An interesting time to be a software dev; the transformer LLM is a first really new development in a long time. AI coding agents have become core, like going from hand tools to power tools. I can only trust a tech if I can take it apart and put it together in my garage. Connecting to a service under someone else's control bugs me—too many ways for it to go wrong. I want a little AI data center in every home.

I need a bunch of GPUs, but anything designed for AI is wildly expensive. AMD overproduced V620 cards for cloud gaming, which flopped. These cards have 32 GB of fast VRAM, no fans, and are cheap on eBay. I got four barely used ones. Motherboard is an X299 from 2017 with four PCIe x16 slots. The CPU is an Intel Core i9 10900 X—Intel's worst chip, but cheap now. RAM and SSD salvaged from other computers. I got a 1600 W power supply (cheaper than 1200 W).

I modeled a 3D-printed fan shroud for two 80mm server fans per card, running at 10,000 RPM—really loud. The box didn't boot at first; I spent an hour in BIOS enabling Resizable BAR and MMIO High Size. Installed Ubuntu 24.04, compiled llama.cpp, and tested Gemma4. It fit comfortably. Then I spun up Deepseek V4 Flash, the real target—slow but it fit. I built a fan controller with an Arduino Nano clone and had the AI server write its own fan control script.