Aircloud Compute is now Lightyard. Aircloud continues as a private inference platformAircloud continues for inference

Hardware

DGX SuperPOD and scalable units explained

How NVIDIA builds DGX SuperPODs from scalable units: SU size, racks, power and fabric counts per platform from the reference architectures, BasePOD vs SuperPOD, partial SUs, and hardware prices from public reporting.

All notesPublished Updated 10 min read
In this note
  1. What a scalable unit is
  2. Why SuperPODs are built from SUs
  3. SU size by platform
  4. Racks and power per SU
  5. Compute fabric switch and cable counts
  6. Storage, management and the other networks
  7. DGX BasePOD vs SuperPOD
  8. Building less than a full SU
  9. What a DGX SuperPOD costs
  10. Sizing a cluster to the workload instead of the SU

A DGX SuperPOD is NVIDIA's reference design for a whole AI cluster, built by repeating a fixed block called a scalable unit (SU). An SU is a set number of DGX systems plus the leaf switches, cabling and rack layout that go with them. On DGX H100 and B200 it's 32 systems (256 GPUs). On DGX B300 with Quantum-X800 InfiniBand it's 72 systems, and on GB200, GB300 and Vera Rubin NVL72 it's 8 racks, both 576 GPUs. NVIDIA rates one GB300 NVL72 SU at 1.2 MW.

Figures below come from NVIDIA's published reference architectures (RAs) unless we say otherwise.

What a scalable unit is

An SU is the smallest piece of a SuperPOD that NVIDIA designs and documents as one unit. It pins down the compute (a set number of DGX systems or NVL72 racks), the leaf layer of the compute fabric, and the physical layout: where the racks go, how much power each draws and how the cables run.

The leaf switches are wired rail-aligned. The same-numbered network port in every system lands on the same leaf switches, so traffic on one rail is a single switch hop from that rail on any other system in the SU. Traffic between SUs, or between rails, goes up to the spine.

Everything else is sized by how many SUs you have: the spine, the storage fabric, in-band and out-of-band Ethernet, and the control-plane servers. NVIDIA's DGX B300 RA (opens in a new tab) is written around 8 SUs (576 systems) and says the design scales "up to and beyond 72 SU with 2000+ DGX B300 nodes".

Why SuperPODs are built from SUs

NVIDIA's own reason is speed. Repeating a tested block, the GB300 RA (opens in a new tab) says, cuts deployment "from months to weeks".

For a buyer, the more useful property is that every SU has the same fabric, so a job sees the same bandwidth wherever the scheduler puts it. If the spine is planned for the final SU count, a new SU plugs into ports that already exist. And because the RAs publish switch, cable and transceiver counts per SU count, you know the network bill of materials before anyone sends a quote.

SuperPOD is also something NVIDIA sells, as well as a drawing. The RAs describe installation led by NVIDIA's installation services team, NVIDIA Mission Control for operations with Slurm or Run:ai for user access, and full-system support (compute, storage, network and software) from NVIDIA Enterprise Support. The customer owns the hardware and runs it in its own data center or a colocation site. It's designed for one tenant with many users, and the B300 RA says NVIDIA doesn't support multi-tenant modifications.

SU size by platform

Platform (RA)One SUGPUs per SUCompute fabricRacks and power
DGX H10032 systems256InfiniBandOver 40 kW per rack in NVIDIA's example layout
DGX B20032 systems256InfiniBandOver 25 kW per rack in NVIDIA's example layout
DGX B300, Quantum-X800, AC power72 systems576800 Gb/s XDR InfiniBand4 systems per rack, about 56 kW per rack, 18 compute racks
DGX B300, Spectrum-4, DC busbar64 systems512Spectrum-4 Ethernet4 systems per MGX busbar rack, over 50 kW per rack, 16 compute racks
DGX GB200 NVL728 racks576InfiniBandLiquid and air hybrid, 1.2 MW per SU
DGX GB300 NVL728 racks576800 Gb/s XDR InfiniBandLiquid and air hybrid, 1.2 MW per SU
DGX Vera Rubin NVL728 racks576800 Gb/s XDR InfiniBand, two planesLiquid cooled, about 2 MW per SU
DGX Rubin NVL872 systems576800 Gb/s XDR InfiniBandLiquid cooled, DC busbar only

The B300 compute rack counts are our division (72 ÷ 4 and 64 ÷ 4). NVIDIA's 8-SU bill of materials for the AC version lists 144 racks for DGX systems, 18 per SU, so it checks out. The Vera Rubin NVL72 RA marks its fabric design "subject to change", so treat those numbers as preliminary.

On Blackwell Ultra and Rubin, 576 GPUs has become the usual block, as 72 eight-GPU systems or 8 NVL72 racks. The Spectrum-4 Ethernet version of B300, at 512, is the exception.

Racks and power per SU

Facilities teams plan in SUs too. Rough figures for one:

One SUCompute racksPower per compute rackCompute power per SUCooling
DGX B300, AC18About 56 kW (NVIDIA)About 1.0 MW (18 × 56 kW)Air, standard EIA racks
DGX B300, DC busbar16Over 50 kW (NVIDIA)Over 0.8 MW (16 × 50 kW)MGX racks with busbar
GB300 NVL728135 kW TDP, up to 155 kW peak (OEM)1.2 MW SU TDP (NVIDIA)Direct liquid plus air
Vera Rubin NVL728Not stated per rackAbout 2 MW SU TDP (NVIDIA)Liquid

The 1.0 MW and 0.8 MW are our multiplication and leave out the network and management racks. The GB300 per-rack figure is from an OEM's GB300 NVL72 product guide (opens in a new tab), which gives 135 kW TDP per rack and up to 155 kW peak depending on workload. Eight racks at 135 kW is 1.08 MW, which sits inside NVIDIA's 1.2 MW for the whole SU.

For GB200, GB300 and Vera Rubin NVL72, NVIDIA also says the hall should meet Uptime Institute Tier 3, TIA-942-B Rated 3 or EN 50600 Availability Class 3, with concurrent maintainability and no single point of failure. If your current room isn't built to that level, bringing it up is part of the project. Our note on building a data center for GPU clusters covers what that involves.

Compute fabric switch and cable counts

DGX B300 on Quantum-X800 InfiniBand

From the scaling table in NVIDIA's DGX B300 RA (opens in a new tab), with 144-port Quantum-X800 Q3400 switches:

SUsSystemsGPUsLeaf switchesSpine switchesSystem-to-leaf cablesLeaf-to-spine cables
17257684576576
21441,1521681,1521,152
42882,30432182,3042,304
85764,60864364,6084,608
181,29610,3681447210,36810,368

NVIDIA's table says 9,216 GPUs in the 18-SU row. But 1,296 systems × 8 GPUs is 10,368, which matches the cable counts in the same row, so that's what we show.

Each GPU gets one 800 Gb/s link, and every link needs optics at both ends. The 8-SU bill of materials lists 4,608 transceivers on the DGX side and another 4,608 twin-port transceivers between leaf and spine. NVIDIA asks buyers to contact it for anything of 4 SUs or more.

NVL72 SuperPODs: GB200, GB300 and Vera Rubin

SUsGPUsGB200 leaf / spine / coreGB300 leaf / spineVera Rubin leaf / spine
157632 / 24 per SU8 leaf per SU16 leaf per SU
21,15264 / 48 / 1816 / 832 / 18
42,304128 / 96 / 3632 / 1864 / 36
84,608256 / 192 / 7264 / 36128 / 72
169,216512 / 384 / 144128 / 72Not listed
1810,368Not listedNot listed288 / 144

The big change from GB200 to GB300 is the switch. GB200's design uses 32 leaf and 24 spine switches per SU, plus a core layer once you go past one SU. GB300 moves to 144-port Quantum-X800 switches, which brings that down to 8 leaf switches per SU (4 rails with 2 leaf switches each) and fits 16 SUs in two layers.

Vera Rubin NVL72 doubles the leaf count again. Each GPU has two ConnectX-9 adapters, split across two independent fabrics NVIDIA calls planes, so a GPU still has a path to every other GPU if one plane fails.

Storage, management and the other networks

A SuperPOD has four physical fabrics, and compute is only one of them. The other three, plus the control plane:

  • Storage runs on InfiniBand or Ethernet, kept separate from compute. NVIDIA's DGX B300 guideline per SU is 80 GB/s aggregate read and 40 GB/s write at its standard level, and 250 and 124 GB/s at the enhanced level (storage architecture (opens in a new tab)). The file system comes from a storage partner certified for SuperPOD, who sizes it.
  • In-band Ethernet carries provisioning, user access, the NFS tier and uplinks. NVIDIA recommends at least two 100 GbE links to the customer edge.
  • Out-of-band Ethernet, on SN2201 switches, connects the management ports of every system, switch and PDU.
  • The control plane in the B300 RA is seven management servers (opens in a new tab): two for Base Command Manager in high availability, three for Kubernetes and two Slurm login nodes. UFM fabric manager appliances sit alongside.

The 8-SU DGX B300 bill of materials (opens in a new tab) gives a sense of scale: 48 racks for switches and management next to the 144 compute racks (about 6 per SU), and 69 out-of-band switches.

GB300 NVL72 folds storage and in-band into one Ethernet fabric, with 4 compute-side leaf switches per SU, at least 2 storage leaf switches, and a pair each of management leaf and out-of-band spine switches.

DGX BasePOD vs SuperPOD

DGX BasePODDGX SuperPOD
Size2 to 8 DGX systems in NVIDIA's current RAWhole SUs of 256 to 576 GPUs, up to thousands of GPUs
SystemsDGX B200, H200, H100DGX H100 through Vera Rubin NVL72
Compute fabricTwo Quantum QM9700 NDR 400 Gb/s InfiniBand switchesRail-optimized leaf and spine per SU
What it isA prescriptive reference architectureA product with a bill of materials, NVIDIA-led installation, Mission Control and full-system support

BasePOD is for a handful of DGX systems on a validated design. Past 8 systems, the SuperPOD SU is the design NVIDIA documents.

Building less than a full SU

Plenty of clusters aren't a multiple of 72 nodes or 8 racks. The B300 RA's guidance is that "the fabric should be designed to support the full SU, including leaf-spine switches and leaf-spine cables, and leave the portion of the fabric unused where these nodes would be located." The GB300 RA says the same about leaf switches and leaf-spine cables.

Here's what that means for 32 B300 nodes (256 GPUs) against the 72-node Quantum-X800 SU:

32 nodesBuilt for a full SU (NVIDIA guidance)Sized to 32 nodes (estimate)
Leaf switches8About 4
Spine switches4About 2
System-to-leaf links used256 of 576256
Leaf-to-spine cables576About 256
Air-cooled compute racks8 at about 56 kW8 at about 56 kW
Growth to 72 nodesAdd nodes and cables to the existing leaf portsAdd switches and recable the spine

The right-sized column is our estimate for a non-blocking two-layer fabric; the real one is sized in the network design.

We'd build for the full SU if the cluster will grow during its life, because the spine is the part that's hard to change later. If the size is fixed for the whole term, size the fabric to the nodes and spend the difference on storage or spares.

Two caveats. A SuperPOD is a DGX product, so an OEM HGX B300 cluster isn't one, though it can use the same rail-aligned layout. And with NVL72 the rack is the smallest unit you can add. Four racks is half an SU, and the same guidance means planning all 8 leaf switches of a GB300 SU with only half the racks in.

What a DGX SuperPOD costs

NVIDIA doesn't publish a SuperPOD price. It's quoted per deal and includes installation, Mission Control and support.

What you can do is multiply reported rack prices by the racks in one SU. The table below is that arithmetic and nothing more: hardware only, from public reporting, not a quote and not a cost of ownership.

One SUReported price per rackSourceHardware per SU (our arithmetic)
8 × GB200 NVL72US$2.8M to US$3.4MTom's Hardware source, March 2026US$22.4M to US$27.2M
8 × GB300 NVL72US$3.7M to US$4.0MLoop Capital estimate, via GPUsmithUS$29.6M to US$32.0M
8 × GB300 NVL72Just under US$5.0MData Gravity, purchase orders, August 2026Just under US$40M
8 × GB300 NVL72, inference configUS$6.0M to US$6.5MTom's Hardware source, March 2026US$48M to US$52M
8 × Vera Rubin NVL72US$5M to US$7MTom's Hardware source, March 2026US$40M to US$56M

Reported prices swing with configuration and with who's buying. Data Gravity says its purchase-order figures are for small orders and that hyperscale buyers do better. Tom's Hardware's source warns that many quotes floating around are ODM prices without a proper warranty, and says the Vera Rubin figure includes about US$1M of storage per rack.

The all-in cost of owning an SU is much higher than the hardware line once you add the team to run it, maintenance, the months to stand it up, power and the facility. Our notes on buying or renting a GPU cluster and GPU depreciation cover those lines.

For a memo:

  • One SU is 576 GPUs on DGX B300 (Quantum-X800), GB300 NVL72 and Vera Rubin NVL72, and 256 GPUs on DGX H100 and B200.
  • NVIDIA rates one GB300 NVL72 SU at 1.2 MW and one Vera Rubin NVL72 SU at about 2 MW. An air-cooled DGX B300 SU is about 1 MW of compute racks.
  • Hardware alone for one GB300 NVL72 SU is roughly US$30M to US$52M from public reporting, before networking, storage, facility, power and staff.

Sizing a cluster to the workload instead of the SU

A SuperPOD assumes you buy NVIDIA's product and staff the operation. We design, source, build and operate NVIDIA GPU clusters sized to the workload rather than to an SU: from 8 HGX nodes (64 GPUs) or one NVL72 rack (72 GPUs), on RoCE v2 Ethernet, spine-leaf. Most clusters are 32 to 128 nodes; larger ones are scoped case by case. Large clusters are handed over in monthly waves of 32 nodes. You can own the cluster and have us build or run it, or reserve one we hold. See cluster deployment, managed clusters, reserved clusters, the HGX B300 and GB300 NVL72 pages, or talk to us.

Questions this note answers

Sources

  1. NVIDIA SuperPOD with DGX B300 and Quantum-X800 InfiniBand RA: abstract (opens in a new tab)
  2. NVIDIA DGX B300 Quantum-X800 RA: DGX SuperPOD architecture (opens in a new tab)
  3. NVIDIA DGX B300 Quantum-X800 RA: network fabrics (opens in a new tab)
  4. NVIDIA DGX B300 Quantum-X800 RA: storage architecture (opens in a new tab)
  5. NVIDIA DGX B300 Quantum-X800 RA: management servers (opens in a new tab)
  6. NVIDIA DGX B300 Quantum-X800 RA: bill of materials (opens in a new tab)
  7. NVIDIA SuperPOD with DGX B300, Spectrum-4 Ethernet and DC busbar power RA: architecture (opens in a new tab)
  8. NVIDIA DGX SuperPOD reference architecture with DGX GB300 (PDF) (opens in a new tab)
  9. NVIDIA DGX SuperPOD reference architecture with DGX Vera Rubin NVL72 (PDF) (opens in a new tab)
  10. NVIDIA DGX SuperPOD reference architecture: DGX GB200, architecture (opens in a new tab)
  11. NVIDIA DGX SuperPOD reference architecture: DGX GB200, network fabrics (opens in a new tab)
  12. NVIDIA DGX SuperPOD Rubin NVL8 reference architecture (May 2026) (opens in a new tab)
  13. NVIDIA DGX SuperPOD reference architecture: DGX B200, architecture (opens in a new tab)
  14. NVIDIA DGX SuperPOD reference architecture: DGX H100, architecture (opens in a new tab)
  15. NVIDIA DGX BasePOD reference architecture: DGX B200, H200 and H100 (opens in a new tab)
  16. OEM product guide: GB300 NVL72 rack-scale AI (opens in a new tab)
  17. GPUsmith: NVIDIA GB300 NVL72 system page (Loop Capital estimate) (opens in a new tab)
  18. Data Gravity: How much does an NVIDIA NVL72 cost? (August 2026) (opens in a new tab)
  19. Tom's Hardware: NVL72 rack prices for GB200, GB300 and Vera Rubin (March 2026) (opens in a new tab)
Get started

Tell us where you're starting.

Capital, hardware, a site or a team: tell us what you have and when you need it running. We usually reply within a few hours and come back with a written plan.

sales@lightyard.ai