Guide
RunPod GPU Math Changes When a Stopped Volume Disk Costs 2x
RunPod bills burst GPU jobs per second, but a stopped volume disk costs $0.20/GB/month—twice its running rate. Include idle storage before switching.
At a glance
Occasional generation jobs can beat a monthly credit plan that resets. On RunPod, a stopped volume disk bills at twice the running rate. That line has to be in the cancel math, or the savings disappear.
A monthly image or video plan is a prepaid bucket. If the bucket resets and you only generate in bursts, you are paying for idle credits. Self-hosting the same jobs on a per-second GPU looks like the clean exit: run for 12 minutes, pay for 12 minutes, stop. That arithmetic is real — and incomplete. On RunPod’s pricing page (footer: Updated July 27, 2026; checked 28 August 2026), a stopped volume disk costs twice as much as a running one. Almost nobody writes that 2× line as the condition for cancelling the subscription. Without it, the “cheap GPU hour” is a storage bill with a GPU attached.
This is not a GPU bake-off, and it is not a multi-cloud price table. It is a total-cost test for one move: drop a resetting credit plan, rent RunPod for occasional image or video jobs, and decide whether the disk you leave behind eats the savings. Confirm every rate on the live page before you cancel anything; GPU and storage prices move.
If the work already fits a laptop you own, that is a different article — see local LLMs in 2026. This one is only the hosted, occasional-job path.
The subscription you are trying to leave
Creator tools often sell a monthly credit or “fast hours” allotment that does not bank. Runway’s help page is explicit: monthly plan credits refresh on the billing day; unused credits are forfeited, except on Max, where up to one month may roll. Cancel, and the balance resets to 0. Extra credits you buy separately start at $10 and do not expire.
That shape is the whole premise. A light month still bills the full fee. A cancelled month zeros the leftover bucket. Before you stand up a GPU, check the in-product exit first: if the vendor sells non-expiring purchased credits, that is usually cheaper ops than a cloud workstation. The image and video guides cover when a subscription is the right shape at all.
Use RunPod only when the remaining job is bursty, you accept self-hosting (ComfyUI, a diffusion container, a short batch), and the monthly credit plan is mostly waste.
What “per-second” actually meters
RunPod bills compute plus local container and volume disks per second, with no data-transfer fees; network volumes are billed hourly. The marketing line is the same: run a job for 3 minutes, pay for 3 minutes. Granularity is not a discount. It only helps if the machine is off when you are off.
Three products sit on the same pricing page. Use the one that matches the job:
- Pods — a dedicated GPU instance. This is the usual ComfyUI / notebook / “open a desktop and generate” path.
- Serverless — inference workers billed while they run. Flex workers scale to zero; Active workers stay up 24/7 (discounts via sales).
- Clusters / Reserved — multi-node and committed capacity. The public page says contact sales. Not an occasional-freelancer product.
A deposit is not the same as a billing minimum. Docs say you can start with as little as $10, and you must hold at least one hour of credits for the GPU you are deploying. At the $0.74 page snapshot below, a three-minute RTX 4090 job still needed about $0.74 sitting in the account to launch, then billed by the second. RunPod credits are non-refundable and cannot be withdrawn. Auto-pay will reload the wallet when the balance drops — which can recreate the subscription you just cancelled. Default spend cap is $80/hour.
Example Secure Cloud Pod rates from the live pricing page, checked 28 August 2026:
| GPU (Pods) | Listed rate |
|---|---|
| RTX 4090 | $0.74/hr |
| A100 PCIe | $1.39/hr |
| H100 SXM | $3.29/hr |
| H200 | $4.59/hr |
Those are dated list examples, not a recommendation of which card to rent. RunPod’s separate August 25 price-cut announcement listed the Secure Cloud RTX 4090 at $0.69, while the live pricing page still rendered $0.74 on August 28. Because those official pages conflict, the deploy console is the final price. The worksheets below deliberately preserve the observed $0.74 page snapshot. A 24 GB 4090 is used only as a reproducible example. If the workflow needs 80 GB-class VRAM, four hours of the page’s $3.29 H100 SXM rate is $13.16 before disk.
Serverless list rates on the same page are higher than Pods for the same class (a 4090 worker is listed at $1.10/hr). You pay the extra for scale-to-zero and request-shaped billing, not for a cheaper hour.
The 2× idle disk is the missing line
Pods have three disks. Only one of them punishes you for stopping.
| Storage | Running | Stopped | What happens to the data |
|---|---|---|---|
| Container disk | $0.10/GB/month | Not charged | Cleared on stop |
Volume disk (/workspace) |
$0.10/GB/month | $0.20/GB/month | Kept until the Pod is terminated |
| Network volume, standard, under 1 TB | $0.07/GB/month | $0.07/GB/month | Independent of the Pod |
| Network volume, standard, over 1 TB | $0.05/GB/month | $0.05/GB/month | Independent of the Pod |
| Network volume, high-performance | $0.14/GB/month on the public pricing page | Same idea — not a 2× idle | Docs say the console figure can vary by data center |
Read the idle column twice. A stopped volume disk is twice the running rate. That is not a rounding quirk. RunPod’s own Pod pricing table prints it. Container disk is the opposite: it goes to zero when you stop, and so does its data. Network volumes do not double when idle; they just keep billing, every month, until you delete them.
Two operational facts make the trap sticky:
- Stop ≠ terminate. Stopping releases the GPU and keeps
/workspace. You still pay volume disk — at the idle rate. Terminating deletes that volume. Docs say it plainly: terminate if you do not need to retain the environment. - Volume disk only grows. You can raise the size; you cannot shrink it. A 200 GB “just in case” models folder is a $40/month idle bill for as long as the stopped Pod exists.
RunPod is not designed for long-term storage. If the account hits $0, Pods stop. With a network volume, data is preserved on the volume while storage charges keep accruing — and the volume can still be terminated if those charges cannot be paid. Without a network volume, the Pod is terminated and the data is gone. Back up off-platform. Treating /workspace as cheap NAS is how the 2× line becomes the whole invoice.
A month of occasional jobs, with the disk included
Use your own hours and disk size. The numbers below are a worksheet, not a quote. Snapshot rates checked 28 August 2026: RTX 4090 at $0.74/hr, volume idle at $0.20/GB/month, standard network volume at $0.07/GB/month. A month is treated as 730 hours. Re-run against the deploy console.
Four GPU hours, 50 GB volume disk, Pod stopped the rest of the month
- Compute: 4 × $0.74 = $2.96
- Volume while running: 50 × $0.10 × (4 / 730) ≈ $0.03
- Volume while stopped: 50 × $0.20 × (726 / 730) ≈ $9.95
- Total ≈ $12.94
The GPU hour was cheap. The parked models were not.
Same four hours, 50 GB standard network volume, Pod terminated
- Compute: $2.96
- Network volume: 50 × $0.07 = $3.50
- Total ≈ $6.46
The 2× tax is gone. You still pay for 50 GB all month, including the weeks you generate nothing. Network volumes must be attached at deploy, they replace the volume disk at /workspace, they are Secure Cloud only, and they cannot be detached later without deleting the Pod. Size them for the models you will actually reload, not for a library you might want.
Same four hours, terminate after each session, no persistent disk
- Compute: $2.96, plus whatever minutes the next boot spends pulling the image and the checkpoints — those minutes are GPU-priced too.
- Storage: ~$0
- Total: a few dollars, if you exported the outputs
Re-downloading a model on the GPU clock looks wasteful until you compare it with $10–$20 of idle volume. Four extra 8-minute cold starts on a 4090 are about $0.39. One 50 GB stopped volume is about $10. For bursty work, deleting the disk is often the cheaper “cache.”
Two GPU hours, 100 GB stopped volume disk
- Compute: 2 × $0.74 = $1.48
- Idle volume: ~$20
- Total ≈ $21.50
This is the case that should have stayed on the subscription — or never kept a 100 GB volume. If idle disk alone is larger than the plan you are cancelling, the switch has already failed. Compute cannot save you.
A compact exit rule:
monthly_idle_volume ≈ GB × $0.20
If that number is greater than or equal to the image/video plan you want to cancel, do not cancel yet. Shrink or delete the volume, move to a network volume, or stay on purchased credits. Then rerun the sum:
GPU_hours × listed_rate + storage + setup_minutes_on_the_clock
Ignore list-price GPU hours that assume the machine is only on while you are generating. Include the boot, the checkpoint load, and the hour you forgot to stop.
Three storage exits that actually stop the 2×
Terminate the Pod when the batch is exported. GPU billing ends. Volume disk ends. Next session starts from a template or a custom image. Best when jobs are rare and the model pull is shorter than a month of idle disk.
Keep a small standard network volume, not a fat volume disk. You pay $0.07/GB/month whether the GPU is up or not, with no idle doubling. Upload checkpoints with the S3-compatible API so the first GPU minute is not a download minute. Do not default to high-performance storage ($0.14/GB/month on the public page) for a weekend batch.
Use Serverless Flex only if the job is an endpoint. Workers bill from start until they fully stop, rounded up to the second: start time (container + model load), execution, then a default 5-second idle timeout. Flex scales to zero, so there is no stopped-Pod volume at $0.20/GB. A network volume attached to the endpoint still bills all month. Active workers are the opposite of this article — they run 24/7. Container disk on Serverless is listed around $0.10/GB/month in 5-minute intervals; it is not a reason to keep a 100 GB workspace “just in case.”
Do not buy a 3- or 6-month savings plan for occasional work. Those plans are prepaid, non-refundable, and they do not cover storage. Storage on a stopped Pod continues at standard rates. That is a committed GPU product wearing a discount, not an escape from a credit plan.
Other exit conditions, same wallet
The idle disk is the line most write-ups skip. These four are the next ones:
- Leftover RunPod credits do not come back. You can recreate the same trapped prepaid balance you just left on Runway. Deposit the next month’s estimated total, not a round number you will not burn.
- Auto-pay re-subscribes you. If the point of leaving the creator plan was “do not bill me on quiet months,” do not enable a reload that fires when a stopped volume quietly drains the wallet.
- $0 is a data event. No network volume: Pod terminated, data gone. Network volume: disk keeps billing until it too can be deleted. Enable a low-balance alert before the first real job.
- A parked Pod is not a reservation. Starting a stopped Pod can come back with zero GPUs if capacity moved. Occasional work has to tolerate “not tonight.”
Who should switch, who should stay
Switch (or trial with $10) if you already measured a light month on a resetting plan, you can name the GPU hours and the disk GB, idle volume at $0.20/GB is smaller than the plan fee, and you will either terminate after export or keep a small network volume. A freelancer who ships two client batches a month and throws away the rest of the credit bucket is the intended reader.
Stay on the creator plan if image or video is a weekly deliverable, you need the vendor’s editor, character consistency, or commercial terms, or you will not actually stop the Pod. Recurring generation is still the case the image and video guides send to a subscription.
Stay, but buy non-expiring credits if the vendor sells them and your volume is bursty inside that product. Runway’s purchased credits are the worked example; check the current help page, not this sentence, before you buy.
Do not use RunPod as the exit if the only GPU that runs your workflow is H100/H200-class and you would need tens of hours, if you need an always-on worker, or if you were about to leave a 100 GB volume disk stopped “so next time is faster.” That last habit is how the 2× line deletes the savings.
Stacked AI seats have their own math in why small teams overpay for subscriptions. A free allowance is not a contract in the free-tier guide. Neither article replaces the disk line.
A one-hour test before you cancel
Do this once. Then cancel, or don’t.
- Open last month’s creator-plan invoice and the leftover credit (or fast-hours) balance. If you used most of the bucket, stop. The plan is doing its job.
- Open runpod.io/pricing today. Write down the GPU you would actually boot and the three storage rates. If the page differs from this article, use the page.
- Time one real job on a Pod: start, generate, export, stop. Record wall-clock minutes, including model load. Convert to money at the listed hourly rate.
- Multiply by the number of sessions you expect in a quiet month. Add GB × $0.20 if you would leave a volume disk stopped. Add GB × $0.07 if you would keep a standard network volume. Add $0 storage if you would terminate.
- Compare that total with the plan fee you would stop paying. Cancel only if the RunPod total is lower and you will follow the storage choice you just priced. If you cannot honestly say you will terminate or keep a small network volume, the idle disk has already voted no.
Quiet months are the only months this exit is for. The 2× idle volume is the condition that decides whether per-second GPU is a cheaper substitute — or a more expensive way to store checkpoints.