Redditor builds dual water-cooled V100 workstation with 128GB VRAM for ~$1,900
a4mula · reddit · 2026-11-01
A Reddit user details their quest to upgrade local LLM inference hardware amid a GPU price crunch:
- Started with a 3060 Ti + 32GB RAM running local models like Qwen, but adding a GPU proved unrealistic across Offerup, Marketplace, and eBay.
- An older C602 workstation + P100 plan came to $1,500 but capped at 48GB VRAM on DDR3/PCIe3.
- Final buy: a Z840 workstation with 128GB DDR4 and dual water-cooled SXM2-to-PCIe 32GB V100s for $1,900 after tax — the cards alone fetch $700–800 each, plus $1,000 in custom water cooling, risers, and adapters.
- Remaining quirks: no native m.2/NVMe (SSD limited to 3,500MB/s via PCIe3 riser) and no video output on V100s, so a headless setup is needed. Benchmarks to follow once it's running.
More from Infra
- Reddit cost breakdown: one agent loop burned 1B tokens/day — on Opus that's hundreds of dollars — OvertaxedOne · 2026-11-01
- Nasdaq 100 on track to lose over $2 trillion in October as AI spending spooks investors — nordicinst · 2026-11-01
- Friedberg: $1 of Nvidia COGS becomes $18 of Anthropic revenue via SpaceX margin — firstadopter · 2026-11-01
- Meta could offload 600K GPUs as 40% of H100/H200 capacity sits idle, insider claims — SumitGup · 2026-11-01
- Beginner Builds a 3-Tier Agent Chain: Grok Plans, Claude Executes, Ollama Does Grunt Work — Previous-Taro-1648 · 2026-11-01
- Google Cloud Assist reply mixes in Japanese text, user jokes it's a Chinese model — freestyler_returns86 · 2026-11-01