Some links in this article are affiliate links. We earn a commission at no extra cost to you when you purchase through them. Full disclosure.
RunPod is a GPU-cloud candidate for development, training and inference. A headline hourly price is not proof of the cheapest usable configuration, and a public listing is not a capacity reservation.
This research-based update checks current first-party pricing and billing documentation on October 2, 2026. We have not measured uptime, setup speed or workload throughput. Earlier promotional, cheapest and guaranteed-availability claims are removed.
Disclosure: We may earn a commission if you sign up through links on this page.
Pods: quote the specific hardware
RunPod’s current public pricing lists the following Pod configurations:
| Listing observed | GPU memory | System memory | vCPUs | Listed hourly compute |
|---|---|---|---|---|
| A100 PCIe | 80GB | 117GB | 8 | $1.59 |
| A100 SXM | 80GB | 125GB | 16 | $1.59 |
| H100 SXM | 80GB | 125GB | 20 | $3.49 |
These are dated public examples, not a promise of stock, regional availability or the exact checkout quote. PCIe and SXM variants are not interchangeable benchmarks. Record GPU count, cloud/location, CPU/RAM allocation, storage, networking and contract mode before comparing providers.
The old $0.19 entry headline and A100/H100 matrix are not retained as current offers. Promotional credits do not reduce the underlying list rate and are not included in these examples.
Pods versus Serverless
Pod billing separates compute and storage lifecycle. Check charges while running, stopped and after termination; deleting a compute resource is not a substitute for reviewing retained storage.
Serverless pricing is a different worker-based billing model. GPU type, worker configuration and active/flex settings affect cost. A Pod hourly rate is not a Serverless request price.
Scale-to-zero requires the relevant worker configuration; active workers or retained storage can still cost money. Do not describe an idle endpoint as unconditionally free.
Match the configuration to the job
Training or fine-tuning: estimate model/optimizer memory, precision, batch size, dataset size and runtime. No universal $5–$20 job budget is supported.
Model inference: benchmark representative input/output lengths, concurrency and cold/warm startup. An 80GB GPU does not guarantee that every 70B model configuration fits.
Image workflows: check template compatibility, model licenses, output storage and necessary dependencies. A prebuilt template is a starting point, not an instant-setup guarantee.
Choose a Pod when the controlled instance workflow fits; evaluate Serverless when managed endpoint/worker scaling is the required job. That distinction is an editorial decision framework, not a performance result.
Compare RunPod Pod and Serverless options for your GPU workload using the current console quote and billing controls.
A safe evaluation sequence
- Select the exact model, license and workload.
- Confirm VRAM, host resources and location/availability.
- Price compute, storage, startup and any idle/active workers separately.
- Run a bounded representative test with a spending limit.
- Verify output, latency, persistence and teardown behavior before scaling.
We do not promise a minimum deposit, promotional credit, two-minute deployment, guaranteed stock or blanket uptime. Contractual availability/SLA terms must be checked for the product actually purchased.
Related: GPU provider comparison, vLLM serving, Ollama Docker setup and serving-framework comparison.