Discount access, not sticker price, determines the real H100 cloud bill
Across 13 named providers, the observed compute floor runs from $1.03 to $4.72 per H100-hour. At 10 GB of public egress per GPU-hour, RunPod remains the lowest measured option at $1.03; hyperscaler network charges can add roughly $0.87–$1.00 per hour.
Three useful buying positions
RunPod
Spot pricing plus a measured zero egress fee produces the lowest modeled hourly cost. Community capacity is variable; Secure Cloud is the more controlled tier.
CoreWeave
The surfaced three-year reserved rate is materially below its $2.23 sticker. The result set describes supply as abundant in North America and Europe.
Voltage Park
Self-serve H100 capacity was described as deployable in about 15 minutes, with no minimum term and zero surfaced egress charge.
What changes the answer
Provider detail
| Provider | Sticker $/GPU-h | Best discount $/GPU-h | Discount / access | Egress $/GB | Effective at 10 GB/h | Availability signal | Source |
|---|---|---|---|---|---|---|---|
| RunPod | 1.47 | 1.03 | Spot | 0.00 | 1.03 | Available; Community variable, Secure Cloud controlled | turion.ai |
| TensorDock | 2.09 | 1.30 | Spot | — | ≥1.30 | Marketplace; on-demand and spot | flevy.com |
| CoreWeave | 2.23 | 1.49 | 3-year reserved | 0.00 | 1.49 | Abundant supply; North America and Europe | cloudpricecalculators.com |
| Vast.ai | 0.90 | 0.60 | Spot / preemptible | 0.09 | 1.50 | Marketplace across many independent hosts | intuitionlabs.ai |
| Voltage Park | 1.99 | 1.99 | No contract | 0.00 | 1.99 | Self-serve, roughly 15-minute provisioning; US only | rfp.wiki |
| FluidStack | 2.10 | 2.10 | Reserved | 0.00 | 2.10 | Claims multi-thousand-GPU deployment in days | inferencehub.org |
| Nebius | 2.95 | 2.15 | Preemptible | 0.00 | 2.15 | EU-north1/2, EU-west1, US-central1 surfaced | spheron.network |
| Lambda | 2.49 | 2.20 | 1-year reserved | 0.00 | 2.20 | On-demand varies; reservation guarantees capacity | gpuadvisor.com |
| Paperspace | 3.09 | 2.24 | 36-month commit | 0.00 | 2.24 | 5–10 instances across regions; high variance | cloudpricecalculators.com |
| Google Cloud | 2.99 | 1.86 | 3-year commitment | 0.10 | 2.86 | A3 regions surfaced; quotas/capacity still apply | deploybase.ai |
| Microsoft Azure | 12.29 | 2.27 | Spot | 0.087 | 3.14 | Active; East US/East US 2 surfaced, 1–4 week signal | sparecores.com |
| AWS EC2 | 4.72* | 4.72* | Capacity Block | — | ≥4.72 | Advance Capacity Block; 8-GPU instance access | aws.amazon.com |
| Crusoe Cloud | 1.04† | 2.94† | Conflicting extraction | 0.00 | — | US-West and US-Central surfaced | computecomparison.com |
Method — Live web scan on 25 Aug 2026; 246 targeted pages plus specialist and hyperscaler result sets were consolidated to 14 normalized rows, then the duplicate AWS brand was merged editorially. “Effective” = lowest surfaced discounted compute price per H100-hour + 10 × first-tier public internet egress $/GB. Unknown egress is shown as ≥, not assumed free. Prices mix H100 SXM and PCIe offerings and may exclude CPU, storage, tax and minimum cluster size. *AWS value is a Capacity Block accelerator rate, not ordinary on-demand sticker pricing. †Crusoe was excluded from ranking because its extracted discount exceeded its sticker.