MacBook Pro M5 Max pushed to limits during transatlantic flight without Wi-Fi, testing local LLMs' viability for engineering work. Using Gemma 4 31B and Qwen 4.6 36B via LM Studio, the experiment focused on running Docker images, programming tasks, and data analysis offline. DuckDB powered a billing analytics tool analyzing two years of cloud spend, revealing hidden patterns.

Hardware constraints emerged immediately: sustained 70-80W draw caused overheating despite 128GB RAM and 40-core GPU. 60W power delivery from a faulty cable throttled performance, with battery draining 1% per minute under load. Model loops and latency spikes above 100k tokens highlighted local inference limitations for complex workflows.

Two custom tools emerged: powermonitor tracked real-time CPU/GPU/adapter metrics (peak 87.3W), while lmstats analyzed token throughput and context-window behavior. These mirrored enterprise-grade instrumentation practices, emphasizing measurement before optimization.

Results confirm local LLMs suit niche tasks like code refactoring and small-scale data work but struggle with high-value inference. The 34W cable discrepancy revealed hidden power constraints, with return flight testing planned using proper 94W charging. This hands-on exploration underscores the need for power-aware AI workflows and Apple Silicon's efficiency advantages.