Most of the outages are actually the unavailability of single AI models, not the core service.