He Turned Discarded Crypto Mining Boards Into a Working Home AI

A hobbyist who had never touched local AI before built himself a working AI computer out of six BC-250 boards — the small, cheap circuit boards that cryptocurrency miners used to run in racks. He had a pile left over after reselling most of them, so he stacked five in the secondhand server case they originally shipped in and started experimenting. After a lot of testing, he landed on a setup he likes, and posted the details asking for more ideas.

The trick is that these boards don't act alone. He links them together over ordinary one-gigabit Ethernet cable — the same kind of cable that connects a home router — and uses free software called llama.cpp to split the work between machines. Each board runs a compressed version of an open AI model, squeezed down so it fits in limited memory. Four boards handle one model; the remaining two run a larger, slightly smarter one.

In plain numbers: the four-board group writes about 28 words per second and can hold roughly 100,000 words of conversation in its memory at once — comfortably a whole book. The two-board group is quicker at 60 words per second. Both are as fast as you can read, though far slower than paid cloud services. Cooling is a cardboard box with three fans taped on top. It's a tinkerer's rig, not a product: expect hours of setup, noise, and electricity costs.

Why care? Because it points at a real shift. Powerful AI is starting to run on cheap secondhand hardware — old mining boards, retired office PCs — instead of rented cloud computers. That means private AI with no subscription and no data leaving your house. It's still a hobbyist project today, but the direction is clear: the hardware is becoming ordinary.