DEPLOYMENT MODELS
Choose an operating model.
Compute strategy includes responsibility, commitment and flexibility as well as hardware.
| Model | Useful starting context | Questions to resolve | Trade-off |
|---|---|---|---|
| On-demand allocation | Experiments, uncertain demand and occasional bursts. | Provisioning, quota, region availability and effective hourly rate. | Variable spend and capacity constraints can complicate deadlines. |
| Reserved or dedicated cluster | Sustained demand with known scheduling and topology needs. | Term, minimum spend, hardware isolation, fabric, support and recovery obligations. | Unused committed hours can raise the cost of useful work. |
| Hybrid or burst model | An established environment with intermittent demand beyond local capacity. | Identity, data movement, network design, software portability and job handoff. | Transfer delays, licensing and operational complexity can dominate. |
| Owned infrastructure | Predictable workloads with facilities and operations capability. | Capital, power, cooling, staffing, maintenance, depreciation and refresh cycle. | A rental calculator is not an ownership-cost comparison. |
Define a representative acceptance run
Use the same application version, dataset, precision, run length and output criteria when comparing proposals. Record useful throughput, failure behaviour, data loading and all billed charges. Keep any hardware or software differences visible.