Running Multiple $10K MRR Products Solo on a $20 per Month Stack and a Used GPU
A veteran solo founder runs several $10K MRR products on a roughly $20 a month software stack, offloading batch AI work to a used $900 GPU instead of paying API fees
Continue exploring this workflow
The Strategy
Steve Hanov is not a newcomer chasing an AI trend; he has run multiple bootstrapped products for years, including websequencediagrams.com, zwibbler.com, and rhymebrain.com, each generating around $10,000 in monthly recurring revenue on its own. What makes his approach worth documenting is not that he uses AI, but how deliberately cheap and self-reliant his infrastructure choices are compared to the typical stack of paid subscriptions assumed throughout most of this directory. The core of his cost-control strategy is running batch AI workloads, such as stock market research tasks, locally on a used RTX 3090 GPU he bought for about $900, using vLLM to serve the models himself instead of paying per-token API fees for high-volume, non-latency-sensitive work. For anything that does need to reach a hosted model, whether Claude or ChatGPT, he routes calls through OpenRouter, which automatically falls back to alternate providers if his first choice is unavailable or degraded, avoiding vendor lock-in and outage risk. Hanov has also built his own lightweight agentic tooling rather than adopting off-the-shelf agent frameworks: a research agent he calls laconic, and an abstraction layer called llmhub that sits between his applications and whichever model backend is actually serving a request. GitHub Copilot Chat handles cross-project code generation, letting him move quickly across several codebases without a large engineering team. The total recurring cost of this approach is strikingly low: around $20 a month for the core software stack, a handful of cheap VPS instances running $5 to $10 a month each, roughly $60 a month for Copilot, and a one-time $900 hardware purchase that keeps ongoing inference costs near zero for batch workloads.
How It Works
Identify which AI workloads are latency-tolerant and high-volume, such as batch research tasks, versus which need real-time hosted model responses.
Buy used, capable consumer GPU hardware, in this case a used RTX 3090 for about $900, to run batch workloads locally instead of paying per-token API costs.
Serve local models efficiently using vLLM, which handles the inference serving layer for self-hosted batch AI work.
Route any hosted-model calls that do need real-time response through OpenRouter, which provides automatic fallback across providers if one is down or degraded.
Build lightweight, purpose-specific agentic tools rather than adopting heavy off-the-shelf frameworks, matching tooling complexity to actual need.
Build an abstraction layer, llmhub in this case, so application code does not need to change when swapping which model backend actually serves a request.
Use GitHub Copilot Chat for cross-project code generation to move quickly across multiple live products without hiring additional engineers.
Run each product on cheap, minimal VPS infrastructure rather than expensive managed platforms, keeping total hosting costs in the single-digit dollars per product per month.
Results
Hanov reports each of his several products, including websequencediagrams.com, zwibbler.com, and rhymebrain.com, generating around $10,000 in monthly recurring revenue individually, while total recurring infrastructure cost sits around $20 a month for the core stack plus roughly $60 a month for Copilot and a handful of cheap VPS instances. These are self-reported figures and the post was highly discussed on Hacker News, though the numbers themselves have not been independently verified.
Our Take
This is a valuable counterweight to the assumption running through much of this directory that a serious AI-powered business needs an expensive stack of paid platform subscriptions. Buying a used GPU to run batch inference locally instead of paying ongoing API fees is a genuinely underused strategy for cost-sensitive, high-volume workloads, and building a thin abstraction layer over multiple model providers is good engineering practice regardless of budget. The tradeoff is that this requires real technical depth: self-hosting inference, building custom agentic tooling, and managing your own provider-fallback logic is not something a non-technical operator can replicate easily. This is most useful as a case study for developer-founders running multiple products who want to minimize recurring costs, not as a template for agencies selling to non-technical local business clients.
Frequently Asked Questions
The practical questions a builder or operator is likely to ask before trying a strategy like this.
What does this saas AI coding workflow actually do?
This saas AI coding workflow is a real workflow where the agent takes on an operational job, not just a brainstorming task. Running Multiple $10K MRR Products Solo on a $20 per Month Stack and a Used GPU shows what that looks like in practice. A veteran solo founder runs several $10K MRR products on a roughly $20 a month software stack, offloading batch AI work to a used $900 GPU instead of paying API fees The practical value comes from the agent handling repeatable business work with enough autonomy that a human only steps in after context has already been gathered.
Who should use a saas AI coding workflow like this?
This example is most relevant for saas operators. It is especially relevant for businesses where speed to lead, after-hours coverage, or consistent intake quality directly affects revenue. The category here is Dev Tools, which means the best fit is a team looking to turn a manual bottleneck into a repeatable system with a saas AI coding workflow.
Which tools are used in this saas AI coding workflow setup?
The source names OpenRouter, Claude, ChatGPT, Copilot. That matters because one of the strongest signals in this directory is whether the operator shared the actual stack. Named tools make a saas AI coding workflow strategy far more useful than vague claims about “an AI system” doing the work.
How hard is it to implement a saas AI coding workflow like this?
Advanced difficulty is the current read. The listing suggests a launch window of months. Startup cost is listed as under $50/mo. We were able to extract 8 concrete workflow steps from the source. We would treat a saas AI coding workflow like this as a workflow that needs real business context, testing, and exception handling rather than something you should copy blindly from one prompt.
What results can a saas AI coding workflow produce?
Hanov reports each of his several products, including websequencediagrams.com, zwibbler.com, and rhymebrain.com, generating around $10,000 in monthly recurring revenue individually, while total recurring infrastructure cost sits around $20 a month for the core stack plus roughly $60 a month for Copilot and a handful of cheap VPS instances. These are self-reported figures and the post was highly discussed on Hacker News, though the numbers themselves have not been independently verified.
How credible is this saas AI coding workflow case study?
Right now the evidence comes from an article from stevehanov.ca. That is enough for us to study and curate the workflow, but not enough on its own to treat this saas AI coding workflow like an audited case study. We look for named tools, concrete results, and enough workflow detail to understand what was actually deployed, then we add our own editorial judgment on top.
Related Strategies
More AI agent strategies you might find useful
Learning to Code With AI and Building a $28K Per Month SaaS Portfolio
A non developer learned to code using AI tools and built a portfolio of SaaS pro…
How a $40K MRR SaaS Growth Playbook Uses Claude Code and Google Search
A SaaS playbook combining Claude Code for development and Google Search for dist…
Running a Company Entirely With AI Agents: What Worked and What Broke
Seven AI agents ran a company on a single VPS for a week. They pivoted the busin…
Pieter Levels Vibe-Coded a Flight Sim to $1M ARR in 17 Days Selling In-Game Ads
Pieter Levels built a browser multiplayer flying game almost entirely with AI co…
A $15K MRR SaaS Built in 12 Hours Using AI Coding Tools
A solo founder built a SaaS that hit $15K MRR with the entire MVP coded in a sin…
A Log Analysis SaaS Built in 30 Days That Hit $1,287 MRR With 41 Paying Customers
A developer tool that turns cluttered production logs into concise incident summ…
Want more strategies like this?
Get weekly AI agent case studies in your inbox.