What to Do with an RTX 4090 to Make Money

If you have an RTX 4090, or any strong consumer GPU, you already own the expensive part of several AI businesses. Open models for images (Stable Diffusion, Flux) and speech (Whisper) run on that card. Every job you run locally is one you do not pay an API for.

A GPU is a cost advantage, not a business. What to do with an RTX 4090 to make money still means finding a buyer. The four stacks below are documented operators who host models, sell the setup, or run a pipeline. Where they used cloud GPUs instead of a single home card, we say so. We do not have 4090-only benchmarks, and we will not invent a dollars-per-hour figure for your card.

1. Let people run their own checkpoints in the cloud

Stack: Run-Your-Custom-Model Stable Diffusion Site

sphuff launched stadio.ai in December 2022 so hobbyists could upload their own Stable Diffusion checkpoints and LoRAs. On Hacker News (Jan 22, 2023) they reported $700 MRR, about seven weeks after launch. That is self-reported, and there is no later update in the source. They moved from a local machine to RunPod and then to their own GPU setup. Startup cost is not disclosed.

Where your 4090 fits: it is enough to build and test the product. Serving paying customers from one home PC makes you the on-call person for uptime. The example did not stay on a single local machine.

Honest catch: Civitai, RunDiffusion and ThinkDiffusion now do parts of this. You also moderate whatever people upload.

2. Turn ComfyUI workflows into an API

Stack: ComfyUI Workflow Hosting / API Platform

ComfyDeploy’s founder reported, in an open letter, getting into Y Combinator at about $2.5K MRR and later $29K MRR ($50K processed in 30 days). They also said growth was slow and that Comfy Org now sells its own cloud. Founder-reported.

Where your 4090 fits: develop and demo workflows locally, then move production to rented GPUs when a client depends on it. The card is the prototype machine, not the whole company.

3. Sell installers to people who do not want to debug CUDA

Stack: Local AI 1-Click Installers & Tutorials Patreon

SECourses (Dr. Furkan Gözükara) sells updated 1-click installers for local generative apps, tested on NVIDIA GPUs, with cloud-GPU variants for RunPod and Massed Compute. Graphtreon estimates $3K–$12K a month from about 1,572 paid members (Oct 2026). Third-party estimate. He has not published revenue.

Where your 4090 fits: you cannot ship an installer you have not run. The card is the test bench. Patreon and YouTube are free to start. The cost that matters is time, because installs break every month.

4. Run Whisper on your own pipeline

Stack: Flat-Rate Unlimited AI Transcription Website

Leif Foged’s TurboScribe launched at $10 a month for unlimited transcription, on an optimized Whisper pipeline he controlled. The founder has not disclosed revenue. A third-party guess of about $257K MRR comes with a $29K–$2.28M range. Treat it as noise.

Where your 4090 fits: transcribe and benchmark locally before you promise “unlimited” to strangers. At real volume you will outgrow one card, and a handful of heavy users is the risk that matters.

Renting the card out

Renting an RTX 4090 on a marketplace (Vast.ai, Salad and similar) is GPU rental, not an AI business. We have no documented operator numbers for it on this site, so we will not quote a rate. Read the renter’s current pricing page if you want to compare that against using the card yourself. A $20 hosted subscription can still beat a weekend of broken drivers if you do not already have a buyer.

How to choose

1. Pick a buyer first. Hobbyists with custom models, studios with a ComfyUI graph, people stuck on an install, or heavy transcription users.
2. Check that the card actually runs the model at a speed you would ship.
3. Run real samples locally and on a hosted tool. Keep local only if the quality holds.
4. Price the outcome, not your electricity.
5. Keep a hosted fallback for deadlines. One home machine is a prototype.

Bottom line

What to do with an RTX 4090 to make money is to remove a cost or a setup pain someone will pay to avoid. The documented results here run from a self-reported $700 MRR image host to a founder-reported ComfyUI platform, and none of them came from the hardware alone. Start with the stack whose buyer you can already reach: custom model hosting, ComfyUI hosting, installers, or flat-rate transcription. More local image products: how to make money with Stable Diffusion.

Every stack in this guide, side by side

This guide links 4 stacks. 0 have payment-verified revenue. The typical stack is about 80% automated, so plan on doing the rest yourself. "MRR" means monthly recurring revenue, what subscribers pay each month. Revenue figures come from each stack's source and date; tap a name for the full breakdown.

Stack	Revenue (as reported)	Startup cost	Difficulty	Automated
Run-Your-Custom-Model Stable Diffusion Site	$0–$700 MRR (example, about 7 weeks after launch)	Not disclosed	Hard	90%
ComfyUI Workflow Hosting / API Platform	$0–$29K MRR (example, founder-reported)	Not disclosed	Hard	80%
Local AI 1-Click Installers & Tutorials Patreon	$3K–$12K (third-party estimate for the example)	Not disclosed	Medium	40%
Flat-Rate Unlimited AI Transcription Website	Not disclosed	Not disclosed	Hard	95%