Reddit user builds 192GB VRAM AI rig with 4 modded RTX 4090s on a dryer line

A resourceful AI enthusiast has built a monstrous local AI rig using four modded RTX 4090 GPUs, each upgraded to 48GB VRAM for a total of 192GB. The system also packs 128GB DDR5 RAM, a Pro WS WRX90E-SAGE SE motherboard, and a 3000W PSU. To power it, they tapped into a 240V/30A dryer circuit—not code compliant—requiring a dual-conversion UPS to handle nuisance GFCI trips. Heat is managed by the laundry room's exhaust fan, automatically kicking in at ~79°F. Noise is barely audible outside the room.

The user's primary use case is a private Jarvis-class assistant with premium voice capabilities: changing voices mid-turn for different characters, voice verification, wake words, continuous conversation, long-term memory, and Home Assistant/Hermes Agent integration. They've experimented with high-quantized models including Qwen 397B, MiniMax M3, Nemotron 3 Ultra, and GLM 4.7/5.2, but found Gemma 4 31B QAT surprisingly effective. More recently, MiMo V2.5 showed promise despite some looping issues. The rig demonstrates the growing trend of local AI powerhouses for serious hobbyist and professional AI development.