GB300 NVL72 is the Blackwell Ultra version of GB200 NVL72: the same 72-GPU rack and 130 TB/s NVLink, with about 278 GB of memory per GPU instead of 186 GB (20 TB per rack against 13.4 TB) and 1.5 times the dense FP4 compute. GB200 keeps far more FP64 and INT8. Pick GB300 for large-model inference and training in FP4 or FP8; GB200 only for FP64 or INT8-heavy work or a lower price. As of 19 October 2026, 8 operators in our pipeline offer GB300 NVL72, with starts from Q1 2027.
Both are liquid-cooled racks of 72 GPUs and 36 Grace CPUs in one NVLink domain. The difference is inside the GPU: Blackwell Ultra trades FP64 and INT8 for more memory and FP4.
| GB200 NVL72 | GB300 NVL72 | |
|---|---|---|
| GPU memory | 13.4 TB HBM3E (186 GB per GPU) | 20 TB (about 278 GB per GPU) |
| Memory bandwidth | 576 TB/s | Up to 576 TB/s |
| FP4 Tensor, sparse | 1,440 PFLOPS | 1,440 PFLOPS |
| FP4 Tensor, dense | 720 PFLOPS | 1,080 PFLOPS |
| FP8 Tensor, sparse | 720 PFLOPS | 720 PFLOPS |
| INT8 Tensor, sparse | 720 POPS | 24 POPS |
| FP64 | 2,880 TFLOPS | 100 TFLOPS |
| NVLink | 130 TB/s | 130 TB/s |
| GPU max power | Up to 1,200 W | Up to 1,400 W |
| Networking | Quantum-X800 InfiniBand or Spectrum-X800 Ethernet | ConnectX-8 at 800 Gb/s per GPU; Quantum-X800 InfiniBand or Spectrum-X |
Sources: NVIDIA GB200 NVL72 (372 GB per two-GPU superchip) and NVIDIA GB300 NVL72 product pages, NVIDIA Blackwell Ultra technical blog (GPU power), seen 14 October 2026. GB300 memory per GPU is the rack's 20 TB over 72 GPUs.
Reserved racks are quoted per request; the rate depends on term, volume, start date and prepayment. We publish GB300 rates only when there are enough priced offers that no single operator's price shows through.
We do not hold GB200 offers at the moment; operators are building GB300. GB300 NVL72 rental.
Two quotes at the same price per GPU-hour are rarely the same price. Before any match we bring every offer to the same eight lines:
Requirements are handled confidentially and shared with operators only on a need-to-know basis, never with your name attached before you approve it. We reply within one business day from dima@ancapex.io.
GB300 NVL72 uses Blackwell Ultra GPUs: 20 TB of GPU memory per rack against 13.4 TB on GB200 (about 278 against 186 GB per GPU), and 1,080 against 720 PFLOPS of dense FP4. NVLink stays at 130 TB/s. GB200 has far more FP64 (2,880 against 100 TFLOPS) and INT8.
For large-model inference and training in FP4 or FP8, yes: more memory per GPU and 1.5 times the dense FP4. For FP64 science workloads, GB200 is the stronger rack.
On public price lists the median on-demand GB200 NVL72 price was $16.00 per GPU-hour across 3 providers on 13 October 2026 (getdeploying). CoreWeave lists a 4-GPU instance at $42.00 per hour.
From a few public clouds on demand. We do not hold reserved GB200 offers at the moment; operators in our pipeline are building GB300.
Yes. Both are liquid-cooled rack-scale systems; the site must support direct liquid cooling.
