Lawyer builds private AI server with eight used Nvidia V100 GPUs

A lawyer on Reddit described a private AI server built from eight used 2018-era Nvidia Tesla V100 data center GPUs so case files never leave the office. The build uses two 4-card SXM2 NVLink boards, four PEX switches and SlimSAS cables, and had one burnt power connector; it now runs the open DeepSeek V4.1 Flash model with every expert in GPU memory at up to 215 tokens per second. The four-GPU boards cost about $2,400 apiece when purchased.