Cheap H100 compute can become expensive once data leaves the cloud
Across 11 providers, the lowest observed compute offer is Vast.ai at $0.50 per GPU-hour, while the lowest fixed-price committed offer is Vultr at $1.68. The effective rate is workload-specific: compute rate + internet egress GB per hour × egress fee.
$0.50
lowest observed GPU-hour
$1.68
lowest fixed commitment
$0–0.12
egress fee per GB
11
providers compared
Comparable scenario: bars show the best public compute rate plus the marginal charge for 100 GB of internet egress during that GPU-hour. Free allowances are not amortized; taxes, storage and CPU/RAM differences are excluded.
Effective USD per H100-hour at 100 GB outbound per hour
blue = purchasable/variable · grey = unavailable or constrained
CoreWeave
$2.23
Vast.ai
$2.50
Vultr
$2.68
Lambda
$3.99
Nebius
$4.15
Crusoe
$5.50+
Oracle
$10.85
RunPod
$11.89
AWS
$12.93
Azure
$20.99
Google Cloud
$23.06
What changes the ranking
CoreWeave’s $2.23 contract price is not an on-demand bargain. Its documentation says the H100 fleet is fully committed and unavailable for on-demand use. coreweave.com
High-egress workloads favor providers with included transfer. Lambda lists $0 egress, whereas Google’s first paid tier is about $0.12/GB; at 100 GB/hour that is a $12/hour difference. lambda.ai
Marketplace lows are volatile. Vast.ai’s observed $0.50 rate is a host asking price, not reserved capacity or a uniform SLA-backed fleet. vast.ai
Hyperscaler access remains a capacity problem. AWS offers reservable Capacity Blocks; Google explicitly describes limited regions/capacity; Azure quota and regional supply vary. aws.amazon.com
Rates and availability
| Provider | Sticker $/GPU-h | Best public spot/commit | Egress $/GB | Current availability | Source |
|---|---|---|---|---|---|
| Vast.ai | 0.50 | 0.50 marketplace | 0.02 | Variable host stock, 30+ data centers | vast.ai |
| Vultr | 1.99 | 1.68, 36 mo prepaid | 0.01 | Listed globally; clusters may need allocation | vultr.com |
| RunPod | 2.89 | — | 0.09 | 31 regions; live stock varies | runpod.io |
| Nebius | 3.85 | 2.15 preemptible | 0.02 | Self-service; Europe/US | nebius.com |
| Lambda | 3.99 | — | 0.00 | Self-serve listing; stock varies | lambda.ai |
| Crusoe | 5.50* | — | — | Available now; VM pricing often quoted | crusoe.ai |
| CoreWeave | 6.16 | 2.23, 2 yr | 0.00 | Documented fully committed; no on-demand | coreweave.com |
| AWS | 6.88 | 3.93 Capacity Block | 0.09 | Multi-region GA; quota/capacity constrained | aws.amazon.com |
| Oracle | 10.00 | — | 0.0085 | GA in many regions; capacity request common | oracle.com |
| Google Cloud | 11.06 | — | 0.12 | Limited regions and capacity | cloud.google.com |
| Microsoft Azure | 12.29 | — | 0.087 | Regional; quota and capacity vary | microsoft.com |
Method: 3,000+ search results across official and comparison sources scanned on 23 Aug 2026; 11 providers retained. Prices normalized to one H100 80GB GPU-hour by dividing multi-GPU node rates. “Effective” chart uses the best public compute rate plus 100 GB/h marginal internet egress. *Crusoe’s $5.50 figure is a self-serve deployment rate; general H100 VM pricing is quote-led. Free monthly egress tiers, storage, CPU/RAM, tax and negotiated enterprise deals are excluded.