Marketplace H100s reach $1.35–$2.04 per hour; dependable capacity costs much more

Ten providers with public H100 prices survived the scan. At a normalized 1 GB of paid internet egress per GPU-hour, Vast.ai is the cheapest published market offer at $1.35; Google preemptible is nearly tied at $1.36 but can be interrupted, while hyperscaler on-demand offers run roughly $4.25–$10.00.

$1.35
lowest effective $/GPU-hour
$2.04
RunPod effective $/GPU-hour
10
providers compared
1 GB
egress normalized per hour

Best public rate after 1 GB of egress

fixed/preemptiblemarketplace
Vast.ai
$1.35
Google Cloud
$1.36
RunPod
$2.04
Nebius
$2.17
Lambda
$2.49
Paperspace
$2.73*
CoreWeave
$6.16
AWS
$6.97
Azure
$7.07
Oracle
$10.00
$0$2.50$5.00$7.50$10 / GPU-h
Best public rate means spot, preemptible, marketplace, or committed price where found. *Paperspace excludes unresolved egress. Free allowances can reduce effective cost below the normalization shown.

What changes the buying decision

01

The two cheapest prices are not equivalent to reserved capacity. Vast.ai is a host marketplace, while Google’s $1.24 compute rate is preemptible and explicitly carries interruption risk. gridstackhub.ai · deploybase.ai

02

Egress matters most on cheap compute. One paid GB adds $0.12 on Google and $0.05 on RunPod, versus $0.09 on AWS; Lambda and CoreWeave are reported free. gpufinder.dev · computecomparison.com

03

Minimum topology can dominate the invoice. AWS P5, CoreWeave HGX and Oracle bare metal are commonly sold as eight-GPU systems; the table divides node price by eight, but the cash burn is still eight times the per-GPU figure. spheron.network · deploybase.ai

04

“Available” is not the same as continuously obtainable. AWS is described as limited/Capacity Blocks, Paperspace as constrained, Nebius as select-region, and marketplace stock varies by host. io.net · gpucloudcost.com

Price, discount, egress and availability

ProviderSticker
$/GPU-h
Best public
$/GPU-h
Egress
$/GB
Effective
at 1 GB/h
Availability and conditionsSource
Vast.ai$1.89 market top$1.35 spot
($1.14 PCIe floor)
$0 published$1.35Live marketplace; host quality, quantity and continuity varygridstackhub.ai
Google Cloud$4.13 on-demand$1.24 preemptible$0.12 Premium$1.36Preemptible interruption/capacity risk; on-demand is quota and region dependentdeploybase.ai
RunPod$1.99$1.99 community$0.05$2.04Marketplace listing verified; stock varies by hostgridstackhub.ai
Nebius$3.85 HGX$2.15 preemptible NVLink$0.015 object storage; compute reported free$2.17 conservativeAvailable in select regions; preemptible supply can disappeardocs.nebius.com
Lambda$2.49 PCIeFree$2.49Offer published; no live stock signal exposed in scanspheron.network
Paperspace$5.95$1,994/mo ≈ $2.73/hUnresolved$2.73*Capacity constrained; dedicated “reserve now” path advertisedgetpulsesignal.com
CoreWeave$49.24 / 8 = $6.16Private contractFree reported$6.16Published 8-GPU offer; exact regional stock not exposeddeploybase.ai
AWS$55.04 / 8 = $6.88Private / Capacity Blocks$0.09 after 100 GB$6.97Limited; Capacity Blocks often requiredspheron.network
Azure$6.98Spot unresolved$0.087 after 100 GB$7.07Quota and region dependent; no live stock signal foundmedium.com
Oracle Cloud$10.00Commit unresolvedFirst 10 TB free$10.00Eight-GPU bare metal; sales, quota and region dependentmedium.com
Method — checked 26 August 2026. The research scan covered 3,220 unique search results; this report cites 12 pricing, egress and availability pages across 11 sites. Prices are normalized to one H100 GPU-hour; node prices are divided by GPU count. “Effective” = best public compute rate + 1 GB of paid internet egress, after free allowances are exhausted; it excludes CPU/RAM/storage premiums, taxes, interruption/restart cost and negotiated enterprise discounts. H100 PCIe, SXM/HGX and NVLink offers are not performance-identical.
Keenable · made with SELECT* · 3,220 pages in 4m 33s · Open the chatShare: XLinkedInReddit