Custom models.
We use this, pretty convenient and less hassle than managing our autoscaling GPU pools.
alt Hacker News