Pricing Comparison
Per-GPU-hour list pricing varies significantly across the cast. Apples-to-apples comparisons require fairness adjustments for reliability, fabric, support, and term.
Per-GPU-hour list pricing (on-demand H100 80GB, illustrative)
| Company | Indicative H100 on-demand |
|---|---|
| Hyperscalers (AWS/Azure/GCP) | $8-12/hour |
| CoreWeave | $3.5-5/hour (heavily discounted on reserved) |
| Crusoe | $3-5/hour |
| Nebius | $3-5/hour |
| Lambda | $2.5-3.2/hour |
| Together.AI (dedicated H100) | Comparable to Lambda |
| RunPod Secure Cloud | $3.5-4.5/hour |
| RunPod Community | $2.5-3.5/hour |
| Hyperbolic marketplace | $2-3/hour |
| TensorDock | $2.3-3/hour |
| Vast.AI on-demand | $2-3/hour |
| Vast.AI interruptible | $1.2-2/hour |
Fairness adjustments
Direct price comparison is misleading because the products differ:
- Reliability variance: Marketplace average is meaningfully lower than dedicated.
- Fabric: Cluster fabric matters for multi-node workloads, available only on dedicated providers.
- Support and SLAs: Bundled in dedicated; absent in most marketplace.
- Compliance: SOC 2 etc. available at dedicated providers; not across marketplace supply.
- Networking, storage: Different ecosystems.
For an apples-to-apples cost, factor these in. A "cheap" Vast instance that requires 1.3x the wall-clock time due to slower DLPerf may not be cheaper end-to-end than a dedicated H100.
Reserved pricing
Reserved discounts off on-demand:
- CoreWeave / Crusoe / Nebius: deepest discounts for multi-year strategic commitments.
- Lambda: 35-45% off for annual.
- RunPod: 30-40% off for annual.
- Marketplaces: typically no formal reserved tier; some providers offer monthly rates.
Takeaway
Headline pricing varies 4-6x across the cast. After fairness adjustments, the spread narrows but real differences persist. The next chapter looks at customer segments.