If it weren’t for the need to write logically and make multiple revisions, this article would have been sent last week.
If it was sent last week, it seems to have “predicted” the sharp drop on Friday.
On that day, MiniMax Hong Kong stock closed at HKD 427, down 6.48% from the previous day. Since its 52 week high of HKD 1330 after listing, it has fallen by over 60%. On the same day, Zhipu AI closed at HKD 2046, a sharp drop of 12.94%, and briefly fell below the HKD 2000 mark during trading. Just four days ago, it hit a historic high of HKD 2980.
But the analysis we are doing is not about the market changes of one or two days or even one or two weeks. We have no intention of playing the role of a stock analyst who predicts a big drop. Last Friday’s decline was influenced by the US stock market, anxiety as the lifting period approached, and a risk aversion of funds taking profits and leaving. These are all short-term shocks. After falling again on Monday and rebounding on Tuesday, no one can accurately predict the short-term level.
What we are trying to answer is a slower, deeper, and more difficult to avoid question: when the story of the “Chinese version of Anthropic” is repeatedly told until the market becomes auditory tired, what is the basis for Zhipu’s nearly trillion dollar market value?
It cannot be denied that Zhipu is a company with genuine original technology, product reputation, and commercial landing capabilities. As early as 2021, its proposed GLM architecture was a third path different from BERT and GPT; Recently, GLM-5.2 won the first globally available model in the Code Arena blind test and has a good reputation in the developer community; Its MaaS platform ARR has increased 60 times in 12 months, and government and enterprise customers are willing to pay for it.
These are all genuine values. These values are enough to sustain an excellent vertical AI service provider, but it is difficult to close a trillion dollar market value platform narrative. A trillion dollar market value requires paradigm definition power, exponential growth, and an irreplaceable moat, and Zhipu has not yet provided market validated answers in these three dimensions.
More subtly, the technological capabilities of Zhipu are being overshadowed by a public opinion field of “internal and external division”. When external developers call the bare API, what they see is that GLM-5.2 has the same output capability as GPT-5.5 on SWE bench and LiveCodeBench; Internal C-end users only see model output text homogenization and do not understand Zhipu’s coding ability.
This kind of division has put the pricing of Zhipu in a swing in the capital market: either overestimating its monetization ability due to external technical praise, or underestimating its training and engineering optimization value due to questioning its “borrowing” from DeepSeek V3.2. But the real mismatch lies in classification. The market is currently pricing Zhipu according to the “Chinese version of Anthropic”, but its technological foundation and financial structure have repeatedly proven that it is closer to a “Chinese version of Palantir”.
It’s not a matter of whether it’s expensive or not, it’s simply not the species that the market thinks it is.
01
The market generally divides AI companies into four completely different species, each corresponding to an independent valuation logic.
The first layer is the base model definer (L1). They discovered the mathematical foundation behind the model, were able to propose and engineer new architectures, and wrote the development direction of the model with its’ syntax ‘. For example, DeepSeek defined “how to train more inexpensively” using MLA and GRPO, while Anthropic defined “how to align more securely” using Constitutional AI. The valuation tolerance of such companies is extremely high, and PS can be given 50-100 times, because once they establish a paradigm, all upper level applications must follow its syntax, and the ceiling is the bottom layer of the entire industry.
The second layer is the infrastructure controller (L2). They control computing power, cloud and distribution pipelines, and are monopolists of the ‘pipeline’. The PS of Azure, AWS, and Google Cloud is usually 20-30 times higher, and the premium comes from economies of scale and resource barriers.
The third layer is the Super Application Dominator (L3). They master user entry, daily activity, and scenarios, and grasp AI “pragmatics” such as Doubao, Qianwen, and WeChat. Such applications are difficult to value technically, and can only be evaluated based on user numbers, daily activity, and monetization ability.
The fourth layer is vertical service providers and physical integrators (L4). They use AI to solve specific problems, such as government and enterprise IT outsourcing, system integration, and industry solutions. Palantir、 SenseTime belongs to this layer. Their valuation is the most traditional, usually only 10-20 times PS, because growth is limited by manpower, budget, and project cycle, and marginal costs do not approach zero as customers increase.
The valuation gap between L1 and L4 is the difference in the quality of the model’s fundamental capabilities. The 100 times PS of L1 implies a “winner takes all” option value, and once the “grammar” is established, the entire industry will have to pay a “tax”. The 10 times PS of L4 reflects the labor value of “remuneration based on work”, and for every additional customer, a batch of engineers needs to be added.
The criterion for determining whether an artificial intelligence company is in L1 is never parameter size or financing valuation, but whether it has proposed a widely adopted underlying mathematical architecture in the industry, or whether it has training efficiency that cannot be replicated by the open source community.
Under this standard, DeepSeek has MLA (Multi Head Latent Attention), GRPO (Group Relative Policy Optimization), and DSA2 (DeepSeek Sparse Attention 2.0). OpenAI has Test time Compute Scaling. Anthropic has Constitutional AI and a mathematical formal framework for alignment. These are not product features, but the ability to define coordinate systems, which determine how the entire industry understands “efficient training,” “long context,” and “safe and controllable.
MiniMax and Zhipu are not in this coordinate system.
But that doesn’t mean they don’t have technology. To be fair, MiniMax is not a ‘pure product company’. As early as early 2025, its MiniMax-01 first implemented linear attention (Lightning Attention) on an industrial scale, and recently, its M3 model further launched its self-developed sparse attention architecture MSA (MiniMax Sparse Attention), supporting 1 million token ultra long context. The Agent product MiniMax Code, specifically designed for M3 training, supports Computer Use, multi file code library refactoring, and up to 12 hours of autonomous programming tasks.
The GLM architecture of Zhipu also has many innovations. In 2021, Zhipu proposed the GLM paradigm of “autoregressive fill in the blank+bidirectional context”, forming a third path with BERT and GPT. GLM-5 retains its self-developed MoE routing and autoregressive fill in the blank framework, while introducing DeepSeek’s open-source DSA component to improve long context efficiency. This is a normal technology absorption in the open-source ecosystem, not a “skin swapping”. GLM-5.2 scored 1595 points in the Code Arena blind test, ranking second on the overall chart and first among available models worldwide; It is also the world’s number one in terms of aesthetic taste in the Design Arena.
However, original technology does not equal the right to define paradigms, and engineering optimization does not equal the ability to rewrite the underlying model.
GLM is a “tributary” rather than a “mainstream”, and the syntax of the global base model is still autoregressive of the GPT route. MiniMax’s linear attention is a component replacement rather than architecture reconstruction, and the industry mainstream is still standard Transformer+various sparse optimizations. Its linear attention has not been widely adopted as the default solution by the global community.
This is different from DeepSeek, whose MLA/GRPO/DSA2 creates a generation gap because innovation occurs at the lowest level of mathematics and is quickly adopted as a standard by the global community. Therefore, DeepSeek has the power to define the base model, while the power to define the spectrum is incomplete.
Without the power to define core technologies, it means there is no pricing power. MiniMax’s API and Code services still face competitive pricing pressure from open-source models such as DeepSeek; Zhipu’s MaaS service can only oscillate between price wars and price increases. In the first quarter of 2026, the price of Zhipu API has increased by 83%, the programming package price has increased by 30%, and the call volume has increased by 400% instead. This may seem like a “pricing power” that is in short supply, but in fact, during the “lobster fever” period, Chinese models attracted a large number of overseas developers and cost sensitive agent workflows due to their low-cost advantages. Manufacturers collectively raised prices for cost considerations.
OpenAI and Anthropic have real technological premiums, although they are priced high, enterprise customers are still willing to pay. And customers of Zhipu and MiniMax will always have alternative solutions.
The problem with Zhipu is that it is priced using L1 logic by the capital market, but its technical foundation can only support L4 positioning before. Even with GLM-5.2, its technical foundation still supports L4 positioning rather than L1 paradigm definers.
More importantly, without the power to define core technologies, the narrative will continue to drift. Zhipu has switched from the early “another Chinese version of OpenAI” to the “open source pioneer”, and then to the current “Chinese version of Anthropic”, each time because the benchmark object in the previous coordinate system has been falsified. The narrative of MiniMax is shifting from “multimodal globalization” and “C-end companionship” to “M3’s coding/agent capabilities”, but commercial validation of the coding/agent market has not yet begun.
In the capital market, the pricing of technology companies is to some extent a positioning narrative, but incorrect positioning narratives will eventually be corrected.
02
The lack of technological premium, if compensated by exponential growth in business models, can still result in high valuations for companies. But both MiniMax and Zhipu’s business models are trapped in the trap of linear growth.
MiniMax’s previous revenue structure was built on C-end companion products. According to its financial report and public disclosure, the annual revenue for 2025 is approximately $79.038 million (about RMB 570 million), a year-on-year increase of 158.9%. Overseas revenue accounts for over 70%, mainly from paid and advertising applications such as Talkie and Xingye. In February 2026, its annual recurring revenue (ARR) exceeded $150 million, doubling in the past two months.
But there is a hidden concern behind these numbers: users of C-end AI companion products do not have migration costs. Users download due to novelty, but without the locking of their social network, they will eventually lose out because of the freshness. Character.AI、ChatGPT、 Even DeepSeek’s open-source model continues to squeeze the user time of MiniMax.
In order to maintain growth, MiniMax must continuously invest high customer acquisition costs, while retention rates are difficult to improve. This is not the network effect of platform companies, but the traffic consumption of content companies. Every penny of growth requires corresponding marketing investment. Its adjusted net loss for the whole year of 2025 is $251 million, and the year-on-year expansion of losses is a direct result of this “burning money for turnover” model.
So, the coding capability of M3 and the release of MiniMax Code products indicate that it is trying to break free from the label of “pure companionship”. The problem is that the competition in the Coding/Agent market is equally fierce (Cursor, Claude Code, GitHub Copilot have already occupied the ecological position), while MiniMax Code’s ranking is not high, and there is currently a lack of independent third-party verification for developer penetration rate, paid conversion rate, and enterprise level customer numbers. It has the technical foundation of the ‘second curve’, but the commercial validation of the ‘second curve’ has not yet begun.
Previously, Zhipu’s business model was also unable to grow exponentially. In 2025, Zhipu’s revenue will be 724 million yuan, of which the revenue from localized deployment will be 534 million yuan, accounting for 73.7%; The revenue from cloud based MaaS and API is only 190 million yuan, accounting for 26.3%. Localized deployment means project-based, engineer on-site, model fine-tuning, data integration, security adaptation, and continuous operation and maintenance. Every government or financial client requires non-standard customization, and human efficiency cannot be improved. The gross profit margin is continuously eroded by labor costs.
Although its MaaS platform ARR reached 1.7 billion yuan in March 2026, a 60 fold increase in the past 12 months, this is mostly an outbreak starting from a very small base, and a sustainable exponential curve has not yet been seen. This also means a shift in sales model. Previously, Zhipu adjusted its internal organization to establish local project teams by region in order to undertake ToG projects. This structure is closer to traditional system integrators rather than platform based technology companies.
True economies of scale require the same product to serve unlimited users at extremely low marginal costs. OpenAI’s ChatGPT Plus is a standardized subscription, serving hundreds of millions of users worldwide with the same model; DeepSeek’s API is a standardized call, with the same set of weights distributed to developers worldwide through cloud computing. Their growth formula is’ user volume x ARPU x retention rate ‘.
In the past, both Zhipu and MiniMax’s monetization methods were trapped in a linear formula of “manpower x time”. The capital market gives SaaS companies 30-50 times PS because it believes that their revenue can double annually in the next three years; And giving 10-15 times PS to project-based or traffic oriented companies is because they know that their growth is limited by manpower and budget. The previous valuations of Zhipu and MiniMax were clearly based on the assumption of the former, but financial data constantly proves that the latter is the reality.
Of course, now Zhipu has GLM-5.2, MiniMax has M3, and both have products that enter the subscription system. If the API calls of GLM-5.2 can maintain high growth and good retention, then it will switch from L4 to L2/L3 hybrid.
相关文章
