The first annual report of the first major AI model company: No. 1 in domestic revenue, Zhang Peng discusses Token economy

Ask AI · How Will Token Economics Change the Artificial Intelligence Ecosystem?

On March 31, “the first global large-model public company,” Zhipu delivered its first financial report after going public. The report shows that as of December 31, 2025, Zhipu achieved full-year revenue of more than 724 million yuan (RMB, the same below), up 131.9% year over year. In terms of revenue scale, Zhipu has become the largest domestic large-model company by revenue volume. By contrast, another Hong Kong-listed large-model company, MiniMax, reported on March 2 that its last year’s total revenue was $79.04M (about 543 million yuan), up 158.9% year over year.

Revenue Growth Up 131.9%, Losses Widen, Gross Margin Under Pressure

According to the financial report, Zhipu generated revenue of 724 million yuan in 2025, up 131.9%; net losses for 2025 were 4.7 billion yuan, up from 2.9 billion yuan in 2024. The report says this was mainly due to the continued increase in R&D spending. On gross margin, the company recorded nearly 300 million yuan in 2025, up 68.7%, but gross margin declined—from 56.3% in 2024 to 41.0% in 2025. The report attributes this mainly to the increased share of cloud-deployment business and a phased decline in gross margin for local deployment business.

Zhipu emphasizes that, among China’s top 10 internet companies, 9 of them use its GLM model deeply. By March 2026, Zhipu’s registered enterprise and user base exceeded 4 million, serving more than 218 countries and regions worldwide.

Using the prior classification method, for all of 2025, Zhipu’s local deployment revenue was 534 million yuan, up 102.3%, accounting for 73.7% of total revenue; cloud deployment revenue was 190 million yuan, up 292.6%, accounting for 26.3%.

In addition, Zhipu said in the report that, because the previous deployment-led classification can no longer cover the company’s current products, it split its offerings by business format and core product lines. Using this classification, its open platform and API business generated 190 million yuan in revenue in 2025, up 292.6%, accounting for 26.3% of total revenue; the enterprise-level agent segment generated 160 million yuan last year, up 248.8%, accounting for 22.9% of total revenue; enterprise-grade general large models generated 360 million yuan in 2025, up 70.5%, accounting for 50.4% of total revenue; technical services and others generated 2.5 million yuan, up 31.6%, accounting for 0.4% of total revenue.

At the earnings release conference, Zhipu CEO Zhang Peng said that the company’s total revenue in 2025 exceeded the target set at the beginning of the year. The core growth driver is the full breakout of the Zhipu MaaS platform (i.e., the open platform and API business). Currently, Zhipu’s MaaS platform ARR (annual recurring revenue) is about 1.7 billion yuan, up by 60x over the past 12 months. In addition, by optimizing engineering on the inference side to an extreme degree, Zhipu has significantly reduced the per-Token unit cost, substantially improving its profitability level; MaaS platform gross margin has increased nearly 5x to 18.9%, far exceeding industry standards.

Based on the revenue mix in the financial data above, although Zhipu has built growth capabilities related to the MaaS platform, Zhipu’s 2025 revenue still mainly comes from local deployment and enterprise-level general large models.

API Pricing Up 83%, Yet Still Out of Supply

In the company’s first financial report after listing, Zhipu put forward two concepts: one is the “TAC era.” Zhipu says that when large models have closed-loop capabilities for long-horizon task execution, the core competitive advantage will be reshaped into TAC (Token Architecture Capability, i.e., Token architecture capability).

Second is the “large model operating system” (LLM-OS). Zhipu believes that a traditional OS (operating system) is a scheduler of hardware resources, while LLM-OS (large model operating system) is an intelligent scheduler. Large models are consuming software; in the future, computing platforms will no longer be stacks of apps, but the collaboration between an API marketplace and an Agent matrix. Whoever’s model enters the system kernel will control the definition of the next generation of computing.

When asked whether Zhipu has already started to plan around these two concepts, Zhang Peng said that, currently, TAC is something that’s happening—for example, token economics, token conversion efficiency, and so on. At the same time, some companies, besides paying cash wages, also distribute Token as part of salaries to employees, which indicates that Token itself already has economic value and intelligent value. Additionally, the large model operating system (LLM OS) is not staying at the conceptual level; Zhipu has been paying attention to this future trend early on and is also discussing with ecosystem partners and operating system vendors whether the base model capabilities can be deeply integrated with the operating system to bring new product experiences.

The financial report shows that as of March this year, Zhipu’s API calling price increased by 83% compared with the end of last year, but demand still exceeds supply. Regarding this phenomenon, Zhang Peng said that in the overall market environment, there are many vendors providing API services based on compute power. Customers’ acceptance of Zhipu’s price increases and their continued choices indicate that the model capability is indeed leading. “I think the essence of long-term pricing is determined by value. Resources that can effectively replace labor, improve conversion efficiency, and enhance intelligence levels are scarce and valuable. We focus more on the value created per unit of Token and the value it brings to customers. Only when customers recognize this value will they be willing to pay higher costs to use it. I believe pricing power is still determined by our technological strength and our leading position driven by long-term trends.”

When discussing Zhipu’s core moat in market competition, Zhang Peng said that from a technical perspective, Zhipu has continued to maintain a leading position in Coding and Agentic capabilities. In addition, Zhipu has already accumulated a large amount of ecosystem resources through commercialization, including more than 4 million enterprise and developer users. These developers and companies build strong TAC capabilities based on Zhipu’s model foundation, APIs, and Tokens, which will gradually influence a broader range of enterprises, developers, and users.

From the angle of competition, Zhang Peng believes, “Big companies will definitely self-develop, but they may not be able to achieve sufficient competitiveness in all scenarios—this is determined by resource constraints. Moreover, big companies themselves are complex ecosystems; they will not rely entirely on themselves at every node. They will also connect with excellent technology suppliers to ensure a fresh supply of new blood in the ecosystem and avoid losing the initiative at any point in time. In a stage where technology iterates quickly and model capability still dominates, independent large model vendors have inherent advantages—this is the core reason why Zhipu can stand firm.”

Reporting/Writing: Lin Wenqi, Nandu N Video reporter