← sources

Went looking for a GPU to rent for ComfyUI and fell into a rabbit hole. Where are you running yours? (runpod, modal, vast etc)

source: reddit (reddit:comfyui)published 2026-09-21 15:54ingested 2026-09-22 02:40

retained text

I don't have a GPU. I tried running ComfyUI on a Mac mini M4 and couldn't get much past Z-Image Turbo, so I went looking for one to rent. The name that kept coming up was RunPod, so that's where I started.

It works. After Civitai, with the failed jobs and the waiting, it felt good to have my own thing.

Two things got in the way though.

1. A pod stays on until you stop it. Obviously... But what I wanted was the opposite: wake the GPU when I generate and put it back to sleep after. A 1k image with Krea 2 takes about 9 seconds on a warm worker, and waking one up takes three to four minutes: roughly 80 seconds to boot, then another 90 or so to move 18 GB of weights from the volume into VRAM.

2. Then... checkpoints and LoRAs have to sit next to the GPU or the boot is even slower, and that storage gets billed every month whether you generate or not. I'm paying roughly €5 a month for a 75 GB network volume. Meh...

So I went looking for something with a faster cold start and cheaper storage. One evening turned into a rabbit hole research type of stuff... below what I actually learned from the ones I looked at properly:

**RunPod** is where I am. It works, it's well documented, the community is the biggest. It's prepaid, so when the balance runs out everything stops. Unpleasant, but never a surprise. Against it: the wake-up, a network volume is $0.07/GB a month (it increases the bill pretty quickly).

**Modal** is where I ended up testing next, and it's the only thing that moved the needle on my actual problem. With GPU snapshot, waking an 18 GB model took 35 seconds against 100 on RunPod serverless with FlashBoot already on. Storage is free up to 1 TiB a month (yup!), so my weights cost nothing. No 4090s though, closest are an L4 at $0.80/h and an A10 at $1.10. Postpaid card, but the spend cap actually stops the workloads (not tested yet, I'm still inside the $30 of free credits on the starter plan), though I'd rather have the prepaid.

**Vast.ai** are four to five times cheaper per hour, $0.14 to $0.16 for a 4090, because they run on other people's machines. No storage pricing (or I did not find it). I did not try them because I read mixed stories about long waits and failed generations (happy to hear I'm wrong).

**Beam** looked good on paper, cheaper per second than Modal and free storage to a terabyte. I never got to try it: there's no free tier and it wants around €20 up front (with runpod you can recharge even 5 bucks) which isn't a test, it's a purchase.

**Google Cloud Run** looked ideal until I read Google's own guidance. There's no persistent disk. That only would make cold starts even longer (not tested but plausible).

**Koyeb and Fly.io** I ruled out on storage alone: $0.50 and $0.15 per GB a month, so 20 GB of weights is $10 and $3.

**Railway, Render and Vercel** have no GPUs as far as I understand.

Honestly, after all that I'm still not sure what the right answer is. I've got a table and no decision.

So I'd rather hear from people who've actually been doing this for a while. What are you running on, and what made you settle there? Is there something obvious I've walked straight past?

document fetched 2026-09-22 00:40 · canonical https://reddit.com/r/comfyui/comments/1wmgqdk/went_looking_for_a_gpu_to_rent_for_comfyui_and