Business Model Comparison
Five distinct business models populate the cast. Reading them side by side clarifies the strategic differences that the company-by-company guides addressed.
The five business model axes
Recapping from the History & Future taxonomy chapter:
- Marketplace. Aggregates third-party supply; takes percentage.
- On-demand dedicated. Owns hardware; rents hourly.
- Reserved capacity. Owns hardware; multi-year customer commitments.
- Managed inference. Per-token API on curated models.
- Energy arbitrage. Owns hardware located at cheap energy sources.
Some companies span multiple axes (RunPod, Hyperbolic, Lambda); some are pure (CoreWeave is dominantly reserved-capacity; Together is dominantly managed-inference).
Mapping the cast
| Company | Primary model | Secondary |
|---|---|---|
| CoreWeave | Reserved | Some on-demand |
| Crusoe | Energy arbitrage / Reserved | Stargate-tied builds |
| Nebius | Reserved | Managed services, inference |
| Lambda | On-demand | Reserved, hardware sales, 1-Click Cluster |
| Together.AI | Managed inference | Fine-tuning, training clusters |
| RunPod | Marketplace + on-demand | Serverless |
| Hyperbolic | Marketplace + managed inference | — |
| Vast.AI | Marketplace | — |
| TensorDock | Marketplace (curated) | — |
Capital intensity
Capital required to scale, ranked roughly:
- CoreWeave (highest — multi-billion-dollar cap-ex).
- Crusoe (very high — energy + GPU infrastructure).
- Nebius (very high — datacenter buildouts + GPU procurement).
- Lambda (high — dedicated cloud infrastructure).
- Together.AI (moderate — hybrid owned/partner).
- RunPod (moderate — Secure Cloud requires capital).
- Hyperbolic (low-moderate).
- Vast.AI (low — marketplace, no owned hardware).
- TensorDock (low — marketplace).
How each scales
- Reserved-capacity players scale by signing more customer commitments.
- Energy-arbitrage players scale by securing more cheap-power sites.
- On-demand players scale by adding capacity that demand fills.
- Managed-inference players scale by growing customer traffic.
- Marketplaces scale by recruiting more providers.
Cyclical exposure
Susceptibility to AI demand cycles, roughly:
- Most exposed: Capital-intensive on-demand and reserved-capacity players (CoreWeave, Crusoe, Nebius, Lambda). Heavy fleet that needs utilization.
- Moderately exposed: Managed inference (Together, Hyperbolic inference). Per-token revenue scales with use but operating cost has fixed component.
- Least exposed: Marketplaces (Vast, TensorDock, RunPod Community). Asset-light; costs scale with revenue.
Takeaway
The business model is the most important single dimension for understanding any neocloud. Capital structure, scaling dynamics, and risk profile all follow from it. The next chapter looks at scale and financial shape.