The cheapest captured H100 rate is $0.80/hr, but it comes with an asterisk — trustworthy floors sit near $1.00–$1.60
Asked:“What is the real effective price per hour for an NVIDIA H100 across GPU cloud providers? Include the sticker price, spot or committed discounts, egress fees, and current availability for each provider.”
This comparison covers 40 named GPU-cloud offerings as captured on 2026-09-06, normalized to US dollars per GPU-hour (often from multi-GPU nodes). Sources span 23 distinct primary price hosts; 17 of the 40 providers rest on a single source host, and those are flagged rather than hidden. Spot, committed, and on-demand rates are distinguished throughout; “effective” rates add a standardized 1 TB/month egress scenario.
best is plain on-demandbest is spot / preemptible (interruptible)best requires a commitmentsticker price▲ warning: single source, mismatched price domain, or non-like-for-like discount
▲ Verda’s $0.80 spot rate is the lowest captured, but its price URL sits on DataCrunch’s domain, not Verda’s — treat it as a low-confidence anomaly, not the winner. datacrunch.io
Among hyperscalers, Azure spot is the outlier at $1.403/GPU-hour (80% off its $6.98 NVL sticker); OCI preemptible is $3.00, AWS spot $3.83, and Google Cloud spot $7.004 against an $11.675 sticker. sepahsalar.org
AWS’s $5.191 “sticker” is an EC2 Capacity Block effective rate, not ordinary on-demand — its true discount versus on-demand is not like-for-like with the other rows. intuitionlabs.ai
Egress moves the ranking: at 1 TB/month a $0.09–$0.12/GB hyperscaler fee adds roughly $0.12–$0.16/GPU-hour, while specialists such as CoreWeave, Lambda, Vast.ai, and TensorDock advertise zero egress; OCI’s first-10-TB-free allowance makes its effective rate $3.000, not $3.012. cloudpricecalculators.com
RunPod’s captured “spot” quote ($2.61) is above its $1.99 PCIe sticker because it covers a different tier — marked non-like-for-like rather than treated as a discount. theplanettools.ai
How to read this
Effective $/GPU-hour = best captured compute rate + (paid egress $/GB × 1,000 GB ÷ 730 GPU-hours): one continuously used GPU exporting 1 TB/month. It is a comparison scenario, not a universal bill. A blank effective value means no defensible egress rate was captured — missing egress is never silently treated as zero; an explicit $0 egress makes effective equal compute. Spot/preemptible capacity can be interrupted and is not guaranteed. H100 PCIe 80GB, SXM/SXM5 80GB, and NVL 94GB are not identical parts. “Availability” is the latest stated page status/region, not a live inventory feed; blank means unresolved, not unavailable. Taxes, storage, add-ons, support, and inter-zone traffic are excluded unless bundled in the instance price.
All 40 providers
Provider
Sticker
Spot
Committed
Term
Best
Egress $/GB
Free egress
Effective @1TB
Variant
GPUs
Availability
Region
Billing
Hosts
Sources
Data: 40 GPU-cloud H100 offerings captured 2026-09-06 from 23 distinct primary price hosts; the underlying search scanned 1,461 comparison pages and 2,709 official-site results, yielding 4,183 broad and 812 official candidate rows before normalization. Values are US dollars per GPU-hour, normalized from instance prices (often 8-GPU nodes). Effective rate assumes 1 TB/month egress over 730 GPU-hours. Blank cells mean the value was not captured. Some long-tail providers are single-source. Availability text is trimmed for space.