How Zhipu’s MaaS Revenue and Gross Margin Are Changing
Zhipu’s first post-listing financial report points to a shift from project-based revenue toward usage-based MaaS API services. This analysis examines revenue growth, the meaning and limitations of ARR, the possible drivers behind the rise in MaaS gross margin, and practical cost-control considerations for developers and enterprise buyers.
From Project-Based Delivery to Productization: Signals from the Revenue Mix
Historically, many large-model companies generated a significant share of their revenue from customized projects. These projects can increase short-term sales, but each one requires separate delivery and is difficult to reuse at scale. Zhipu’s first financial report after listing suggests that its revenue mix is beginning to shift toward usage-based API services.
According to the filing, the company generated approximately RMB 724 million in revenue for the year ended December 31, 2025, representing year-over-year growth of 131.9%. Revenue was approximately RMB 57 million in 2022, RMB 125 million in 2023, and RMB 312 million in 2024, with growth accelerating in recent years.
Total revenue reflects recognized sales. To determine whether the revenue mix is becoming more stable, it is also necessary to examine the share of recurring revenue.
Understanding ARR: What the Metric Does and Does Not Show
The metric that attracted the most attention in the financial report was the annual recurring revenue (ARR) of the company’s MaaS API platform. Zhipu reported ARR of approximately RMB 1.7 billion, an increase of about 60 times year over year.
ARR generally refers to recurring revenue at a specific point in time or over a particular period, annualized to estimate a full year. Its purpose is to reduce the impact of one-time contracts and show whether customers continue calling the models and consuming tokens on a stable basis.
Three points are important when interpreting this figure.
First, ARR is an annualized measure and is not the same as recognized revenue for the full year. The two figures cannot be directly subtracted or divided to infer profit or completion rates.
Second, the calculation date and methodology can materially affect the result. For example, annualizing revenue from a single month at the end of the period by multiplying it by 12 may produce a very different figure from annualizing quarterly revenue by multiplying it by four.
Third, rapid ARR growth indicates that usage is increasing quickly, but it does not mean that this usage has already translated into stable profits. Readers should rely on the definition provided in the prospectus or annual report and verify whether the figure has been audited.
ARR growth also has practical implications for the business. Once enterprise customers embed model capabilities into their internal workflows, the cost of switching providers tends to increase, which may strengthen usage stickiness. Whether this stickiness is sustainable requires longer-term observation of renewal rates and usage retention. A single year of data is not enough to prove it.
Why MaaS Gross Margin May Have Recovered
The financial report shows that the gross margin of the MaaS business increased from 3.3% in 2024 to 18.9% in 2025. This change may have resulted from several factors. The following three areas are worth examining separately.
- Improved inference efficiency: Technologies such as better KV Cache hit rates, dynamic batching, model quantization, and distillation can reduce the computing resources required for each token request. As usage increases, lower unit costs are more likely to be reflected in gross margin. Relevant indicators include cost per token, cache hit rate, and average batch size. These metrics are not usually disclosed in financial reports, so they must be assessed using the company’s technical announcements or industry research.
- Better compute-resource scheduling: As platform traffic grows, GPU resource pooling and elastic scaling may improve. Scheduling workloads with different latency requirements across the same hardware pool can help reduce hardware idle time.
- Changes in the product mix: Early API usage may have been concentrated in long-form text generation, which generally consumes more computing resources. If lower-cost, higher-priced services such as vectorization, embeddings, and structured information extraction account for a larger share of revenue, the overall gross margin may rise.
Not all of these factors can be verified directly from the financial report. Revenue-recognition practices, cost-allocation methods, and the impact of one-time projects may also cause gross-margin fluctuations. Therefore, when citing the 18.9% figure, it is important to identify the relevant filing and explain the accounting basis used.
The prospectus indicates that the company’s overall gross margin remained above 50% from 2022 through 2024. Overall gross margin and MaaS gross margin are calculated on different bases and should not be compared directly. However, the overall figure provides useful background for understanding changes in the business structure.
A Cost-Governance Checklist
For teams building or using MaaS services, the following checks can help determine whether costs and pricing remain under control.
- Token measurement: Record the number of input and output tokens for each request, along with the model version, customer, and interface type. This ensures that billing data can be reconciled with computing-resource consumption.
- Budget alerts: Set usage thresholds for each customer or project. Send alerts when usage reaches a predefined percentage, and provide visibility into daily and monthly consumption trends.
- Cost-allocation granularity: Allocate GPU or cloud-resource costs by model version, request batch, and interface type. At a minimum, the system should show which types of requests consume the most compute and which generate the highest margins.
- Billing transparency: Provide customers with clear usage details, explain the components of the price, and offer cost-reduction recommendations, such as shortening prompts, enabling caching, or selecting a smaller model.
- Pricing reviews: Compare unit costs and prices for each interface every quarter. If the gross margin of a service continues to decline, adjust pricing or optimize the underlying implementation promptly.
Considerations for Developers and Enterprise Decision-Makers
When selecting a large-model service, benchmark scores and marketing claims are not enough. Buyers can also evaluate whether API pricing has remained stable over time, whether inference latency is reliable during peak periods, whether the provider regularly explains efficiency improvements or price changes, and whether billing is transparently based on actual usage.
Enterprise customers should also review data-security and compliance terms. They need to understand whether request data may be used for model training and where the data is stored. These provisions can directly affect procurement decisions in industries such as finance, government, and manufacturing.
Teams planning to integrate large models into business workflows should first run a small-scale pilot and estimate the cost of each task. They can then use actual token consumption to forecast costs at scale. This approach generally provides more useful cost data than relying on industry averages.
Data Limitations and Further Verification
The data in this analysis comes primarily from the company’s first post-listing financial report and related disclosures. It covers one company and one fiscal year, so the evidence is not sufficient to generalize the conclusions to the entire Chinese market.
The ARR, gross-margin, and revenue-growth figures in the financial report should be checked against the original announcements item by item. Particular attention should be paid to the calculation basis, audit status, and treatment of non-recurring items.
If more companies disclose comparable metrics, or if third-party organizations publish research on the MaaS market, those data points can be compared with this analysis to assess whether its conclusions remain valid.
