HGX B300 8-GPU
Eight B300s on one baseboard. 2.3 TB of HBM3e behind NVLink 5 for the largest single-node models.
Quote. Lead time 6 to 10 weeks.
Volume pricing
| Units | Discount | Unit price |
|---|---|---|
| 1 unit | List | $420,000 |
| 2 to 7 units | 3% off | $407,400 |
| 8 to 31 units | 6% off | $394,800 |
| 32+ units | 10% off | $378,000 |
Unit price at 1: $420,000
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
- 2,304 GB
- Memory
- HBM3e
- Bandwidth
- 64 TB/s
- FP8
- 72 PFLOPS
- FP4
- 144 PFLOPS
- Board power
- 11,200 W
- Form factor
- HGX baseboard
- NVLink
- Yes
- Architecture
- Blackwell Ultra
- GPUs
- 8
- Per GB of VRAM
- $182.29
- Availability
- Quote, 6 to 10 weeks
- Sale
- Quote only
- SKU
- HGX-B300-8GPU
What it is for.
Workloads, sizing and what ships in the box.
2,304 GB of HBM3e on one baseboard.
Eight B300 GPUs on one HGX baseboard, linked by NVLink 5. 2,304 GB of HBM3e and 72 PFLOPS of FP8 in a single node, at 11,200 W for the GPUs alone. Buy the baseboard or a full 8U node, decided at quote. Lead time is 6 to 10 weeks and we date it in writing. Indicative $420,000, about $182 per GB of VRAM.
Good for
- DeepSeek-V3 671B at FP16 (1,610 GB with headroom) on one node, no cross-node traffic.
- Llama 3.1 405B at FP16 (972 GB) with more than 1 TB left for KV cache and batch.
- Fine-tuning 70B models with full weights and optimizer state on one box.
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
- B200 SXM 192GB, as serial-matched spares.
- GB300 NVL72, when one node stops being enough.
Baseboard only or integrated 8U node, your choice at quote.
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.