The first question most companies ask when buying GPU capacity is "which option is cheapest?" A more useful question is: how is your workload spread across the month?
On-demand: paying for uncertainty
On-demand instances are billed per GPU-hour actually used, with no minimum term. They suit:
- Models still in the experimental stage, with irregular training runs
- Temporary inference peaks around launches or campaigns
- Benchmarking before a longer commitment
On-demand rates are usually the highest, because the provider has to keep idle capacity ready. If a GPU is busy only a few hours a day, it is still the cheapest option overall.
Reserved: trading commitment for rate
Reserved capacity locks in dedicated nodes at a fixed monthly or annual rate. Once a workload is steady — continuous training of a large model, or an inference service with predictable traffic — reserved capacity is typically far cheaper than on-demand.
A simple rule of thumb: if a GPU is working more than half the time in a month, reservation is worth considering.
Wholesale: for resale and large clusters
Wholesale contracts suit companies that need a whole cluster or resell compute to their own customers — cloud providers, systems integrators and channel partners. Terms are longer, rates lower, and contracts usually include delivery schedules and service levels.
Mixing the three
Most mature teams combine them: reserved capacity for the steady baseline, on-demand for peaks, and a wholesale contract once volume justifies it.
Whichever model you choose, Qovane prices in US dollars and delivers under a service contract — no compute "shares" and no promised returns. For a quote, tell us the GPU model, count, region and term.