Running ComfyUI on a cloud GPU
If your own graphics card is too small, you can rent one. There are three ways to do it, and they differ in what you pay for while nothing is running.
Updated October 2026
Three ways to rent
| Way | How it works | What you pay for | Where your files are |
|---|---|---|---|
| A GPU computer by the hour | You start a cloud computer with a graphics card, often with ComfyUI already on it, and use ComfyUI in your browser. | Every hour it's switched on, busy or idle, plus storage for your models. | On that computer and its storage, in your cloud account. |
| Serverless, by the second | ComfyUI and the model are packed into a worker. A GPU starts when a job arrives and switches off seconds after. | Only the seconds a job runs. | In your cloud account, if the account is yours. |
| A hosted ComfyUI website | A company runs ComfyUI and shows it to you in the browser. | Credits or a subscription, at the site's prices. | On the company's servers. |
A GPU computer by the hour
This is the closest thing to having the card at home. On RunPod it's called a pod: you pick a graphics card, start a template with ComfyUI already installed and open it in your browser. You download models and add-ons onto a storage volume so they're still there next time.
The meter runs for as long as the computer is on, whether a job is running or you're reading the next tutorial. Forgetting to stop it is the classic surprise bill. It suits long sessions and building workflows.
Serverless, by the second
With serverless GPUs (RunPod Serverless, for example), you don't rent a computer. You set up an endpoint that holds a worker image, a package with ComfyUI, the model and its workflow. When you send a job, RunPod starts a GPU, runs it and switches the GPU off a few seconds after it's done. You pay for those seconds and nothing else.
Two things to know:
- Setup is technical. Someone has to build the worker image, create the endpoint and write the code that sends jobs and fetches results.
- Starts take time. The very first job downloads the model, which can take 10 to 30 minutes. After a break, a start takes seconds to several minutes while the model loads.
A hosted ComfyUI website
The simplest way to try ComfyUI: sign up and use it in your browser. You pay the site's prices, your prompts and results are stored on its servers, and you get the models and add-ons it offers.
What the cards cost on RunPod
RunPod's serverless prices for the cards CozyGPU uses, checked October 21, 2026. Serverless bills by the second; hourly GPU computers have their own, usually lower, hourly rates.
| Card | Per hour | Per second |
|---|---|---|
| Ada 24 GB | $1.10 | $0.00031 |
| Ada 32 GB Pro | $1.58 | $0.00044 |
| Ampere 48 GB | $1.22 | $0.00034 |
| Ada 48 GB Pro | $1.75 | $0.00049 |
| Ada 80 GB Pro | $4.79 | $0.00133 |
| Blackwell 96 GB | $3.49 | $0.00097 |
Because only working seconds count, a result costs very little: about $0.0024 for a Z-Image Turbo picture and about $0.146 for a 5-second MiniMax H3 clip.
How CozyGPU uses this
CozyGPU does the serverless setup for you. When you connect your RunPod account, it creates one endpoint per model in that account, each with a tested worker that runs ComfyUI. The app on your Mac or PC sends your jobs straight to your account and saves the results on your computer. Nothing runs and nothing is charged between jobs, and your prompts and files never pass through CozyGPU. The setup guide lists everything it creates in your account, and what leaves your computer lists every request the app sends.
Common questions
What is the best cloud GPU for ComfyUI?
For pictures, an NVIDIA card with 24 to 48 GB is plenty for nearly every model. For video at full quality, 80 to 96 GB cards such as the H100 or the RTX PRO 6000 Blackwell. A bigger card isn't faster for small models, so match the card to the model.
How much does it cost to run ComfyUI on RunPod?
On an hourly GPU computer, the hourly rate for as long as it's switched on, plus storage. On RunPod Serverless, only the seconds a job runs. With CozyGPU's settings, a Z-Image Turbo picture costs about $0.0024 of GPU time and a 5-second MiniMax H3 video clip about $0.146.
Do I pay for a cloud GPU while it's idle?
On an hourly GPU computer, yes, until you stop it. On a serverless GPU, no: it switches off seconds after the last job. Storage you keep in your account is charged per gigabyte per month either way.
Can I use my own ComfyUI workflows on a cloud GPU?
On an hourly GPU computer, yes, any workflow and any add-on. Serverless works with any workflow too, but you have to pack it into a worker yourself. CozyGPU runs its own tested workflows and doesn't load yours, though you can add your own LoRAs.
Is a cloud GPU private?
With your own cloud account, your prompts and files go to that account and nowhere else. On a hosted website they're on the company's servers. CozyGPU uses your own RunPod account, and none of your prompts or files reach CozyGPU.
Why does the first job on a cloud GPU take so long?
The GPU has to download the model, often 20 to 70 GB, the first time, and load it into memory after each break. With serverless GPUs the download happens once per model; after that a start after a break takes seconds to several minutes, depending on the model's size.
CozyGPU rents the graphics card for you, sets 12 open-source models up on it and gives you a simple app on your Mac or PC. It runs ComfyUI for you on a GPU in your own RunPod account, billed by the second, so nothing is charged while you're not using it. $20 a month, plus the GPU time.