Before becoming "China's Anthropic," Zhipu still needs to solve three variables.

Produced by/Future Tech World Author/Li Yan Editor/Yang Yu

In 2025, for Zhipu AI, it is a milestone year. This AI company, formed through the transformation of technological achievements from the Knowledge Engineering Laboratory at Tsinghua University, officially listed on the Hong Kong Stock Exchange on January 8, 2026, becoming the “world’s first major-model stock.” However, its first annual financial report shows a confusing “two-sidedness”:

One side is the encouraging growth data: In 2025, Zhipu’s revenue grew 131.9% year-over-year to 724 million yuan; its MaaS platform annual recurring revenue (ARR) surged 60 times to 1.7 billion yuan; registered users exceeded 4 million; and paid developers reached 242,000.

The other side is the alarming loss figures: The company incurred a loss of 4.718 billion yuan during the year, expanding 59.5% year-over-year; the adjusted net loss was 3.182 billion yuan, expanding 29.1% year-over-year; shareholders’ equity was -8.111 billion yuan, putting it in a technical state of “liabilities exceeding assets”; its overall gross margin fell from 56.3% in 2024 to 41.0%.

From the financial report, it’s clear that what Zhipu raised at the beginning of the year was money to buy time. But in the “money-burning” race of AI large models, Zhipu is still standing at the crossroads between life and death.

Is the cloud business starting to become the main line?

Zhipu AI’s revenue growth mainly relies on two business lines.

One is localized deployment. In 2025, this business generated revenue of 534 million yuan, up 102.3% year-over-year, accounting for 73.7% of total revenue. This business mainly serves government and large enterprise customers, providing large-model solutions with private deployment. Because it involves data security and customized needs, the average transaction price is relatively high and customer stickiness is stronger, making it the most stable revenue source for the company at its current stage.

The financial report shows that this “core base” is under pressure. On one hand, vendors such as Baidu, Alibaba, and Huawei have stepped up enterprise large-model solutions, rapidly intensifying competition in the localized deployment market; on the other hand, pricing pressure is starting to show, with this business’s gross margin falling from 66.0% in 2024 to 48.8% in 2025.

By contrast, the cloud deployment business represents the growth direction. In 2025, this business’s revenue was 190 million yuan, up 292.6% year-over-year, far higher than the growth rate of localized deployment, and its share also increased to 26.3%. Its core carrier is the MaaS (Model-as-a-Service) platform, which provides model capabilities to developers and small and medium-sized enterprises via API.

From a business model perspective, compared with localized deployment, cloud deployment has stronger scalability potential. Once model capabilities mature, they can be replicated through APIs at low marginal cost, creating room for exponential growth driven by “user scale × call frequency.”

And that is exactly why, even though this business’s profitability is currently clearly weak, Zhipu AI continues to intensify investment in this direction.

In comparison, MiniMax next door relies on AI-native products as its main source of revenue, with revenue from its open platform and enterprise services accounting for nearly one-third. Its scale is 25.96 million USD (approximately 180 million yuan). Last year, its overall gross margin was 25%, far lower than Zhipu’s overall gross margin, but higher than Zhipu’s cloud deployment gross margin.

The good news is that Zhipu’s cloud deployment gross margin improved from 3.3% in 2024 to 18.9% in 2025. This change indicates that as call volumes expand, compute power costs are being gradually spread out, and scale effects are starting to emerge.

Massive losses, the soaring stock price

For large-model vendors’ massive losses, the industry and capital markets are already not surprised.

In 2025, Zhipu’s loss during the year was 4.718 billion yuan. By contrast, the day after the financial report was released, the stock price surged 31.94%, with the market value jumping by nearly 100 billion yuan in a single day.

Breaking it down, the most core reason for the losses is the sustained ramp-up of R&D investment—annual R&D spending reached 3.18 billion yuan, up 44.9% year-over-year, equivalent to 4.4 times revenue.

But such investment is the norm in the large-model industry. OpenAI’s annual R&D spend has already exceeded 3.0 billion USD, and Anthropic is also above 2.0 billion USD. Top-tier vendors generally operate in a stage of high investment and high losses. In other words, competition in large models is fundamentally a “race to buy a technology ceiling with capital.” Against this backdrop, Zhipu’s high R&D intensity is a necessary cost to maintain its technical seat.

From the output side, such high investment has indeed led to technical breakthroughs. GLM-5 achieved 77.8% on the SWE-bench-Verified programming benchmark, ranking first among open-source models; on the customer side, the company has penetrated 9 of China’s top ten internet companies, including ByteDance, Alibaba, Tencent, and other leading enterprises. These indicators show that Zhipu has entered the first tier in both model capability and enterprise deployment.

Apart from R&D, there are two other changes on the expense side worth paying attention to.

One is the abnormal rise in general and administrative expenses. In 2025, this expense item reached 505 million yuan, up 278.3% year-over-year, mainly driven by IPO-related expenditures, including audit, legal, underwriting fees, and equity incentive costs before listing. These types of expenses have clear one-off characteristics, and after the completion of the listing, they are expected to decline.

The other is the “restraint” in sales expenses. Full-year sales and marketing expenses were 391 million yuan, up only 0.9% year-over-year, far below revenue growth. On the one hand, this reflects improved sales efficiency; on the other hand, it also shows the company’s trade-offs in resource allocation— in the large-model track that is extremely “money-burning,” Zhipu chooses to put more resources into the R&D side instead of driving growth through marketing.

What should be kept in mind is that even with a successful listing, Zhipu’s safety margin is still not sufficient. As of the end of 2025, the company held 2.259 billion yuan in cash and cash equivalents. Based on the current loss pace, without considering financing, cash would only support about 8 to 9 months. But as it successfully entered the capital market and raised about 50 billion HKD, the company’s overall cash reserves increased to about 7.0 billion yuan, corresponding to a “survival period” of about 2 to 2.5 years.

Against Anthropic, can Zhipu’s high valuation hold up?

At Zhipu’s performance meeting on its financial results, founder Zhang Peng compared Zhipu to “Anthropic in China.”

From a business model standpoint, Zhipu and Anthropic are highly similar. Anthropic’s Claude provides services to enterprises via API, and its revenue growth depends entirely on the pricing power of model capability; Zhipu’s MaaS platform works the same way—after the pricing for API calls rose 83% in the first quarter, demand in the market still outstripped supply.

But the similarity in business models cannot automatically translate into a rationale for a similar valuation. According to multiple overseas media outlets such as Reuters and The Wall Street Journal, Anthropic’s valuation is about 350 billion USD, and OpenAI’s valuation is about 830 billion USD. As of the time of writing, Zhipu’s market capitalization is about 347.3 billion HKD (about 44.3 billion USD), and its price-to-sales ratio is close to 500 times—far higher not only than OpenAI, but also higher than Anthropic. This valuation implies an aggressive assumption: that Zhipu’s future growth rate will be significantly higher than that of international peers.

The reality is that Zhipu faces a competition landscape far more complex than that of Anthropic. Anthropic’s main competitors in the United States are OpenAI, creating a relatively stable duopoly structure; while in China, Zhipu has to face many players such as Alibaba, ByteDance, Baidu, DeepSeek, and others, and the tech giants have the ability to fund themselves with traffic, cloud infrastructure, and “cash cow” businesses. In this environment, whether Zhipu can maintain its current pricing power largely depends on whether its technological leadership can be sustained.

At the performance meeting, Zhang Peng also proposed the concept of TAC (Token Architecture Capacity), attempting to paint a bigger picture for Zhipu. The essence of TAC is to upgrade large models from a “dialogue tool” to a “task execution system,” enabling enterprises and developers to build complex Agent workflows. This vision is undoubtedly grand, but it is still far from large-scale commercialization. At present, the vast majority of users remain in the simple stage of API calls, and the proportion of customers that truly have TAC capabilities is extremely low. The path from “available” to “deep usage” is longer than people might imagine.

Returning to the annual report itself, what Zhipu AI presents is a real snapshot of an industry still in its early stage. Overall, whether Zhipu can become “Anthropic in China” depends on three key variables: first, whether technological leadership can be sustained; second, whether the API business model can penetrate from “early adopters” into the mainstream market; and third, whether it can maintain independence amid a siege by major players.

The price that capital has assigned already anticipates a bet on “outcomes.” But each of the above variables will be magnified for scrutiny by the market in the future.