Hybrid is becoming the practical model. Apple describes a family of Foundation Models that spans on-device and server-based models. GSMA Intelligence similarly frames many AI use cases as likely to use a hybrid set-up between device and cloud.
What a hybrid approach can look like
Short, immediate or offline-tolerant tasks can be designed to run closer to the user where a device supports them. More demanding tasks can use server-side resources. The exact split is product-specific, and it should be explained plainly rather than treated as a blanket privacy or performance promise.
A useful way to read AI announcements
When evaluating a mobile AI feature, ask three questions: What runs locally? What requires a network? And what happens when the connection is unavailable? Those answers are more useful than an AI label on its own.