Real-time notifications at scale: fan-out on write vs fan-out on read

we're designing a notification system for a social platform, and the fan-out strategy is a big debate. for fan-out on write, we'd pre-generate notifications for all followers whenever a user posts something. this gets incredibly expensive for celebrity users with millions of followers. fan-out on read, where we generate notifications when the user fetches their feed, means slower load times for the notification feed itself. we're looking at a hybrid approach, maybe fan-out on write for a subset of 'active' followers, and fan-out on read for the rest. has anyone successfully implemented a hybrid fan-out model for a system with 10 million users and a power-law distribution of followers? what were the complexities and scaling hurdles?