Local runs of models like Gemma 4 31B on M5 Max hardware incur amortized costs of $0.40–$4.79 per million tokens after hardware depreciation, versus $0.38–$0.50 via OpenRouter. The M5 Max draws 50–100 W under load at $0.18–$0.20 per kWh, yielding electricity costs of roughly $0.02 per hour, yet the $4,299 device price dominates when spread over three to ten years of service life. Token throughput measured at 10–40 tokens per second produces 36,000–144,000 tokens per hour, insufficient to offset capital recovery. Primary data from the source and EIA residential electricity tables show hardware depreciation at five-year life equals $0.098 per hour, exceeding electricity by a factor of five. Cloud providers achieve higher sustained throughput of 60–70 tokens per second on the same model class, cutting effective cost per token while shifting energy load to hyperscale facilities with reported PUE below 1.2. The original analysis understates utilization variance: consumer laptops rarely sustain 100 % inference duty cycles, inflating per-token hardware allocation beyond the modeled range. Real-world benchmarks from Apple ML Performance Reports and EIA 2025 averages confirm that efficiency gains in silicon do not translate to lower total ownership cost when inference demand is intermittent.