A lawyer built a custom AI server with eight Nvidia V100 GPUs to keep legal files secure
A lawyer has built an offline AI server using eight second-hand Nvidia Tesla V100 GPUs to process confidential legal documents without taking them off-site. The system is capable of running the DeepSeek V4.1 Flash model at speeds of up to 215 tokens per second.
Seeking to ensure the security of legal files, a lawyer shared a project on Reddit detailing how they assembled their own private AI server using eight second-hand Nvidia Tesla V100 data center GPUs.
The system was built using two 4-card SXM2 NVLink boards, four PEX switches, and SlimSAS cables. Despite technical challenges during the setup, such as a power connector burning out, the server is now able to run the DeepSeek V4.1 Flash model while keeping all expert parameters in GPU memory. This configuration offers processing speeds of up to 215 tokens per second.
According to the lawyer, each set of four 2018-era GPUs cost approximately $2,400. This custom setup aims to maximize data privacy by allowing sensitive legal documents to be processed locally without leaving the office network.
Sources
- tomshardware.com · Link