The latest articles on DEV Community by erniou86 (@ernioiu8). https://dev.to/ernioiu8 https://media2.dev.to/dynamic/image/width=90,height=90,fit=cover,gravity=auto,format=auto/https:%2F%2Fdev-to-uploads.s3.us-east-2.amazonaws.com%2Fuploads%2Fuser%2Fprofile_image%2F4117644%2F806358ce-f4cb-4fa7-b1e6-76d47fbf2859.png https://dev.to/ernioiu8 en erniou86 Wed, 09 Sep 2026 13:50:56 +0000 https://dev.to/ernioiu8/i-put-a-paid-ai-video-generator-on-my-own-gaming-gpu-no-cloud-bill-24hk https://dev.to/ernioiu8/i-put-a-paid-ai-video-generator-on-my-own-gaming-gpu-no-cloud-bill-24hk <h1> I Put a Paid AI Video Generator on My Own Gaming GPU — No Cloud Bill </h1> <p>A few weeks ago I asked myself a dumb question: <em>what would it take to run a real, paid AI video service from the GPU that used to play Elden Ring?</em></p> <p>Turns out the answer is: <strong>a lot of nights, one stubborn graphics card, and a pile of quantized weights.</strong></p> <p>Here's the honest walkthrough of how I built <a href="https://ai.jxzhihi.com/intl" rel="noopener noreferrer">Marvis AI Studio</a> — a tiny paid AI video generation service running entirely on a single RTX 5060 Ti 16GB at home.</p> <h2> Why would anyone do this? </h2> <p>Cloud GPU video generation costs real money per second. For a hobby project I wanted to:</p> <ul> <li>Own the full stack (queue, billing, generation) instead of renting it</li> <li>Learn how quantization and local inference actually behave under production-ish load</li> <li>Keep it cheap enough that "free trials" don't bankrupt me</li> </ul> <p>Spoiler: the hardware choice matters more than the model choice.</p> <h2> The stack </h2> <div class="table-wrapper-paragraph"><table> <thead> <tr> <th>Layer</th> <th>Choice</th> </tr> </thead> <tbody> <tr> <td>GPU</td> <td>RTX 5060 Ti 16GB (single card, no SLI magic)</td> </tr> <tr> <td>Generation engine</td> <td> <a href="https://github.com/comfyanonymous/ComfyUI" rel="noopener noreferrer">ComfyUI</a> on localhost:8188</td> </tr> <tr> <td>Video model</td> <td>MiniMax H3 (pruned INT8 quantized local weights)</td> </tr> <tr> <td>VLM</td> <td>qwen3-VL (32B, NVFP4/AWQ) for prompt understanding</td> </tr> <tr> <td>API gateway</td> <td>Python (FastAPI-style) on :8088 — queue, quota, redemption codes</td> </tr> <tr> <td>Public access</td> <td>Cloudflare tunnel → <code>ai.jxzhihi.com</code> (no port forwarding, no static IP needed)</td> </tr> </tbody> </table></div> <p>Everything except the public domain runs on one Windows box in my apartment.</p> <h2> The model that actually fits </h2> <p>The critical moment was finding <strong>MiniMax H3 weights quantized/pruned for local VRAM</strong>. 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".</p> <p>Real timings on this card:</p> <div class="table-wrapper-paragraph"><table> <thead> <tr> <th>Resolution / duration</th> <th>Wall-clock per video</th> <th>File size</th> </tr> </thead> <tbody> <tr> <td>864×480 / ~3s</td> <td>~6 min</td> <td>~0.6 MB</td> </tr> <tr> <td>1344×768 / ~5s</td> <td>~21 min</td> <td>~2.5 MB</td> </tr> <tr> <td>1664×928 / ~5s</td> <td>~40 min</td> <td>~2.6 MB</td> </tr> </tbody> </table></div> <p>Yes, it's slow. That's the trade-off of a $300-ish consumer card doing what data-center GPUs do. The <strong>single serial queue</strong> became a feature: nobody can hammer it, because physics won't let them.</p> <h2> Gotchas that cost me weekends </h2> <ol> <li> <strong><code>HostBuffer.read_file_slice failed</code> on ComfyUI</strong> — a known bug pattern. Fix: launch with <code>--cuda-device 0 --disable-pinned-memory</code>. Without that flag the workflow randomly dies mid-generation.</li> <li> <strong>Default gateway timeouts</strong> — 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.</li> <li> <strong>Exposing the wrong port</strong> — keep ComfyUI strictly on localhost and only expose the gateway. One route mistake and strangers are queuing jobs on your raw engine.</li> <li> <strong>Abuse protection before marketing</strong> — free tier is 3 generations per IP per day with a 90-second minimum gap. Real humans try it; scripts bounce.</li> </ol> <h2> The business part (tiny but real) </h2> <ul> <li>Free trial: 3 generations/IP/day — enough to feel the quality</li> <li>Paid: redemption codes, because wiring Stripe to a home server is a story for another post</li> <li>A public <a href="https://ai.jxzhihi.com/intl" rel="noopener noreferrer">/intl landing page</a> with honest stats: real PV, tips, and shares (no fake numbers)</li> </ul> <p>One trick I recommend: <strong>log tip/share intent before you build payment integration.</strong> Knowing that strangers <em>want</em> to pay you is better market research than any survey.</p> <h2> Would I do it again? </h2> <p>Yes — but I'd warn my past self about three things:</p> <ol> <li>Pick the quantization first, then pick the GPU. I got lucky.</li> <li>Assume every long job will be interrupted. Design the queue to survive restarts.</li> <li>A single GPU is a hobby-scale business. Treat it as a learning lab with a URL, not a cloud competitor.</li> </ol> <p>If you're curious, you can poke the live thing at <strong><a href="https://ai.jxzhihi.com/intl" rel="noopener noreferrer">ai.jxzhihi.com/intl</a></strong> — free trials included. I genuinely want feedback from people who run local AI stacks: what would you change about the architecture?</p> <p><em>Built solo at home. One GPU, one tunnel, zero cloud bill.</em></p> ai gpu selfhosting machinelearning