H100 SXM 80GB
The part the whole industry trained on. Still the best price per FLOP in SXM, allocated in 1 to 2 weeks.
Allocation. Lead time 1 to 2 weeks.
Volume pricing
| Units | Discount | Unit price |
|---|---|---|
| 1 unit | List | $27,500 |
| 2 to 7 units | 3% off | $26,675 |
| 8 to 31 units | 6% off | $25,850 |
| 32+ units | 10% off | $24,750 |
Unit price at 1: $27,500
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
- 80 GB
- Memory
- HBM3
- Bandwidth
- 3.35 TB/s
- FP8
- 3.96 PFLOPS
- FP4
- Not supported
- Board power
- 700 W
- Form factor
- SXM
- NVLink
- Yes
- Architecture
- Hopper
- GPUs
- 1
- Per GB of VRAM
- $343.75
- Availability
- Allocation, 1 to 2 weeks
- Sale
- Quote only
- SKU
- H100-SXM
What it is for.
Workloads, sizing and what ships in the box.
The SXM part most training code was tuned on.
80 GB of HBM3 at 3.35 TB/s, 3.96 PFLOPS of FP8 at 700 W. The software stack, kernels and cluster recipes most teams use were tuned on it. Allocation, 1 to 2 weeks, quote only. Sets of eight for HGX, or spares for an existing fleet. From $27,500, about $344 per GB of VRAM.
Good for
- Qwen2.5 32B at FP16 (77 GB with headroom) on one GPU.
- Llama 3.1 70B at 4-bit (42 GB) with room for long context.
- Llama 3.1 405B at FP8 (486 GB) across 8 in one HGX node.
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 SXM 141GB, same socket with 76% more memory.
- B200 SXM 192GB, for a new Blackwell training node.
Sets of 8 for HGX, also as spares.
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.