lowest effective
Nebius
$1.50 + egress
Preemptible H100 plus $0.015/GB internet egress gives $1.65 at the 10 GB workload. Evidence says eu-north1, self-service and no waiting list.
Discounted H100s cluster near $1.50–$2.25 per GPU-hour; hyperscaler egress erases part of the discount
A live scan covered 1,683 pages and retained nine providers with concrete H100 pricing or capacity evidence. On a workload exporting 10 GB per GPU-hour, Nebius is lowest at about $1.65; AWS, Google Cloud and Azure rise by roughly $0.87–$1.20 per hour before free allowances.
$1.65
lowest effective, 10 GB out / GPU-hour
$1.50
lowest discounted compute only
9
providers compared
1,683
pages in primary-source search set
Best public price plus 10 GB internet egress per GPU-hour, USD
known egress egress unknown
$0$3$6$9/hr
“Best public” mixes spot, preemptible, community-marketplace and term commitments; interruption risk and contract length are therefore not identical. A plus sign means the provider’s egress could not be reduced to one defensible public rate.
The three clearest value cases
lowest committed
Lambda
$1.89
The low rate is tied to a three-year reserved cluster. Egress was reported free; on-demand H100 pricing is materially higher.
lowest flexible market
RunPod
$1.99
Community Cloud H100 PCIe is the floor, versus about $2.69 for a more controlled cloud tier. Capacity is marketplace-dependent despite broad regional coverage.
What changes the invoice
1
Egress matters only when data leaves. At 10 GB per GPU-hour it adds about $0.90 on AWS, $1.20 on Google and $0.87 on Azure; training that checkpoints internally pays none of this marginal internet charge. aws.amazon.com
2
The cheapest rates buy different risks. Nebius is preemptible, RunPod is a third-party community market, CoreWeave is spot, while Lambda’s $1.89 is a three-year reservation—not equivalent service guarantees. coreweave.com
3
Hyperscaler sticker prices remain far above specialist clouds. The scan found about $12.29 per H100-hour on AWS and $11.68 on Google, against $2.95–$6.16 among most specialist providers. tech-insider.org
4
“Available” is not the same as guaranteed capacity. AWS Capacity Blocks can be reserved, Google may require quota, CoreWeave lists zones subject to capacity, and Vast.ai inventory changes with host offers. gpucost.org
Provider detail
| Provider / offer | Sticker $/GPU-h | Best discount $/GPU-h | Discount condition | Egress $/GB | Effective at 10 GB/h | Availability now | Source |
|---|---|---|---|---|---|---|---|
| AWS p5 H100 | 12.29 | 5.19 | Capacity Block | 0.09 | 6.09 | Reservable; regional inventory | aws.amazon.com |
| Google A3 H100 | 11.68 | 7.00 | Spot | 0.12 | 8.20 | Quota approval may be required | tech-insider.org |
| Azure ND H100 v5 | ≈10.00 | 8.00 | Lowest public rate found | 0.087 | 8.87 | East US and South Central US cited | tech-insider.org |
| CoreWeave HGX H100 | 6.16 | 2.23 | Spot | 0 | 2.23 | Several zones; subject to capacity | coreweave.com |
| Crusoe self-serve | 5.50 | — | No public discount found | — | 5.50+ | Self-serve described as available now | gpucost.org |
| Lambda H100 | 2.99 | 1.89 | 3-year reserved cluster | 0 | 1.89 | On-demand first-come; clusters by quote | gpucost.org |
| Nebius H100 SXM5 | 2.95 | 1.50 | Preemptible | 0.015 | 1.65 | Self-service in eu-north1 | thehost.cloud |
| RunPod H100 | 2.69 | 1.99 | Community Cloud PCIe | 0 | 1.99 | Instant where marketplace stock exists | deploybase.ai |
| Vast.ai H100 PCIe | 3.69 | 2.21 | Lowest marketplace offer found | — | 2.21+ | Variable host inventory; no SLA | gpucost.org |
Method: live web scan on 24 Aug 2026; 1,683 unique pages from nine provider-focused searches, consolidated across nine providers. Prices are normalized to one H100 GPU-hour by dividing multi-GPU instances by GPU count. “Effective at 10 GB/h” = best public compute rate + 10 × first paid internet-egress rate, before taxes, storage, support, free-transfer allowances and interruption losses. Blanks mean no single defensible public rate was extracted; availability is a page-level claim, not a capacity guarantee.