Discount capacity makes H100 compute cheap; egress and access decide whether it is real
Across ten providers, the lowest observed effective rate is Vast.ai at $0.44 per GPU-hour, but it is marketplace capacity and therefore variable. RunPod is the lowest clearer-availability option at $0.68; hyperscaler sticker rates become much less competitive once 10 GB of hourly internet egress is included.
Best purchasable rate versus sticker rate
Effective rate = compute price + (egress price per GB × 10 GB/hour). Spot prices are observed floors, not guaranteed capacity. Committed prices require their stated term.
What changes the ranking
- Marketplace floors are cheapest but not fungible. Vast.ai’s $0.44 result comes with variable inventory, host quality and location; it should not be treated like contracted cluster capacity.
- Zero-egress specialists preserve their compute advantage. RunPod, CoreWeave, Crusoe and Lambda remain close to their compute quote under this transfer scenario.
- Hyperscaler spot can be cheap but quota and interruption risk are material. AWS, Azure and Google show large discounts, yet H100 capacity is region-, quota- and interruption-dependent.
- Workload data movement matters. At 10 GB/hour, $0.09–$0.12/GB adds roughly $0.90–$1.20 to every GPU-hour; at zero egress, subtract that amount.
Provider detail
| Provider | Compared GPU | Sticker | Best compute | Discount | Egress / GB | Effective | Availability signal |
|---|---|---|---|---|---|---|---|
| Vast.ai | H100 SXM | $1.87 | $0.34 | Spot | $0.010 | $0.44 | Variable marketplace inventory |
| RunPod | H100 SXM | $2.69 | $0.68 | Spot | $0 | $0.68 | Available; US / EU signal |
| CoreWeave | H100 | $4.76 | $1.03 | Spot | $0 | $1.03 | US signal; access can require approval |
| Crusoe | H100 80GB | $3.90 | $1.39 | 1-year reserved | $0 | $1.39 | US / Texas signal |
| Nebius | H100 | $3.45 | $1.25 | Preemptible | $0.015 | $1.40 | EU / Finland signal |
| Lambda | H100 SXM | $2.99 | $1.49 | Reserved | $0 | $1.49 | US signal |
| Google Cloud | H100 | $9.00 | $0.30 | Spot floor | $0.120 | $1.50 | us-central1 signal; quota / capacity constrained |
| AWS | H100 | $6.88 | $1.14 | Spot floor | $0.090 | $2.04 | us-east-1 signal; capacity and quota apply |
| Microsoft Azure | H100 | $8.25 | $1.23 | Spot floor | $0.087 | $2.10 | East US signal; quota gated |
| DigitalOcean | H100 80GB | $5.49 | $2.24 | Reserved | $0.010 | $2.34 | GPU regions advertised; exact inventory unresolved |
Method: live-web scan on 20 Aug 2026 across provider pages, pricing comparisons and availability references; 10 normalized provider rows assembled from hundreds of source mentions. Prices are per single H100 GPU-hour. Multi-GPU instance prices were normalized by GPU count where stated. “Best” uses the lowest observed spot, preemptible, reserved or committed quote; it is not a promise that capacity can be acquired at that instant. Taxes, storage, CPU/RAM overages, inter-zone transfer and interruption/restart overhead are excluded.