{
  "id": 20901766,
  "title": "Lawyer’s $9,450 private AI server with 12 used Nvidia V100s now runs every DeepSeek V4.1 Flash expert",
  "url": "https://urgent.news/2026/11/20/lawyers-9-450-private-ai-server-with-12-used-nvidia-v100s-now-runs",
  "topic": "ai",
  "section": "AI",
  "published": "2026-11-20T12:30:00.000Z",
  "source": {
    "name": "Tom's Hardware",
    "slug": "tom-s-hardware",
    "url": "https://www.tomshardware.com/tech-industry/artificial-intelligence/lawyers-usd9-450-private-ai-server-of-12-used-nvidia-v100s-now-runs-every-deepseek-v4-1-flash-expert-at-4-bit-nvfp4-up-to-184-tokens-per-second-on-code-and-long-documents-start-answering-in-about-half-the-time"
  },
  "original_language": "en",
  "account": "A lawyer has upgraded his private AI server by replacing eight used Nvidia V100s with twelve, boosting its capacity to run every DeepSeek V4.1 Flash expert at 4-bit NVFP4 precision. This configuration can process up to 184 tokens per second for code, while handling long documents approximately half as quickly as before. Priced at $9,450, the upgraded server outperforms published DGX Spark stacks for decoding a single request, delivering twice the performance of three Sparks at half the price. The server can read a 128K file in 28 seconds and a 256K file in 185 seconds. The Nvidia V100 GPUs lack FP4 or FP8 tensor cores, so every NVFP4 number must be unpacked into half precision, which, despite the inefficiency, allows the cards to handle larger data sets. The experts for DeepSeek V4.1 Flash model require 278.9 GiB of memory, surpassing the 15,360 experts that the original eight V100s could accommodate. To meet this demand, Mermiges upgraded his server to include twelve V100 GPUs, with four on each board, arranged in three groups. Each section is distributed across the GPU boards via Nvidia's fast link, NVLink, with handoffs between boards occurring over the host's PCIe via two separate PCIe switches. The power settings were adjusted, raising each card's power cap from 150 W to 175 W and lowering its clock cap from 1,522 MHz to 1,452 MHz, yielding noticeable performance improvements. Despite optimizations like a conversation store, tighter expert packing, and precise arithmetic, some improvements could not be implemented due to unmet targets. Mermiges has been using this optimized DeepSeek machine since Monday, with the model performing as well as Opus 4.6.",
  "summary": "A lawyer’s private AI server now runs DeepSeek V4.1 Flash on 12 used Nvidia V100 GPUs, with every expert at 4-bit NVFP4.",
  "key_points": [],
  "editors_take": null,
  "illustration": null,
  "coverage": {
    "outlets": 1,
    "also_reported_by": []
  },
  "ai_generated": true,
  "disclaimer": "Summaries, key points and the editor’s take are written by software from other outlets’ reporting and may contain errors — always check the linked original."
}