Home-built 4x RTX 3090 rig runs GLM-5.2, MiniMax, and more locally
Solo dev builds $6000 AI workstation with 4 used 3090s, power-capped to 200W each
A solo forward-deployed engineer (Reddit user Important_Quote_1180) spent 40 hours and approximately $6,000 building a custom home AI workstation. The rig is built around 4 used RTX 3090 GPUs, each purchased in local transactions from gamers upgrading to 4090/5090 models. Each GPU is power-capped to 200W in Linux for efficiency. The system uses 192GB of DDR5 RAM overclocked from 5200MHz to 5600MHz, mounted on a budget motherboard and CPU from an Aegis prebuilt (PSU upgraded to 1250W platinum). The cooling loop from the prebuilt was retained.
The machine runs multiple models simultaneously: GLM5.2 as a planner at 7 tokens/sec, MiniMax 2.7 entirely on VRAM at 45 tokens/sec for coding, Qwen3.6 27B at Q8 quantization as a checker/testing loop model at 50 tokens/sec, and Flux2Klein for diffusion (1 image per 6 seconds when batching with 2 GPUs). The build is intended for enterprise automation workflows that the engineer builds for a dozen companies. By keeping it on consumer hardware and powering it with solar energy, he gains full independence from cloud AI providers like OpenAI and Anthropic. Future upgrades will focus on GPUs, possibly adding a dedicated server for GLM.
- 4 used RTX 3090s all power-capped to 200W each, bought from gamers for ~$1,500 total
- 192GB DDR5 RAM overclocked to 5600MHz combined with 1250W platinum PSU
- Runs GLM5.2 (7 t/s), MiniMax 2.7 (45 t/s), Qwen3.6 27B Q8 (50 t/s), and Flux2Klein (1 img/6s batched)
Why It Matters
Consumer hardware can replace cloud API for serious AI workloads, enabling local, solar-powered, cost-effective automation.