Lawyer builds 256GB local LLM lab

Hot take: this is less a "PC build" and more a private AI lab, and the real bottlenecks are power, cooling, and operational complexity, not raw VRAM.

• –256GB of VRAM is legitimately impressive, but for legal workflows the bigger win is usually retrieval quality, citation discipline, and reproducibility.
• –Trying to chase the biggest distributed model can turn into an engineering hobby of its own; the simplest reliable stack often beats a more ambitious cluster.
• –QLoRA could help domain adaptation, but only after the user has a strong eval set and a clean local RAG pipeline.
• –The power and cabling problems are not cosmetic here; they are core reliability risks that will affect uptime, heat, and maintenance.
• –For privacy-sensitive work, the "keep it local" strategy makes sense, but model choice should be guided by latency, context length, and ease of serving rather than hype.