54 GB of VRAM for $35...CHEAPEST LOCAL AI RIG?

A developer reportedly built a nine-GPU local AI machine out of discarded crypto-mining cards for 35 dollars: nine Nvidia P106-100s with 54 GB of VRAM between them. This video checks three things before you copy it: the real price, the actual speed, and the problems a GPU benchmark doesn't show. What's in it: - what P106-100 and P104 cards cost where old mining farms dumped them, and what a nine-card rebuild costs in the US - why 54 GB across nine cards isn't the speed of one 54 GB GPU (layer splitting, ~192 GB/s per card) - why a dense 70B model crawls and a mixture-of-experts model like Qwen3.5-35B-A3B doesn't - idle power, prompt processing, booting from USB, x1 risers, and the end of driver support for Pascal - what to look at instead: Tesla P100, Tesla P40, AMD BC250, and what happened to the CMP 170HX ## Chapters 0:00 Nine cards, 54 GB, $35 0:21 The real price 1:05 What a P106 actually is 1:35 Splitting a model across nine cards 2:10 Mixture of experts changes the math 2:39 Power, prompts and boot times 3:14 Driver and software limits 3:29 What to buy instead 4:15 Should you copy it? #llm #localai #qwen3 #vram #nvidia