What AI changes about self-hosting
The strongest argument for renting software was never features. It was staffing. "You don't have an ops team" ended most self-hosting conversations for fifteen years, and it deserved to — keeping a service healthy meant patching, backups, certificates, disk space, log-watching, restore drills, and being the person who gets paged. Renting made all of that someone else's job. Worth paying for.
That argument depended on an assumption: routine operations require a person. It quietly stopped being true.
What routine actually is
Strip the mystique off running one well-built service and you get a short list that repeats forever: apply security updates. Verify backups actually restore. Renew certificates before they expire. Watch disk usage. Read the logs for anything new. Restart the thing that leaked memory. Almost none of it requires judgment; all of it requires showing up. The cost of self-hosting was never the difficulty — it was the discipline, every week, forever.
This is precisely the shape of work AI agents are good at. Not because they're brilliant, but because they're relentless. An agent doesn't skip the backup check because the quarter ended. It doesn't let a certificate lapse over a long weekend. Given a runbook, it runs the book.
The arithmetic that changed
Self-hosting used to have a high fixed cost: a competent person's attention. Renting converted that into a subscription. The subscription made sense as long as competence was expensive.
What's changed is that the fixed cost collapsed. The routine 95% of operations can now be done by agents supervised in minutes a week, not hours a day. The judgment 5% — "should we upgrade the database major version," "is this alert real" — still needs a human, but it always did; vendors just hid it behind an account manager. When the fixed cost of competence drops below the recurring cost of rent, the arithmetic flips. For infrastructure software, it has flipped.
A year of one instance, honestly
What does keeping one NetworkGX instance healthy actually involve? Boot-time checks that fail loudly instead of silently. Certificates that renew themselves and page only on failure. Nightly backups with a weekly automated restore test — a backup that's never been restored is a hope, not a backup. An agent that reads the logs each morning and tells you the one line worth reading. Updates pulled on your window, with the previous version a rollback away.
That is the whole job, most weeks. It is not nothing. It is also no longer a hire.
The honest caveats
Agents fail. The design response is the same as for any operator you'd employ: fail closed, keep the last-good state, make every automated action loud and reversible. And responsibility genuinely is yours now — when it breaks at 2am it's your 2am. The difference from a decade ago is that the thing waking you up can also be the thing that already tried the runbook, and tells you exactly where it stopped.
You built the network because owning beat leasing. The same logic now reaches the software — because the reason it didn't is gone.
Now piloting with a small number of network owners — if this is your network too: