Exploring User Heterogeneity-Aware Differentiated Token Pricing for On-Premises Large Language Models

With the growing adoption of locally deployed large language models in companies, designing an effective token pricing strategy that simultaneously incentivizes engagement across heterogeneous user groups and maximizes platform profit has become an urgent challenge. This work addresses this issue by developing a theoretical model for token-based billing that incorporates user heterogeneity and platform costs to systematically analyze the principles for designing an optimal pricing strategy. By introducing a tiered package mechanism that combines fixed access fees with stepwise per-unit token rates, we propose an incentive-compatible, differentiated token pricing strategy and a practical pricing process. Extensive simulation experiments are conducted across centralized and log-normal user type distributions. The results demonstrate that under high user heterogeneity, the proposed strategy can increase platform profit by more than 15% compared with uniform linear pricing, while maintaining the participation of low-demand users. Furthermore, this work analyzes the impact of the cost functions, capacity constraints, and adaptive update policies on pricing performance, validating the robustness of differentiated pricing strategy across various scenarios. These findings offer both a theoretical foundation and a practical pathway for providers of on-premises LLM services to develop commercial pricing strategies.