Start with the billing unit
Four GPUs running for three hours consume 12 GPU-hours. That says nothing about utilization: the GPUs could be busy throughout, or spend half the session waiting for a dataset. Check whether a displayed price covers one GPU or an entire instance with several GPUs.
Lambda, for example, documents hourly instance prices billed in one-minute increments, including time when a running instance is idle. Other products use different billing boundaries. Save the billing terms alongside the quote so the team estimating the job uses the same assumptions as the person paying for it.
References: Lambda billing
A worked budget for a small experiment
Consider an illustrative four-GPU experiment at $2 per GPU-hour. Six hours of training costs $48. One hour of setup adds $8; repeating two hours after a failed checkpoint adds $16. Add an assumed $4 for storage and transfers and the run costs $76.
Those are planning assumptions, not provider quotes. The calculation exposes a useful question: would spending an hour fixing the checkpoint process save more than finding a slightly cheaper rental?
| Item | Assumption | Cost |
|---|---|---|
| Training | 4 GPUs × 6 hours × $2 | $48 |
| Setup | 4 GPUs × 1 hour × $2 | $8 |
| Repeated work | 4 GPUs × 2 hours × $2 | $16 |
| Storage and transfer | Illustrative allowance | $4 |
| Total | One completed experiment | $76 |
Use an output unit your team recognizes
For embeddings, measure cost per million documents that pass validation. For image generation, use accepted images at a fixed resolution and sampling setup. For interactive inference, include the latency target: cheap tokens that arrive after the user has left are not equivalent to timely answers.
Record compute time, setup time and failed attempts separately. A spreadsheet with those columns makes the next estimate more useful than a single total from the previous invoice.
Before you approve the rental
Confirm the GPU count, currency, billing increment, minimum commitment and stop-versus-delete behavior. Record which storage survives termination and which resources continue billing after compute stops. Add a budget alert and an owner responsible for cleanup.
Price two scenarios: an ordinary run and a run with one restart. If the second exceeds the budget, reduce the experiment size or improve recovery before adding a larger GPU allocation.
Questions readers ask
Does a GPU-hour mean an hour of active computation?
Usually it describes allocated or billable time. The exact billing definition belongs to the product, so idle time can still cost money.
How do I compare two hourly quotes?
Match GPU count and variant, then compare the total cost of the same completed workload, including storage, transfers and failures.
Sources and editorial notes
Primary documentation reviewed October 9, 2026. Provider descriptions reflect published documentation. Recommendations are our analysis; worked figures are illustrative unless explicitly stated otherwise. No independent provider benchmark or uptime ranking is claimed.
Read the editorial policy. NeuraMesh is a prototype, not live verified GPU inventory.