I Put a Paid AI Video Generator on My Own Gaming GPU — No Cloud Bill

A few weeks ago I asked myself a dumb question: what would it take to run a real, paid AI video service from the GPU that used to play Elden Ring?

Turns out the answer is: a lot of nights, one stubborn graphics card, and a pile of quantized weights.

Here's the honest walkthrough of how I built Marvis AI Studio — a tiny paid AI video generation service running entirely on a single RTX 5060 Ti 16GB at home.

Why would anyone do this?

Cloud GPU video generation costs real money per second. For a hobby project I wanted to:

• Own the full stack (queue, billing, generation) instead of renting it
• Learn how quantization and local inference actually behave under production-ish load
• Keep it cheap enough that "free trials" don't bankrupt me

Spoiler: the hardware choice matters more than the model choice.

The stack

Layer	Choice
GPU	RTX 5060 Ti 16GB (single card, no SLI magic)
Generation engine	ComfyUI on localhost:8188
Video model	MiniMax H3 (pruned INT8 quantized local weights)
VLM	qwen3-VL (32B, NVFP4/AWQ) for prompt understanding
API gateway	Python (FastAPI-style) on :8088 — queue, quota, redemption codes
Public access	Cloudflare tunnel → ai.jxzhihi.com (no port forwarding, no static IP needed)

Everything except the public domain runs on one Windows box in my apartment.

The model that actually fits

The critical moment was finding MiniMax H3 weights quantized/pruned for local VRAM. The INT8 pruned build fits inside 16GB with room for ComfyUI overhead. Quality at 864×480 is surprisingly usable; 1344×768 is where I'd call it "good".

Real timings on this card:

Resolution / duration	Wall-clock per video	File size
864×480 / ~3s	~6 min	~0.6 MB
1344×768 / ~5s	~21 min	~2.5 MB
1664×928 / ~5s	~40 min	~2.6 MB

Yes, it's slow. That's the trade-off of a $300-ish consumer card doing what data-center GPUs do. The single serial queue became a feature: nobody can hammer it, because physics won't let them.

Gotchas that cost me weekends

1. HostBuffer.read_file_slice failed on ComfyUI — a known bug pattern. Fix: launch with --cuda-device 0 --disable-pinned-memory. Without that flag the workflow randomly dies mid-generation.
2. Default gateway timeouts — 20-minute timeouts are not enough when one job takes 40. I set the queue wait ceiling to 60 minutes and surface queue position to the user.
3. Exposing the wrong port — keep ComfyUI strictly on localhost and only expose the gateway. One route mistake and strangers are queuing jobs on your raw engine.
4. Abuse protection before marketing — free tier is 3 generations per IP per day with a 90-second minimum gap. Real humans try it; scripts bounce.

The business part (tiny but real)

• Free trial: 3 generations/IP/day — enough to feel the quality
• Paid: redemption codes, because wiring Stripe to a home server is a story for another post
• A public /intl landing page with honest stats: real PV, tips, and shares (no fake numbers)

One trick I recommend: log tip/share intent before you build payment integration. Knowing that strangers want to pay you is better market research than any survey.

Would I do it again?

Yes — but I'd warn my past self about three things:

1. Pick the quantization first, then pick the GPU. I got lucky.
2. Assume every long job will be interrupted. Design the queue to survive restarts.
3. A single GPU is a hobby-scale business. Treat it as a learning lab with a URL, not a cloud competitor.

If you're curious, you can poke the live thing at ai.jxzhihi.com/intl — free trials included. I genuinely want feedback from people who run local AI stacks: what would you change about the architecture?

Built solo at home. One GPU, one tunnel, zero cloud bill.

Comments

[Luis Cruz]

Your approach to leveraging a consumer GPU for AI video generation is fascinating, especially the emphasis on quantization and how it can impact performance on limited hardware. The lessons learned about error handling and design trade-offs are incredibly valuable, especially for anyone looking to build scalable AI solutions on a budget. It sounds like your queue management strategy is crucial in maintaining a smooth user experience; perhaps implementing a more sophisticated prioritization system could enhance it further. If you’re considering expanding the functionality or tackling any additional challenges, I’m open to discussing potential collaborations to help make that happen. What other features or improvements do you envision for Marvis AI Studio?