Self-hosted software runs on servers you control, whether that is your own hardware or a cloud account in your name, rather than as a service a vendor operates for you. You install it, update it, secure it, and keep it running. In the AI context, it often means running an open-weight model on your own machines instead of calling someone else's API.
The pull toward self-hosting is usually data, cost, or control. Sensitive information never leaves your environment, which can matter for privacy rules or nervous enterprise buyers. At steady high volume, owning the hardware can undercut per-request API pricing. And you are insulated from a vendor changing terms, deprecating a model, or going down. Each of these is a legitimate driver on its own.
The cost is operational weight. Someone has to size the servers, apply security patches, handle scaling and outages, and stay on call when something breaks at an inconvenient hour. A managed service folds all of that into the price; self-hosting moves it onto your team, which is real work an early startup may not have the hands for.
The pragmatic path is to start managed and revisit self-hosting once a specific pressure appears, such as a compliance requirement or a bill that clearly exceeds what owning the stack would cost. Choose it for a concrete reason, not for the feeling of control alone. That reason will be obvious when it arrives.