During the period the model was offline, a company employee publicly stated it was serving essentially no traffic from that model.
The statement is a direct measure of the damage. A model launched as flagship, distributed across several clouds, with customers starting to integrate, reduced to residual traffic within days.
There's an additional, less visible effect. A customer who has to deliver doesn't wait: they migrate to an alternative, adapt the code, and once adapted rarely switch back simply because the original supplier returned.
That is exactly what appeared the same week elsewhere in the market, with a large institutional buyer testing competing models to replace the one that had gone offline in sensitive systems.
For anyone selling AI services, the lesson is about where the risk lives. It isn't only technical availability: a prolonged interruption pushes customers to competitors, and the return isn't automatic.