Apple Silicon LLM Inference Costs ~3x More Than OpenRouter — Counter-Intuitive Energy and Cost Analysis — 119 Points

William Angel / Hacker News·medium signal

A detailed cost analysis published today (May 17) shows that running LLMs locally on Apple Silicon amortizes to approximately $1.50 per million tokens when accounting for hardware depreciation, while OpenRouter offers equivalent models (Gemma4 31b) at $0.38–0.50 per million tokens at 3-7x the speed (60-70 tok/s vs 10-20 tok/s). The Apple M3 Ultra delivers 23x more tokens per joule than an RTX 5090, but hardware cost dominates the equation. Privacy and offline use remain the primary justifications for local inference.