MaaS Revenue Growth and Commercialization: Insights from an LLM Company’s First Post-IPO Annual Report
This article examines the first annual results disclosed by a China-based large language model company after going public, focusing on Model-as-a-Service (MaaS) annual recurring revenue (ARR) and commercialization rather than headline revenue alone. It explains the shift from project-based delivery to API-based model consumption, why ARR can be a useful indicator of business quality, and how inference utilization, model tiering, token efficiency, and call reliability affect gross margin. It then outlines the access, scheduling, inference, observability, and operations layers of a sustainable MaaS platform; reviews what enterprise customers prioritize in production; and identifies the metrics to watch as growth gives way to operating quality. The article concludes by considering how LLM providers can combine implementation services with standardized model usage to build a durable commercialization path.
On March 31, a China-based large language model company released its first annual results since going public. For the year ended December 31, 2025, revenue exceeded RMB 724 million, up 131.9% year over year. Historical figures show revenue of approximately RMB 57 million in 2022, RMB 125 million in 2023, and RMB 312 million in 2024. The company has thus expanded its revenue base for several consecutive years, with growth accelerating noticeably.
Compared with revenue alone, the more revealing indicators in this report concern its MaaS platform. Its ARR was approximately RMB 1.7 billion, about 60 times higher year over year. ARR is not equivalent to revenue recognized in the current period, but it can help observers assess whether customers continue to subscribe, continue to call the models, and whether enterprises have formed relatively stable usage habits.
From Project-Based Revenue to Model-Usage Revenue
In the early stage of LLM commercialization, businesses often relied on projects. A customer would define a specific requirement, and the service provider would deliver around model deployment, knowledge base construction, application development, and scenario adaptation. Revenue could come from one-time project contracts, model fine-tuning services, or system integration fees.
This model has advantages: average deal sizes can be relatively high, and it can quickly validate industry demand. But its shortcomings are also clear:
- Project cycles are long, and revenue recognition can be volatile.
- Each customer may require separate adaptation, increasing R&D and delivery costs.
- Whether a project generates further revenue after completion depends on maintenance and renewals.
- Scaling the business often requires expanding the delivery team in a roughly linear fashion.
The MaaS model packages model capabilities as standardized services. Enterprise customers and developers call models through APIs and use the services according to call volume, resource consumption, or subscription period. In this way, model capability changes from a project deliverable into continuously running infrastructure.
For customers, MaaS can reduce upfront investment. They do not need to procure large-scale compute or build their own model training and inference systems, nor do they need to handle complex version upgrades. For service providers, standardized APIs allow the same model, inference cluster, and operations system to be reused, improving delivery efficiency.
Why ARR Can Reflect Business Quality
Traditional revenue mainly answers how much money a company recognized during a given period. ARR focuses more on how much recurring revenue may continue over the next year under current customer conditions.
In the LLM business, ARR is typically related to the following factors:
- The number of enterprise customers continuously subscribing to model services;
- Customer API call frequency;
- The scale of input and output token consumption;
- Account renewals and expansion;
- Stable calls in production environments, not short-term trials during testing.
If a customer is only conducting one-time validation, usage may rise quickly in the short term but is unlikely to become stable revenue. Only when models enter production workflows such as customer service, office productivity, search, R&D assistance, content moderation, and data analysis does usage become long-term consumption.
For this reason, a MaaS business should not focus only on registered customers or trial users. It also needs to track the conversion rate from trial to production, monthly consumption after launch, account retention, and expansion among major customers. For enterprise customers, the real value is not simply “having integrated a model,” but whether the model is embedded in daily workflows and continues to produce quantifiable business results.
Core Mechanisms Behind Gross Margin Improvement
The costs of LLM services are concentrated mainly in training, inference, storage, network transmission, R&D, and customer delivery. The larger the model and the higher the usage, the more evident compute resource consumption tends to be. To improve gross margin, a MaaS provider cannot rely only on raising prices; it must reduce unit usage costs through both technology and operations.
1. Increase Inference Resource Utilization
Inference services usually experience significant traffic fluctuations. If resources are reserved for peak request volumes over the long term, much capacity may sit idle. If resources are configured too low, latency may rise and requests may queue.
A more mature resource scheduling system dynamically allocates compute based on request volume, response latency, concurrency, and task priority. It can improve equipment utilization through batching, request merging, and elastic scaling. For non-real-time tasks, asynchronous processing can shift work away from peak periods and reduce overall resource costs.
2. Adopt a Model Tiering Strategy
Different business tasks have different requirements for model capability. Complex reasoning, long-document analysis, and high-risk decision-making require stronger models, while simple classification, information extraction, and text rewriting can use lighter-weight models.
A MaaS platform can therefore adopt model tiering:
- Use high-capability models for high-complexity tasks;
- Use general-purpose models for standard question answering;
- Use lightweight models for high-frequency, low-complexity tasks;
- Use rules, caching, or specialized models for fixed-format tasks.
This routing approach prevents every request from entering the most expensive model and thereby lowers average token cost.
3. Reduce Wasteful Token Consumption
Tokens are an important unit of measurement for LLM services. Overly long system prompts, repeated context, irrelevant historical conversations, and unnecessary retrieved content all increase input costs.
When designing applications, enterprises need to control context length and compress and rank knowledge base retrieval results. For repeated questions, caching can be introduced to return valid results for frequent requests directly. For structured tasks, output formats should be constrained as much as possible to reduce redundant content.
4. Improve Model-Call Success Rates
If a request frequently times out, fails, or produces unusable results, customers may repeat the call. This increases costs and degrades the user experience. Timeout control, retry strategies, fallback models, result validation, and exception monitoring can reduce invalid calls.
It is important to note that retries should not continue indefinitely. Reasonable retries should depend on the type of error: temporary network errors can be retried, while parameter errors or content that violates constraints should enter a correction process; otherwise, call volume can spiral out of control.
Technical Architecture Priorities for a MaaS Platform
A sustainably operated MaaS platform typically includes not only the models themselves but also comprehensive service governance capabilities.
The first layer is the access layer, responsible for authentication, rate limiting, request routing, and protocol compatibility. The platform needs to configure access permissions, usage quotas, and concurrency limits for different customers so that a single account does not affect overall service stability.
The second layer is the scheduling layer, responsible for distributing requests according to model type, task priority, real-time load, and regional resources. The scheduling system also needs to support failover, switching requests to other available nodes when an inference node becomes abnormal.
The third layer is the inference layer, responsible for actually generating results. This layer needs to monitor metrics such as first-token latency, overall response time, concurrent throughput, and resource utilization.
The fourth layer is the data and observability layer, which includes call logs, usage statistics, cost accounting, quality evaluation, and security auditing. Without a well-developed observability system, it is difficult for a provider to determine which customers truly contribute revenue or to locate cost anomalies.
The fifth layer is the operations layer, responsible for customer lifecycle management. It creates a closed loop from trial, evaluation, and production to expansion and renewal. Only by combining the technical platform with commercial operations can model calls be converted into stable revenue.
What Enterprise Customers Care About Beyond Model Capability
Model capability is an important factor in customer choice, but it is not the only one. Once models enter production environments, enterprises usually care more about the following issues:
- Whether the service is stable and can handle peak requests;
- Whether data processing meets security and compliance requirements;
- Whether usage costs are predictable;
- Whether model version upgrades will affect existing applications;
- Whether log tracing, permission control, and usage management are supported;
- Whether anomalies can be diagnosed and recovered quickly.
A MaaS platform therefore needs to provide clear service-level commitments, usage statistics, and failure-handling mechanisms. For enterprise developers, integration cost also matters. Standardized documentation, unified request formats, stable response structures, and well-designed error codes often have more influence on long-term adoption than the best result achieved in a single test.
From Revenue Growth to Operating Quality
Rapid revenue growth does not mean the business model has matured. LLM companies also need to continuously monitor several key indicators:
- New-customer conversion rate: how many trial customers progress to production environments;
- Customer retention rate: whether customers continue to call models rather than stopping after a short trial;
- Revenue per customer: whether key customers increase usage as their business expands;
- Unit inference cost: whether costs decline as model call volume increases;
- Gross margin: whether scaling genuinely brings operating efficiency improvements;
- Customer concentration: whether revenue depends excessively on a small number of large customers;
- Cash flow: whether contract revenue is converted into cash collection in a timely manner.
Customer expansion and declining unit costs are especially important. If the number of customers increases but each customer’s usage remains at the experimentation stage, the business may still lack sustainability. If rising call volume brings proportional or even higher compute spending, a larger revenue scale may not improve profitability.
Implications for LLM Commercialization
Moving from project-based services to MaaS does not mean project delivery will disappear entirely. In practice, enterprise customers often still need knowledge base construction, permission systems, data governance, workflow orchestration, and application integration. A more reasonable path is to use one-time implementation services as an entry point for customers, then generate long-term revenue through standardized model calls, subscription services, and continuous operations.
This requires LLM companies to possess three capabilities at the same time. First, they must continuously improve foundation model quality and ensure performance on complex tasks. Second, they must build stable, low-cost, observable inference infrastructure. Third, they must establish clear product and operations systems for developers and enterprise users.
Overall, the value of MaaS is not simply placing a model behind an API. It lies in turning model capability into a metered, operable, and reusable digital service. The rapid growth of ARR suggests that the market is gradually shifting from “whether large models are needed” to “how large models can be embedded in production workflows.” In the future, truly competitive companies will need not only stronger models, but also the ability to support continuous customer usage with lower costs, higher stability, and clearer commercial rules.
