Five ways to pay less
- Reserve what runs all month. If a GPU is busy more than about 70% of the time, a reserved term is cheaper than on demand.
- Pay per minute for the rest. Experiments and short fine-tunes should not pay for idle hours.
- Right-size the GPU. Models up to 70B in 4-bit run on one RTX PRO 6000; you may not need an H100 at all.
- Fix utilization first. Batching and a proper serving engine often double throughput per GPU. See getting more inference out of every GPU.
- Use marketplaces only for restartable jobs. Cheap hours lose their value if a host goes offline mid-run.

