Listen to this post (in German), created with OmniVoice:
Welcome back to the Brain Dump! Today I'm tackling a topic for everyone who runs into their hardware limits with local AI models: cloud GPUs, or more precisely the RunPod platform.
What is RunPod and why would you need it?
In short: RunPod is a service where you can rent powerful graphics cards (GPUs) in the cloud by the hour. That makes a lot of sense when you run complex AI workflows that eat huge amounts of VRAM (video memory).
Locally, my main PC (HauptHorst) has an RTX 3060 Ti running under CachyOS. It's a rock-solid card, but when it comes to demanding video generation, things get sluggish fast. That's exactly where RunPod comes in: instead of sinking thousands of euros into new hardware, you simply rent the compute power exactly when you need it.
Getting started and paying the obligatory tuition
Right now I use RunPod exclusively for ComfyUI together with the video model MiniMax H3. Since I'm still at the stage where I need to understand the whole system and the workflows, a fair bit of time (and money) went into testing at first.
A huge thank-you at this point to my sponsor Ascot Media, who gave me $20 of starting credit at the beginning and has now put up another $30 for some serious "work"! I really needed it, too, because I'm still paying my tuition.
Why does he do it? With this, Ascot Media is investing specifically in no0bz Production. I make music videos for him, and we've got quite a bit more planned for the future. Our shared goal is to become completely independent of commercial video generators. We want to keep full creative control and have the freedom to show a bit more "skin" now and then or add some "gore" – content that the usual paid services block immediately because of their strict policies. With open-source models on RunPod, we have exactly that freedom.
Watch the video: https://www.youtube.com/embed/0bGOJNcGhXk?feature=oembed
Made with an LTX 2.3 workflow
Recently I blew $3 through a silly mistake: I assumed a pod migration was finished just because I could reach the ComfyUI interface again. Nope. Apparently data was still being copied in the background. Because my folders looked empty, I (once again, sigh) started downloading the models again. The end of the story? Massive data chaos where nothing worked anymore. So I had to set up the pod from scratch – which, to be fair, was done surprisingly quickly.
New setup: a network volume instead of frustration
You learn from your mistakes. So that I don't have to re-download my models and workflows every time or lose them during migrations, I changed my strategy. I now use a network volume on RunPod for a cheap $0.024 per hour.
The big advantage: the storage is decoupled from the actual pod. When I now start a new pod (usually from a MiniMax H3 template), I simply attach this volume. All my models, custom nodes and workflows are back instantly, no long wait. With the pod running on a rented RTX 5090, I end up at a total of about $1.03 per hour, network storage included. (Whether those are Community or Secure instances, I honestly don't know exactly 😉.)
Before that, download times were a lottery. Everything from agonisingly slow 80 KiB/s to lightning-fast 800 MiB/s, which made the initial setup really tedious at times.
Keeping costs under control, and my file management
At around a dollar an hour, you can easily fall into a cost trap if the pod keeps running unnoticed. So my simple but effective rule is: I only keep the pod running while I'm actually sitting in front of it. As soon as I'm done with my generations, it gets switched off.
As for file management, I keep things simple. I download the generated videos straight through the browser and then back them up to my Google Drive. That way everything is in one central place and I don't risk losing anything.
The time saved: is it worth it?
Once the pod is running, the real benefit shows. Pure render time drops enormously. Here's a direct comparison for generating a 10-second clip (note the different resolutions!):
| Hardware | Resolution | Render time | Note |
|---|---|---|---|
| RTX 3060 Ti (my local machine) | 720p | approx. 20 minutes | My daily driver |
| RTX 5070 Ti (local at Ascot Media) | 720p | approx. 10 minutes | Ascot Media's local PC |
| RTX 5090 (RunPod) | 1080p | approx. 9 minutes | Very fast, but there's more to come! |
Future plans: quality vs. time
At the moment I mainly use the RTX 5090 for practice. I want to make sure my workflows run stably and that I've fully internalised how ComfyUI works.
But the real plan for the future looks different: once everything is down pat, I want to make the jump from the 5090 to an absolute monster of a card – an RTX 6000 with a whopping 96 GB of VRAM. Only then are we really talking about "brute force". The goal then is not just to save time, but to squeeze the absolute maximum in quality and resolution out of MiniMax H3.
My verdict: Despite the sometimes fiddly setup and the fluctuating download speeds, RunPod is a real game changer for demanding video workflows. If you want to spare your local hardware and see results in a reasonable amount of time, there's hardly any way around a solution like this.