H100 NVL 94GB
94 GB HBM3 on PCIe at 400 W. Eight in a 4U gets you 752 GB without liquid cooling.
In stock. Lead time 72 hours.
Volume pricing
| Units | Discount | Unit price |
|---|---|---|
| 1 unit | List | $26,000 |
| 2 to 7 units | 3% off | $25,220 |
| 8 to 31 units | 6% off | $24,440 |
| 32+ units | 10% off | $23,400 |
Unit price at 1: $26,000
Request a bulk quote
The full sheet.
Figures from the manufacturer datasheet. Price is an indicative street price, shown as from.
- VRAM
- 94 GB
- Memory
- HBM3
- Bandwidth
- 3.9 TB/s
- FP8
- 3.34 PFLOPS
- FP4
- Not supported
- Board power
- 400 W
- Form factor
- PCIe dual-slot
- NVLink
- Yes
- Architecture
- Hopper
- GPUs
- 1
- Per GB of VRAM
- $276.60
- Availability
- In stock, ships 72h
- Sale
- Buy online
- SKU
- H100-NVL
What it is for.
Workloads, sizing and what ships in the box.
94 GB of HBM3 at 400 W, air-cooled.
Hopper on PCIe with 94 GB of HBM3 at 3.9 TB/s and 3.34 PFLOPS of FP8, at 400 W. Passive cooling, so it needs a chassis with real airflow. Eight in a 4U is 752 GB without liquid cooling. NVLink bridges pair cards. In stock, ships in 72 hours. $26,000, about $277 per GB of VRAM.
Good for
- Llama 3.1 70B at FP8 (84 GB with headroom) on a single card.
- Qwen2.5 32B at FP16 (77 GB) with 17 GB spare.
- Qwen2.5 72B at FP16 (173 GB) across two bridged cards.
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
- H200 NVL 141GB, when 94 GB per card is short.
- L40S 48GB, for mixed inference and video encode fleets.
Passive cooling, needs chassis airflow.
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.