MaaS and ARR: How Large Model Companies Build Recurring Revenue from AI Services
This article examines the shift in large model commercialization from one-off project delivery to MaaS-based continuous usage. It explains why ARR can be a useful indicator of recurring business quality, how MaaS improves gross margins through standardization and efficient inference scheduling, what enterprise buyers evaluate beyond model quality, and how development teams should integrate MaaS APIs with proper engineering controls. It concludes with the key tests for turning revenue growth into long-term competitiveness.
MaaS and ARR: How Large Model Companies Build Recurring Revenue from AI Services
1. From One-Off Delivery to Continuous Invocation: The Business Model Is Changing
A large model company disclosed its first annual financial information after listing. Its 2025 revenue exceeded RMB 724 million, up 131.9% year over year. In the previous three years, annual revenue was RMB 57 million, RMB 125 million, and RMB 312 million, respectively, showing a clear acceleration in growth.
What matters more than a single year of revenue is how that revenue is generated. Traditional artificial intelligence projects often rely on custom development, private deployment, or phased delivery, with revenue usually recognized after project acceptance. This model can scale quickly, but it has clear limitations: project cycles are long, delivery depends heavily on manpower, revenue is volatile, and customer repurchase and continued usage are not easy to observe directly.
MaaS, or Model-as-a-Service, packages large model capabilities into standardized services. Enterprises and developers do not need to train or deploy a complete model themselves. Instead, they submit requests through an API or platform console to obtain capabilities such as text generation, content understanding, structured extraction, intelligent question answering, and code assistance. The platform bills based on invocation volume, computing resources, or service period. The more consistently customers use the service, the stronger the continuity of platform revenue.
This means the growth logic of large model companies is gradually shifting from “how many projects were signed” to “how much model capability customers continue to invoke.”
2. Why ARR Can Better Reflect Business Quality Than Annual Revenue
An important indicator disclosed in the financial report is that the MaaS API platform ARR was approximately RMB 1.7 billion, up about 60 times year over year. ARR is generally understood as an annualized estimate of annual recurring revenue, used to measure the current sustainable business scale of an enterprise.
It is important to note that ARR is not completely equivalent to realized accounting revenue. It is more of an operating metric used to observe the annualized revenue that customer contracts, subscriptions, or continuous invocations can generate under current conditions. Therefore, a high ARR does not mean future revenue will necessarily grow at the same rate. It still needs to be assessed together with customer retention, invocation frequency, contract duration, revenue recognition method, and cost structure.
For MaaS businesses, the value of ARR is mainly reflected in three areas.
First, it helps show whether customers continue to use the model. A one-off project may generate high revenue, but it does not prove that the customer will retain the service over the long term. API invocations directly correspond to actual usage behavior, and continuous growth usually means the product has entered the customer’s business processes.
Second, it helps distinguish “trial demand” from “production demand.” Developer testing, short-term campaigns, and proof-of-concept work often generate temporary traffic. In contrast, production scenarios such as customer service, moderation, search, office automation, and knowledge question answering usually have more stable invocation cycles.
Third, ARR can help platforms plan resources. Model inference requires computing resources, network bandwidth, storage, and operations and maintenance investment. Stable recurring revenue provides a clearer basis for capacity expansion, model optimization, and service quality assurance.
3. How the MaaS Model Improves Gross Margin
In the early stage of large model businesses, gross margin is often under pressure, usually because training costs, inference costs, and delivery costs are high. Especially under the project-based model, each customer may require independent data adaptation, deployment environment configuration, permission systems, and business process integration. This requires significant technical personnel input and has low reuse.
MaaS platforms improve gross margin not simply by raising prices, but by reducing unit service costs through standardization and scale.
3.1 Standardized Model Services
The platform packages model capabilities into a unified invocation entry point, authentication mechanism, rate-limiting policy, logging, and billing system. Different customers can share the same infrastructure. The platform only needs to configure a small number of differentiated requirements instead of repeatedly building complete systems.
Standardization also reduces maintenance complexity. When model versions, security policies, or monitoring capabilities are upgraded, platform-level changes can be completed centrally and then gradually opened to customers, reducing the cost of project-by-project maintenance.
3.2 Centralized Scheduling of Inference Resources
Large model inference costs are closely related to request volume, input and output length, concurrency scale, and response latency. Platforms can improve computing resource utilization through request queuing, batch processing, cache reuse, elastic scaling, and model tiering.
For example, requests with high real-time requirements can use low-latency resources, while batch tasks that can tolerate delay can use scheduling methods with higher resource utilization. Different tasks no longer occupy the same fixed configuration, making overall costs easier to control.
3.3 Model Tiering and Task Routing
Not every business task needs the largest model. Tasks such as classification, summarization, keyword extraction, and format conversion can often be handled by lower-cost models. Complex reasoning, long-text analysis, and high-reliability tasks can then invoke more capable models.
Therefore, the platform can route requests based on task type, input length, customer tier, and accuracy requirements. Reasonable model tiering can reduce unnecessary high-cost computation while maintaining output quality.
3.4 Reducing Repeated Delivery
Once customers access models through a unified API, the platform can consolidate common capabilities into reusable components, such as prompt template management, structured output validation, content safety filtering, invocation monitoring, and failure retries. One development effort can serve multiple customers, and marginal delivery costs decline as scale expands.
4. Enterprise Customers Care About More Than Model Capability
From the perspective of enterprise technology procurement, model quality is only one condition for choosing a MaaS platform. Whether the service can be reliably integrated into production systems often depends on the following capabilities.
First is service stability. Enterprise applications usually require clear availability targets, timeout handling mechanisms, and failure degradation plans. If API services fluctuate frequently, they may directly affect customer service, moderation, or office workflows.
Second is data security. Enterprises need to understand how request data is transmitted, stored, and processed, and configure access permissions, audit records, and data isolation policies according to business sensitivity. Applications involving internal knowledge bases, contract texts, and user information also require additional data masking and permission controls.
Third is observability. Platforms should provide metrics such as invocation volume, error rate, latency, input and output size, and resource consumption to help development teams locate problems and help managers judge business value. Without observability, enterprises can hardly distinguish model quality issues from network issues or business logic issues.
Finally, there is migration and compatibility. Enterprises usually do not focus on a single model only; they combine models based on cost, quality, and stability. Therefore, a unified invocation abstraction, clear version management, and replaceable model routing mechanisms help reduce long-term lock-in risk.
5. Engineering Practices for Developers Integrating MaaS
For development teams, integrating large model services should not mean simply writing an HTTP request into business code. It should involve building a complete engineering wrapper layer.
A general invocation flow usually includes:
User request↓Business parameter validation↓Prompt and context assembly↓Model routing and permission check↓Call MaaS API↓Response validation and safety filtering↓Cache, log, and metric recording↓Return to business system
During implementation, teams can focus on the following issues.
First, set timeout, retry, and circuit breaker mechanisms. Model services may experience network jitter, resource congestion, or abnormal responses, and business systems should not wait indefinitely. Retries should be limited and use gradual delays to avoid traffic amplification during failures.
Second, control context length. The longer the input text, the higher the computing cost and response latency usually are. Developers should trim historical conversations, retrieved content, and attachments, and prioritize information directly relevant to the current task.
Third, perform structured validation of results. For outputs that need to be written to a database or drive subsequent processes, natural language format cannot be fully relied upon. Field checks, type validation, and exception branches should be used to prevent unexpected model output from contaminating business data.
Fourth, establish cost budgets. Invocation volume and token consumption can be tracked by department, application, user, or business line, with alerts set for abnormal growth. For low-value, high-frequency requests, caching, batch processing, or rule-based systems should be considered as alternatives.
6. From Revenue Growth to Long-Term Competitiveness
Rapid MaaS ARR growth shows that the market is gradually moving from “trying large models” to “embedding models into daily business.” But ARR itself is not the final result. Platforms still need to answer three questions.
First, can customers be retained over the long term? If customers invoke models only during the proof-of-concept stage, ARR may rise quickly and fall quickly. A truly healthy business needs to form stable production traffic and quantifiable business benefits.
Second, does the unit economics model hold? Revenue growth must match inference costs, research and development investment, sales expenses, and service costs. If greater invocation volume leads to more severe losses, then scale does not equal improved commercial quality.
Third, does the platform have the ability to iterate continuously? Model capabilities, inference efficiency, security requirements, and enterprise needs are all changing. Only by combining model optimization, resource scheduling, development tools, and operations systems can MaaS evolve from a simple API service into stable infrastructure.
Overall, the commercialization of large model companies is undergoing a transition from project-driven to platform-driven. Revenue scale reflects the speed of market expansion, ARR reflects sustained usage capability, and gross margin reflects whether the platform can turn technical capability into replicable and scalable business results. For enterprise users and developers, understanding this logic helps them evaluate large model services more rationally: they should pay attention not only to model quality, but also to cost, stability, security, and long-term maintainability.
