Reserved vs on-demand GPUs: calculate the utilisation break-even
A lower committed hourly rate only helps when the workload uses enough of the capacity you pay for.
· Arvica Cloud · Analysis & buying guide
Start with comparable offers
Use offers for the same complete configuration, billing currency and cost boundary. Include storage, transfer, support and tax consistently. Confirm whether a reservation is a commitment to pay, a capacity guarantee, or both. A discount without the capacity you need may not solve the operational problem; a capacity guarantee without flexibility may create a different budget risk.
An explicit hypothetical calculation
Assume a 30-day period of 720 hours. A hypothetical reserved rate of A$3 per GPU-hour charged for all 720 hours costs A$2,160 per GPU. An otherwise equivalent on-demand rate of A$5 breaks even at 432 billable hours: 2,160 ÷ 5. That is 60% of the period. These invented rates demonstrate the formula; they are not Arvica prices or a current market survey.
What the simple formula leaves out
Add setup and migration costs, unused capacity between jobs and any minimum commitment. Check whether on-demand resources are actually obtainable when needed, and whether stopped instances keep storage charges. If a reservation has a fixed monthly fee plus usage charges, use that actual tariff instead of the simplified hourly model. Run a conservative, expected and high-usage case; avoid making the commitment depend on an optimistic forecast alone.
Arvica analysis: reserve a dependable baseline
A useful starting point is to consider a reservation for a predictable baseline and separately quote peak demand, if the contract and application permit that split. Record who can approve expansion, the notice period and how cancellation or renewal works. Send Arvica your GPU specification, expected billable hours, peak requirement and term to compare concrete options. The lowest hourly number is only one part of the decision.
Original procurement analysis using the explicitly stated arithmetic assumptions.
Apply this to your workload
Further reading
How much GPU memory does a 70B model really need? →
Your GPU is busy waiting: investigate the pipeline before renting more →
Eight GPUs are not automatically eight times faster →
Before the long reservation: a GPU server acceptance checklist →
AI inference economics: why the GPU-hour is only the starting point →
B300 and Rubin: plan the workload before chasing the roadmap →
AI data-centre power demand: what compute buyers should ask →