“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.”
Sixteen H100 offers across 14 providers, compiled from 38 provider-page records plus 2,366 hyperscaler pricing records, as found on 2026-09-03. All prices are USD per H100 GPU-hour; 8-GPU node rates are divided by eight, and PCIe vs SXM/HGX variants are labeled because they are not equivalent hardware.
| Provider | Variant / shape | Sticker $/GPU-hr | Best discount | Egress $/GB | Effective @10 GB/hr | @100 GB/hr | Availability signal | Source |
|---|
Effective hourly price = compute $/GPU-hour + marginal internet-egress $/GB × outbound GB per GPU-hour (0, 10 or 100 GB). This is a usage model, not a claim that every workload exports that much data. Free allowances were applied only where a source states one; where no public egress fee was found the cell stays blank and the effective figure is marked compute-only / egress not verified — unknown egress is never treated as zero. Likewise, blank discounts mean no publicly documented rate was found, not that none exists.
Availability here is the freshest explicit public capacity or access signal found on 2026-09-03 — instant/self-service, first-come, spot/interruptible, reservation, or sales-contact — not a live launch test. Actual inventory is region-, account- and quota-dependent and changes minute to minute.
Normalization. Node rates are divided by stated GPU count (e.g. CoreWeave 8× HGX H100 $49.24/node = $6.155/GPU-hr; AWS p5.48xlarge $55.04/node = $6.88). PCIe and SXM/HGX rows are kept separate. Rows sourced only from independent trackers rather than official provider pages are marked secondary.
Data: 38 provider-page price records and 2,366 structured hyperscaler pricing records collected up to 2026-09-03, plus one aggregated sanity-check row. Prices in USD per H100 GPU-hour; instance rates normalized by stated GPU count. Cut for space: serverless per-second endpoints, non-H100 GPUs, decentralized marketplaces, regional price variants beyond the cheapest cited, and marketing pages with no explicit rate.