Rack elevation to scale: 41U of 42U used, 1U free.
NVIDIA · priced at Broadberry, Insight
Rack of 4 × DGX B200
Four DGX B200 systems (40U) under one Quantum-2 QM9700 switch (1U) in an APC 42U rack. The switch's 64 ports at 400 Gb/s connect all 32 of the systems' ConnectX-7 cards.
Key features
- 32 Blackwell GPUs in one rack: 5.8 TB of GPU memory
- 72 PFLOPS of dense FP16 across four DGX B200
- One QM9700 switch, 64 ports at 400 Gb/s: every GPU gets its own port
- About 59 kW at full load in 41U of a 42U rack
- Room for a ~2.3-trillion-parameter model at 16-bit across the rack
Parts and prices
| Part | Qty | Price |
|---|---|---|
| NVIDIA DGX B200 | 4 | $562,154.09 × 4 = $2,248,616.36 Broadberry, 25 Oct 2026 |
| NVIDIA Quantum-2 QM9700 InfiniBand switch MQM9700-NS2F | 1 | $39,250.99 Insight, 25 Oct 2026 |
| APC NetShelter SX 42U rack, 750 × 1200 mm AR3350B2 | 1 | $4,306.99 Insight, 25 Oct 2026 |
| Total | $2,292,174.34 | |
Not included: cables and optics, power distribution, cooling. At full load the rack draws about 59 kW.
Specifications
NVIDIA DGX B200 × 4
- GPUs
- 8 × B200 (180 GB each)
- CPU
- 2 × Intel Xeon Platinum 8570, 112 cores in total
- System memory
- 4 TB DDR5 as configured (NVIDIA: 2 TB, up to 4 TB)
- Storage
- 2 × 1.92 TB M.2 + 8 × 3.84 TB NVMe
- Networking
- 8 × ConnectX-7 400 Gb/s InfiniBand/Ethernet; 2 × BlueField-3
- Form factor
- 10U, 897 mm deep, 142 kg
- Note
- 1,440 GB of GPU memory in total; 3-year support included in the price
- Power
- 14.3 kW (maximum, NVIDIA)
Sources: NVIDIA DGX B200 specifications ↗ · NVIDIA DGX B200 user guide (size) ↗ · NVIDIA HGX B200 (FP16) ↗ · Broadberry configuration ↗
NVIDIA Quantum-2 QM9700 InfiniBand switch
- Networking
- 64 ports of 400 Gb/s NDR InfiniBand in 32 OSFP cages, 51.2 Tb/s
- Form factor
- 1U
- Power
- 1.7 kW (maximum, with active cables)
Sources: NVIDIA QM9700 user manual ↗ · NVIDIA QM9700 specifications ↗
APC NetShelter SX 42U rack, 750 × 1200 mm
- Form factor
- 42U enclosure, 1991 mm tall, 750 mm wide, 1200 mm deep
Sources: APC AR3350B2 product page ↗
Power adds each part's maximum rated power, before cooling and power losses. Model sizes are memory arithmetic: 2 bytes per parameter at 16-bit, half a byte at 4-bit, keeping a fifth of the memory for the context cache. They say what fits, not how fast it runs.
Can't afford it? Rent it for real.
Cloud providers rent the NVIDIA B200 from $7.99 per GPU an hour, or $2.50 on spot, compared on GlobeGPU (rates of 25 Oct 2026).
Compare NVIDIA B200 prices on GlobeGPU ▸