First Post-IPO Earnings Report from an LLM Company: MaaS Revenue Growth and Gross Margin Improvement
This article examines the first annual earnings report from an LLM company after its IPO, focusing on the shift from project-based delivery to usage-based MaaS revenue. It reviews the company’s revenue growth, the role of ARR as an operating metric, the cost structure behind model services, and the factors that contributed to gross margin improvement. It also outlines the engineering capabilities and operating metrics that may determine whether MaaS can evolve into a stable enterprise infrastructure business.
On March 31, an LLM company disclosed its first annual earnings report after going public. For the year ended December 31, 2025, the company reported revenue of more than RMB 724 million, up 131.9% year over year. Revenue was approximately RMB 57 million in 2022, RMB 125 million in 2023, and RMB 312 million in 2024, indicating continuous expansion and an accelerating growth trend.
Compared with traditional software businesses, revenue growth at LLM companies tends to carry greater uncertainty. On one hand, model training, inference, and deployment require sustained investment in computing resources. On the other hand, customer procurement models are still evolving, spanning one-time projects and API services billed by usage. In this earnings report, changes in MaaS revenue and gross margin reflect a broader commercialization shift from project delivery toward continuous usage.
1. From Project Revenue to Model Invocation Revenue
Early commercialization of large models often depended on customized projects. Enterprise customers might purchase services for knowledge base Q&A, intelligent customer service, content generation, code assistance, or internal office workflows. Such projects typically have higher contract values and longer delivery cycles, but revenue recognition and renewals can be affected by customer budgets, project progress, and changes in customization requirements.
The MaaS model offers another commercial path. Model capabilities are exposed through standardized APIs, allowing developers and enterprises to use models based on call volume, input and output scale, or service period. The basic chain can be abstracted as follows:
Customer application↓API gateway and authentication↓Model routing and task scheduling↓Inference service cluster↓Token metering, billing, and usage analytics
The core change is that revenue is no longer mainly tied to one-time project acceptance but to customers’ continuous usage. Once a customer embeds the model into business processes, API calls can continue, creating a relatively stable recurring revenue source.
MaaS does not mean project-based business disappears entirely. For large customers, model integration, data governance, access control, private deployment, and business process transformation still require professional services. A more mature commercial structure is usually “standardized model services plus necessary industry implementation,” where standardized calls drive scale while professional services improve customer deployment efficiency.
2. Why ARR Can Reflect Business Quality
The earnings report shows that the company’s MaaS API platform had ARR of approximately RMB 1.7 billion, up about 60 times year over year. ARR refers to annual recurring revenue and is typically used to estimate the recurring revenue a company may generate over the next year based on its current business scale and customer usage.
It is important to note that ARR is not equivalent to recognized accounting revenue, nor does it guarantee that the company will fully realize the same amount in the future. It is better understood as an operating metric for observing revenue continuity and customer usage stability.
For MaaS businesses, ARR is useful in three main ways:
First, it helps assess whether customers continue to invoke models. A one-time purchase does not fully demonstrate that a customer has embedded the model into core processes, whereas continuous calls suggest that model capabilities are becoming part of specific business workflows.
Second, it helps observe whether the revenue structure is improving. If revenue mainly comes from customized projects, annual fluctuations may be larger. If API call revenue increases as a share of total revenue, business forecasting usually becomes clearer.
Third, it helps measure the activity of the developer and enterprise ecosystem. Growth in call volume may indicate not only more customers but also deeper usage per customer, such as moving from testing to production or expanding from a single scenario to multiple business functions.
ARR should still be analyzed together with retention rate, customer concentration, average revenue per customer, call growth rate, and actual collection performance. If ARR mainly comes from short-term promotions, low-price trials, or a small number of large customers, its stability requires further validation.
3. Why MaaS Often Shows Revenue Growth Without Profit Growth
Revenue and cost in MaaS are closely correlated. The more customers call the model and the more input and output content the model processes, the greater the demand for computing resources, storage, network bandwidth, and operations systems.
The cost structure broadly includes:
- Model inference cost: computing overhead from accelerated compute resources, memory usage, request concurrency, and context length.
- System operation cost: gateways, authentication, traffic scheduling, logging, monitoring and alerting, and failure recovery.
- R&D and iteration cost: model training, evaluation, data processing, version upgrades, and security governance.
- Customer service cost: technical support, solution design, deployment adaptation, and troubleshooting.
If a company pursues call volume growth without optimizing cost per token, revenue may rise while costs rise even faster. This is especially true during periods of intense price competition, when lower model service prices can further compress margins.
The key for MaaS is therefore not simply to pursue more calls, but to improve “value per call” and “output per unit of compute.” Scale effects become real only when revenue grows faster than variable costs.
4. The Core of Gross Margin Improvement: Efficiency, Pricing, and Customer Mix
The data shows that the gross margin of the company’s MaaS API platform rose from 3.3% in 2024 to 18.9% in 2025, a notable improvement. Gross margin changes are usually not caused by a single factor but by the combined effect of multiple operating actions.
4.1 Improve Inference Efficiency and Reduce Cost per Token
Cost control in model services depends first on inference efficiency. Common optimization directions include:
- using models of different sizes based on task complexity;
- using lightweight models for high-frequency, low-complexity requests;
- using higher-capability models for complex reasoning tasks;
- optimizing batch processing and request scheduling to improve compute utilization;
- reducing invalid context and controlling input length;
- using caching mechanisms to reduce compute overhead from repeated requests;
- improving elastic resource scheduling to reduce idle capacity during off-peak periods.
In real systems, not every request requires the highest-specification model. Companies can establish model routing mechanisms to select appropriate inference resources based on task type, latency requirements, accuracy requirements, and budget constraints. This can maintain service quality while lowering average cost.
4.2 Optimize Pricing and Avoid Low-Price Competition
Pricing for a MaaS platform cannot be based simply on “number of calls.” It should also be tiered by model capability, response speed, context length, concurrency scale, service level, and data security requirements.
For experimental calls, a standard service with a lower entry barrier can be offered. For production-grade applications, pricing should be designed more reasonably around stability, service guarantees, concurrency capacity, and governance capabilities. This can prevent all customers from concentrating in the lowest price tier and improve overall revenue quality.
Whether a price increase affects call volume depends on whether customers can obtain sufficient business value from the model service. If the model can reduce manual review, improve content production efficiency, shorten R&D cycles, or improve customer response speed, enterprise customers usually focus more on return on investment than on pursuing the lowest unit price.
Therefore, a healthy business model is not simply about raising prices. It is about improving model capabilities, service stability, and business value so that customers are willing to pay for higher-quality services.
4.3 Increase the Share of High-Value Customers
Customer mix is another important factor affecting gross margin. Small-scale testing customers usually have limited call volume and require more technical support. Large production customers may have larger call volume, but if their requirements are more standardized, the unit service cost may be lower.
High-value customers typically have the following characteristics:
- they have integrated models into stable production processes;
- their call volume is continuous and predictable;
- they have clear requirements for service stability and security governance;
- they can expand from a single scenario to multiple business scenarios;
- they are relatively less price-sensitive to model quality and service quality.
When customers move from trial to production, both revenue predictability and resource utilization on the model platform improve. At the same time, the service provider can reduce delivery costs through unified interfaces, standardized deployment, and automated operations, contributing to gross margin improvement.
5. From a Technical Platform to an Operating Platform
A MaaS platform is not only a model invocation interface; it also needs complete engineering capabilities. A sustainably operated platform should cover at least the following modules:
- Identity and access management: distinguishing permissions for organizations, projects, applications, and users;
- Usage metering: recording request volume, token consumption, response latency, and errors;
- Resource scheduling: allocating inference resources based on load and task type;
- Quality monitoring: tracking response accuracy, refusal rate, and abnormal outputs;
- Cost analytics: calculating unit costs across models, customers, and scenarios;
- Security auditing: recording access logs and controlling the spread of sensitive data;
- Operational analytics: identifying active customers, dormant customers, and high-growth scenarios.
Among these, cost observability is particularly important. The platform needs to know how many resources each customer, each task type, and each model version consumes. Only with fine-grained accounting can a company determine which businesses truly create profit and which merely generate surface-level call volume.
6. Key Metrics for Future MaaS Competition
As model services become increasingly standardized, competition among companies will not be limited to parameter scale or single benchmark results. It will shift toward comprehensive operating capabilities.
Key metrics may include:
- Effective call volume: real production calls after excluding tests, repeated requests, and low-value traffic.
- Customer retention rate: whether customers continue using the service and move from trial to long-term procurement.
- Gross profit per token: measuring the balance between model capability and inference cost.
- Share of production-grade applications: whether calls truly enter core business processes.
- Customer expansion rate: whether a single customer expands from one scenario to multiple scenarios.
- Service stability: including availability, response latency, and failure recovery capability.
- Cash flow quality: whether revenue can be collected in a timely manner rather than remaining as booked revenue.
From this perspective, rapid ARR growth is a positive signal, but it is not the endpoint. Companies still need to prove that this revenue has a stable renewal base and that unit economics can continue to improve while call volume scales.
Conclusion
The core change reflected in this earnings report is that LLM companies are beginning to move from competition in “model R&D capability” to competition in “model productization and operating efficiency.” Rapid full-year revenue growth indicates expanding market demand. The sharp increase in MaaS ARR shows that continuous usage is gradually becoming an important revenue source. The improvement in platform gross margin indicates that inference efficiency, pricing strategy, and customer mix are working together.
For LLM companies, a truly sustainable business model should satisfy three conditions at the same time: model capabilities can solve real problems, API services can be continuously used by customers, and platform operations can keep costs controllable as scale expands. Only by completing the transition from project delivery to standardized services, from one-time revenue to recurring revenue, and from call growth to profit growth can MaaS evolve further from a technical service model into stable enterprise-grade infrastructure.
