Lawyer Builds Private AI Server With Eight Nvidia V100s

A lawyer builds a private AI server around eight used Nvidia Tesla V100 GPUs capable of reaching up to 118 tokens per second on legal writing, according to a Reddit build described by Tom’s Hardware on Nov. 9. By that builder’s estimate, 50% to 60% of their existing large-language-model workflows now run locally instead of going to frontier models.

The server earns its keep daily: extracting facts, dates and figures from case files, summarizing records, drafting routine material. Files stay inside the office. Every output still gets a human review before it goes anywhere.

This is a practical experiment in keeping legal AI work off external services, not a commercial product launch. The builder called it mostly a hobby, floated a possible business idea, but offered no timeline. The server is already handling daily tasks; running multiple requests at once remains unfinished work.

What the Lawyer Builds the Private AI Server to Do

The system runs the open DeepSeek V4.1 Flash model. Its workflow pairs document review with retrieval databases and tool calls; the lawyer reaches it through a Discord-style interface over SSH from a Windows PC. One request at a time.

In practice, it reads a folder of case records, pulls out facts, dates and numbers, summarizes what it finds, and produces an initial draft for routine work. The retrieval databases let the model work with relevant material rather than a bare prompt. The current setup is strictly sequential — it cannot serve several lawyers or jobs in parallel.