First Annual Report After Going Public: How a Large Model Company Grows MaaS Revenue and Improves Gross Margin

    A Chinese large model company released its first annual report after going public, reporting over RMB 724 million in revenue for 2025 (up 131.9% year over year), with its MaaS API platform's ARR reaching about RMB 1.7 billion (roughly 60x growth) and gross margin rising from 3.3% to 18.9%. This article breaks down what these numbers reveal about the shift from project-based delivery to recurring API consumption, why ARR is a meaningful metric for model businesses, the technical and operational factors behind gross margin improvement, and how enterprises should evaluate MaaS offerings.

    On March 31, a Chinese large model company published its first annual report since going public. According to the report, the company’s revenue for the fiscal year ending December 31, 2025 exceeded RMB 724 million, a year-over-year increase of 131.9%. Historical figures show revenue of roughly RMB 57 million, RMB 125 million, and RMB 312 million in 2022, 2023, and 2024 respectively—a pattern of continuous growth with clearly accelerating momentum.

    Beyond the headline revenue, what matters most to technical practitioners and enterprise users is that the company’s business model is changing: its revenue is shifting from early-stage large model project delivery toward a recurring, consumption-based Model as a Service (MaaS) model.

    The company’s MaaS API platform now generates an annual recurring revenue (ARR) of approximately RMB 1.7 billion, up about 60x year over year. Meanwhile, the platform’s gross margin improved from 3.3% in 2024 to 18.9% in 2025. These two metrics reflect both the sustainability of the business and the improvement in unit economics—signaling that large model companies are moving from the question of “can you provide model capabilities” to “can you provide model capabilities at a sustainable cost.”

    1. From Project Revenue to Recurring Consumption Revenue

    Early-stage large model companies typically rely on customized projects for revenue, such as model adaptation, private deployment, industry solutions, and one-time technical services. These engagements validate customer demand quickly, but they come with several structural problems.

    First, project revenue is inherently discontinuous. Companies must constantly find new customers and sign new contracts; projects completed last year don’t automatically convert into next year’s revenue.

    Second, project delivery carries high labor costs. Beyond model R&D, teams must invest in solution design, data processing, deployment, customer communication, and ongoing support. Even a large contract can end up with thin margins once delivery timelines and customization scope are factored in.

    Third, the project model poorly reflects actual model usage. Whether a customer keeps calling the model, whether call volume grows, and whether the model has made it into production systems—none of this shows up directly in a one-time contract amount.

    MaaS works differently. The platform exposes model capabilities to developers and enterprises through standardized APIs, and customers pay based on call volume, input/output token counts, or other resource consumption. Once a customer has integrated the model into customer service, office productivity, search, data analytics, code generation, or workflow automation, usage tends to persist.

    The core of MaaS revenue, therefore, is not “selling a model service once,” but embedding the model into the customer’s daily business processes, creating stable usage frequency and consumption volume.

    2. Why ARR Measures the Quality of a Model Business

    Annual recurring revenue is typically used to measure businesses with subscription or sustained consumption characteristics. For a MaaS platform, ARR is not the same as recognized annual revenue; it is an estimate of the revenue capacity likely to carry over into the next year, based on current customer usage, contract size, or sustained consumption levels.

    In the large model context, ARR mainly reflects three things.

    1. Whether customers have moved from testing to production

    Many enterprises start with model evaluations—accuracy tests, cost testing, security validation, and concurrency testing. But test-time usage doesn’t represent real commercial value.

    Only when a model is integrated into production systems does usage become stable business traffic. For example, when a model powers internal knowledge Q&A, employees generate queries every day; when it powers customer service, user requests continuously trigger inference; when it handles content moderation or data extraction, business data keeps flowing into the pipeline.

    Sustained ARR growth therefore usually signals that more customers have completed production deployments.

    2. Whether customers are expanding their usage

    MaaS growth comes not only from new customers but from existing customers expanding their usage. A customer may start with a trial in one department and later roll out across multiple business lines; they may begin with text generation and then add structured extraction, long-document analysis, intelligent retrieval, and automated workflows.

    This kind of expansion drives higher token consumption and more stable call volumes, and it is a major source of platform revenue growth.

    3. Whether revenue is predictable

    Compared with a single large project, API calls spread across many enterprises and use cases tend to form a more diversified revenue mix. Even if one customer’s usage declines, others may keep growing, reducing revenue volatility.

    That said, ARR is not a perfectly stable metric. Customer budget changes, model migrations, rising inference costs, and seasonality can all create gaps between projected and actual revenue. Companies still need to evaluate ARR alongside customer retention, net revenue retention, customer concentration, and usage growth curves.

    3. Why MaaS Often Delivers Revenue Growth Without Profit Growth

    MaaS revenue is directly tied to compute costs. More customer calls mean more requests to process and higher GPU or other compute consumption. If a platform simply scales up call volume without improving resource utilization, revenue growth doesn’t necessarily translate into profit growth.

    The main cost categories typically include:

    1. Inference compute costs

    Model inference consumes compute continuously. Longer inputs, longer outputs, and larger model sizes all increase the compute required per request.

    For high-concurrency workloads, platforms must also reserve headroom to avoid latency spikes during traffic bursts. Idle resources during off-peak hours and insufficient capacity during peaks both hurt cost efficiency.

    2. Storage and data processing costs

    Enterprise applications often require persisting conversation history, knowledge base content, logs, prompt templates, and evaluation results. Long-context scenarios may also involve document chunking, vectorization, index building, and data cleaning.

    These steps don’t always show up as model call fees, but they add to storage, compute, and operations spending.

    3. Network and system reliability costs

    An enterprise-facing MaaS platform must deliver a stable API service, including rate limiting, authentication, permission management, failover, monitoring and alerting, and security auditing.

    If customers rely on the model for mission-critical workloads, the platform must meet higher availability requirements—multi-region deployment, redundant resources, and dedicated ops teams all weigh on gross margin.

    4. Customer onboarding and support costs

    Although MaaS emphasizes standardization, enterprise customers often still need help with model selection, prompt optimization, knowledge base integration, evaluation, and workflow redesign. If the platform applies deep customization to every customer, service costs climb quickly and erode the economies of scale that make API businesses attractive.

    4. The Technical Drivers Behind Gross Margin Improvement

    The platform’s MaaS gross margin rising from 3.3% to 18.9% doesn’t mean compute costs have disappeared. It more likely reflects combined improvements in technical efficiency, product mix, and customer scale.

    1. Better inference efficiency

    Per-unit serving cost is closely tied to inference efficiency. General techniques such as batching, request coalescing, cache reuse, dynamic routing, and load balancing can raise compute utilization.

    For example, when requests arrive in concentrated bursts, the platform can process them in combined batches to reduce idle compute; for highly repetitive requests, response caching can eliminate redundant inference.

    These optimizations don’t change the API surface customers see, but they lower the average resource cost per call.

    2. Multi-model routing to reduce serving costs

    Different tasks demand different model capabilities. Complex reasoning, long-document analysis, and difficult generation tasks may require more capable—and more expensive—models; simple classification, summarization, rewriting, and information extraction can be handled by cheaper, lighter models.

    Through a unified gateway with policy-based routing, the platform can select the right model based on task type, latency requirements, text length, and customer budget. This satisfies quality requirements while avoiding the waste of routing every request to the most expensive model.

    3. Lightweight models expand servable use cases

    Model compression, quantization, distillation, and training task-specific models are important levers for reducing inference costs. For tasks with fixed formats and clear objectives, specialized small models are often more cost-effective than general-purpose large models.

    For example, enterprise tasks like intent recognition, field extraction, and text classification don’t necessarily require open-ended generation. Shifting these workloads to lighter models reduces the share of calls served by expensive large models.

    4. Scale improves resource utilization

    When API volume is small, compute resources sit underutilized and fixed operations costs are hard to amortize. As customer count and request volume grow, the platform can forecast capacity more accurately and improve utilization through elastic scaling.

    From a business perspective, scale growth brings not just more revenue, but potentially lower infrastructure cost per request.

    5. From “Selling a Model” to “Managing Token Consumption”

    A MaaS platform’s core asset is not a single model, but the complete service chain built around model usage. To improve retention and revenue quality, the platform must continuously focus on the actual value of token consumption.

    A mature MaaS system typically includes:

    • Unified API access that lowers the barrier for developers;
    • API key, tenant, and permission management to meet enterprise security requirements;
    • Usage analytics and cost reporting to help customers control budgets;
    • Retry, rate limiting, and circuit breaking to improve service stability;
    • Model evaluation to continuously monitor output quality;
    • Logging, auditing, and sensitive data handling to reduce data risks;
    • Version management and gradual rollout to prevent model upgrades from breaking business workflows;
    • Multi-model orchestration so different tasks can use different models.

    For developers, the decisive factors in platform selection go beyond model size. They include the end-to-end experience after integration: response speed, output stability, context handling, documentation quality, failure recovery, and cost predictability.

    MaaS competition, therefore, is gradually expanding from single-model capability to engineering platform capability and developer ecosystem capability.

    6. How Enterprises Should Evaluate MaaS Offerings

    When choosing a MaaS provider, enterprises shouldn’t compare per-call prices alone. They need a complete unit economics model.

    First, measure the true cost of each business workflow. Beyond API fees, account for data processing, retrieval, storage, system integration, monitoring, and human review costs.

    Second, assess whether model output actually reduces manual work. If calls are cheap but still require heavy human verification, the overall return may be disappointing. Conversely, even if per-call costs are somewhat higher, total ROI can still be favorable if the model significantly reduces repetitive labor.

    Third, understand the boundaries of usage growth. Enterprises should verify in advance whether the system runs reliably under peak traffic, long-context inputs, and high concurrency, and confirm the mechanisms for rate limiting, timeouts, and failure handling.

    Finally, establish joint monitoring of quality and cost—for example, tracking request success rate, average latency, token consumption, human correction rate, and business conversion rate together. Only by combining technical metrics with business metrics can you judge whether a MaaS application truly creates value.

    7. Key Challenges for the Next Stage of MaaS

    Despite strong growth potential, MaaS faces several long-term challenges.

    First, as model capabilities converge, price competition may intensify. Platforms will need to differentiate through tooling, industry adaptation, service reliability, and developer experience—not just the models themselves.

    Second, customers may adopt multi-model strategies. To reduce vendor risk and control costs, enterprises may integrate several model providers simultaneously. Platforms will need compatibility, routing capabilities, and easy migration paths to strengthen customer stickiness.

    Third, compute costs remain uncertain. Changes in model scale, request length, and traffic patterns can all reshape the cost structure. Platforms must keep optimizing inference efficiency and build fine-grained cost accounting systems.

    Fourth, data security and compliance requirements keep rising. Enterprise customers increasingly care about data isolation, access control, log auditing, and sensitive information protection. Only by baking security into the standard product can MaaS platforms expand into more mission-critical workloads.

    Overall, the revenue growth, ARR gains, and gross margin improvement in this report show that large model commercialization is shifting from “project delivery” to “platform-based services.” The key to this transition is not just having a stronger foundation model, but the ability to continuously lower per-call costs, deepen customer usage, and reliably embed model capabilities into real business workflows.

    For large model companies, the metrics that will truly matter going forward are no longer limited to parameter counts or fundraising totals. What counts is whether customers keep calling, whether the platform can scale stably, whether unit economics improve, and whether every token consumed creates measurable business value.