GB300 NVL72
Seventy-two Blackwell Ultra GPUs, one NVLink domain, 20 TB of HBM3e in a single liquid-cooled rack.
Quote. Lead time 8 to 12 weeks.
Volume pricing
| Units | Discount | Unit price |
|---|---|---|
| 1 unit | List | $3,900,000 |
| 2 to 7 units | 3% off | $3,783,000 |
| 8 to 31 units | 6% off | $3,666,000 |
| 32+ units | 10% off | $3,510,000 |
Unit price at 1: $3,900,000
Quote only. Price shown is indicative; we confirm price and date in writing before you pay.
Request quote
The full sheet.
Figures from the manufacturer datasheet. Price is an indicative street price, shown as from.
- VRAM
- 20,736 GB
- Memory
- HBM3e
- Bandwidth
- 576 TB/s
- FP8
- 720 PFLOPS
- FP4
- 1,440 PFLOPS
- Board power
- 132,000 W
- Form factor
- Rack
- NVLink
- Yes
- Architecture
- Blackwell Ultra
- GPUs
- 72
- Per GB of VRAM
- $188.08
- Availability
- Quote, 8 to 12 weeks
- Sale
- Quote only
- SKU
- GB300-NVL72
What it is for.
Workloads, sizing and what ships in the box.
20,736 GB of HBM3e in one rack, one NVLink domain.
Seventy-two Blackwell Ultra GPUs and 36 Grace CPUs in a single liquid-cooled rack, joined by NVLink switch trays into one memory domain. That is 20,736 GB of HBM3e and 720 PFLOPS of FP8 behind one scheduler. Power draw is 132 kW, so a site survey comes before any quote. Lead time runs 8 to 12 weeks. Indicative price $3,900,000, about $188 per GB.
Good for
- DeepSeek-V3 671B at FP16 (1,610 GB with headroom) with room for twelve copies or one very long context.
- Llama 3.1 405B at FP16 (972 GB) serving many tenants from one NVLink domain.
- Pretraining and full fine-tunes of models up to 8,600B parameters at FP16 weights by the rule above (optimizer state needs more, plan for sharding).
Fit rule: FP16 weights take 2 bytes per parameter, FP8 1 byte, 4-bit 0.5 bytes, plus 20% headroom for KV cache at short context.
Pairs with
- HGX B300 8-GPU, for a staging node while the rack is built.
- H200 SXM 141GB, for inference capacity that ships sooner.
Liquid-cooled rack. Includes 36 Grace CPUs, NVLink switch trays, power shelves. Site survey required.
How much VRAM per box.
Card count is capped by slots and by the chassis power budget.
Parts bought alongside.
Same workload, a different trade between VRAM, bandwidth and price.