How Much Does a Local LLM Actually Cost to Run? I Measured Every Watt on Apple Silicon

Running a large language model on your own computer reveals that cost is driven by processing throughput, not simply the model's size. A detailed analysis on Apple Silicon hardware showed a 120-billion-parameter model was five times cheaper per token than a much smaller 27-billion-parameter model. This surprising efficiency occurs because Mixture-of-Experts (MoE) models only use a small fraction of their total parameters for any given token, resulting in faster speeds and lower power use. Conversely, dense models must read all their weights for every token, which dramatically slows them down and increases electricity consumption, making them the most expensive option.