H100 Price per Hour: Every Major Cloud, Dated
H100 rental prices per GPU-hour on 26 October 2026: on-demand from $2.50, spot from $1.54, reserved from $1.75. RunPod, Lambda, Hyperstack, Azure and more.
H100 on-demand, per GPU-hour
| Provider | Variant | On-demand | Spot | Reserved |
|---|---|---|---|---|
| Hyperstack | PCIe | $2.50 | $2.00 | $1.75 |
| Hyperstack | NVLink 94 GB | $2.60 | – | $1.82 |
| RunPod | PCIe | $2.89 | – | contact sales |
| RunPod | NVL 94 GB | $3.19 | – | contact sales |
| Hyperstack | SXM | $3.20 | – | $2.72 |
| Jarvislabs | SXM | $3.49 | $1.54 | – |
| Modal | SXM, serverless | $3.95 | – | – |
| RunPod | SXM | $3.99 | – | contact sales |
| Lambda | SXM | $3.99 | – | contact sales |
| Together AI | cluster | $3.99 | $1.99 | $3.19 (91–180 days) |
| DigitalOcean | HGX | $4.41 | – | $3.26 (12 months) |
| CoreWeave | HGX, 8-GPU node | $6.16 per GPU | $2.46 | up to 60% off with commitment |
| Azure | ND H100 v5, 8-GPU VM | $12.29 per GPU | $2.27 | – |
Sources: each provider's pricing page, and Azure's retail price API, read on 26 October 2026. Node prices divided by eight.
SXM, PCIe or NVL?
- SXM (80 GB, about 3.35 TB/s): the version in 8-GPU HGX servers, linked by NVLink. Best for multi-GPU training.
- PCIe (80 GB, about 2 TB/s): cheaper; fine for single-GPU inference and fine-tuning.
- NVL (94 GB, about 3.9 TB/s): more memory, sold in pairs linked by NVLink.
Specs: NVIDIA H100.
What an H100 job costs
Ten hours on one H100 SXM costs $32 to $40 on-demand at the rates above, before storage and egress. On a provider that only rents 8-GPU nodes, the same hours cost eight times as much. Use the job cost calculator for your case.
Spotted an error or a figure that has changed? Let us know