Systems & racks
HGX and DGX platforms and rack-scale NVL systems: node-level specifications and the facility requirements they impose.
NVIDIA DGX B200
DGX B200 is NVIDIA's fully integrated, factory-built Blackwell system: eight B200 SXM GPUs, NVLink 5 fabric, dual Xeon hosts and eight ConnectX-7 NICs in one 10U chassis with a single support contract from NVIDIA. It is the reference appliance the DGX SuperPOD architecture is built from. Teams that want a validated, drop-in node with NVIDIA-direct support buy DGX; teams with an existing OEM relationship, in-house cluster ops, and a preference to tune the CPU, storage and networking BOM buy the equivalent HGX B200 platform from Supermicro, Dell, HPE, Lenovo or similar and integrate it themselves. The delta is mostly integration risk and support terms, not raw GPU performance.
NVIDIA DGX H200
DGX H200 is NVIDIA's factory-integrated Hopper-generation system: eight H200 SXM GPUs (141 GB HBM3e each) on NVLink 4, dual Xeon hosts, and ConnectX-7 fabric in an 8U chassis with NVIDIA-direct support. It is the prior-generation counterpart to DGX B200, sharing the H100 chassis and board layout but with the higher-capacity, higher-bandwidth H200 GPU swapped in. Buy DGX H200 for a validated, single-vendor-supported node, particularly where an existing H100/H200 DGX or HGX fleet needs a drop-in match; buy an HGX H200 OEM server (Supermicro, Dell, HPE, Lenovo) when you want to control the CPU, storage or networking BOM or already run your own cluster ops.
NVIDIA HGX B200
HGX B200 is the 8-GPU Blackwell baseboard NVIDIA sells to OEMs, not a server you can rack yourself. Supermicro, Dell, Lenovo, HPE, ASUS and ASRock Rack each bolt their own CPUs, chassis, power delivery, storage and networking around the same baseboard and NVSwitch fabric, then sell the result as a complete server. Most private HGX builds go through one of those OEM boxes rather than NVIDIA's own DGX B200, because the OEM path lets a buyer choose CPU vendor, memory footprint, NIC mix and air vs liquid cooling to fit an existing rack and budget, while DGX B200 ships as a fixed configuration at a materially higher price for the same GPU count.
NVIDIA GB200 NVL72
The GB200 NVL72 is a liquid cooled, 48U rack that wires 72 Blackwell B200 GPUs and 36 Grace CPUs into a single NVLink domain, so software addresses one accelerator with a 13.4 TB HBM3e pool instead of 36 separate servers stitched together over InfiniBand. It targets teams training or serving frontier scale models that need that pool at NVLink speed rather than sharded across PCIe or network hops. Hosting one draws 125 kW to 135 kW of continuous IT load through direct to chip liquid cooling with zero rack fans, a facility class that only purpose built AI data centers and a small number of retrofitted colocation halls can deliver today.
NVIDIA GB300 NVL72
The GB300 NVL72 is the Blackwell Ultra revision of NVIDIA's rack scale platform: the same 48U, 72 GPU, single NVLink domain layout as GB200 NVL72, but each B300 GPU carries 288 GB of HBM3e instead of roughly 186 GB, lifting the rack's NVLink-domain memory pool to 20 TB and its dense FP4 throughput to about 1.1 EFLOPS. It is aimed at teams running large scale reasoning and long context inference alongside training, where the extra memory per GPU and doubled 800 Gb/s scale out networking matter more than raw FLOPS growth. Facility demands go up along with the compute: reported operating power sits at 132 kW to 140 kW continuous, again 100 percent direct to chip liquid cooled with no rack fans.
NVIDIA DGX Spark
DGX Spark is a desk-side box built around NVIDIA's GB10 Grace Blackwell superchip, 128 GB of unified memory, and the full CUDA stack, aimed at a single developer prototyping and fine-tuning models locally before moving to real GPU cluster capacity. It is not a cluster node: its 273 GB/s memory bandwidth is well behind a datacenter GPU, and while two units can be linked over ConnectX-7 to handle models up to 405B parameters, that is a two-box desk setup, not a scalable rack architecture. Buy it to develop against CUDA and iterate on model code without renting cloud GPU time; don't buy it expecting DGX B200-class throughput.