Author benchmarks Qwen3.6-35B-A3B local inference power cost

An author benchmarked running a large local language model, the Qwen3.6-35B-A3B, on an RTX 4070 Ti Super PC paired with a Ryzen CPU to determine the real-world power cost. By tracking peak power draw and assuming worst-case continuous generation rates, the author found that local AI inference costs a fraction of a single $20 monthly cloud subscription. Open article →