2 ms·

Capacity for inference isn't a cost issue, it's an availability issue. There just isn't enough hardware out there.