Local inference economics: $60/month power bill for slow speeds

Thin_Pollution8843 · reddit · 2026-08-19

A real-world cost analysis of local inference on a Threadripper 3975WX with 4x V620 GPUs reveals a monthly electricity cost of $55-60 for 6 hours daily use. Despite this, performance with Qwen3.8-27B is sluggish (1-1.3k prefill, 30ts). The author argues that cloud services like OpenRouter or ChatGPT offer better value unless privacy is paramount or solar power is available.

Original post →

More from Infra

Infra channel →