How MaaS Revenue Growth Is Shifting Toward Recurring API Usage

    A large model company reported annual revenue of more than RMB 724 million in 2025, while its MaaS API platform’s annual recurring revenue (ARR) reached about RMB 1.7 billion. This article explains why recurring usage is a meaningful measure of MaaS growth, how API-based model consumption changes the revenue model, and why profitability still depends on inference costs, customer retention, and efficient production use.

    1. The Changing Structure Behind Rapid Revenue Growth

    A large model company’s latest annual operating data shows that revenue exceeded RMB 724 million for the year ended December 31, 2025, an increase of 131.9% year over year. Annual revenue was approximately RMB 57 million in 2022, RMB 125 million in 2023, and RMB 312 million in 2024. Revenue has grown steadily, with growth accelerating over time.

    Revenue totals alone may suggest that commercialization is progressing quickly. But revenue by itself does not fully describe the business quality of a large model company. Traditional software revenue may come from one-time licenses, project delivery, subscriptions, or implementation services. For model providers, revenue is also closely tied to the number of model calls, input and output tokens, inference complexity, concurrency, and how frequently customers use the service.

    A more complete assessment of a large model company should therefore consider at least three dimensions:

    1. Revenue scale: the company’s current level of commercialization.
    2. Revenue growth: whether market demand is expanding.
    3. Revenue sustainability: whether customers continue to call models rather than making a one-time project purchase.

    From this perspective, annual recurring revenue (ARR) from a Model-as-a-Service (MaaS) platform is one of the operating metrics worth watching alongside annual revenue.

    2. Why ARR Matters for MaaS

    ARR estimates the recurring revenue a business could generate over the next year. For subscription software, platform services, and usage-based businesses, it helps distinguish ongoing revenue from one-off income.

    Under a traditional project-based model, a large contract can significantly increase revenue in the short term. After the project ends, further revenue depends on renewals, maintenance, or additional projects. MaaS, by contrast, is primarily billed according to model usage. As long as customers continue to use the APIs, the platform can continue generating revenue.

    The company reported ARR of approximately RMB 1.7 billion for its MaaS API platform, about 60 times higher than a year earlier. ARR is not the same as revenue recognized during the reporting period, and it does not guarantee that the projected revenue will be fully realized. It is better understood as a run rate that indicates the revenue level the current scale of customer usage could support if that usage continues.

    For MaaS, ARR is useful in several ways:

    1. It shows whether customers are becoming repeat users

    Frequent model calls suggest that a model may be embedded in a business workflow rather than used only for testing. Common applications include customer service, content generation, knowledge retrieval, coding assistance, data analysis, and office automation.

    Once a model becomes a foundational component of a business system, customers may be less likely to discontinue it. Revenue can then shift from one-time purchases toward ongoing usage.

    2. It indicates developer ecosystem activity

    API platform growth depends not only on a small number of large customers, but also on whether developers can integrate the service at low cost, validate ideas quickly, and deploy reliably. Developer numbers, active projects, call frequency, and the conversion rate from testing to production can all affect the quality of ARR growth.

    If ARR comes primarily from a small number of projects, the business may face concentration risk. A broader base of customers making ongoing calls generally indicates a healthier revenue mix.

    3. It reflects the fit between model capabilities and business needs

    Model size alone does not determine commercial value. Customers care about whether a model can balance accuracy, response speed, reliability, controllability, and cost.

    A model creates more stable business value when it is integrated into production systems and used continuously in real workflows. API usage and ARR can therefore provide some indication of whether model capabilities are translating into practical productivity.

    3. From Project Delivery to Token Consumption

    Large model companies often begin commercializing through customized projects, such as model adaptation, knowledge base development, industry-specific application development, and deployment. These projects provide access to real customer needs, but can involve long delivery cycles, substantial staffing requirements, and fluctuating revenue.

    MaaS turns model capabilities into standardized services. Businesses and developers call models through APIs and pay according to usage. The business logic can be summarized as follows:

    Customers integrate model capabilities, applications generate API requests, and the platform earns revenue based on usage.

    The key shift is that revenue growth depends less on the value of individual projects and more on the number of customers, call frequency, and depth of usage.

    For example, a company may first test a model internally, then integrate it into a customer service system. Once that system is fully deployed, call volume can rise with the number of customer inquiries. The model platform’s revenue may grow alongside the scale of the business using it.

    From an operating perspective, this model can offer economies of scale:

    • Standardized APIs reduce repetitive delivery work.
    • The same model capabilities can serve customers across multiple industries.
    • Customers may generate ongoing calls after deployment.
    • Marginal service costs may decline as new customers are added.
    • Product improvements can benefit many API users at once.

    This does not mean MaaS is inherently more profitable than project-based work. It changes how revenue is generated, but also changes the cost structure and operating requirements.

    4. The Core MaaS Challenge: Does Revenue Growth Lead to Profit Growth?

    A defining challenge for MaaS is the close relationship between revenue and inference costs. Each additional customer call typically consumes computing resources. Higher usage can generate more revenue, but it also increases costs for compute, storage, networking, scheduling, and operations.

    If prices fall faster than the cost of inference per unit, revenue may grow without a corresponding improvement in profitability.

    The basic economics of an individual call can be expressed as:

    Contribution margin per call = Revenue per call − Direct cost per call

    Revenue per call depends on factors such as input and output token counts, model type, and billing method. Direct costs are affected by compute prices, inference efficiency, request concurrency, context length, and resource utilization.

    Providers can improve contribution margins in several ways:

    1. Improve inference efficiency

    General techniques such as model compression, batch inference, cache reuse, request aggregation, and dynamic scheduling can reduce the computing resources required for each request. Caching is particularly useful for repetitive workloads because identical or similar content may not need to be processed from scratch every time.

    2. Use a tiered model strategy

    Different tasks require different levels of model capability. Simple classification, information extraction, and format conversion may be handled by lower-cost models, while complex reasoning, long-document analysis, and demanding generation tasks may require more capable models.

    Task routing and model tiers can help balance output quality and cost.

    3. Increase resource utilization

    Inference resource utilization has a direct effect on costs. Insufficient capacity during peak periods can require expansion, while idle resources during quieter periods can go to waste. Elastic scheduling, load forecasting, and capacity planning can help improve overall utilization.

    4. Help customers optimize their usage

    Enterprise customers also need to manage token costs. Overly long contexts, repeated transmission of conversation history, ineffective prompts, and unnecessary high-precision outputs can all increase API usage costs.

    Application developers can reduce avoidable consumption through summarization, context trimming, result caching, staged calls, and structured outputs.

    5. What Businesses Should Evaluate When Choosing a MaaS Provider

    Developers and enterprise users should assess more than model capabilities and per-call prices. They need a broader set of technical and operational measures.

    1. Reliability

    Evaluate API availability, response latency, rate-limiting policies, recovery capabilities, and performance under high concurrency. Even a capable model is unlikely to create lasting value if it cannot reliably support production workloads.

    2. Cost predictability

    Usage-based billing is flexible, but costs can fluctuate. Businesses should monitor call volumes, set budget alerts, and detect unusual usage to prevent runaway charges caused by software errors, call loops, or sudden traffic increases.

    3. Data security and access controls

    Businesses should understand how data is transmitted, how logs are retained, how access is isolated, and how sensitive information is handled. When calls involve internal knowledge, customer information, or business data, organizations should apply the principle of least privilege and use appropriate redaction and auditing.

    4. Portability and replacement options

    Application logic should not be tightly coupled to a single model service. A sound engineering practice is to provide a unified layer for model calls and manage prompts, retries, timeouts, logging, and model-switching logic centrally. This can reduce the cost of migrating to another service later.

    5. Evaluation processes

    Model quality should not be judged by subjective impressions alone. Businesses should maintain offline evaluation sets and track metrics such as accuracy, recall, format compliance, hallucination rates, refusal rates, and business conversion rates.

    In production, teams should also combine real user feedback with human spot checks and use gradual rollouts to validate model-version upgrades.

    6. What Will Define MaaS Competitiveness

    As more companies move from experimentation to production, competition among large model providers is likely to shift from simply offering a model to supporting real workloads reliably and at lower cost.

    Key capabilities may include:

    • Continued improvement in foundation models.
    • Greater API reliability and concurrency capacity.
    • Lower inference costs.
    • More mature multi-model coordination and intelligent routing.
    • Better developer tools and enterprise management capabilities.
    • Stronger data security, access controls, and observability.
    • Higher conversion rates from customer trials to production use.

    Rapid ARR growth is a positive signal, but it is only a starting point for assessing commercialization quality. Further analysis should consider recognized revenue, customer retention, customer concentration, the sources of call growth, gross margin trends, and cash flow.

    7. Conclusion: From Revenue Growth to Operating Quality

    The large model industry is moving from a period of technical competition toward a test of commercialization. In the past, attention focused largely on model size, benchmark results, and release speed. Today, the ability to generate recurring revenue, reduce unit costs, and retain long-term customers is becoming more important.

    MaaS uses APIs and token-based consumption to deliver model capabilities as services that customers can call continuously. The model offers the potential for recurring, scalable revenue, but also brings risks related to compute costs, price competition, and customer retention.

    Assessing whether a large model company has built a healthy business therefore requires more than looking at annual revenue or ARR growth multiples. It also requires asking:

    1. Does revenue come from ongoing model usage?
    2. Are customers moving from testing to production?
    3. Does growth in usage lead to higher gross margins?
    4. Are unit inference costs continuing to decline?
    5. Has the company built a stable developer and customer ecosystem?

    MaaS can become a core foundation for long-term growth—not merely a billing method for model services—when model capabilities, customer demand, usage scale, and cost efficiency reinforce one another.