someone measured the real electricity cost of running AI at home. it's almost nothing

A developer wall-calibrated an Apple Silicon power meter against a smart plug and found the real number: fractions of a cent per million tokens. Aliteq Lena Fischer · AI & Local Compute Editor

Electricity cost per million output tokens, sustained generation: $0.063 (Qwen3.5-4B), $0.087 (Qwen3.6-35B MoE), $0.109 (gpt-oss-120B MoE), $0.554 (Qwen3.6-27B dense) — at a $0.31/kWh electricity…

The 120-billion-parameter MoE model is cheaper to run than the 27-billion-parameter dense model, because it only activates a fraction of its weights per token.

Over 30 days of real mixed traffic, the MoE models averaged $0.057-0.071/million tokens; the dense model averaged $0.78/million.

Cloud API comparison at mid-2026 prices: GPT-5.4-nano $1.25/million tokens, Gemini 2.5 Flash $2.50, Claude Haiku 4.5 $5.00 — 10 to 90 times the local electricity cost.

This measures marginal electricity only. It does not include the price of the Mac itself, which the author flags as 'by far the larger number.'

No. Whatever GPU or Mac you're running, electricity is very likely the smallest number in the whole equation — smaller than the hardware, smaller than your time, smaller than a single cloud API…

Is electricity a reason to worry about running local AI?

Read the full story

Aliteq someone measured the real electricity cost of running AI at home. it's almost nothing Read the full story on Aliteq