What is self-hosted AI?

It is more than the model

Running a local model still leaves conversations, files, memory and account data in whatever application you used. Self-hosting moves all of that onto your own infrastructure.

The distinction matters for anyone whose obligations are about stored data rather than about inference. Where the model runs and where the history lives are separate questions, and only self-hosting answers the second.

What it costs operationally

You take on upgrades, backups, monitoring, TLS, and the debugging when something breaks at an inconvenient hour. That is a real, ongoing cost measured in attention rather than money.

It is worth it when the data genuinely cannot sit elsewhere, or when you need the deployment to outlive any vendor relationship. It is not worth it as a general precaution.

Self-hosted does not mean disconnected

A self-hosted deployment can still call hosted models. Many do: the platform and its data are yours, and cloud providers are used where their capability is worth the data leaving.

The combination that removes external processing entirely is self-hosting plus local models, and it is a deliberate configuration rather than the default.

Questions people ask

Is self-hosting the same as local AI?

No. Local AI is about where the model runs. Self-hosting is about where the application and its data live. You can have either without the other, and the strongest privacy position needs both.

Does self-hosting make us compliant?

No. It can be a component of a compliance story, but compliance is about contracts, controls, evidence and audits. Where the software runs is one input among many.

What does it take to run?

For most platforms, containers, a database, and somewhere to run them — plus a person who owns the upgrade path. The last part is the one that gets underestimated.