User builds $3k inference server with 128GB VRAM

A Reddit user posted a DIY inference server built for about $3k featuring 4×V620 GPUs (128GB VRAM), 256GB DDR4 RAM, EPYC 7452 CPU, and a Samsung 970 EVO SSD; it runs Qwen3.8-next with usable performance but high power draw (500–900W).

• Qwen3.8 27b game dev post2026-09-16
• Hope for KVCache + Engram in Upcoming Models2026-09-14
• CyberTiel 35B-A3B uncensored quant excels at coding2026-09-11