MaaS and ARR: LLM Commercialization Shifts from Project Revenue to Recurring Revenue
A first annual report from an LLM company shows rapid revenue growth and a clear shift toward MaaS API platform revenue. This article examines why ARR, production usage, gross margin, cost structure, and customer conversion matter more than one-off project revenue, and what enterprises should evaluate when choosing a model service.
MaaS and ARR: LLM Commercialization Shifts from Project Revenue to Recurring Revenue
Industry Changes Reflected in an Annual Report
An LLM company’s first annual report after listing shows a clear acceleration in growth. The company reported revenue exceeding RMB 724 million for the year ended December 31, 2025, up 131.9% year over year. Historical data shows revenue of approximately RMB 57 million in 2022, RMB 125 million in 2023, and RMB 312 million in 2024. Revenue has continued to expand, and the growth curve has gradually moved from an early capability-building phase into a commercialization and scale-up phase.
Revenue alone can show the scale of commercialization, but it is not enough to explain growth quality. For an LLM company, more important questions are where revenue comes from, whether it is repeatable, and whether model calls are embedded in customers’ real business processes. Compared with one-off project delivery, ongoing model calls are closer to the revenue logic of infrastructure-type businesses and provide stronger evidence that customers have formed stable usage habits.
The core observation from this report is therefore not only how much revenue grew, but that the company is shifting from a project-based delivery model toward a recurring-revenue model centered on a model service platform.
From Project Revenue to MaaS Platform Revenue
Model-as-a-Service (MaaS) generally refers to providing model capabilities to developers and enterprise customers through standardized interfaces. Customers do not need to procure large amounts of hardware or build training and inference systems from scratch. Instead, they use model capabilities based on call volume, computing resource usage, or service packages.
The basic flow of this model can be summarized as follows:
- A customer connects to the model service through an interface.
- The application system sends input content to the model.
- The model performs text generation, classification, extraction, reasoning, or multimodal processing.
- The platform meters usage based on actual consumption or an agreed service arrangement.
- The customer continuously generates calls as business operations run.
Compared with traditional project-based models, the biggest change in MaaS is that revenue no longer depends mainly on acceptance of individual projects. Instead, it is tied to customers’ long-term usage behavior. As long as model capabilities remain embedded in scenarios such as customer service, office work, R&D, knowledge management, marketing, and data processing, the platform has an opportunity to generate stable recurring revenue.
Of course, MaaS does not mean that simply opening an API will naturally lead to scale. When enterprise customers connect to a model service, they usually consider the following factors:
- Whether model quality meets business requirements;
- Whether call latency is stable;
- Whether the service offers continuous availability;
- Whether data security and permission management are reliable;
- Whether it supports growth in call volume;
- Whether costs are predictable;
- Whether it is easy to integrate with existing systems.
Therefore, MaaS is not simply selling a model. It provides a technical infrastructure that includes model capabilities, inference services, access control, monitoring and statistics, and operational support.
Why ARR Deserves More Attention Than One-Time Revenue
The report shows that the company’s MaaS API platform had annual recurring revenue (ARR) of approximately RMB 1.7 billion, up 60x year over year. ARR is typically used to estimate the recurring revenue a business could generate over a year if it continues operating at its current business level.
It is important to note that ARR is not completely equivalent to recognized accounting revenue. It is more of an operating metric used to observe the stability of customer contracts, subscriptions, or ongoing consumption behavior. In model service scenarios, ARR is valuable mainly in three areas.
First, whether customers form ongoing call behavior
If revenue comes mainly from one-time consulting, model customization, or project delivery, a company may achieve high revenue in one year, but it is uncertain whether that revenue will continue the next year.
Ongoing API platform calls mean the model has entered the customer’s production environment. Customer systems may automatically send requests every day, and model capabilities have become part of the business process. In this case, revenue sustainability is usually higher than that of one-time projects.
Second, customer usage depth
The number of model service calls, input and output length, task complexity, and concurrency scale all affect platform consumption. After customers move from testing to production, call volume often increases significantly.
For example, an internal knowledge Q&A system may be used by only a small number of employees during testing. Once it is connected to a customer service system or office system, call behavior shifts from manual triggering to automatic system triggering, and both usage frequency and stability increase.
Third, whether the business model has scale effects
After a platform accumulates a large number of customers, new customers can usually reuse existing models, inference clusters, monitoring systems, and development tools, creating an opportunity to reduce marginal delivery costs. If model inference efficiency continues to improve, the platform may improve gross margin while revenue grows.
However, ARR should not be judged separately from customer retention, net revenue retention, customer concentration, and actual collections. Only when recurring revenue can be continuously renewed, consumed stably, and converted into cash flow can it indicate higher commercialization quality.
Why MaaS Can Grow Revenue Without Growing Profit
Model service revenue is directly related to Token consumption, and Token consumption corresponds to computing resources, storage, network, scheduling, and operational costs. Growth in call volume will certainly expand revenue, but if unit call costs do not decline fast enough to offset price declines, greater scale may actually increase profit pressure.
This issue usually comes from four areas.
High model inference costs
LLM inference continuously consumes computing resources. Model parameter scale, context length, concurrent request volume, and output complexity all affect unit request cost. Tasks that require long-text understanding, complex reasoning, or multi-turn dialogue usually consume more resources.
Intensifying industry price competition
When multiple service providers offer similar capabilities, customers often compare unit Token prices, response speed, and service stability. If a platform relies solely on price cuts to acquire customers, revenue may grow while unit gross margin declines.
Fluctuating customer call volume
Enterprise customer call volume may be affected by business seasonality, campaign cycles, and system upgrades. If a platform builds resources according to peak demand but cannot ensure full utilization during low periods, fixed costs will be diluted.
Customization requirements increase delivery costs
Some large customers require private deployment, dedicated model tuning, data isolation, permission management, or complex system integration. If these services are not reasonably priced, project delivery costs may erode the profit of the model service itself.
Therefore, the key for a MaaS platform is not simply to pursue call volume, but to balance unit Token cost, service price, and customer value.
What Gross Margin Improvement Indicates
The company reported that the gross margin of its MaaS API platform rose from 3.3% in 2024 to 18.9% in 2025. This change suggests that the platform may have improved inference efficiency, resource scheduling, customer mix, or pricing strategy.
From a technical perspective, gross margin improvement is usually related to the following capabilities.
Model inference optimization
By reducing unnecessary computation, improving batching strategies, increasing cache hit rates, and optimizing context management, the resource consumption of each request can be reduced. For high-concurrency scenarios, reasonable request scheduling can also improve computing resource utilization.
Tiered model services
Different tasks do not necessarily require models of the same scale. Simple classification, information extraction, and format conversion tasks can use lighter models, while complex reasoning and high-quality generation tasks can use more capable models.
This tiered strategy can avoid using the highest-specification model for every request, thereby reducing overall costs.
Improved resource utilization
The platform can dynamically adjust computing resources based on business peaks and troughs to reduce idle resource occupation. At the same time, it needs to establish capacity forecasting and elastic scheduling mechanisms to ensure that service quality is not affected by resource shortages during peak periods.
Optimized customer mix
Large enterprise customers usually have more stable call volume and clearer business budgets, and they are more willing to pay reasonable fees for stability, security capabilities, and service assurance. When the share of high-value customers increases, revenue stability and bargaining power usually improve.
What More Than 4 Million Users Means
The report shows that more than 4 million enterprise users and developers are continuously calling model capabilities in production environments, with services covering 218 countries and regions.
User scale itself does not equal commercial value, but it can indicate that the platform has a certain degree of developer reach and application diffusion capability. For a model service platform, as the number of users expands, a layered structure usually forms from trial and development to testing and production deployment.
The key areas to focus on are the proportion of production users, the proportion of active calling users, and the contribution of large customers. A large number of registered users who do not call the service for a long time cannot generate actual revenue. Conversely, a smaller number of high-frequency production customers may contribute higher revenue and more stable growth.
Platform operations can build an analysis system around the following metrics:
- Conversion rate from registered users to active users;
- Conversion rate from active users to production deployment users;
- Paid conversion rate of trial customers;
- Monthly and annual customer retention rates;
- Average revenue per customer;
- Customer call volume growth rate;
- Revenue share of large customers;
- Unit Token cost;
- API platform gross margin.
Only by combining these metrics can the health of a MaaS business be fully assessed.
What Enterprises Should Evaluate When Choosing a Model Service
For developers and enterprise users, choosing a model platform should not depend only on model quality or unit price. A more reasonable approach is to validate model capabilities within the complete business workflow.
1. Evaluate task success rate
Use real business samples to test the model, including long text, structured output, professional Q&A, complex instructions, and abnormal inputs. Demonstration cases alone are rarely enough to determine whether a model is suitable for production environments.
2. Evaluate stability and latency
Enterprise applications usually need to consider average latency, tail latency, concurrency capability, and error rate. Even if answer quality is high, frequent timeouts during peak periods can affect the final user experience.
3. Evaluate cost controllability
Estimate monthly costs based on actual call volume, and distinguish additional consumption caused by input Tokens, output Tokens, context length, and retry requests. For automated systems, pay particular attention to abnormal loops and invalid calls.
4. Evaluate data and permission security
Enterprises should clarify whether data will be used for training, how retention periods are set, how access permissions are controlled, and whether sensitive information may appear in logs. When internal materials are involved, mechanisms for desensitization, auditing, and tiered authorization should also be established.
5. Evaluate migration and disaster recovery capabilities
Core business should not become overly dependent on a single model service. By abstracting a unified calling layer, configuration layer, and monitoring layer, migration costs can be reduced when switching models, handling service fluctuations, or scaling the business.
The Next Stage of LLM Commercialization
The financial data shows that the LLM industry is gradually moving from a “model capability race” to a “continuous usage race.” Model performance remains important, but the factors that determine commercial results now extend to inference cost, service stability, customer retention, ecosystem development, and industry deployment.
For model companies, three things need to be addressed at the same time. First, continue improving foundation model capabilities and expand the range of tasks that can be handled. Second, reduce unit call costs through engineering optimization and improve platform gross margin. Third, embed models into customers’ production systems to create sustainable, reusable, and scalable usage scenarios.
For enterprise customers, the focus should also shift from “which model has more parameters” to “which service is more suitable for the business.” Model quality, cost, latency, security, observability, and long-term stability need to be placed within the same evaluation framework.
Overall, the value of the MaaS model is not simply putting a model behind an API. It lies in packaging complex training, deployment, inference, and operational capabilities into a sustainable foundational service. Only when a platform can achieve continuous customer calls, continuously declining costs, and continuously improving service quality can revenue growth have an opportunity to translate into healthy commercial growth.
