Source | Bohu Finance (bohuFN)

Author | All too well

On March 31, Zhipu, touted as the "pioneer of large model IPOs," unveiled its inaugural annual financial results post-listing.

The financial report revealed that Zhipu's total revenue for 2025 surpassed 700 million yuan, marking a staggering year-on-year surge of approximately 132%. However, this growth was accompanied by a widening net loss, reaching 4.7 billion yuan, up 59.5% year-on-year, primarily due to escalating R&D investments. The adjusted net loss stood at 3.18 billion yuan, up 29.1% year-on-year.

Despite these substantial losses, Zhipu's performance in the capital market has been nothing short of remarkable.

Following the release of the financial report, Zhipu's stock price soared by over 30%, and by April 1, its market capitalization had surpassed the 400 billion Hong Kong dollar threshold.

One cannot help but marvel at the fact that the valuation of large model companies seems to hinge more on imagination than on the models themselves.

01 Why the Excitement?

What is driving the market's enthusiasm? Let's delve into the financial report.

Zhipu's revenue primarily stems from two business segments.

The first segment involves local deployment, where AI systems are directly installed on clients' servers. This approach caters to customers with stringent data security requirements, such as governments, large state-owned enterprises, and financial institutions. Although this "project-based" model boasts high customer loyalty, it is characterized by long delivery cycles and significant resource investment.

In 2025, revenue from this segment surged from 260 million yuan to 530 million yuan, accounting for 74% of total revenue, with a growth rate of 102.3%. It currently serves as Zhipu's primary revenue generator.

However, the growth potential of this segment is clearly capped.

The project-based model's growth is constrained by the efficiency of the sales team and project delivery capabilities, with renewal rates subject to uncertainty. Moreover, high implementation costs and resource-intensive operations make large-scale expansion challenging.

Notably, the gross margin for this segment declined from 66.0% in 2024 to 48.8% in 2025, indicating a significant drop in profitability. The financial report attributed this decline to increased delivery resources invested to meet customer demands, leading to a temporary margin squeeze. Meanwhile, the revenue share of local deployment is gradually diminishing.

In contrast, the second segment, cloud deployment, although currently smaller in scale, offers greater growth potential.

Zhipu's cloud deployment operates under the MaaS (Model as a Service) model, providing model invocation services to developers and enterprises through public cloud APIs on a pay-per-use basis. This standardized and lightweight delivery model grows independently of human expansion and relies on model capabilities and invocation scale.

In 2025, revenue from this segment reached 190 million yuan, surging by 292.6% year-on-year, becoming Zhipu's fastest-growing business segment.

More remarkably, in February 2026, Zhipu consecutively raised API service prices, with cumulative increases of 83% in the first quarter. Despite these price hikes, invocation volume still grew by 400%, indicating that market demand far outstrips supply.

This surge is attributed to temporary tightness in computing power supply (driven by the explosion of Agent scenarios brought by OpenClaw, where single-task token consumption in Agent scenarios far exceeds ordinary conversations and usage frequency is more stable) and the model's proven strength. For instance, GLM-5, released on February 11, scored 50 points in Artificial Analysis's Intelligence Index, ranking fifth globally and first among open-source models.

This is also why the market is so focused on Zhipu's cloud business—because the path Zhipu is taking has already been validated overseas by Anthropic, which serves as Zhipu's benchmark.

Anthropic is a quintessential example of the API business model among U.S. AI companies, with its core strategy being delivering the most powerful models to enterprises and developers via APIs.

Currently, Anthropic serves over 300,000 enterprise clients. Over the past year, the number of clients spending over $100,000 annually has grown sevenfold, with over 500 clients spending over $1 million annually. Dubbed the "fastest-growing enterprise software company in history" by The Wall Street Journal, Anthropic has experienced annual growth exceeding tenfold over the past three years.

Following SpaceX's filing, Anthropic has emerged as the second-highest-valued AI unicorn globally, second only to OpenAI.

When model capabilities are strong enough, Model as a Service becomes the core business model.

Zhipu is eager to test these waters.

Currently, nine of China's top ten internet companies are deeply integrating Zhipu's GLM model. As of March 2026, Zhipu's platform has registered over 4 million enterprises and users, with services spanning more than 218 countries and regions globally.

During the annual earnings call, Zhipu CEO Zhang Peng revealed that the full-scale explosion of its open platform and API business was the core growth driver in 2025, with the MaaS platform's annual recurring revenue reaching approximately 1.7 billion yuan, a 60-fold increase over the past 12 months.

More notably, through engineering optimizations on the inference side, Zhipu significantly reduced the unit cost of tokens, markedly improving business profitability. The gross margin of the MaaS platform nearly quintupled to 18.9%, clearly above the industry average.

Buoyed by the explosion of Agent scenarios like OpenClaw, cloud revenue is highly likely to witness a genuine surge in the next one to two years, which is why everyone is excited.

But things aren't that simple.

02 Optimists Forge Ahead

The question remains: Is this growth sustainable?

The financial report indicates that a significant portion of Zhipu's API revenue comes from Chinese internet giants. Although these companies have developed their own models, they opt to leverage external model capabilities in specific business scenarios. This "multi-model invocation" model provides a stable source of demand but does not equate to true large-scale growth. Relying solely on top clients—who are also competitors—poses significant risks.

Zhipu is not the only player eyeing this path.

Compared to MiniMax, Zhipu does not hold an advantage in model invocation volume. On the OpenRouter leaderboard, MiniMax dominated for five consecutive weeks, with monthly invocations reaching 6.9 trillion tokens, while Zhipu and Yuezhiyuan's main models both had around 2.7 trillion tokens in invocations.

Meanwhile, cloud giants like Alibaba, Baidu, and Tencent are ramping up their MaaS efforts, boasting larger cloud infrastructures and richer enterprise client resources. Competitors like ByteDance and DeepSeek are also closing in on model capabilities. The competition to win over developers and enterprise clients is far from over.

Correspondingly, R&D costs remain a concern.

The financial report revealed that Zhipu incurred a full-year loss of 4.718 billion yuan in 2025, up 59.5% year-on-year, with R&D expenditures reaching 3.18 billion yuan, up 44.9% year-on-year. This means the company spends 4.4 yuan on R&D for every 1 yuan earned.

Meanwhile, capital expenditures dropped from 460 million yuan in 2024 to 74.7 million yuan, down 83.8% year-on-year. In simpler terms, Zhipu shifted its computing power procurement from a relatively fixed leasing model to a combination of leasing and service procurement, thereby reducing capital expenditures.

However, enhancing model capabilities still requires sustained increases in R&D investment. R&D and computing power costs do not naturally decline as invocation scale expands. On the contrary, the premise of revenue growth itself drives up costs.

This places large model companies in a common dilemma. To secure more invocations, they must continually enhance model capabilities; yet to improve model capabilities, they must keep increasing investment.

The faster the growth, the greater the cost pressure.

Of course, this is not just Zhipu's problem but a challenge facing the entire large model industry. Before these issues are truly resolved, MaaS can drive growth but will struggle to deliver profits.

But Anthropic's growth is too rapid to ignore.

In mid-February 2026, Anthropic announced an annualized revenue of approximately $14 billion; just three weeks later, this figure soared to nearly $19 billion—a nearly $5 billion increase in three weeks. According to internal documents disclosed by The Wall Street Journal, Amazon-backed AI company Anthropic is expected to achieve its first profit in 2028.

However, the issue remains: Zhipu is not Anthropic. The gap between the two is comprehensive.

In terms of revenue scale, Anthropic's annualized revenue nears $19 billion, while Zhipu's MaaS platform ARR is approximately 1.7 billion yuan (about $250 million), nearly 80 times smaller. In terms of business model, around 80% of Anthropic's revenue comes from enterprise-level API invocations, whereas Zhipu's cloud API revenue accounts for only 26.3% of total revenue. In terms of cost environment, Anthropic is backed by Amazon and Google, with abundant computing power and lower costs, while Zhipu faces tight computing power supply and high adaptation costs for domestic chips.

Even so, Zhipu has managed to ride Anthropic's coattails.

But whether it can truly hold its ground depends on Zhipu's own capabilities.

Reference Sources:

1. Chaoqian Laboratory: Ban, Hype, and a $380 Billion Valuation: Anthropic's 2026 Fantastic Voyage

2. China Entrepreneur Magazine: A Lobster Became the God of Wealth for MiniMax, Yuezhiyuan, and Zhipu

3. Dingjiao One: Zhipu, MiniMax: The More Money Burned, the Higher the Market Value

4. AGI Interface: Zhipu Sprints While Bleeding

5. Dolphin Research: Zhipu: Earning 700 Million, Losing 3.2 Billion? Dreams Are Big, Talking About Losses Is 'Small-Minded'

Disclaimer: The cover image and accompanying illustrations in the article are owned by their respective copyright holders. If the copyright owners believe their works are unsuitable for public browsing or should not be used free of charge, please contact us promptly, and our platform will make immediate corrections.