MaaS Monetization: How Foundation Model Companies Can Grow Revenue and Profit

    Foundation model providers are moving from capability demonstrations to commercial validation. Using the latest annual results of a listed model company as a starting point, this article examines the shift from project-based delivery to Model-as-a-Service, why annual recurring revenue (ARR) is a useful but incomplete signal of durable demand, and why higher usage can still pressure margins. It then explains how inference efficiency, resource utilization, customer mix, outcome-based services, developer tooling, and platform reliability affect unit economics, and outlines the capabilities that will shape long-term MaaS competitiveness.

    On March 31, a foundation model company released its first annual financial report since going public. For the year ended December 31, 2025, the company reported revenue of more than RMB724 million, representing year-over-year growth of 131.9%. Its reported revenue was RMB57 million in 2022, RMB125 million in 2023, and RMB312 million in 2024. Revenue has therefore increased every year, and the growth rate has accelerated.

    From an industry perspective, these figures reflect more than one company’s operating changes. They also indicate that the foundation model industry is moving from a capability demonstration phase into a commercial validation phase. In the past, the market focused more on parameter scale, benchmark scores, and release speed. Today, whether companies can generate recurring revenue, control inference costs, and improve customer retention has become a key measure of commercial quality.

    1. From Project Revenue to Model-Call Revenue

    The common commercialization paths for foundation model companies generally fall into two categories.

    The first is project-based delivery, including model deployment, industry solutions, custom development, private deployment, and technical services. This type of business often has high contract values, but revenue recognition tends to depend on the project cycle and can be affected by procurement processes, delivery schedules, and customer budgets. Projects may also lack continuity, requiring companies to keep finding new orders to sustain growth.

    The second is Model-as-a-Service, commonly referred to as MaaS. The company provides model capabilities to developers and enterprise customers through standardized APIs, and customers pay based on call volume, input and output tokens, or other service dimensions. Compared with one-off projects, MaaS is closer to a software service model: as long as customers continue using the model, the platform can generate relatively stable recurring revenue.

    In this financial report, what matters is not only the full-year revenue scale, but also the MaaS API platform’s annual recurring revenue of approximately RMB1.7 billion. ARR is not equivalent to revenue recognized in the current period, but it can be used to observe customers’ current capacity for continued use. For a model platform, ARR generally corresponds to stable call frequency, an active customer base, and ongoing token consumption.

    This means that the growth logic is changing: revenue no longer depends mainly on a small number of large projects, but is increasingly built on sustained model use by a large customer base.

    2. Why ARR Is Especially Important for MaaS

    Traditional software companies often measure business continuity through subscription fees, license fees, or renewal rates. Although MaaS platforms use different billing methods, they still need to answer three questions:

    1. Are customers continuing to call the model?
    2. Is usage scale expanding?
    3. Are customers willing to move more business processes onto the model platform?

    If customers call the model only during testing, or use it only briefly within a single project, platform revenue is likely to fluctuate significantly. By contrast, if the model is embedded in production workflows such as customer service, knowledge retrieval, content moderation, code assistance, data analysis, and intelligent office applications, usage has a better chance of forming long-term growth.

    Therefore, ARR for a MaaS platform is not only a financial metric; it can also be seen as a combined result of customer usage depth. It reflects not “how many customers have registered,” but “how many customers are continuously consuming model capabilities.”

    However, ARR should not be viewed in isolation. Companies also need to evaluate it alongside the following indicators:

    • The number of paying customers and its growth trend;
    • Monthly or quarterly customer retention rates;
    • Average revenue per customer;
    • Average usage volume and call frequency per customer;
    • Revenue concentration among large customers;
    • Conversion rate from trial to paid usage;
    • The proportion of model-call revenue versus project-delivery revenue.

    Only when these indicators improve together is ARR growth more likely to be sustainable.

    3. Why MaaS Can Grow Revenue Without Growing Profit

    The core cost of MaaS is directly related to model call volume. Every additional model call requires the platform to bear corresponding computing, storage, network, and system operations costs. Complex reasoning tasks, in particular, consume more compute resources.

    This creates a typical problem: revenue growth does not necessarily equal profit growth.

    Suppose a platform lowers prices to attract customers and call volume rises quickly, but the computing cost per request does not decline at the same pace. Total revenue may grow, while unit gross margin continues to narrow. If model service prices fall faster than inference costs, the result can be a situation in which higher usage increases loss pressure.

    MaaS businesses usually need to manage three variables at the same time:

    • Revenue per token;
    • Inference cost per token;
    • Effective call volume per customer.

    The first two variables determine the unit economics model, while the third determines economies of scale. Only when unit gross margin is positive and scaling up does not materially reduce service quality can a platform achieve healthy growth.

    4. The Key to Higher Gross Margin: Model Efficiency and Customer Mix

    The financial report shows that the MaaS API platform’s gross margin rose from 3.3% in 2024 to 18.9% in 2025, a relatively notable improvement. Combined with previously disclosed overall operating data, the company’s overall gross margin remained above 50% during the relevant period.

    The improvement in MaaS platform gross margin usually comes from several areas.

    1. Improving Inference Efficiency

    Cost optimization for model services starts with reducing the compute required per request. Common technical directions include model compression, quantization, cache reuse, batch inference, dynamic routing, and resource scheduling.

    Not all requests need the highest-specification model. For simple classification, summarization, rewriting, and basic question answering, lighter models can be used. For complex reasoning, long-text analysis, and high-requirement generation tasks, stronger models can be called. Through task tiering and model routing, the platform can reduce unnecessary compute consumption while maintaining performance.

    In addition, appropriately setting context length, optimizing prompt templates, reusing fixed prefixes, and reducing duplicate input can also help lower the number of tokens per request.

    2. Better Resource Utilization

    Model inference resources have significant peak and trough characteristics. If the platform keeps a large amount of idle resources during off-peak periods, overall costs are spread across fewer requests. Elastic scheduling, request aggregation, priority management, and multi-tenant resource isolation can improve compute resource utilization.

    For enterprises, resource utilization is not purely an infrastructure issue; it is also related to product design. For example, planning asynchronous tasks and batch-processing non-real-time requests can reduce dependence on peak compute resources.

    3. A Higher Share of High-Value Customers

    Beyond technical optimization, customer mix is also an important factor in gross margin improvement.

    Different customers vary in requirement complexity, price sensitivity, and service expectations. Large enterprises often place greater emphasis on stability, security, service levels, and business outcomes, and they tend to rely more deeply on model capabilities. Compared with customers focused primarily on low prices, these customers usually have higher willingness to pay and more stable call volume.

    When the revenue share of high-value customers increases, the platform can spread fixed R&D and infrastructure costs across a more stable usage scale, thereby improving overall gross margin.

    5. From Selling Models to Selling Verifiable Business Outcomes

    MaaS commercialization cannot stop at providing a model API. What enterprise customers really care about is whether the model can be embedded in business workflows and deliver measurable results.

    For example, in intelligent customer service, customers care not about model parameter scale but about issue resolution rate, human handoff rate, and average response time. In knowledge Q&A, customers care more about answer accuracy, source citations, and data security. In intelligent office applications, customers care about content generation efficiency, review costs, and employee adoption.

    Therefore, model platforms need to build service systems around business outcomes, including:

    • Providing clear call monitoring and usage statistics;
    • Supporting model selection and cost control for different tasks;
    • Providing permissions, auditing, and data isolation capabilities;
    • Establishing quality evaluation, error tracking, and feedback loops;
    • Helping customers measure the efficiency improvements brought by model applications.

    When customers can clearly calculate return on investment, model calls are more likely to shift from experimental consumption to long-term production expenditure.

    6. The Developer Ecosystem Is the Infrastructure for MaaS Growth

    The long-term value of a MaaS platform depends heavily on whether developers can integrate at low cost and use it reliably.

    A mature platform should maintain consistency in onboarding, documentation quality, error messages, version management, usage monitoring, and incident handling. Developers do not want to rewrite large amounts of business code every time a model is upgraded, nor do they want to rely solely on manual support channels when API calls fail.

    Therefore, platform development should emphasize the following engineering capabilities:

    • Providing stable, clear, and version-controlled APIs;
    • Clearly defining request limits, error types, and retry policies;
    • Supporting call logs, cost analysis, and quota management;
    • Providing separation between test and production environments;
    • Monitoring the safety and quality of model outputs;
    • Reducing migration and integration costs through standardized tools.

    For enterprise users, platform observability is equally important. Only when teams know how many resources each type of task consumes and how much cost it generates can they formulate reasonable model usage policies.

    7. The Future Core Competitiveness of MaaS

    At the current stage of industry development, competition among foundation model companies will not focus only on model capabilities. It will gradually extend across four levels.

    The first is foundation model capability, including understanding, generation, reasoning, long-context processing, and multimodal handling.

    The second is the unit economics model: whether revenue per call can cover inference and operating costs and continue to improve as scale increases.

    The third is the depth of customer embedding. Whether a model truly enters an enterprise’s core processes determines switching costs and long-term retention.

    The fourth is platform capability, including the developer ecosystem, service stability, operational tools, data security, and business monitoring.

    For model companies, the ideal state is not simply to pursue call volume, but to ensure that every call carries reasonable commercial value while maintaining service quality.

    Overall, the signal from this financial report is that the foundation model industry is shifting from a “model capability contest” to a “commercial loop contest.” For the MaaS model to achieve true scale, it must address four issues at the same time: revenue durability, inference cost, customer value, and platform efficiency. Only when model calls become part of customers’ daily business can ARR growth, revenue growth, and gross margin improvement form a mutually reinforcing long-term cycle.