Apple Silicon Local LLM Inference Costs 3x More Than Cloud APIs
William Angel published a data-driven cost comparison of running local LLM inference on Apple Silicon versus using cloud APIs through OpenRouter. The results challenge a common assumption among local-first advocates: cloud is both cheaper and faster.
Using a $4,299 M5 Max MacBook Pro with 64GB RAM, Angel amortizes hardware costs across 3, 5, and 10-year lifespans. On a Gemma 4 31B model, local inference runs at 10-40 tokens/second, costing $0.40 to $4.79 per million tokens when combining electricity and hardware depreciation. Meanwhile, comparable models on OpenRouter cost just $0.38-0.50 per million tokens, with inference speeds of 60-70 tokens/second.
The real multiplier, however, is human time. For an engineer earning six figures, slower local inference extends development cycles by enough to dwarf the marginal token cost. As Angel puts it: “Throwing money at Anthropic makes more sense.”
This piece is valuable because it replaces instinct with data. For AI infrastructure procurement decisions, developer toolchain design, and model deployment strategy, this kind of cost transparency is critical — especially as agentic applications increasingly depend on high-frequency inference calls.