Cheapest 8xH100 node per hour
A full 8xH100 (SXM) node — the standard HGX building block for distributed LLM training — costs from about $15.44/hr on Vast.ai ($1.93/hr/GPU x 8), or roughly $11,271/month run continuously. Hyperscalers cost several times more for the same node. Ranked below by per-GPU H100 SXM rate x 8.
Source: Provider pricing pages. Data as of June 2026.
8xH100 node cost ranked cheapest first
| Rank | Provider | Per GPU /hr | 8-GPU node /hr | Est. /month |
|---|---|---|---|---|
| #1 | Vast.ai | $1.93/hr | $15.44/hr | ~$11,271/mo |
| #2 | Vultr Cloud GPU | $2.30/hr | $18.40/hr | ~$13,432/mo |
| #3 | Hyperstack | $2.40/hr | $19.20/hr | ~$14,016/mo |
| #4 | Verda (formerly DataCrunch) | $3.25/hr | $26.00/hr | ~$18,980/mo |
| #5 | RunPod | $3.29/hr | $26.32/hr | ~$19,214/mo |
| #6 | Lambda | $3.29/hr | $26.32/hr | ~$19,214/mo |
| #7 | Paperspace (DigitalOcean) | $3.39/hr | $27.12/hr | ~$19,798/mo |
| #8 | Nebius | $3.85/hr | $30.80/hr | ~$22,484/mo |
| #9 | Crusoe | $3.90/hr | $31.20/hr | ~$22,776/mo |
| #10 | Together AI | $4.79/hr | $38.32/hr | ~$27,974/mo |
| #11 | CoreWeave | $6.16/hr | $49.28/hr | ~$35,974/mo |
| #12 | AWS EC2 | $6.88/hr | $55.04/hr | ~$40,179/mo |
| #13 | Google Cloud | $10.98/hr | $87.84/hr | ~$64,123/mo |
| #14 | Microsoft Azure | $12.29/hr | $98.32/hr | ~$71,774/mo |
Source: Provider pricing pages. Data as of June 2026.
Node rate = published H100 SXM per-GPU rate x 8; monthly = node rate x 730 hours (continuous). Excludes storage, egress and inter-node InfiniBand. Snapshot June 2026 — cloud GPU prices change weekly; verify on the provider's pricing page before relying on a figure.
Frequently asked questions
How much does an 8xH100 node cost per hour?
The cheapest 8xH100 (SXM) node we track is about $15.44/hr ($1.93/hr/GPU x 8) on Vast.ai — roughly $11,271/month if run continuously. Hyperscalers cost several times more for the same node. Snapshot June 2026 — cloud GPU prices change weekly; verify on the provider's pricing page before relying on a figure.
Why rent a full 8xH100 node instead of single GPUs?
An 8-GPU HGX node has fast intra-node NVLink/NVSwitch bandwidth, which large-model training needs for tensor/pipeline parallelism. Eight separate single-GPU instances would not have that high-bandwidth interconnect. For single-GPU fine-tuning or inference you do not need a full node.
Is the cheapest 8xH100 node always the best?
No. For distributed training the inter-node fabric (InfiniBand) and the provider's reliability matter as much as price. A slightly pricier provider with non-blocking InfiniBand can finish a multi-node run faster and cheaper overall. The per-node price here is a starting point, not the whole story.
How is the node price calculated?
We take each provider's published per-GPU H100 SXM on-demand rate and multiply by 8. The monthly figure is that node rate x 730 hours (continuous use). It excludes storage, egress and inter-node networking. See the methodology page.
Keep exploring
Last updated: 2026-06-21