NVIDIA uses a handful of prefixes and suffixes across its data center hardware, and contracts, quotes and operator proposals all use them without explanation. Here is what each one means, and what it tells you about the hardware you are contracting around.
HGX and DGX: who built the server
HGX is NVIDIA's name for the baseboard, the slab of silicon and interconnect that holds the GPUs. Server manufacturers, companies like Supermicro, Dell or HPE, take that baseboard and build a complete server around it. The finished product carries the manufacturer's name and the HGX label.
DGX is NVIDIA's own complete server. NVIDIA designs the baseboard, builds the system and sells it under its own brand. When an operator quotes you DGX, the whole box came from NVIDIA. When they quote you HGX, a third party assembled the server around NVIDIA's baseboard.
For most capacity contracts, the distinction matters less than the GPU generation inside. But it can affect firmware support, warranty terms and what the operator's engineers are trained on, so it is worth knowing which you are getting.
NVL8 and NVL72: how many GPUs in the unit
NVL stands for NVLink, the high-speed interconnect NVIDIA uses to let GPUs share memory and communicate directly. The number after NVL tells you how many GPUs are linked together in the unit.
- NVL8 is an 8-GPU server. Each GPU connects to the others over NVLink inside a single chassis. The DGX Rubin NVL8, for example, is NVIDIA's own 8-GPU Rubin server.
- NVL72 is a rack-scale system. Seventy-two GPUs and, in the Blackwell generation, 36 Grace CPUs, are linked across a full rack. The GB200 NVL72 and GB300 NVL72 are the two Blackwell rack-scale systems. The Vera Rubin NVL72 is the rack-scale Rubin system, with 72 Rubin GPUs and 36 Vera CPUs.
The jump from NVL8 to NVL72 is not just more GPUs. It is a different class of hardware, a different power envelope and a different facility requirement. A rack of GB300 NVL72 draws up to 142 kW. Ten 8-GPU B300 servers draw about 140 kW in total, roughly the same load, but they sit in standard rack positions rather than a single integrated rack, and the cooling approach differs.
The NVL144 rename
Before December 2025, NVIDIA called its rack-scale Blackwell system the NVL144. That name counted GPU dies, because each Blackwell GPU package contains two dies. NVIDIA then renamed it NVL72, counting GPU packages instead. The hardware did not change. If you see NVL144 in an older document or an operator's older proposal, it refers to the same rack as NVL72.
SuperPOD: a cluster, not a server
A SuperPOD is NVIDIA's reference architecture for a multi-rack cluster. It defines how many racks go together, how they connect, and what the facility must support. NVIDIA's DGX SuperPOD reference architecture for GB300 is built around scalable units of 8 racks, each unit drawing about 1.2 MW including networking. The architecture requires the data center to meet Uptime Institute Tier 3 design standards, meaning concurrent maintainability and no single point of failure.
When an operator says they are running a SuperPOD, they mean their cluster follows that reference architecture. It is a signal about how the site is engineered, not just how many GPUs are present.
Trade shorthand that is not an NVIDIA name
A few labels circulate in the market that NVIDIA did not create.
- VR200, R100 and R200 appear in operator quotes and press coverage as shorthand for Rubin-generation hardware. They are not NVIDIA product names. If you see them in a contract, ask the operator to confirm the exact NVIDIA product designation.
- Hopper, Blackwell and Rubin are NVIDIA's architecture generation names, not product names. Hopper covers H100 and H200. Blackwell covers B200, B300, GB200 and GB300. Rubin covers the Rubin NVL8 servers and the Vera Rubin NVL72 rack.
Questions to take into a call with an operator
- Is this an HGX system from a server manufacturer, or a DGX system from NVIDIA directly?
- Which GPU generation and which NVL configuration, NVL8 or NVL72?
- If the proposal says NVL144, can you confirm this is the same hardware now called NVL72?
- Does your facility meet Tier 3 standards, and is that relevant to the configuration you are quoting?
- What is the exact NVIDIA product name for any hardware listed under trade shorthand like VR200 or R100?
For a broader look at the GPU generations these names refer to, see From Hopper to Rubin: a plain-English guide to NVIDIA data center GPUs.
Sources
Product naming and architecture details are drawn from NVIDIA product announcements, NVIDIA's DGX SuperPOD reference architecture documentation, Supermicro's GB300 NVL72 datasheet and NVIDIA's CES January 2026 announcements for the NVL8 family. The NVL144-to-NVL72 rename is from NVIDIA's December 2025 product naming update. Trade shorthand classifications reflect industry usage as reported across the infrastructure press.
Share your requirements. We'll canvas the market.
Tell us what you need (for colocation space and power: market, kW or MW, cooling and timing; for reserved GPU capacity: region, GPU type, cluster size and timing) and we'll canvas colocation and neocloud operators in parallel. Shortlist in 48 hours. We don't sell hardware. We broker colocation space and power, and reserved GPU compute capacity.
Earlier conversations get better terms. When you engage early, we have time to negotiate with vendors before you need to commit. We're paid only when a deal closes, and the fee structure is agreed in writing before any introduction.