GB200 NVL72 / GB200 NVL72 vs HGX B200
GB200 NVL72 vs HGX B200
Both are Blackwell. One is a 72-GPU liquid-cooled rack, the other an 8-GPU server. Which to rent depends on how far your model needs to stretch.
The short version
| GB200 NVL72 | HGX B200 | |
|---|---|---|
| Unit | One rack, 72 GPUs | One server, 8 GPUs |
| NVLink domain | 72 GPUs, 130 TB/s | 8 GPUs, 14.4 TB/s |
| GPU memory | Up to 13.4 TB per rack | 1.4 TB per server |
| FP4 (sparse) | 1,440 PFLOPS per rack | 144 PFLOPS per server |
| Host CPU | NVIDIA Grace (Arm) | x86 |
| Cooling | Liquid | Air or liquid, by server design |
| Smallest sensible rental | A rack, or a 4-GPU slice | One node |
From NVIDIA's GB200 NVL72 and HGX pages (GB200 NVL72, HGX). Per GPU the two are close: about 20 vs 18 PFLOPS of sparse FP4. The difference is how many GPUs share one NVLink domain.
Pick GB200 NVL72 when
- Your model needs tensor or expert parallelism beyond 8 GPUs.
- You serve large mixture-of-experts or long-context models where the KV cache spans GPUs.
- You're renting at rack scale anyway and want the most work per rack.
Pick HGX B200 when
- Your jobs fit in 8 GPUs, or you scale out with data parallelism.
- Your stack assumes x86 hosts and you'd rather not port to Arm.
- You want a small start, or more provider choice. HGX servers fit far more data centers than 120 kW liquid-cooled racks.
Ask for both
Quotes for the same GPU generation vary a lot between the two systems and between operators. If you're flexible, say so in the form and we'll come back with options for each. For a wider list of B200 clouds, see Best Neocloud.
Questions
Is DGX B200 the same as HGX B200?
DGX B200 is NVIDIA's own 8-GPU server built on the HGX B200 board. Clouds and OEMs build their own HGX B200 servers too.
Is HGX B200 the same GPU as GB200?
Same Blackwell generation. GB200 pairs two Blackwell GPUs with a Grace CPU on one superchip; HGX B200 puts eight Blackwell GPUs on a board with x86 hosts.
RackShore runs this site and may be paid a referral fee by the provider you choose. Buyers pay us nothing. Spot a mistake? Tell us.