nvidia-smi provides instantaneous card-level readings and cannot distinguish which process or request is consuming VRAM; per-tenant attribution requires instrumentation at each service layer.