Training is often more demanding than inference, all else being equal. AI training with “whatever gpu” is gonna end in tears and frustration. Have you tried your workload against some per-second billing instance first before making a longer term commitment?
Choking on lobster could be deadly.
Source link