DCGM can only provide hardware-level monitoring metrics and cannot perceive upper-level scheduling logic, user identity, or business intent—that is, it cannot associate an idle GPU with the reserver and their needs.