A lawyer posted on Reddit about their private AI server built from eight used 2018-era Nvidia Tesla V100 data center GPUs so case files never have to leave the office. With two 4-card SXM2 NVLink boards, four PEX switches, and SlimSAS cables everywhere — plus one burnt power connector — the builder 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 the builder bought them.
Get Tom's Hardware's best news and in-depth reviews, straight to your inbox.