At what point does running your own AI model become worth the extra work?

API costs can get surprisingly high as an app grows so self-hosting open-source models starts to look attractive. But cheaper inference doesn't always mean cheaper overall. You still have to think about GPU costs server maintenance traffic spikes and whether the model gives users the same quality of results. For a small app paying per request feels easier. But once usage grows I wonder if managing your own infrastructure becomes the better option or just creates another problem to maintain. For those who have tried both what made you switch and did self-hosting actually reduce your total costs?