Zhipu's 2025 revenue soars, losses widen

How can AI · API mode drive the growth of Zhipu token calls?

Reporter Qu Zhongfang, China Business Journal, Beijing

On the evening of March 31, “the first stock in large models” Zhipu (02513.HK) released its first financial report since listing. The 2025 financial report shows that Zhipu’s total annual revenue was 724 million yuan, a year-on-year increase of 131.9%; gross profit was 297 million yuan, a year-on-year increase of 68.7%. The loss for the year was 4.718 billion yuan, expanding by 59.5% year-on-year; adjusted net loss was 3.182 billion yuan, expanding by 29.1% year-on-year.

The financial report shows that, in terms of deployment approach, Zhipu’s revenue from cloud deployment since 2025 was 190 million yuan, accounting for 26.3% of total revenue. This figure is up 10.8 percentage points compared with 2024. Localized deployment business still remains the major segment, with revenue of 534 million yuan. However, its contribution ratio to revenue fell from 84.5% in 2024 to 73.7% in 2025.

It needs to be pointed out that cloud deployment has the characteristics of standardization, high reusability, and low marginal costs, while localized deployment often requires on-site implementation and has characteristics of customization and heavy delivery. Based on the gross margin data disclosed in the annual report, Zhipu’s gross margin from cloud deployment increased from 3.3% in 2024 to 18.9% in 2025; the gross margin from localized deployment business fell from 66.0% last year to 48.8%. This suggests that Zhipu is continuously optimizing its business structure, expanding the scale of cloud deployment business, and transitioning to the high-growth, high-reusability MaaS (Model-as-a-Service) cloud model. Zhipu disclosed that its current MaaS platform’s ARR (annual recurring revenue) is about 1.7 billion yuan, which has increased 60-fold over the past 12 months. It has significantly lowered the unit cost of tokens and improved profitability.

In the financial report, Zhipu breaks down its business into 4 major categories based on different business forms and product lines. Enterprise-level general-purpose large models generated revenue of 366 million yuan in 2025, with a year-on-year growth rate of 248.8%, contributing 50.4% of the total; open platform and API, enterprise-level intelligent agents, technical services, and others contributed 26.3%, 22.9%, and 0.4% of revenue, respectively.

According to materials provided by Zhipu’s official team to a reporter from China Business Journal, CEO Zhang Peng repeatedly mentioned overseas AI star company Anthropic—parent company of the large model Claude—during the annual performance telephone conference, using it as a benchmark. Zhang Peng said: “Anthropic’s ARR grew from $1 billion at the end of 2024 to $9 billion by the end of 2025, and 80% of its revenue comes from enterprise-level API calling services. This proves that when the model is strong enough, the API (Application Programming Interface) itself is the best business model. The quality of intelligence creates ‘pricing power,’ and deep usage by enterprises and users creates growth in Scalling (referring to scaled expansion). The same business model and growth are happening at Zhipu.”

As of March 2026, the number of registered enterprises and users on Zhipu’s MaaS platform has exceeded 4 million, and its services cover more than 218 countries and regions globally. In addition, the GLM Coding Plan (programming package) launched first by Zhipu in 2025 attracted 242,000 paying developers worldwide. Over the past 6 months, the token calling volume increased by 15 times. This year in February, even after prices were raised by 30%, the programming package continued to grow rapidly. In March, when OpenClaw’s “lobster farming” went viral, Zhipu launched the ClawPlan earlier in China. Within 20 days of launch, the number of subscribed users exceeded 400,000.

Zhang Peng said in an interview with the media that API is a way to convert AI’s infrastructure capabilities into economic operating resources, not a one-time windfall. At present, AI capabilities have shifted from “available and playable” to truly solving problems that are increasingly complex and important, so that token API calls and consumption can genuinely be converted into economic value.

In Zhang Peng’s view, the intelligent paradigm in 2026 will evolve from lightweight programming into industrial-grade intelligent agent engineering, and further into digital engineers with capabilities for autonomous planning, environmental perception, and self-iteration. Ultimately, this will achieve closed-loop execution of long-range tasks with multi-step iteration and logic consistency, which will further bring exponential growth in token calls.

(Editor: Zhang Jingchao; Review: Li Zhenghao; Proofreading: Zhai Jun)