Make the unit explicit.
A rate per GPU-hour is different from a rate per instance-hour. One instance may contain multiple GPUs, CPU, RAM and disk. Compare equivalent configurations and avoid multiplying the GPU count twice when the quoted price already covers the whole instance.
Count billable time, not just busy time.
Your service may bill while capacity is provisioned, during startup, or when waiting for requests. Scale-to-zero and stopping rules vary. Check the official billing conditions for the exact product rather than assuming no requests means no cost.
Keep unknown extras visible.
Storage, data transfer, networking, support and tax may sit outside the compute rate. Where an amount is unknown, label it unknown. A known compute subtotal can still help you compare scenarios, but it is not a complete quote.