linux server, the boundary of the system is the boundary of your infrastructure. Documents, the search index, conversations and the audit log live and die on hardware you control. Backups follow your policy. Network access follows your firewall. The vendor's obligations shrink to shipping good software, which is the healthiest possible arrangement between a company and its most sensitive data.
The strongest posture goes one step further: a self-hosted deployment answering on a self-hosted, open-weight AI model. Then generation never leaves the building either, and the sub-processor list for the whole system reads "none". Deployments that prefer a managed AI instead use the Naxis AI service under a data-processing agreement with zero-retention terms: the question and the permitted excerpts go out, the answer comes back, nothing is stored and nothing trains any model.
What self-hosting should not change
The historical price of self-hosting was operational: you got the control and inherited the maintenance. That trade is obsolete when the software updates itself. A Naxis deployment installs releases in a quiet window you choose, takes a safety backup first, and rolls back on its own if a health check fails, on your server exactly as on managed hosting. Same product, same automatic updates, different landlord.
The feature set should not change either. Self-hosting that costs you channels, permissions or the audit record is a demo, not a deployment. The entire product travels: every source, every channel, the permission model, the compliance tooling.
Managed is the same product with a different landlord
Managed hosting remains the fast path: a deployment comes online at your own subdomain minutes after checkout, in an EU data centre, still single-tenant, still isolated. Plenty of organisations start managed and move on-premise later, or run managed forever with a clear conscience. The point of self-hosting is not that everyone should do it. The point is that the option existing keeps the architecture honest.
Install guide, self-hosted See the plans
Full self-hosting is a flat monthly add-on on any plan. If your DPO or IT lead wants to inspect what actually ships before anything runs, the security record is public and the reviewer documents are one account request away.
