H100 Rental Cost Calculator
Estimate NVIDIA H100 rental price by hourly rate, GPU count, active hours, utilization, storage, and network overhead. Use it before committing to a cloud GPU provider or comparing rent vs buy.
Tracking 14 H100 providers where pricing is available. Last pricing refresh: Jan 15, 2026.
Calculate H100 GPU Rental Cost
Estimated cost
How to use the H100 rental cost calculator
Choose a live pricing source
Start from the cheapest tracked H100 rate, a provider row, or a custom quote from your vendor.
Set GPU count and schedule
Enter the number of H100 GPUs, active hours per day, days in the run, and expected utilization.
Add overhead
Include storage, bandwidth, orchestration, support, and idle-time overhead when the provider bills more than pure GPU runtime.
Compare with hardware
Use the break-even estimate to decide whether renting H100 GPUs or buying hardware fits the workload.
Example H100 rental cost scenarios
| Scenario | GPUs | Schedule | Rate | Estimated compute cost | Hardware break-even signal |
|---|---|---|---|---|---|
| Prototype run | 1 | 6 hr/day x 3 days | $1.38/GPU-hr | $24.89 | Rent is likely safer |
| Fine-tuning sprint | 8 | 10 hr/day x 14 days | $2.99/GPU-hr | $3.3k | Rent is likely safer |
| Always-on inference test | 2 | 24 hr/day x 30 days | $1.38/GPU-hr | $2.0k | Rent is likely safer |
| Training cluster month | 64 | 18 hr/day x 30 days | $2.99/GPU-hr | $103.3k | Rent is likely safer |
What changes the real H100 rental price?
On-demand vs spot
Spot H100 prices can be lower, but interruption risk matters for training jobs without reliable checkpointing.
Single GPU vs 8x node
An 8x H100 node may include NVLink, more CPU, RAM, storage, and networking, so per-GPU cost can differ from bare rental rows.
Utilization
Idle notebooks, stopped pods that still bill storage, and queue time can make realized cost higher than the headline hourly rate.
Networking and storage
Large datasets, checkpoint storage, egress, and persistent volumes can be meaningful for multi-day jobs.
Reserved commitments
AWS, Azure, GCP, and some GPU clouds may discount committed use, but the lock-in only helps predictable workloads.
Buy vs rent depreciation
H100 resale value, power, cooling, rack space, support, and future Blackwell pricing all affect the hardware side of the comparison.
Tracked H100 provider rates
| Provider | GPU | Instance | On-demand / GPU-hour | Spot / GPU-hour | Availability |
|---|---|---|---|---|---|
| Vast.ai | H100 NVL | vastai-h100-nvl | $1.38 | $1.07 | - |
| Fluidstack | H100 SXM | fluidstack-h100-sxm | $2.10 | - | - |
| Genesis Cloud | H100 SXM | genesis-h100-sxm | $2.19 | - | - |
| Paperspace | H100 SXM | paperspace-h100-sxm | $2.24 | - | - |
| TensorDock | H100 SXM | tensordock-h100-sxm | $2.25 | - | - |
| Datacrunch (Verda) | H100 SXM | datacrunch-h100-sxm | $2.29 | - | - |
| RunPod | H100 PCIe | NVIDIA H100 PCIe | $2.39 | $1.25 | - |
| Lambda Labs | H100 PCIe | lambda-h100-pcie | $2.49 | - | - |
| RunPod | H100 SXM | NVIDIA H100 80GB HBM3 | $2.69 | $1.75 | - |
| Jarvis Labs | H100 SXM | jarvislabs-h100-sxm | $2.99 | - | - |
| Lambda Labs | H100 SXM | lambda-h100-sxm | $2.99 | - | - |
| RunPod | H100 NVL | NVIDIA H100 NVL | $3.07 | $1.65 | - |
For all GPUs and provider filters, use the cloud GPU pricing comparison.
When renting H100 GPUs is better than buying
Rent H100 GPUs when
- The workload is temporary, seasonal, or uncertain.
- You need 8x or 64x H100 scale for a short training window.
- You want managed networking, images, and faster provisioning.
- Your utilization is below the hardware break-even point.
Model hardware purchase when
- H100s will run near 24/7 for many months.
- Data locality, compliance, or low-latency networking requires on-prem hardware.
- You can manage power, cooling, rack space, failures, and resale risk.
- Your total rental cost approaches the hardware-equivalent budget.
Useful pricing and specification references
GPU Cost tracks provider rows for comparison, but final purchase or rental decisions should also check official source pages: NVIDIA H100 specifications, AWS EC2 P5 instances, Google Cloud GPU pricing, and Azure VM pricing.
H100 Rental Cost FAQ
H100 rental price depends on provider, region, instance shape, networking, and availability. Specialized GPU clouds and marketplaces can be cheaper than hyperscalers, while enterprise clouds may include stronger compliance, networking, and support.
Multiply the per-GPU hourly rate by GPU count, active hours per day, days per month, and expected utilization. Then add storage, bandwidth, support, idle time, and platform fees that are not included in the GPU-hour headline rate.
Spot can reduce cost, but interruption risk is real. Use spot for training only when checkpoints are frequent, datasets can resume cleanly, and a restart will not waste too much expensive GPU time.
Buying starts to make sense when usage is steady and near the break-even hour count after including power, cooling, rack space, maintenance, financing, and depreciation. Short projects usually favor renting.
Multi-GPU nodes include CPU, RAM, local storage, high-speed networking, NVLink or InfiniBand, and provider allocation constraints. Some providers price the complete node, so the effective per-GPU rate can differ from bare single-GPU rental.