NVIDIA H100 SXM rental price per hour
80GB HBM3 VRAM · Hopper architecture · released 2023
The cheapest way to rent an NVIDIA H100 SXM by the hour in June 2026 is about $1.93/hr on Vast.ai. We track 14 providers publishing an on-demand price for it , and spot/community rates reach as low as $0.95/hr. Large-scale LLM training and high-throughput inference. The de-facto training workhorse; SXM form factor with NVLink for multi-GPU scale-out.
Source: Provider pricing pages. Data as of June 2026.
NVIDIA H100 SXM price by provider
| Provider | Type | On-demand /hr | Spot/community /hr | Notes |
|---|---|---|---|---|
| Vast.ai | marketplace | $1.93/hr | $0.95/hr | Marketplace median (~$1.53-2.27); interruptible ~$0.75-1.10. Host-set. |
| Vultr Cloud GPU | specialist | $2.30/hr | — | HGX H100, per GPU. |
| Hyperstack | specialist | $2.40/hr | — | On-demand (a price tracker shows $1.90 — likely an unreflected drop; verify on page). |
| Verda (formerly DataCrunch) | specialist | $3.25/hr | — | SXM on-demand; spot ~35-65% lower. |
| RunPod | specialist | $3.29/hr | $1.99/hr | Secure Cloud on-demand; Community Cloud from ~$1.99/hr. |
| Lambda | specialist | $3.29/hr | — | On-demand; 8x config also seen near $3.99/GPU. |
| Paperspace (DigitalOcean) | specialist | $3.39/hr | — | DigitalOcean GPU Droplet (HGX, per GPU); 12-mo reserved $2.50. |
| Nebius | specialist | $3.85/hr | $2.15/hr | HGX per GPU; reserved/preemptible $2.15. |
| Crusoe | specialist | $3.90/hr | — | On-demand; spot lower. |
| Together AI | specialist | $4.79/hr | — | GPU Cluster per-GPU; Dedicated Inference 1x $6.49; reserved from $3.29. |
| CoreWeave | specialist | $6.16/hr | — | HGX H100, per GPU from 8-GPU node. |
| AWS EC2 | hyperscaler | $6.88/hr | $3.37/hr | Derived: p5.48xlarge $55.04/hr / 8 GPUs. us-east-1. |
| Google Cloud | hyperscaler | $10.98/hr | $4.98/hr | Derived: a3-highgpu-8g $87.83/hr / 8 GPUs. us-central1. |
| Microsoft Azure | hyperscaler | $12.29/hr | $2.27/hr | Derived: ND H100 v5 $98.32/hr / 8 GPUs. East US. |
Source: Provider pricing pages. Data as of June 2026.
Sorted cheapest on-demand first; "—" means the provider does not publish a price for this GPU (see the note). Snapshot June 2026 — cloud GPU prices change weekly; verify on the provider's pricing page before relying on a figure.
About the NVIDIA H100 SXM
Large-scale LLM training and high-throughput inference. The de-facto training workhorse; SXM form factor with NVLink for multi-GPU scale-out.
- VRAM: 80GB HBM3
- Architecture: Hopper
- Released: 2023
- Cheapest on-demand: $1.93/hr on Vast.ai
- Cheapest spot/community: $0.95/hr on Vast.ai
Frequently asked questions
How much does it cost to rent an NVIDIA H100 SXM per hour?
The cheapest on-demand NVIDIA H100 SXM we track is $1.93/hr on Vast.ai, with 14 providers publishing a price. Spot/community rates go as low as $0.95/hr. Snapshot June 2026 — cloud GPU prices change weekly; verify on the provider's pricing page before relying on a figure.
Which provider is cheapest for the NVIDIA H100 SXM?
Vast.ai at $1.93/hr is the cheapest on-demand option in our June 2026 snapshot. Specialist neoclouds and marketplaces usually beat the hyperscalers (AWS/GCP/Azure) by 2-4x on the same GPU.
What is the NVIDIA H100 SXM good for?
Large-scale LLM training and high-throughput inference. The de-facto training workhorse; SXM form factor with NVLink for multi-GPU scale-out. It has 80GB HBM3 of VRAM and uses the Hopper architecture (released 2023).
Is spot or on-demand cheaper for the NVIDIA H100 SXM?
Spot, community and interruptible tiers can cut the NVIDIA H100 SXM price roughly in half — the cheapest discounted rate we track is $0.95/hr versus $1.93/hr on-demand. The trade-off is your instance can be reclaimed, so use spot for fault-tolerant or checkpointed jobs.
Related GPUs
- AMD Instinct MI300X (192GB HBM3)
- NVIDIA B200 (192GB HBM3e)
- NVIDIA H100 PCIe (80GB HBM2e)
- NVIDIA H200 (141GB HBM3e)
Keep exploring
Source & accuracy
Figures are a dated snapshot from each provider's published pricing page (as of June 2026). Hyperscaler per-GPU prices are derived from 8-GPU instance prices (instance price / 8) and include bundled vCPU/RAM; marketplace (Vast.ai) figures are moving medians, not list prices. We do not invent numbers — where a value is unknown we show "—" with a reason. See our methodology. This is an informational comparison, not a quote.
Last updated: 2026-06-21