GB200 NVL72 vs HGX B200
Same GPU, different system: a 72-GPU NVLink rack against an 8-GPU server.
The difference
Both use the Blackwell GPU. GB200 NVL72 pairs it with Grace CPUs and joins 72 GPUs in one NVLink domain; HGX B200 joins 8. Per GPU, the GB200 runs slightly faster (5 vs 4.5 dense FP8 PFLOPS) with 186 GB against 180 GB. It is liquid-cooled only, at about 120 kW per rack.
Side by side
| Per GPU | HGX B200 | GB200 NVL72 | GB200 NVL72 / HGX B200 |
|---|---|---|---|
| Architecture, released | Blackwell, Early 2025 | Blackwell, Early 2025 | |
| Memory | 180 GB (1.4 TB per 8-GPU board) | 186 GB (13.4 TB per rack) | 1.03x |
| Memory bandwidth | 8 TB/s | 8 TB/s (576 TB/s per rack) | 1.00x |
| BF16, dense TFLOPS | 2,250 | 2,500 | 1.11x |
| FP8, dense TFLOPS | 4,500 | 5,000 | 1.11x |
| FP4, dense TFLOPS | 9,000 | 10,000 | 1.11x |
| FP64 (Tensor / matrix), TFLOPS | 37 | 40 | 1.08x |
| Max power per GPU | 1,000 W air-cooled, 1,200 W liquid | Not published per GPU; about 120 kW per rack | |
| Power per 8 GPUs | About 14 kW | About 120 kW per 72-GPU rack (13.3 kW per 8 GPUs) | 0.95x |
| GPU-to-GPU link | NVLink 5, 1.8 TB/s per GPU | NVLink 5, 1.8 TB/s per GPU, 130 TB/s per rack | |
| Scale-up domain | 8 GPUs (HGX board) | 72 GPUs (one NVLink domain) | |
| Form factor, cooling | SXM module, 8-GPU HGX board; air or liquid | Grace Blackwell Superchip (1 Grace CPU + 2 GPUs), 72-GPU rack; liquid only | |
| Price per GPU, October 2026 | $45,000-$58,750 (New 8-GPU node, per GPU) | - | |
| Launch price reported in the press | $30,000 to $40,000 per unit, per NVIDIA’s CEO (2024) | $60,000 to $70,000 per Superchip (1 CPU + 2 GPUs); $3M per NVL72 rack, HSBC (2024) | |
| Residual band, 2027 (conservative) | 40-54% of original price | - |
Tensor Core figures are dense. Prices as of October 2026.
Sources: American Compute, NVIDIA, Managed GPUs (American Compute), CNBC, Tom's Hardware.
The difference is the system
The difference is the system, not the chip. In an HGX B200, eight GPUs share NVLink on one board, and anything larger crosses the network. In GB200 NVL72, switch trays join all 72 GPUs at 130 TB/s of aggregate bandwidth, so a model split across the rack talks over NVLink.
Power, cooling and resale
An NVL72 rack needs a liquid-cooled site. It draws roughly 120 kW, and past about 40 kW per rack air cooling stops working. An HGX B200 server draws about 14 kW and goes into far more data centers, so it has more buyers when it changes hands.
We publish no separate rack-scale band. Our residual report groups GB200 with B200 by technology age; the B200 conservative band is 40-54% of original price in 2027.
Sources: NVIDIA, American Compute, Managed GPUs (American Compute).
Residual value and appraisals
| Year | HGX B200 | GB200 NVL72 |
|---|---|---|
| 2026 | 59-67% | - |
| 2027 | 40-54% | - |
| 2028 | 30-42% | - |
| 2029 | 16-26% | - |
| 2030 | 6-11% | - |
What a used unit is expected to fetch, as a share of its original price: the conservative band from our residual value report.
HGX B200 appraisal · HGX B200 residual value · GB200 NVL72 appraisal
Discuss a transaction
Send us the equipment list and the loan or lease you have in mind. We will come back with an indication and the questions we would need answered.
Prefer email? hello@amcompute.com