LMaaS: Exploring Pricing Strategy of Large Model as a Service for Communication

Panlong Wu, Qi Liu, Yanjie Dong, Zhaorui Wang, Fangxin WangOpenReview 2026 · OpenReview

TL;DR

This paper introduces "Large Model as a Service" (LMaaS) for semantic communication, proposing a Stackelberg game framework to optimize pricing and rental strategies between a service provider and heterogeneous, dynamic customers. They develop novel algorithms, IMP for the seller and RSR for customers, to maximize profits under uncertainty, achieving near-optimal seller profits and significant gains for customers.

What Problem Does This Paper Address?

The paper addresses the emerging field of semantic communication (SC), which aims to transmit condensed semantic information rather than raw data for increased efficiency. This is enabled by powerful large language models (LLMs). The authors envision a "Large Model as a Service" (LMaaS) model where LLMs are offered as a pay-as-you-go service. However, a significant challenge lies in designing effective pricing strategies for this service. This is complicated by:

  • Heterogeneous Customers: Users have varying needs, budgets, network conditions (distance, energy), and application requirements.
  • Dynamic Environments: Customer environments are not static; factors like transmission distance and energy availability can change, making profit guarantees difficult.
  • Complex Market Dynamics: The interaction between a single seller and multiple buyers, each with their own optimization goals, creates a complex game theory problem.

Existing pricing strategies often fail to adequately account for these environmental dynamics and customer heterogeneity, leading to suboptimal profits for both parties.

Methods

The core of the paper's contribution lies in its novel algorithmic approaches within a Stackelberg game framework:

  • Stackelberg Game Formulation: The interaction is modeled as a two-step game where the seller (leader) sets prices first, and then customers (followers) make their selection and rental decisions based on these prices. This captures the market's hierarchical nature.
  • Iterative Model Pricing (IMP) Algorithm (Seller): This algorithm is designed to optimize the seller's pricing strategy. It iteratively adjusts prices by considering customers' predicted rental decisions and their own profit feedback. The goal is to find a near-optimal pricing strategy that maximizes the seller's revenue while accounting for model training and operational costs.
  • Robust Selecting and Renting (RSR) Algorithm (Customers): This algorithm addresses the customers' profit maximization problem under uncertain environmental conditions. It employs a two-stage robust optimization approach to guarantee optimal selection and rental duration even in the worst-case scenarios. The RSR algorithm decomposes the problem into a master problem (MP) and a subproblem (SP), iteratively solving them.

Datasets

Not discussed in the paper. The paper focuses on theoretical modeling and simulation-based experiments using synthetic data generated according to specified parameters.

Key Insights and Ideas

  • LMaaS as a Viable Model: The paper strongly advocates for the LMaaS model, highlighting its suitability for the "train once, use widely" nature of LLMs in communication.
  • The Importance of Robustness: For customers, simply optimizing for average conditions is insufficient. A robust approach that accounts for worst-case environmental dynamics is crucial for guaranteeing profit.
  • Interplay of Pricing and Selection: The seller's pricing directly influences customer behavior, and the seller must anticipate this to set optimal prices. Conversely, customers' ability to adapt their choices based on prices is key to their profit maximization.
  • Complexity of Heterogeneity: The experiments demonstrate that customer profit is significantly impacted by factors like the number of customers, available models, transmission distances, and energy budgets, underscoring the need for tailored strategies.
  • Seller's Profit Sensitivity: The seller's profit is shown to be sensitive to the number of customers and, to a lesser extent, the number of available models and transmission distances.

Key Results

  • Customer Profit: The RSR algorithm consistently outperforms baseline methods (Greedy and StaticEnv-Opt) in maximizing customer profit. It achieves significant gains, up to 43.96% higher profit than Greedy and 10.18% higher than StaticEnv-Opt under certain conditions (e.g., longer transmission distances).
  • Seller Profit: The IMP algorithm achieves near-optimal seller profits, with only a small loss (e.g., 0.67% with one customer, 1.01% with five customers) compared to a computationally intensive Near-Opt baseline.
  • Efficiency: The IMP algorithm offers substantial computational savings, reducing execution time by up to 93.05% compared to the Near-Opt baseline while maintaining comparable profit.
  • Robustness: Experiments show that both RSR and IMP algorithms are robust to variations in parameters like random seeds and the number of initial prices.

Limitations and Future Work

  • Assumptions: The paper makes several simplifying assumptions, such as each customer owning a single device (though it notes this can be extended) and specific functional forms for utility and cost.
  • Computational Complexity: While IMP is efficient, the robust optimization in RSR can still be computationally intensive, especially with a large number of customers and models.
  • Future Work: The authors suggest extending the model to more complex scenarios, such as multiple sellers, different types of customers, and more sophisticated SC model customization processes. Further research into dynamic pricing adjustments and real-world deployment challenges would also be valuable.

Summary updated

Abstract

One of the most important features of next-generation communication is to incorporate intelligence towards semantic communication, where highly condensed semantic information considering both source and channel features will be extracted and transmitted. The recent popular large models such as GPT4 and the boosting learning techniques are envisioned to accelerate its practical implementation in the near future. Given the characteristics of “training once and widely use” of those multimodal large language models, we argue that a pay-as-you-go service mode will be suitable in this context, referred to as Large Model as a Service (LMaaS). However, the trading and pricing problem is quite complex with heterogeneous and dynamic customer environments, making the pricing optimization problem challenging in seeking on-hand solutions. In this paper, we optimize the profit of both the seller and customers. We formulate the LMaaS market trading as a Stackelberg game with two steps. In the first step, we optimize the seller's pricing decision and propose an Iterative Model Pricing (IMP) algorithm that optimizes the prices of large models iteratively by reasoning customers’ future rental decisions, which achieves a near-optimal pricing solution. In the second step, we optimize customers’ selection decisions by designing a robust selecting and renting (RSR) algorithm, which is guaranteed to be optimal with rigorous theoretical proof. Extensive experiments confirm the effectiveness and robustness of our algorithms, outperforming the state-of-the-art solution by 43.96% in profit at the customer side and achieving the near-optimal profit at the seller side.
BibTeX
@misc{wu2026lmaas,
  author = {Panlong Wu and Qi Liu and Yanjie Dong and Zhaorui Wang and Fangxin Wang},
  title = {{LMaaS: Exploring Pricing Strategy of Large Model as a Service for Communication}},
  year = {2026},
  url = {https://openreview.net/forum?id=jLe7JekuXh},
  howpublished = {OpenReview},
}
Figure