AI BriefingI run multiple companies with $10k monthly MRR on a $20/month tech stack
Key point
A practical guide to cutting operating costs to the extreme using VPS, Go, SQLite, and local AI.
Details
Start with just one $5~10/month VPS, and instead of complex AWS/EKS, manage services, logs, and restarts all in a simple way on a single server. 1GB RAM is enough, and if needed, supplement it with a swapfile.
The backend is written in Go. There's no need to run an interpreter, and without dependency hell, you build it as a single static binary and put it on the server—done. The post presents a very small http.ListenAndServe example, explaining that this structure simplifies deployment.
Large batch jobs and long inference are handled by local AI. Using a GPU at home, run vLLM; use Ollama during the experimentation phase, vLLM in production, and Transformer Lab when training/fine-tuning is needed. For research purposes, they also build their own laconic for compressing context and llmhub for simplifying model access.
Top-tier models for real-time users are bundled together with OpenRouter. Using a single OpenAI-compatible interface, you can use multiple frontier models, and automatically switch to another provider on failure to reduce errors.
IDE/agent costs are cut mainly with GitHub Copilot. The key point is that even long tasks are handled on a per-request basis, so the perceived cost is low—give it detailed prompts and clear success criteria and let it fix things through to completion.
For the database, SQLite is used by default. Local file access is much faster than a remote DB that goes through TCP, and turning on WAL greatly reduces read/write contention.
PRAGMA journal_mode=WAL;PRAGMA synchronous=NORMAL;
With this setting, they claim that a single .db file plus NVMe can handle many concurrent users.
Lastly, authentication is simplified with a self-built auth library. It supports sign-up, sessions, password resets, and Google/Facebook/X/SAML login, reducing the hassle of authentication infrastructure for early-stage products.
The core message is clear. Not complex cloud architecture and big fixed costs, but a simple stack and a low burn rate create a longer runway.
This summary was generated automatically by AI. Check the original for the author's claims and context. Copyright belongs to the original author.
Our guide explains how the AI works. Report summary errors, attribution issues, or removal requests via Contact.