NVIDIA · Data centre
H200
Rented by 2 providers, though not in the same exact variant — the figures below describe different hardware and should not be read as one spread. 3 exact variants are priced separately, because they are different silicon.
RunPod · NVL · 143 GB
Every published rate
cheapest first · USD per GPU per hour| Provider | What they call it | Their own price, and the arithmetic | Per GPU-hour |
|---|---|---|---|
| RunPod
NVL · 143 GB |
H200 NVL
NVIDIA H200 NVL |
$0.5000 per hour = $0.5000 per GPU-hour
region Secure Cloud · 1 offer Low stock |
$0.5000 |
| RunPod
SXM · 141 GB |
H200 SXM
NVIDIA H200 |
$3.59 per hour = $3.5900 per GPU-hour
region Secure Cloud · 1 offer High stock |
$3.5900
$3.63 per PFLOP-hr |
| Azure
141 GB |
H200
Standard_ND96isr_H200_v5 |
$84.80 per hour ÷ 8 GPUs = $10.6000 per GPU-hour
region US West 2 · 138 offers low-priority $2.12 |
$10.6000 |
The silicon
vendor-published| Variant | Architecture | Memory | FP16 dense | Board power |
|---|---|---|---|---|
| H200 SXM | Hopper 2024 | 141 GB HBM3e | 989.4 TFLOPS | 700 W |
Vendor-published figures. FP16 throughput is dense tensor throughput with no sparsity assumed; vendors often headline the sparse figure, which is twice as large.
Vendor page ↗
2 of this chip's 3 variants carry no vendor specification here — they are priced all the same, and simply say nothing where a figure would go.
What they call it
Providers write the same hardware down differently. Every name below is one this index has seen for this chip; nothing was renamed to make the table tidier.
H200 NVLH200 SXMH200