LLM MaaS: The Technical Logic Behind Revenue Growth and Gross Margin Improvement

    This article explains why the shift from project-based delivery to Model-as-a-Service (MaaS) can drive recurring revenue growth, and why that growth does not automatically improve profit margins. It covers ARR as a business health metric, the main cost drivers of model inference, and two paths to higher gross margin: inference efficiency and customer structure. It also outlines practical challenges and recommendations for teams building MaaS platforms.

    1. From One-Time Delivery to Continuous Usage

    Large model commercialization is moving from early project-based delivery toward Model-as-a-Service (MaaS). Project-based work centers on the needs of a single customer. It covers model adaptation, knowledge base construction, application development, system integration, and deployment and operations. Revenue is usually recognized upon contract acceptance or staged delivery.

    This model has clear strengths. Each contract carries a high average value, and it helps teams validate customer needs quickly. However, it also has limits. Project revenue is discontinuous, delivery cycles are long, and staffing demands are heavy. Revenue growth can depend heavily on a few large customers and large projects. When a company wants to scale, relying only on custom projects often requires expanding pre-sales, R&D, implementation, and operations teams in parallel. Marginal costs do not fall naturally.

    MaaS packages model capabilities as standardized services. Developers and application systems access them through APIs, online consoles, or other general-purpose service entry points. Capabilities include text generation, content understanding, vector representation, code assistance, and knowledge Q&A. Customers pay based on call volume, token consumption, or an agreed service period. The provider earns recurring revenue from continuous usage.

    The core change is that revenue comes from ongoing usage rather than from completing a project. Whether customers keep calling the model, whether call volume keeps growing, and whether applications go deeper into business processes have become key indicators of business quality.

    2. Why ARR Reflects Business Quality

    Annual recurring revenue (ARR) estimates the stable revenue a company may earn over the next year, assuming current customers and subscription relationships continue. It is not the same as the total revenue recognized in a given year. For platform businesses, however, it is a valuable metric.

    For MaaS companies, ARR can reflect business health in several ways:

    1. Whether customers keep using the service

    If a customer runs only a one-off test or a short project, revenue may spike in a given month but is hard to forecast. Continuous calls mean the model has been embedded in the customer’s products, operating processes, or internal systems. Switching costs and business dependency are usually higher in that case.

    1. Whether call volume can expand

    Enterprise customers often start by using the model in one department or one application. After the results are validated, usage may spread to customer service, marketing, office work, R&D, and data analysis. Growing call volume usually signals that the platform has room to scale horizontally.

    1. Whether revenue depends less on any single project

    As recurring usage revenue grows as a share of total revenue, the revenue mix becomes more diversified. When a large project ends, ongoing usage from other customers can still support revenue, which strengthens operational stability.

    According to public disclosures, a leading large model company reported 2025 revenue of over RMB 724 million, up 131.9% year over year. Its MaaS API platform reached ARR of about RMB 1.7 billion, roughly 60 times the prior-year level. ARR is an operating metric and should not be equated with revenue reported in financial statements for the period. Still, this change shows that platform call volume and customer usage depth are growing quickly.

    3. Why MaaS Can Grow Revenue Without Growing Profit

    The cost structure of MaaS differs from traditional software. After traditional software is developed, the marginal cost of serving an additional user is relatively low. Each inference call in a large model service, however, can consume compute, storage, and network resources, as well as the operations capacity needed to run them.

    Tokens are the main unit for measuring the size of text input and output. Larger call volumes usually mean more text is processed, which requires more computation for inference. If a company scales call volume without improving resource utilization, revenue can grow faster than cost optimization.

    The main factors affecting MaaS gross margin include:

    • Inference cost: Model parameter scale, context length, output length, and concurrency all affect the resources used per request.
    • Hardware utilization: Idle compute raises unit cost. Overloaded resources can slow response times and affect service quality.
    • Request mix: Short requests, long requests, batch jobs, and peak-hour traffic can differ significantly in cost.
    • Service tier: Low-latency, high-concurrency, and high-reliability services usually require reserving more resources.
    • Customer bargaining power: Large customers often have stronger pricing leverage and may push down unit prices.
    • Model version mix: High-performance and lightweight models differ in capability, price, and compute cost.

    Therefore, a MaaS platform should not focus only on revenue scale. It also needs to track metrics such as cost per million tokens, gross profit contribution per customer, resource utilization, request success rate, and customer retention.

    4. The First Path to Higher Gross Margin: Inference Efficiency

    When model services reach large-scale commercial use, algorithm optimization and engineering optimization both matter. Model quality determines how much customers are willing to pay. Inference efficiency determines how much of that revenue turns into profit.

    Common optimization directions include:

    1. Model Compression and Quantization

    Reduce compute consumption through parameter compression and low-precision computation, while preserving model quality as much as possible. For tasks that need real-time responses but have relatively limited complexity, a lighter model can handle the work. This avoids sending every request to an expensive model.

    2. Request Batching

    The service can group multiple requests received within a time window and process them together, improving compute utilization. Batching suits scenarios with flexible latency requirements, such as offline content generation, bulk document analysis, and data preprocessing.

    3. KV Cache and Prefix Reuse

    In multi-turn conversations, fixed system prompts, and repeated context, some input content is processed again and again. Caching reusable intermediate computation results reduces redundant computation and lowers the resource cost of long-context requests.

    4. Dynamic Routing

    The platform can route requests based on task complexity, customer service tier, latency requirements, and how well each model fits the task. Simple classification, summarization, or format conversion tasks do not need the highest-end model. Complex reasoning and high-value tasks can go to more capable models.

    5. Elastic Resource Scheduling

    Allocate compute dynamically according to peak and off-peak demand to reduce long-term idle capacity. Combined with container platforms, monitoring and alerting, and auto-scaling, this helps control infrastructure costs while maintaining service stability.

    5. The Second Path to Higher Gross Margin: Customer Mix

    Cutting costs alone cannot sustain gross margin improvement. Customer structure is another important variable.

    Low-value customers often have unstable call volumes, high price sensitivity, heavy support needs, and high levels of customization. Their actual revenue contribution is limited. These customers can consume significant pre-sales, support, and operations resources, so nominal revenue and actual profit may not match.

    High-value customers usually focus more on model stability, data security, service continuity, and business results. Once a model is embedded in core processes, these customers may be less sensitive to price than to availability and reliability. Large customers also provide more stable call volumes, which help the platform raise resource utilization and lower unit service costs.

    Therefore, a MaaS platform needs a customer tiering system:

    • For trial customers, provide standardized capabilities with clear usage limits.
    • For growing customers, provide complete development tools, call monitoring, and usage analytics.
    • For key enterprise customers, prioritize stability, access control, data isolation, and service continuity.
    • For highly customized, low-reuse projects, evaluate the real gross margin to avoid overinvestment.

    Customer tiering does not simply mean raising prices. It means designing a sound commercial mix based on service tier, call volume, support investment, and business value.

    6. From Project Revenue to Platform Revenue: Changes in Business Operations

    In the project phase, companies typically focus on contract value, delivery timelines, and project acceptance. After moving to Ma