Hourly billing looks flexible: you only pay for what you use. Monthly billing looks rigid: you pay for the whole month. Which one is cheaper depends on a single number, your utilisation, the share of the month the GPU is actually busy or reserved for you.
The break-even formula
A month has about 730 hours. If the hourly price is h and the monthly price is m, monthly is cheaper as soon as you keep the GPU more than m ÷ h hours.
Because CoreGPU prices are set at about 75% of the market median (m ≈ 0.75 × 730 × h), the break-even sits at roughly 75% utilisation, about 18 hours a day, against median on-demand prices.
Break-even by GPU
| GPU | Median $/h | CoreGPU / month | Break-even | Per day |
|---|---|---|---|---|
| H100 | 3.72 | $1,999 | 537 h | 17.7 h |
| H200 | 4.94 | $2,699 | 546 h | 18.0 h |
| B300 | 8.98 | $4,899 | 546 h | 17.9 h |
Why real utilisation is higher than it looks
Most teams underestimate how long they hold a GPU. On hourly clouds you rarely release a flagship GPU between experiments, because getting it back is not guaranteed. In practice the meter runs through:
- nights and weekends between training runs or evaluation jobs,
- debugging, profiling and data-loading fixes,
- checkpoint uploads and restarts after a failure,
- inference services that must stay up around the clock.
A server that is held for the month is, from a billing point of view, at 100% utilisation, whatever the GPU is doing.
When hourly is the better choice
Be honest about your workload. Hourly billing wins for short bursts: a two-day experiment, a weekly batch job, or a quick benchmark. If you need a GPU for less than about two thirds of the month and can release it in between, an hourly provider will usually be cheaper.
When monthly wins
- Training runs of a week or more, and iterative fine-tuning.
- Production inference that runs 24/7.
- Teams that keep a dedicated development node.
- Any budget that must be known in advance, such as a grant or a startup runway.
Use our savings calculator with your own numbers, or check monthly prices on the pricing page.