--- title: "A 990% Loss Rate and a Global No. 6 Ranking: Zhipu AI’s Real Financials Are Startling" author: "Unique Research" sourcePublication: "Unique Research Substack" originalPublishedAt: "2025-12-25T08:18:13+00:00" canonical: "https://ffcap.cn/en/research/src-20251225-01html" source: "https://uniqueresearch.substack.com/p/src-20251225-01html" language: "en" --- # A 990% Loss Rate and a Global No. 6 Ranking: Zhipu AI’s Real Financials Are Startling _Original · Unique Research · 2025-12-25_ _Historical edition: This complete translation preserves the December 25, 2025 author’s analysis of the prospectus, company statements and then-current model rankings. Financial figures, user counts and claims about customers and model performance are reported as stated in that source, not newly audited results or current rankings. Monetary figures are expressed in yuan as in the source. The headline’s “990%” is a rounded source characterization: dividing the displayed 1.899 billion operating loss by 191 million revenue gives approximately 994%, not exactly 990%; the source later describes the signed margin as close to -1000%. Gross margin is not net profit or cash flow. The claim that coding and agents dominate GLM usage is the author’s inference, not a provider-level usage breakdown established by the cited report. LMArena supplies evaluation rankings, not traffic or token-use measurements; the source’s later reference to “usage” there is retained as its wording, not treated as measured utilization. Z.ai and Qingyan figures are not deduplicated across users, and no combined unique-user total is claimed._ Over the past two days, discussion surrounding Zhipu AI’s prospectus has displayed a peculiar sense of disconnection. On one side are alarming interpretations such as “an operating loss rate of 990%” and “losing money to provide APIs,” accompanied by shocking media headlines; on the other are the strong results achieved by GLM-4.6 and GLM-4.7 on LMArena and OpenRouter, which have drawn widespread acclaim from the developer community. If you look at only one side, the company seems like a mystery. So I placed its hundreds of prospectus pages, third-party evaluation data, and 5.27 million monthly active users on the same chart. Many apparent contradictions then began to make sense. The real shape of revenue: it is not what you think Begin with the most basic figures. Zhipu AI’s revenue rose from RMB 57.409 million in 2022 to RMB 124.538 million in 2023, then reached RMB 312.414 million in 2024. In the first half of 2025, it had already generated RMB 190.877 million, meaning that its half-year revenue was already 61% of the previous full year’s total. This is not a company with “no business”; its business is accelerating and becoming more substantial. More important, however, is the composition of that revenue. It contains a fact that overturns conventional assumptions: In the first half of 2025, 84.8% of Zhipu AI’s revenue—approximately RMB 162 million—came from on-premises deployment, while only 15.2%—approximately RMB 29.10 million—came from cloud deployment. This directly contradicts many people’s expectations. At a time when OpenAI has swept the world with APIs and subscriptions, one of China’s leading large-model companies derives most of its revenue not from cloud services, but from ToB projects that deliver “deployed” large-model systems to enterprises. Zhipu AI is not a typical SaaS company; it is more like a “high-end software delivery provider.” Which business line makes money? Look at gross margin, not only losses The next figures are even more interesting. Zhipu AI’s overall gross margin is hardly poor: it was 54.6%, 64.6%, and 56.3% in 2022–2024, respectively, and 50.0% in the first half of 2025. This indicates that “model capabilities + engineering delivery” is not inherently a business that loses money on every sale. The key is the difference between segments: ● Gross margin for on-premises deployment: 59.1% (first half of 2025) ● Gross margin for cloud deployment: -0.4% (first half of 2025) One is positive and the other negative; the dividing line could not be clearer. The customer mix for on-premises deployment services may differ from what many people imagine. Zhipu AI senior vice president Wu Weijie recently disclosed a key figure on a social platform: internet and high-tech customers account for 50%, with no denominator specified in the quoted source; it works with 9 of the top 10 internet companies; and government clients account for less than 20%. Its MaaS platform has 2.9 million enterprises and developers, of whom more than 10% are paying customers, while the annual repurchase rate among paying customers is 70%—a very strong achievement in such an intensely competitive MaaS market. This overturns the stereotype that large-model on-premises deployment relies mainly on government projects. The reality is that major internet companies and telecommunications operators—organizations with extremely demanding requirements for data security, compliance, and deep customization—are the core customers of Zhipu AI’s on-premises business. They are willing to pay high fees for “dedicated deployment + dedicated models + on-site support.” A single project can easily be worth several million or even more than RMB 10 million. To put it bluntly, large projects are currently “supplying the blood,” while the cloud business is still “growing up.” Behind the 990% loss rate: where did the money go? In the first half of 2025, Zhipu AI recorded an operating loss of approximately RMB 1.899 billion against revenue of RMB 191 million, producing a loss rate close to 990%. At first glance, that is indeed frightening. But if you continue into the detailed expense breakdown, you will find that the source of the losses is not mysterious: ● Research and development expenses: - 2022: RMB 84.377 million - 2023: RMB 528.884 million - 2024: RMB 2.195436 billion - First half of 2025: RMB 1.594661 billion ● Sales and marketing expenses: - 2022: RMB 15.139 million - 2023: RMB 101.198 million - 2024: RMB 387.475 million - First half of 2025: RMB 208.57 million As the figures show, in 2024 alone Zhipu AI spent more than RMB 2.1 billion on research and development and nearly RMB 400 million on sales. This is not a case of “the business collapsing and costs spiraling out of control,” but a deliberate decision to invest heavily in model research and development, computing capacity, the sales organization, and ecosystem construction. You can certainly question whether this approach is too aggressive or whether the company expanded too early, but it is not the same story as “business contraction.” The prospectus states the conclusion plainly: cumulative losses came mainly from “significant investments in research and development and sales and marketing.” The scorecard from the coding battlefield: GLM is proving itself through performance After finishing the financial statements and returning to the models themselves, the picture looks completely different. ● GLM-4.6, released in September 2025, has a very clear position: an open-source model focused on real-world development scenarios and agent workflows. It supports a 200K context window, emphasizes coding, long-context reasoning, search, and Agent use cases, and is available under the MIT license. The keywords from third-party evaluations are clear: it improves distinctly on GLM-4.5 across multiple benchmarks, approaches Claude Sonnet 4 in practical coding, and can be integrated directly into Coding Agent tools such as Cline, Roo Code, and Claude Code. ● In December, Zhipu AI released GLM-4.7. The LMArena WebDev leaderboard at the time showed GLM-4.7 ranked No. 6 globally for Web development tasks, making it the highest-scoring open-source model on the leaderboard. Even more important data came from “State of AI: An Empirical 100 Trillion Token Study,” jointly released by OpenRouter and a16z: ● Open-source models now account for close to one-third of total usage ● The weekly token share of Chinese open-source models reached 30% during some periods ● Programming and role-playing are the two most important application scenarios Although the report did not provide the exact token volume of any single company, the prospectus disclosed that “the GLM family has entered the top 10 by token consumption on OpenRouter and ranks among the leading Chinese model providers.” Taken together, this supports a reasonable inference: The place GLM is used most today is not casual conversation, but coding and Agent workflows. This corroborates both the LMArena ranking and the support GLM receives from various open-source tools. ToC reach, ToB collections: a dual-entry strategy According to Unique Research data from November 2025: ● Z.ai has 1.59 million monthly active users on the Web, primarily overseas ● Zhipu Qingyan has 5.27 million monthly active users across the internet, primarily in China Together, these two entry points have moved beyond the scale of a “niche developer toy.” When combined with the revenue structure disclosed in the prospectus, the logic becomes clear: ● ToC products—Z.ai and Zhipu Qingyan—serve reach, education, and conversion functions. Through free tiers, low-priced subscriptions, and developer ecosystems, they naturally direct some users toward APIs, enterprise editions, and on-premises projects. ● In the short term, the real cash flow still comes primarily from enterprise and institutional deployments and project contracts. This also matches actual user behavior. When Claude was inaccessible in China during certain periods, many developers and heavy users naturally looked for alternatives, and GLM met the demand during that window. Zhipu AI even introduced a dedicated migration program for Claude users. Going global is not only about traffic; it also produces real project revenue The prospectus contains an easily overlooked detail: ● In the first half of 2025, revenue from on-premises deployments in Southeast Asia was RMB 17.927 million, accounting for 11.1% of total on-premises deployment revenue. This means Zhipu AI’s global expansion is not merely “launching an English website and watching the traffic.” It has already begun delivering on-premises large-model projects in Malaysia, Singapore, and other markets. Enterprises are willing to pay locally for GLM-family projects, rather than merely making a few remote API calls through OpenRouter and leaving. If this business line develops smoothly, it will directly change the geographic composition of Zhipu AI’s revenue—from “a single Chinese market” to a combination of “China + Southeast Asia + other regions.” Its share is not yet high, but this is an important signal. Customer concentration is falling: risk is becoming more distributed There is another key indicator: ● The share of revenue from the five largest customers fell from 61.5% in 2023 to 40.0% in the first half of 2025. In the author's reading, this indicates that Zhipu AI is not relying on a handful of major customers; its customer base is expanding. From the perspective of commercial health, this is a positive signal—the more diversified the sources of revenue, the greater the ability to withstand risk. So are these financial indicators “abnormal”? If you look only at several isolated percentages, Zhipu AI’s numbers are indeed frightening: an operating loss rate close to -1000%, a negative cloud gross margin, and sales- and R&D-expense ratios that appear extraordinarily high. But when you examine the revenue structure, margin structure, model performance, and international expansion path together, another picture emerges: 1\. On-premises deployment sustains the entire business. This line has a gross margin close to 60%, customer concentration is declining, and revenue is growing rapidly. 2\. Cloud MaaS is still “finding the right posture.” It currently resembles a growth engine more than a profit engine. In its prospectus, Zhipu AI explicitly states that it plans to optimize the profitability of cloud services through measures such as tiered pricing and restricted access to resource-intensive models. 3\. GLM’s performance in coding and Agent scenarios, together with its actual use on OpenRouter and LMArena, proves that the model itself is already being “put to use,” rather than remaining on PPT slides and Benchmarks. 4\. International expansion is no longer merely a concept; real project revenue has already been delivered in Southeast Asia. Therefore, instead of simply saying “the financial indicators are abnormal,” it is more accurate to recognize that this is a model company heavy in assets, research and development, and sales, deliberately pressing the accelerator hard while racing to capture a window of opportunity. Risk and opportunity coexist Of course, the risks are real: ● If, over the next 2–3 years, the growth of on-premises deployment slows, cloud gross margin fails to improve, and overseas projects do not achieve scale, these investments will quickly become a burden. Everyone reading the prospectus should remain alert to that possibility. But if you connect the prospectus with GLM’s model trajectory, the 5.27 million monthly active users, and its LMArena ranking, one point can at least be made: what Zhipu AI is “losing” today is not only money, but also the cost of an entire experiment in commercializing models in China. What it truly wants to “earn” is threefold: large on-premises projects, high-margin cloud tiers, and global developer mindshare. The specific developments worth continuing to monitor are therefore clear: ● The speed of geographic and industry expansion for on-premises deployments ● Whether cloud gross margin can rise from -0.4% back into positive territory ● How many real paying users Z.ai and Zhipu Qingyan can convert ● Whether GLM’s usage curves on OpenRouter and LMArena continue to rise ● Whether the share of Southeast Asian on-premises projects can continue growing from 11% As these lines gradually become clearer, looking back at the prospectus’s enormous loss figures may evoke more than the single word “surprise.” This is Zhipu AI’s real ledger: earning the respect of developers around the world in code repositories on one side, while using high-margin ToB projects to support its capital-intensive technological ambitions on the other. It is not a perfect business story, but it is at least a real experiment unfolding before us. --- Original publication: https://uniqueresearch.substack.com/p/src-20251225-01html On-site reading page: https://ffcap.cn/en/research/src-20251225-01html