Lawyer builds private AI server with eight used Nvidia V100 GPUs
Reported by Tom's Hardware1 source ↓
A lawyer runs DeepSeek V4.1 Flash on a locally built server to process case files without sending them to cloud models. The system is used for first-pass research and routine drafting, with human review before work is sent out.
Key points
- The server handles first-pass work on case files, including extracting facts and preparing routine drafts
- The lawyer estimates 50–60% of existing LLM workflows can run on the server
- DeepSeek V4.1 Flash reaches up to 215 tokens per second for code and 118 for legal writing
- The setup handles one request at a time and can take about a minute to start responding to a 128K-token prompt
- The system uses two GPU boards, and a burnt power connector and hardware errors required troubleshooting
Details
Tom’s Hardware reports that the lawyer assembled the server from eight used Nvidia Tesla V100 GPUs and runs DeepSeek V4.1 Flash with the model’s experts in GPU memory. It is used daily to review folders of case records, extract facts, dates and figures, summarize material, and draft routine documents. The lawyer says nothing is sent out without human review.
In the builder’s measurements, code generation reached up to 215 tokens per second and legal writing up to 118. A 128K-token prompt took about 59 seconds to produce its first token; the system supports context up to 524K tokens. The lawyer says the setup currently handles one request at a time.
The system divides the model’s layers between two four-GPU boards, which communicate over PCIe. The builder reported that spreading layers across all eight cards had previously reduced performance, and that the server experienced fatal PCIe errors and a burnt power connector during setup.
Written from coverage as of
Images
- Tom's Hardware
Stocks in this story
Change since open: 20 Nov to 21 Nov 2026
Last refresh