Zhipu MaaS ARR reaches 1.7 billion yuan, highlighting the value of the intelligent foundation

Ask AI · How will Token Architecture power reshape the standards by which enterprise value is measured?

Zhipu’s goal is to ensure that every drop of Token can be converted into deliverable economic incremental value.

According to an IPO early-warning report, Zhipu (2513.HK) released its annual performance announcement for the year ended December 31, 2025 on March 31, which is also its first performance announcement since becoming the “world’s first publicly listed large model company” on January 8 of this year.

The financial report shows that Zhipu’s 2025 revenue exceeded RMB 724 million, up 131.9% year over year, and continues to hold the top spot among domestic large model companies by revenue.

Among this, the full-scale surge of its MaaS platform (i.e., the open platform and API business) has become the core factor driving Zhipu’s performance growth—currently, Zhipu’s MaaS platform ARR is about RMB 1.7 billion, up 60x over the past 12 months. Through extreme engineering optimization on the inference side, Zhipu significantly reduced Token unit costs, resulting in a major improvement in business profitability. The MaaS platform’s gross margin rose nearly 5x to 18.9%, far exceeding industry benchmarks.

Worth noting is that Zhipu for the first time defined the value of AI productivity—Token Architect Capability (TAC), i.e., “intelligent usage volume × intelligent quality × economic conversion efficiency.” As enterprise TAC demand continues to grow, Zhipu’s MaaS platform is becoming the infrastructure connecting foundation models and industry applications.

In Zhipu CEO Zhang Peng’s view, improving the ceiling of intelligence is the only “first principle” in the era of large-model AGI. In 2026, Zhipu will further explore breakthroughs in raising the ceiling of intelligence, leading paradigm innovation, and continuously outputting high-quality Tokens as a foundation.

The China-based large-model company that’s most like Anthropic right now

There’s no denying that Anthropic is one of the most watched companies in the global AI industry over the past year, and its growth logic is very clear— focusing on delivering the strongest model via API to enterprises and developers so that intelligence participates in creating economic value. From the end of 2024 to the end of 2025, Anthropic’s ARR quickly jumped from 1 billion USD to 9 billion USD, which also indicates: when the model is strong enough, the API itself is the best business model.

In fact, Zhipu is also the China-based large-model company most like Anthropic today—Anthropic’s early growth relied on the model being used at high frequency by developers worldwide, and Zhipu’s growth clearly follows the same pattern. One data point as evidence is that in 2025, Zhipu became the first in China to launch the programming package “GLM Coding Plan,” which rapidly covered the globe. The number of paying developers quickly surpassed 242k; Token usage volume increased 15x in six months; and developers are precisely the most sensitive group in perceiving the ceiling of intelligence. This year in February, even after raising prices by 30% and canceling the first-purchase discount, Zhipu’s programming package still remained in a state of supply not meeting demand, becoming one of the AI programming services with the fastest growth rate globally.

On the other hand, around 80% of Anthropic’s revenue comes from enterprise-level API call services. Similar to this, leveraging BigModel.cn, Zhipu’s MaaS platform has now become the hub connecting foundation models with 4 million enterprise applications and developers.** Among the top 10 internet companies in China, 9 already make deep daily calls to GLM; after each model release, within 24 hours it is recognized by leading companies.** For example, within 24 hours after the release of Zhipu’s flagship foundation model GLM-5, it received official integrations from leading platform product offerings such as ByteDance TRAE, Alibaba Qoder, Tencent CodeBuddy, Meituan CatPaw, Kuaishou 万擎, Baidu Intelligent Cloud, and WPS Office.

From a global perspective, Zhipu has also achieved value monetization of Tokens worldwide, and each time it releases models, international markets have shown strong attention. Currently, the GLM models have been fully deployed across top global cloud service providers such as Google Vertex AI, AWS Bedrock, Fireworks, and Cerebras, and have been onboarded onto major international model aggregation platforms such as OpenRouter and Vercel. It is also ranked as OpenRouter’s NO.1 paid model. In addition, GLM has become the default model for internationally well-known Coding platforms (such as Windsurf) and well-known CodingAgent platforms (such as OpenCode).

“Both volume and price rising” reflects the pursuit of “the ceiling of intelligence”

Of course, behind the commercial breakthrough of Zhipu’s MaaS, lies Zhipu’s extreme pursuit of the first principle of the AGI era: “the ceiling of intelligence.”

In 2025, Zhipu completed a leap in industry paradigm—from Vibe Coding (atmosphere-based programming) to Agentic Engineering (agent-oriented engineering)—with high-frequency releases ranging from GLM-4.5, 4.6, 4.7 to GLM-5 and 5-Turbo. It has continuously topped global open-source models as well as China models. Among global models, it ranks only behind GPT, Claude, and Gemini, and stays firmly in the top tier. Meanwhile, GLM-5 scored 50 points on the Artificial Analysis Intelligence Index list—this is the first time an open-weight model has reached this level of score.

In addition, Zhipu’s proprietary Slime framework enables a revolution in asynchronous reinforcement learning efficiency. Coupled with its self-developed algorithms, the model learns efficiently from more than 10k real software engineering environments, supporting the birth of the world’s first OpenClaw foundation model, GLM-5-Turbo. At the same time, GLM-5 completes co-design of software and hardware for domestic chips. Through innovative quantization strategies, it compresses VRAM usage to the extreme; a single domestic server can be stably deployed. Costs are reduced by 50%, and it achieves inference efficiency comparable to top international chips on domestic chips, building a technical closed loop of “ceiling of intelligence + autonomous computing foundation.”

Backed by confidence brought by higher-order intelligence, Zhipu’s API call pricing increased by 83% in this year’s first quarter. Even so, the market still shows a situation of supply not meeting demand: call volume still grew by 400%, again confirming that high-quality Tokens are a scarce resource today. Whoever controls the ceiling of intelligence controls pricing power.

From the above, it’s not hard to see that, “using the ceiling of intelligence as a barrier and API as the main product form” is the commercial path that Anthropic and Zhipu are both carrying out.

Zhang Peng emphasized that, after becoming one of the domestic manufacturers with the highest paid Token consumption, breakthroughs in the ceiling of intelligence will further drive an exponential increase in Token consumption— as models become stronger and users’ usage scenarios become deeper, more diversified, and more complex, Token call volume will grow more and more.

“Both Anthropic’s 10x growth in ARR over the past year and Zhipu’s 60x growth in its API platform reflect that growth is no longer linear. Meanwhile, the positive feedback loop in business supports us to invest in more computing power and R&D, further raising the ceiling of intelligence—this flywheel has already started turning,” Zhang Peng said.

The long-term value of the “intelligent foundation” is expected to keep being released

It can be said that Zhipu is essentially not a traditional software company in the conventional sense, but a native intelligence laboratory built on a belief in AGI.

“Our moat isn’t built by stacking up computing power, but by deconstructing the underlying essence of intelligence, and by the resolve to convert that understanding into social productive forces.” Zhang Peng said, “Looking ahead to 2026, the intelligent paradigm will evolve from lightweight Vibe Coding into industrial-grade Agentic Engineering, and further into digital engineers with the ability to autonomously plan, perceive the environment, and self-iterate. Ultimately, it will achieve a Long-horizon Task (long-range task) closed-loop execution with multi-step iteration and logical consistency. This will further bring breakthroughs in the ceiling of intelligence and exponential growth in Token calls.”

Also, according to Zhang Peng, in the era of large models, when large models have closed-loop capabilities for long-horizon task execution, the core competitiveness will be reshaped into TAC (Token Architecture Capability, Token architecture capability).

From this perspective, today’s Zhipu customers—whether enterprises, developers, or product users—are essentially individuals using AI to create productivity and generate economic value. When large models have closed-loop capabilities for long-horizon task execution, the future standard for measuring the value of organizations and employees will be their ability, as Token architects, to deliver tasks with agents. And as a foundation-model provider, Zhipu’s enterprise services deliver not only the models themselves, but also the capability for enterprises to truly make good use of the models and build TAC digital employees.

“TAC = intelligent usage volume × intelligent quality × economic conversion efficiency. In the future, the standard for measuring the value of an individual or an organization will no longer be how much information they possess, but rather their ability, as a Token architect, to build complex Agent systems under a given budget and drive the large model to run autonomously to complete complex Agent system operations. Zhipu’s goal is to become the infrastructure for improving TAC across the entire society, so that every drop of Token can be converted into deliverable economic incremental value.” Zhang Peng added.

Looking further into the long term, market attention will ultimately return to model performance. When the model gets better, the business model will have more opportunities for scaling up.

As mentioned earlier by JPMorgan, model capability is the fundamental element that determines long-term competitiveness. The long-term economic benefit of large-model vendors mainly depends on whether they can keep model capabilities at a globally leading position across multiple rounds of technical cycles. The business model form, deployment approach, and short-term profit margin structure, to a large extent, reflect downstream performance derived from that capability.

And Zhipu, clearly, has already secured its position in dimensions such as technical strength, commercial progress, and ecosystem building. Its long-term value as an “intelligent foundation in the AGI era” is expected to keep being released.