The plans scale down far enough that they can't possibly cover the cost of a private model-loaded-to-vram instance at the low end.