The lowest hourly rate is only useful if the machine can finish your job. Use this checklist to compare two actual offers, then run a small test before committing more money or sensitive data.
Before checkout
- Save the exact offer. Record GPU model, GPU count, memory per GPU, CPU/RAM allocation, disk capacity, region, rental type, hourly price and capture time. Confirm whether the price is per GPU or per machine.
- Price the whole job. Include installation and model download time, storage, transfers, idle time, retries and operator time. Check minimum funding, billing increments, taxes and cancellation terms in checkout.
- Check the software requirements. Match the model’s license and serving engine requirements to the GPU architecture, driver, CUDA version and available memory. Do not infer fit from parameter count alone.
- Set a test budget and stop condition. Decide how much you will spend and what constitutes failure before creating the instance.
Stopping is not always the end of the bill
As checked on November 17, 2026:
| Provider documentation | What matters for the bill |
|---|---|
| Runpod Pod pricing | Compute and storage are separate. Its pricing table distinguishes running and stopped storage; volume disks and network volumes can remain chargeable when compute stops. Pods currently have no ingress/egress fee. |
| Vast instance pricing | Charges include compute, storage and bandwidth. Storage continues while the instance exists, including when stopped. Transfer rates vary by host. |
Check your particular product and quote again before purchase. These Pod/instance rules do not establish the terms of every serverless or other product from the same company.
Write down which disks are temporary, which survive a stop, what deletion destroys and which resources must be removed to end charges. Back up anything you need before deleting it, then inspect the billing/resource list afterward.
Estimate setup time, then measure it
For a transfer of 120 decimal GB at an assumed sustained 0.5 Gbit/s:
Ideal transfer seconds = 120 × 8 / 0.5 = 1,920
Ideal transfer minutes = 1,920 / 60 = 32
This is a unit-conversion example, not an observed download speed. Protocol overhead, the source, disk writes and unpacking add time. A network specification is not a promise that your model download will achieve it.
Test with non-sensitive data
Use a small representative workload first. Record:
| Check | Evidence to keep |
|---|---|
| Model loads and remains stable | Exact model revision, engine version, quantization, peak memory and errors |
| Requests meet your target | Input/output lengths, concurrency, completed requests and latency distribution |
| Enough useful work per billed hour | Setup time, completed work, aggregate throughput and total cost |
| Restart/recovery works | Recovery steps, time required and verified backup |
| Spending ends as intended | Final resource list and billing check |
If the offer can be interrupted, test whether the job can resume and include lost work in the estimate. Do not assume a cheap offer or a high marketplace score guarantees availability.
Runpod specifically documents that a stopped Pod releases its GPU and may not have a GPU available on restart. That makes restart availability worth checking even when the stored data survives. Runpod restart troubleshooting
Check access and data requirements separately
A hardware listing is not a security assessment. Before using sensitive data, verify the applicable provider agreement, data location, access controls, retention/deletion terms and any required contractual assurances. Do not infer tenant isolation, disk sanitization or an SLA from a marketplace label.
Use scoped credentials rather than a personal master key; avoid baking secrets into reusable images. Limit exposed ports, authenticate services, and revoke test credentials afterward. If you cannot establish that a service meets your data-handling requirements, do not upload that data.
Keep or reject the offer
Keep it only if the test meets your workload targets and the complete cost is acceptable. Otherwise stop the test, preserve necessary output and remove chargeable resources according to the provider’s lifecycle rules.
Compare offers in the GPU listings, then use the API-versus-rental worksheet to check whether renting beats your API bill. No paid test or provider security audit was performed for this checklist.