DGX Systems are purpose-built AI supercomputing platforms from NVIDIA, integrating multiple high-density GPUs—such as H100, H200, or B200 Tensor Core GPUs—with high-bandwidth NVLink and NVSwitch interconnects, fast NVMe storage, and pre-installed CUDA, cuDNN, and NEMO or other AI framework software stacks into a single validated appliance. The lineup spans the DGX H100 (8× H100 SXM5), DGX B200 (8× B200 SXM), and the multi-node DGX SuperPOD reference architecture, each rated for specific FP8/FP16/BF16 throughput figures published in NVIDIA's product briefs.

Primary buyers are research institutions, university HPC centers, financial modeling teams, and enterprise AI/ML engineering groups that need to train large language models, run multi-modal inference at scale, or execute GPU-accelerated simulation workloads in-house rather than relying on cloud bursting. Key selection criteria include GPU generation, aggregate GPU memory (e.g., 640 GB HBM3e on DGX H100), network fabric readiness for InfiniBand or Ethernet scale-out, rack power draw (10–21 kW per node), and cooling infrastructure compatibility.