Short answer: RunPod charges by the second for GPU time, with big differences between community cloud (cheaper, less guaranteed) and secure cloud (pricier, enterprise-grade), plus a serverless option that scales to zero. Real cost depends on utilization, not the headline hourly rate.
Not sure which platform fits? Nythral helps teams pick and deploy AI infrastructure. See private AI models.
How RunPod Bills
RunPod meters GPU usage per second. You choose a GPU type and a cloud tier, and pay for the time a pod runs. Serverless endpoints bill only while requests execute and scale to zero between them, which changes the math for bursty workloads.
The Pricing Tiers
| Option | What it is | Best for |
|---|---|---|
| Community cloud | Lower-cost capacity from vetted hosts | Cost-sensitive, interruption-tolerant jobs |
| Secure cloud | Data-center-grade, higher reliability | Production and sensitive workloads |
| On-demand | Guaranteed instance while running | Steady, predictable usage |
| Spot | Cheaper, interruptible | Fault-tolerant training and batch |
| Serverless | Per-second, scale to zero | Bursty inference and APIs |
What Drives Your Real Cost
RunPod vs the Alternatives
RunPod is strong on raw GPU price; managed platforms trade some cost for developer experience. Compare in Modal vs RunPod and Modal vs AWS Lambda. For GPU selection, see H100 vs H200.