What’s stopping small AI startups from building their own models?

Feels like everyone just plugs into existing APIs instead of training anything new. Is it cost, data access, or just practicality?