NVIDIA AI Enterprise is a per-GPU software subscription that bundles supported drivers, operators, frameworks and the NIM inference microservices. NIM packages a model with an optimized serving backend behind an OpenAI-compatible API, so a team can stand up a production endpoint in minutes, but the details matter: model profiles, engine selection, cache volumes, NGC credentials, release branches, and the difference between Developer Program access and a production entitlement.
This course starts with licensing and lifecycle so you know what you are buying and supporting. You will run an LLM NIM with Docker, select profiles for your GPUs, and call it with standard OpenAI clients. On Kubernetes you will layer the GPU Operator, Network Operator and NIM Operator, cache models with NIMCache, deploy NIMService resources, and see where NVIDIA Dynamo adds disaggregated prefill and decode and KV-cache-aware routing. You will assemble retrieval with NeMo Retriever embedding and reranking NIMs, enforce policies with NeMo Guardrails, use AI Blueprints as reference architectures, and fine-tune or serve LoRA adapters. The last module covers benchmarking (TTFT, inter-token latency, throughput under concurrency), Prometheus and DCGM monitoring, and an honest build-versus-buy comparison with self-managed vLLM or SGLang.