The cheapest H100s rent for $1.60–$1.79 an hour, while Oracle's list price sits at $10 — a 6× spread for the same silicon
Thirteen cloud providers compared on the price of one NVIDIA H100 GPU-hour in USD, normalized per GPU where the instance's GPU count is known. Research date 28 August 2026; source dates range from October 2025 to August 2026, drawn from 50 provider and analyst pages.
These floors are not equivalent products. H100 PCIe versus SXM, single-GPU versus 8-GPU instances, included CPU and RAM, regions, minimum cluster sizes, and interruption or commitment terms all differ between the quoted rates. Spot prices are interruptible; committed prices trade away cash flexibility and are not directly comparable with on-demand. Current availability is a point-in-time signal, not a capacity guarantee.
Real effective price is defined as:
effective $/H100-hour = compute charge per H100-hour + (billable public-internet egress GB generated that hour, after free allowance) × egress $/GB
Verified per-GB internet egress prices were not consistently recovered for any provider in this evidence set, so the primary comparison below is the lowest compute-only floor. A missing egress fee is not treated as zero: where the fee is unverified, the all-in effective cost is left blank rather than shown as equal to the compute rate.
Sticker price vs lowest qualified compute floor, USD per H100 GPU-hour · sorted by lowest floor · hover a row for detail
floor is spot / interruptible
floor is committed / reserved
floor is plain on-demand (no cheaper tier found)
sticker on-demand price
Among hyperscalers the extracted floors were Google Cloud spot $2.10 and Microsoft Azure spot $2.46 — Azure showing the widest sticker-to-floor gap, $6.98 down to $2.46 (
tech-insider.org,
cantech.in) — followed by AWS committed/reserved $2.97 (
aiwiki.ai) and Oracle Cloud at a flat $10.00 with no cheaper tier extracted (
blogs.oracle.com).
Egress is the unresolved half of the effective price: only Lambda's extraction states free/zero internet egress (
spheron.network), while for CoreWeave "$0.10" appears in an egress-note field but not in the verified numeric-fee field, so it is not counted as a charge (
deploybase.ai). All other providers' internet egress fees are unverified in this dataset.
Availability signals were extracted for only six providers: AWS capacity blocks across listed US, Asia Pacific, Australia, Europe and South America regions (
aiwiki.ai), CoreWeave strong capacity across North America and Europe (
gridstackhub.ai), Google a supported zone with stock unverified (
tech-insider.org), Nebius EU plus US (
aiwiki.ai), Hyperstack EU/UK (
cantech.in) and Vultr fluctuating supply (
saastweaks.com). The rest are unknown — a blank is not "available".
All 13 providers, row by row
Method: 13 normalized provider rows form the primary comparison, cross-checked against 62 provider/product-level price rows and 193 official-provider documentation extractions; research date 2026-08-28, source publication dates 2025-10 to 2026-08. Each number is USD per one NVIDIA H100 GPU-hour, normalized by GPU count where the instance's count is known. Excluded unless embedded in the quoted rate: taxes, attached storage, idle VM time, CPU/RAM differences, cross-zone transfer, and minimum-billing or commitment-breakage costs. Precomputed 1/10/100 GB egress scenario columns are not shown as all-in totals because per-GB internet egress fees are unverified for every provider; those scenarios equal the compute floor and are labelled egress-unverified instead. Product-level evidence rows were condensed for space; unresolved fields are left blank rather than dropped.