Stop measuring AI training costs in GPU hours

Why idle time, checkpointing, and cluster failures are quietly inflating your training budget