The question: what is the real effective price per hour for an NVIDIA H100 across GPU cloud providers, counting sticker price, spot or committed discounts, egress fees, and current availability? This scan, dated 2026-08-31, normalizes 18 providers to a lowest published compute-only $/GPU-hour, backed by 384 source-derived evidence rows. Specialist clouds and marketplaces cluster at $1.80–$3.90; hyperscaler list rates here run $10.34–$12.29. Rates are compute-only / egress unresolved unless an egress fee was explicitly published, and the rows mix H100 PCIe, SXM/HGX and 8-GPU bundles — so this is a shortlist, not a ranking.
The five cheapest published compute-only rates — Vast.ai $1.80, CUDO Compute $1.82, Hyperstack $1.90, TensorDock $2.09, Genesis Cloud $2.19 — are not the same H100 form factor, region or bundle, so the 6.8× gap to AWS/Azure is not a like-for-like verdict. intuitionlabs.ai
Discounts are real but conditional: evidence shows CoreWeave spot near $2.44 and Azure 3-year-reserved near $4.13 per GPU-hour — spot carries eviction risk, reserved needs multi-year terms, and CoreWeave sells H100 as 8-GPU nodes only. spheron.network · jimmyresearch.com
Egress is the silent tax: AWS internet egress at $0.09/GB adds about $0.90 per GPU-hour at 10 GB/hour, while Crusoe, Fluidstack, Lambda, RunPod and CoreWeave show $0.00/GB in evidence rows; most other providers left egress unresolved in this scan. io-net.ghost.io
Availability diverges as much as price: AWS H100 capacity shows months-long waitlists and Lambda is reported out of stock regularly, while marketplace listings (Vast.ai 86%, RunPod 88%, CUDO 100% over 7 days) were live but can disappear without notice. oxmaint.ai · gpufinder.dev
effective $/GPU-hour = compute $/GPU-hour + (internet egress GB per GPU-hour × $/GB) + separately billed storage / CPU / networking
The 10-GB-per-hour scenario is fully loaded only where the egress fee is explicitly published; everywhere else the numbers on this page are compute-only. Sensitivity: every $0.01/GB of egress adds $0.10 per GPU-hour at 10 GB/hour, and $1.00 at 100 GB/hour — enough to erase most of the specialist-cloud discount for data-heavy inference. Spot/preemptible rates carry eviction risk, marketplace capacity can vanish, and reserved rates can require long terms or large minimum clusters.
| Provider | H100 variant | Sticker $/GPU-h | Spot / commit $/GPU-h | Egress $/GB | Effective read | Availability | Caveat | Source |
|---|
Method: normalized shortlist of 18 GPU cloud providers, one row each, dated 2026-08-31, from a scan of thousands of web results reduced to 384 extraction rows (up to 20 per provider). Each bar is the lowest published compute-only H100 rate in USD per GPU-hour; spot, committed, egress and availability fields are shown only where a source row states them explicitly — blank means no verifiable public rate found in scan, never zero. Prices mix PCIe and SXM/HGX form factors, regions and 8-GPU minimums; live market rates change. Oracle Cloud and Latitude.sh appear in the evidence set but not in the normalized shortlist and were cut for space.