Per-Second Billing for RunPod GPUs | Generated by AI
Question: On RunPod, for a single B200 GPU, is billing based on minutes or hours? What’s the minimum unit used to calculate cost?
Answer:
RunPod bills per-second, not per-minute or per-hour. The advertised rate is quoted per-hour, but the actual charge accrues by the second for the time your pod is running.
A few specifics that matter for a B200 single-GPU pod:
- Secure Cloud / Community Cloud pods: per-second billing while the pod is running. You’re charged for the wall-clock time the pod exists in a running state, not for compute utilization. An idle-but-running pod still bills.
- Stopped pods: GPU billing stops, but you still pay for the persistent volume/storage (charged per-hour, prorated, even when stopped — and storage on stopped pods can still be billed, so delete the pod if you don’t need the disk).
- Serverless: billed per-second of actual execution (active workers), which is the more cost-efficient mode for bursty inference.
- Credit deduction granularity: balance is deducted continuously; if your balance hits zero, the pod gets stopped/terminated.
So the minimum unit is 1 second. Practical implication for your workflow: spin up, run the job, and tear down via the API/CLI rather than leaving it parked — a B200 idling for an hour between REPL iterations is real money. Script the lifecycle (runpodctl or the GraphQL API) so the pod only exists during active work.
Verify the current B200 hourly rate on the pricing page since GPU rates shift, but the billing granularity (per-second) is stable RunPod policy.
References: