Can NVIDIA NIM Really Operate Disconnected? An Enterprise Guide to Air-Gapped Private AI

This article provides a comprehensive enterprise guide on operating NVIDIA NIM in air-gapped or disconnected environments. It explains that simply pulling a container once does not make it air-gap ready; a complete software and artifact chain including OCI images, model weights, profiles, runtime manifests, GPU drivers, Kubernetes operators, and certificates must be internally mirrored and validated. The guide details a two-phase workflow: prepare assets on a connected system, then transfer and run locally without outbound access. It emphasizes separating artifact planes (container images, model artifacts, platform dependencies, entitlement records) and testing cold starts, restarts, scaling, and rollbacks under blocked egress. Common failure points like missing profiles, untrusted certificates, or hidden calls to external registries are highlighted. The article also distinguishes between connected, restricted egress, and fully air-gapped operating models, noting their different controls and operational burdens.