If the goal is actual productivity (especially coding assistance), cloud subscription services far outperform running local large models on pure CPU hardware in terms of cost-effectiveness