GB200 vs B200, GB300 vs B300: what Grace Blackwell and NVL72 mean
Short answer: the B200 and B300 are GPUs. The GB200 and GB300 are "superchips" that put the same Blackwell GPUs on one module with NVIDIA's Grace CPU, two GPUs to one CPU. You normally meet a B200 or B300 as eight GPUs in an HGX server, and a GB200 or GB300 as 72 GPUs in an NVL72 rack. Same GPU family, very different machines.
What the "G" stands for
The G is Grace, NVIDIA's own Arm-based data-centre CPU. A GB200 superchip joins one Grace CPU and two Blackwell GPUs with a fast chip-to-chip link called NVLink-C2C, so the CPU and GPUs can share memory far faster than over PCIe. The GB300 does the same with two Blackwell Ultra GPUs, the chip inside the B300. The earlier GH200 paired one Grace with one Hopper GPU.
A B200 or B300 server, by contrast, pairs eight GPUs with ordinary x86 server CPUs, such as the two Intel Xeon processors in NVIDIA's own DGX B200.
What NVL72 means
NVL72 is a whole rack built as one machine: 36 Grace CPUs and 72 Blackwell GPUs, all on a single NVLink fabric with 130 TB/s of bandwidth across the rack, according to NVIDIA. Every GPU can talk to every other GPU in the rack at full NVLink speed, 1.8 TB/s each. In an 8-GPU HGX server, that fast fabric stops at eight GPUs, and anything bigger goes over the slower InfiniBand network between servers.
That rack is also a heavy load for a data centre. NVL72 racks are liquid-cooled and draw in the region of 120 to 140 kW each (HPE lists the GB200 NVL72 at 132 kW nominal; NVIDIA calls the GB300 NVL72 a 120 kW rack). An 8-GPU DGX B200, for comparison, draws about 14.3 kW. Most data-centre halls, in India and elsewhere, were not built for 120 kW racks, which is why NVL72 capacity is concentrated in a small number of purpose-built sites. Our India data centre guide covers what that means for power and space.
Side by side
B200 / B300 in an HGX server | GB200 / GB300 in an NVL72 rack | |
|---|---|---|
What it is | 8 GPUs on an NVIDIA HGX board in one server | 72 GPUs and 36 Grace CPUs in one rack |
CPU | Standard x86 server CPUs | NVIDIA Grace (Arm), one per two GPUs |
GPUs on one NVLink fabric | 8 | 72 |
NVLink per GPU | 1.8 TB/s | 1.8 TB/s |
Memory per GPU | B200: 180 GB; B300: 288 GB | GB200: about 186 GB; GB300: up to 288 GB |
Power | About 14 kW per 8-GPU server | About 120 to 140 kW per rack |
Cooling | Air or liquid, depending on the server | Liquid only |
Best for | Training, fine-tuning and serving that fit in 8 GPUs, or scale out over InfiniBand | The largest frontier-model training and serving, where 72 GPUs must act as one |
Figures are NVIDIA's published specifications as of October 2026, plus the power ratings named above. The B200 chip carries 192 GB, of which 180 GB per GPU is usable in HGX and DGX B200 systems.
Is the GB200 better than the B200?
GPU for GPU, they are the same Blackwell generation. The GB200's advantage is the system around it: 72 GPUs on one NVLink fabric, plus the Grace CPU's fast shared memory. That matters for the very largest models, especially big mixture-of-experts models whose experts are spread across many GPUs and constantly exchange data. For models and jobs that fit in eight GPUs, an HGX B200 or B300 server does the same work without the rack-scale power, liquid cooling and minimum size.
The same holds for the GB300 vs the B300: same Blackwell Ultra GPU, with up to 288 GB each. Choose the rack only if your workload needs more than eight GPUs to behave as one.
What comes next: Vera Rubin NVL72
NVIDIA's next generation, Vera Rubin, follows the same pattern: the Vera Rubin NVL72 rack pairs 72 Rubin GPUs, with 288 GB of HBM4 each, with 36 Vera CPUs on sixth-generation NVLink. It was first announced as "NVL144", when NVIDIA briefly counted each GPU's two dies separately, and then renamed. Our guide to GPU form factors explains those names, along with PCIe, SXM, HGX and DGX.
Where FlexiCloud fits
We do not rent NVL72 racks. What we rent are dedicated 8-GPU B200 and B300 HGX servers, the same Blackwell and Blackwell Ultra GPUs, as single-tenant bare metal in India, the EU and the US, with multi-server clusters joined over InfiniBand on request. For most training, fine-tuning and inference, that is the practical way to run Blackwell. See GPU server prices, or tell us about your workload and we will tell you honestly whether 8-GPU servers fit it.