Physix Frontier · News Briefing Card (IT Home · Nov 21, 2026)
Lawyer builds AI cluster from 12 used V100s to run DeepSeek
KEY FACTS
- A lawyer posted on Reddit about building a private AI server from 12 used Nvidia V100 GPUs.
- The full hardware purchase cost about $9,450, roughly 63,239 yuan at current exchange rates.
- The cluster runs DeepSeek-V4.1-Flash locally, reaching a peak throughput of 184 tokens per second.
- The system uses a 3-stage pipeline parallel structure with 4-card tensor parallelism per stage, plus 4-bit NVFP4 quantization.
- In single-request scenarios, code generation is about 15% to 45% faster than a setup of four DGX Spark units combined.
KEY DATA
about $9,450Hardware purchase price
184 tokens per secondPeak throughput
about 2100WPeak power draw
about 15% to 45% fasterCode generation speed advantage
PHYSIX OBSERVATION
The used V100 cluster, at roughly half the cost of three DGX Spark units, delivers higher throughput in single-request scenarios, showing that old cards combined with quantization and parallelism strategies still offer good value. But its 2100W power draw and noise keep it out of office environments; this path suits technical users with space and power, not ordinary consumers.
Source: IT Home report
Physix Frontier