MaaS Revenue Growth and Gross Margin Expansion: Technical Paths to Sustainable Profitability

    Model-as-a-Service (MaaS) is shifting AI commercialization from one-off delivery projects to recurring API consumption. This article examines why recurring revenue and ARR matter, why MaaS platforms can experience revenue growth without profit growth, and how inference optimization, customer segmentation, cost accounting, and platform governance can improve unit economics and long-term profitability. It also provides practical guidance for developers, enterprise users, and MaaS platform operators.

    As of December 31, 2025, a leading large-model company reported annual revenue of more than RMB 724 million, representing year-over-year growth of 131.9%. Its revenue increased from approximately RMB 57 million in 2022 to RMB 125 million in 2023 and RMB 312 million in 2024, maintaining rapid growth over several consecutive years.

    Compared with one-time project delivery, model and service offerings are becoming a major path for large-model commercialization. Enterprises and developers continuously call models through APIs, while platforms charge based on input and output tokens, request volume, or service tiers. Revenue is therefore gradually shifting from one-off contracts to recurring consumption.

    Notably, the company’s model-service API platform generated approximately RMB 1.7 billion in annual recurring revenue, or ARR, representing an approximately 60-fold year-over-year increase. ARR is not the same as revenue recognized during the current accounting period. Instead, it is used to estimate the future recurring revenue capacity of existing customers and subscription-oriented businesses.

    For a MaaS platform, ARR is particularly useful because it indicates whether customers have embedded models into their business processes and whether model usage has become stable and repeatable.

    1. From Project Revenue to Recurring Model Consumption

    In the early stages of commercialization, large-model companies often rely on customized projects, including model adaptation, industry knowledge-base development, private deployment, and application integration. These projects may have relatively high contract values, but they also have three clear limitations.

    First, project revenue is largely one-time revenue. After acceptance, additional revenue often requires a new contract. The provider must continue investing in sales and delivery resources to maintain its business volume.

    Second, project delivery can create significant fluctuations in resource utilization. A company may need to concentrate algorithm engineers, implementation staff, and computing resources during a particular period. Once the project ends, some of these resources may become underutilized.

    Third, project-based pricing does not accurately reflect the value created by model usage. Customers pay for implementation and construction rather than for the actual business value generated by the model. The platform also has fewer opportunities to continuously improve its products through usage data.

    MaaS works differently. Customers connect to models through standardized interfaces, and applications continuously generate requests during operation. As long as the model continues to support customer service, content generation, coding assistance, data extraction, intelligent review, or enterprise knowledge retrieval, the platform can generate recurring revenue.

    The key change is that growth is no longer driven primarily by “how many projects have been signed.” Instead, it depends on “how many customers continuously use the model.” Platforms therefore need to monitor metrics such as active customer count, customer retention, request frequency, token consumption per customer, depth of model usage, and the proportion of high-value customers.

    2. Why ARR Provides a Better View of Business Quality

    Annual revenue primarily reflects operating results recognized during a completed accounting period. ARR is more closely related to business continuity and predictability.

    In a MaaS business, ARR can be understood as an annualized estimate based on the current usage level of a group of stable customers. It is generally based on subscriptions, contractual commitments, or recurring API consumption. ARR is not a formal accounting revenue metric and should not be treated as equivalent to cash flow. However, it is useful for evaluating customer stickiness and the maturity of a business model.

    If revenue growth depends mainly on a small number of large projects, revenue may rise sharply in certain quarters while customer count, usage frequency, and renewal rates remain unchanged. By contrast, sustained ARR growth generally suggests that:

    1. Customers have embedded models into real production workflows rather than limiting usage to testing;
    2. Model calls have shifted from experimental use to routine, systematic use;
    3. Customers have developed a degree of dependence on model quality and reliability;
    4. The API platform has stronger standardized delivery capabilities;
    5. Revenue forecasting and compute-resource planning have become more manageable.

    However, ARR should not be evaluated separately from gross margin and customer quality. If a platform uses aggressive discounts to increase usage, ARR may grow while compute, bandwidth, and operations costs rise at the same time.

    Healthy MaaS growth should therefore satisfy three conditions simultaneously: revenue growth, improved usage retention, and better unit economics.

    3. Why MaaS Can Produce More Revenue Without More Profit

    MaaS revenue is linked to model usage, and model usage directly creates inference costs. Every request may consume computing resources. Long contexts, complex reasoning, multi-turn conversations, and high-concurrency workloads can significantly increase the cost of each request.

    If a platform focuses only on increasing request volume without improving inference efficiency, several problems may occur:

    • More customer usage leads directly to higher compute consumption;
    • Low pricing increases revenue but fails to cover marginal costs;
    • Long text and repeated context create large amounts of unnecessary computation;
    • Additional resources must be reserved during peak periods, reducing utilization;
    • Differences in cost across customers are not accurately measured, resulting in cross-subsidization.

    The central question for a MaaS platform is therefore not simply whether it can sell more tokens. It is whether each unit of token consumption can generate a reasonable gross margin while maintaining service quality.

    Gross margin can be simplified as follows:

    Gross margin = (Revenue − Direct costs) ÷ Revenue

    For model services, direct costs generally include inference compute, storage, network transfer, operations and maintenance, and certain platform infrastructure costs. If the revenue generated by a request is P and the unit cost is C, the platform must keep P consistently above C and widen the gap through engineering improvements.

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

    Inference efficiency is the foundation of MaaS profitability. Platforms can reduce unit costs through model architecture, service scheduling, and application engineering.

    4.1 Provide Tiered Model Services

    Different tasks do not require the same level of model capability. Simple classification, summarization, format conversion, and structured extraction may not require the largest available model. A platform can provide multiple capability tiers, allowing lower-cost models to handle simple requests while routing complex tasks to more capable models.

    This is not simply a matter of replacing a high-end model with a low-end model. The objective is to balance cost and quality through task routing. For example, the system can first determine whether a request involves complex reasoning, a long context, or multi-step planning, and then select the appropriate model. Accurate routing can reduce unnecessary calls to high-cost models.

    4.2 Improve Batching and Resource Utilization

    Online inference often experiences significant traffic fluctuations. If every request occupies compute resources independently, resources may sit idle during off-peak periods and queues may form during peak periods. Dynamic batching, request aggregation, and elastic scheduling can improve hardware utilization.

    The basic idea behind dynamic batching is to process multiple requests together while meeting latency requirements. For throughput-oriented workloads, this approach reduces idle capacity and increases the useful output generated by each unit of compute.

    4.3 Reduce Unnecessary Context

    Long contexts increase computational requirements. Repeatedly sending conversation history, fixed prompts, and redundant documents also increases token consumption. Platforms can reduce unnecessary content through context trimming, summary compression, cache reuse, and segmented retrieval.

    In an enterprise knowledge-answering scenario, the system should not send an entire knowledge base to the model. It should retrieve relevant information first and provide only the content related to the current question. This approach reduces inference costs and helps prevent irrelevant information from affecting output quality.

    4.4 Control Output Length

    Longer outputs generally cost more, but longer responses do not always provide more value. Platforms can set appropriate output limits at the application layer and control token consumption through structured outputs, result compression, and staged generation.

    For example, a data-extraction task may require only fields and values rather than a long explanation. A classification task may need only a category and confidence information. Matching the output format to the actual business requirement can directly improve request-level economics.

    5. The Second Path to Higher Gross Margin: Increase the Share of High-Value Customers

    Customer composition also has a major impact on gross margin. Although different customers may all use an API, their service costs, willingness to pay, and business value can differ substantially.

    Lower-value customers often have unstable usage, high price sensitivity, and significant support requirements. The platform may need to invest considerable presales and operations resources while generating limited revenue. High-value customers usually have more stable usage, stronger renewal intentions, and greater business dependence on the service, allowing them to generate longer-term revenue.

    Platforms should therefore establish customer segmentation based on at least the following factors:

    • Monthly and annual request volume;
    • Usage growth trends;
    • The importance of model services to the customer’s business;
    • Requirements for reliability, latency, and compliance;
    • Technical support and customization requirements;
    • Contract duration and renewal probability;
    • Service cost relative to revenue.

    Based on these factors, customers can be grouped into trial, growth, scale, and strategic segments. Each segment should receive an appropriate resource-allocation model.

    For trial customers, the priority should be standardized documentation, sample code, and automated diagnostics. For high-volume customers, the focus should be capacity planning, service levels, disaster recovery, and cost transparency.

    The purpose of customer segmentation is not simply to raise prices. It is to align pricing, service levels, and resource investment. Customers that require high reliability, low latency, or dedicated resource guarantees should receive more appropriate service tiers and commercial pricing. Otherwise, high-cost services may remain permanently covered by low-priced packages.

    6. MaaS Platforms Need More Granular Cost Accounting

    Many platforms initially track only total compute spending. They cannot answer questions such as which model, customer, or request type generates the highest profit. As the business scales, this simplified accounting approach can limit operational decision-making.

    A more effective approach is to establish a request-level cost allocation system that links the following information:

    • Customer identifier;
    • Model type;
    • Input and output token counts;
    • Request latency and retry count;
    • Compute resources used;
    • Peak and average concurrency;
    • Service tier;
    • Actual revenue;
    • Direct cost per request.

    With this data, the platform can calculate gross margin at the customer, model, and use-case levels. For example, a customer with very high usage may generate lower actual gross margin if most requests involve long contexts and low pricing. Another enterprise customer with lower volume but higher unit pricing may be more profitable.

    Cost accounting can also help identify abnormal consumption. If an application generates duplicate requests, unnecessary retries, or unusually large prompts, the platform should issue alerts. It can then use rate limiting, caching, or request-strategy optimization to reduce waste.

    7. From Technical Platform to Sustainable Business Infrastructure

    A MaaS platform is more than an API gateway. A mature platform also needs comprehensive governance capabilities, including authentication, access control, usage tracking, quota management, audit logging, service monitoring, incident alerts, and version management.

    Observability is especially important. The platform must know more than whether a request succeeded. It should also identify:

    • Which stage caused a request failure;
    • Whether latency came primarily from queuing, inference, or networking;
    • Which model versions increased costs;
    • Whether customer usage shows abnormal fluctuations;
    • Which business scenarios generate the most unnecessary tokens;
    • Whether a model upgrade improves both quality and cost.

    Enterprise environments also require data isolation, access control, and sensitive-information protection. Customers will be more willing to move core workloads to model services only when the platform can balance security, reliability, cost, and output quality.

    8. Practical Implications for Developers and Enterprise Users

    Developers evaluating a MaaS platform should not compare model prices alone. They should also assess cost per completed task, response stability, output quality, rate-limiting policies, and failure-recovery capabilities. A service with a lower listed price may have a higher actual cost if it requires frequent retries.

    Enterprise users should establish token budgets and usage-governance mechanisms during the system-design stage:

    1. Set usage quotas for different business functions;
    2. Separate test, production, and high-priority requests;
    3. Preprocess and compress long text;
    4. Use caching to reduce duplicate calls;
    5. Track model output quality and business-conversion results;
    6. Regularly evaluate model performance against unit cost;
    7. Avoid routing every task to the highest-capability model by default.

    For platform builders, the priority should not be request volume alone. The goal is to create a closed loop connecting model capability, inference efficiency, customer retention, revenue quality, and cost control.

    ARR growth has stronger commercial meaning only when customers continue using the platform, unit costs decline, and service quality improves consistently.

    Conclusion

    Large-model commercialization is moving from a phase focused primarily on model capability to one that gives equal weight to technical capability and operating efficiency. Revenue exceeding RMB 700 million and API-platform ARR of approximately RMB 1.7 billion indicate that recurring, consumption-based services are beginning to scale.

    However, the long-term profitability of MaaS still depends on three core capabilities:

    1. Continuously improving model quality;
    2. Reducing the cost of every inference request;
    3. Attracting and retaining high-value customers willing to pay over the long term.

    Only platforms that improve both technical efficiency and commercial governance can avoid the trap of growing usage while allowing costs to consume the resulting profit.