The GB300 NVL72 is the GB200 NVL72 rebuilt around Blackwell Ultra: 279 GB per GPU instead of 186, 1.5 times the dense FP4 and twice the network speed per GPU, with most of the FP64 and INT8 removed. Both are liquid-cooled racks of 72 GPUs and 36 Grace CPUs in one NVLink domain. For inference at long context the GB300 is the better rack. For double-precision HPC the GB200 is the better one, and for FP8 training the two do the same Tensor Core math. QuantaCloud builds either on request as reserved capacity, and we discuss rack-scale builds before we quote them.
GB200 vs GB300 NVL72 specs#
The rack, the CPUs and the NVLink fabric are the same, and the GPU is what changed. These are NVIDIA's figures from the two NVL72 pages, the Blackwell and Blackwell Ultra datasheets (October 2025) and the two DGX datasheets, re-read on 2026-11-12. FP8 and INT8 are quoted with sparsity, and dense is half.
| Per rack unless noted | GB200 NVL72 | GB300 NVL72 | What changes |
|---|---|---|---|
| Configuration | 72 Blackwell GPUs, 36 Grace CPUs | 72 Blackwell Ultra GPUs, 36 Grace CPUs | The GPU |
| GPU memory | 13.4 TB HBM3E | 20 TB HBM3E | 1.5x |
| Memory per GPU | 186 GB HBM3E | 279 GB HBM3E | 1.5x |
| GPU memory bandwidth | 576 TB/s | Up to 576 TB/s | Same |
| CPU memory | 17 TB LPDDR5X at 14 TB/s | 17 TB LPDDR5X at 14 TB/s | Same |
| CPU cores | 2,592 Arm Neoverse V2 | 2,592 Arm Neoverse V2 | Same |
| FP4 Tensor Core | 1,440 PFLOPS sparse, 720 dense | 1,440 PFLOPS sparse, 1,080 dense | Dense 1.5x, sparse the same |
| FP8/FP6 Tensor Core | 720 PFLOPS sparse | 720 PFLOPS sparse | Same |
| INT8 Tensor Core | 720 POPS sparse | 24 POPS sparse | Cut |
| FP64 | 2,880 TFLOPS | 100 TFLOPS | Cut |
| Attention layers | Baseline | 2x, per NVIDIA | Faster softmax |
| NVLink | 130 TB/s across the rack | 130 TB/s across the rack | Same |
| Network port per GPU (DGX) | 400 Gb/s ConnectX-7 | 800 Gb/s ConnectX-8 | 2x |
| BlueField-3 DPUs per rack (DGX) | 36 | 18 | One per tray instead of two |
| Max power per GPU | 1,200 W | 1,400 W | 200 W more |
| Rack power | About 120 kW at full load | Up to 142 kW | See below |
| Compute capability | 10.0 (sm_100) | 10.3 (sm_103) | New compile target |
Memory: 186 GB against 279 GB per GPU#
Memory is the first difference to check, because it decides how many GPUs a model occupies. At vLLM's default --gpu-memory-utilization of 0.92 the engine gets 170.2 GiB of a GB200 GPU and 255.3 GiB of a GB300 GPU (our calculation from the 189,471 and 284,208 MiB that nvidia-smi reports, the second from a single public report). The table counts weights and KV cache only, before activations.
| Model and precision | Weights | On GB200 | On GB300 |
|---|---|---|---|
| Llama-3.3-70B, BF16 | 141.1 GB (131.4 GiB) | 38.8 GiB left on one GPU: about 127,000 tokens of KV cache, just short of one full 131K context | 123.9 GiB left: about 406,000 tokens, three full contexts |
| DeepSeek-V4-Flash, FP4 and FP8 | About 167 GB on disk (155.5 GiB) | 14.7 GiB left on one GPU | 99.8 GiB left |
| DeepSeek-V4-Pro, FP4 and FP8 | About 893 GB on disk | More than one four-GPU tray holds (744 GB) | Fits one four-GPU tray (1,116 GB), as in DeepSeek's own example |
The Llama row uses 320 KiB of KV cache per token at BF16 (our calculation: 2 x 80 layers x 8 KV heads x 128 x 2 bytes). Across the rack the same ratio holds: 13.4 TB of HBM3E against 20 TB. The KV cache guide has the formula for other models.
Compute: dense FP4 up, FP64 and INT8 down#
The GB300's compute gain is dense FP4 and attention, and the other Tensor Core figures do not move. Dense FP4 rises from 720 to 1,080 PFLOPS per rack while sparse FP4 stays at 1,440, and NVIDIA's DGX GB300 datasheet sums the change up as 1.5x the dense FP4 and 2x the attention performance of DGX GB200. FP8 stays at 720 sparse PFLOPS per rack, so for FP8 training the GB300's advantages are memory and networking.
The losses are just as specific. FP64 drops from 2,880 to 100 TFLOPS per rack and INT8 from 720 to 24 POPS. If any part of the work is double-precision simulation or INT8 inference, the GB200 is the better rack. For NVFP4 inference it is the other way round.
Power and cooling#
Both racks are liquid-cooled, and the GB300 needs the larger power budget. NVIDIA's DGX rack guide, which covers both, puts rack power consumption at about 120 kW, and its Mission Control documentation gives the same figure for a GB200 rack at full load. For the GB300, NVIDIA's NVL72 reference architecture says a full rack can require up to 142 kW, so that is the figure I provision for. Both racks are fed from eight power shelves of 33 kW, each with six 5.5 kW supplies. The GPUs' own allowance rises from 86.4 kW to 100.8 kW (our calculation: 72 x 1,200 W and 72 x 1,400 W).
Cooling is the same design in both. Coolant runs through manifolds to cold plates on every CPU and GPU, and NVIDIA's DGX rack guide lists fans for the network cards and drives, although NVIDIA describes the GB300 as fully liquid-cooled. The GB300 power shelves also carry energy storage that charges when GPU demand drops and discharges when it spikes. NVIDIA reports a 30% lower peak grid demand when training the Megatron LLM, and said in the same post that the feature was coming to GB200 racks too.
Networking#
Inside the rack nothing changes: every GPU has 1.8 TB/s of fifth-generation NVLink, nine switch trays tie all 72 into one domain, and the total is 130 TB/s. Between racks the GB300 doubles the port. NVIDIA's DGX GB300 gives each GPU an 800 Gb/s ConnectX-8 port, against a 400 Gb/s ConnectX-7 port on a DGX GB200, with BlueField-3 DPUs carrying storage and management traffic in both. That matters as soon as a job spans more than one rack, because every byte that leaves the NVLink domain goes through those ports. The InfiniBand vs NVLink guide covers which traffic belongs on which link.
Software: sm_100 against sm_103, both on Arm#
The host is the same Grace CPU, so both racks need arm64 builds of every container image and Python wheel. The GPUs differ. A GB200 is compute capability 10.0 and runs on CUDA 12.8 or newer with an R570 driver. A GB300 is 10.3 and needs CUDA 12.9 or newer and an R580 driver, first listed in 580.82.07. Both need the open kernel modules on Linux. Kernels built for sm_100a run on a GB200 and not on a GB300, while the sm_100f family target covers both. vLLM states CUDA 12.8 as its minimum for GB200 and recommends CUDA 13 for GB300. The driver and CUDA version guide shows how to confirm what a node is running.
What each rack costs#
The published estimates put a GB200 NVL72 at about $3 million and disagree about the GB300. HSBC analysts estimated about $3 million per GB200 rack in 2024, a Wolfe Research note covered by Investing.com on 2026-01-30 cited reported prices of about $3 million for a GB200 NVL72 and $4.3 million for a GB300 NVL72, and on 2026-03-24 a Tom's Hardware source put them at $2.8 million to $3.4 million and $6 million to $6.5 million. The same source warned that many quotes online are manufacturer prices without a proper warranty. I cannot reconcile the two GB300 figures from public information, so treat both as dated estimates.
NVIDIA has never confirmed a list price for its NVL72 racks, and none of these figures covers the facility work around the rack. QuantaCloud quotes either build in writing.
Which rack to build#
The Ultra rack is worth its extra power when memory per GPU, FP4 or attention limits the job.
| Workload | Build | Why |
|---|---|---|
| Reasoning inference at long context | GB300 | 279 GB per GPU for weights and KV cache, attention 2x |
| FP4 inference | GB300 | 1,080 dense FP4 PFLOPS against 720 |
| FP8 or BF16 training that fits in 186 GB per GPU | GB200 | Same FP8 Tensor Core throughput |
| Training limited by memory, or spread across several racks | GB300 | 1.5x the memory and 800 Gb/s per GPU between racks |
| FP64 simulation or INT8 inference | GB200 | 2,880 against 100 TFLOPS, 720 against 24 POPS |
| A site that cannot give a rack more than about 120 kW | GB200 | The GB300 is planned for up to 142 kW |
There is a third NVL72 now. NVIDIA lists the Vera Rubin NVL72 as available, with 20.7 TB of HBM4, 2,520 dense FP4 PFLOPS and 227 kW at its highest power setting. If the model and its KV cache fit in one eight-GPU server, a B300 or B200 server is far easier to place than any rack.
Rack-scale builds are discussed first#
Either rack starts with where it will run. Your brief should cover the workload, the number of racks, how they connect to each other and to storage, the location and the start date.
The configuration, lead time and terms come back in writing before you commit.
Talk to us about an NVL72 buildGB200 vs GB300 questions#
Is the GB300 faster than the GB200?
For dense FP4 and for attention layers, yes: 1.5x and 2x on NVIDIA's figures. For FP8 the Tensor Core throughput is the same, and for FP64 and INT8 the GB200 is far ahead.
Will a GB300 rack fit where a GB200 rack does?
Check the power and cooling margin first. NVIDIA's documents give both racks the same eight 33 kW power shelves, but its reference architecture plans a GB300 rack for up to 142 kW, against about 120 kW for a GB200.
Can I rent a GB200 or GB300 by the hour?
No. Both are reserved capacity, built on request. The GPUs page shows what is on demand right now.
The rule I follow: choose the GB300 NVL72 when FP4 inference at long context is most of the work and the site can give each rack up to 142 kW, and the GB200 NVL72 when the job is FP8 training that fits in 186 GB per GPU, or needs FP64 or INT8. Either way the first step is a brief on the reserved capacity page with the models, rack count, network, storage, location and start date.
Talk to us about an NVL72 build