Zhipu AI is selling tokens, while MiniMax is seeking its proper market identity.
On June 22, 2026, Zhipu's intraday price surged to HK$2,980, with its market capitalization once hitting HK$1.33 trillion, making it the first domestic AI company with a market value exceeding HK$1 trillion. On the same day, MiniMax's share price was still hovering around HK$300.
The two companies went public almost on the same day half a year ago, with their revenue scales in the first half of 2026 differing by less than 20% and their loss scales being almost identical — yet the capital market priced them nearly 7 times apart. As of September 10, Zhipu's total market capitalization stood at HK$381.35 billion, while MiniMax's was around HK$101.98 billion, the former being nearly 4 times the latter.
However, the financial data released by Zhipu and MiniMax in the first half of 2026 look almost like twin brothers when placed side by side.
Zhipu recorded a revenue of RMB 954 million, while MiniMax posted a revenue of USD 117 million (RMB 782 million); Zhipu's adjusted net loss was RMB 1.964 billion, and MiniMax's adjusted net loss was USD 293 million (RMB 1.965 billion). Their revenue sizes are close, and so are their loss scales.
When you break down the revenue structure, you will find that the two companies are using completely different approaches to answer the same question: How on earth should the large model business be run? Zhipu puts 86.5% of its revenue on API calls, becoming an almost pure Token economy company; MiniMax, on the other hand, has pushed the proportion of enterprise service revenue from 30% to over 60%, completing the core shift from the C-end to the B-end.
Both companies are also doing the same thing: shifting their revenue focus from project-based and localized deployment to more standardized, more scalable API and platform services.
This is not a story about the superiority or inferiority of technology, but a story about "how the market evaluates a company".
The answer does not lie in the financial figures themselves, but in the capital market's judgment on the "certainty" of the two companies. The market does not buy current profits, or even current revenues, but the expectation of whether a business path is viable and how far it can go. The difference between Zhipu and MiniMax falls exactly on this expectation.
01. At the Time of Listing: The C-end Story Was More Eye-catching
In January 2026, Zhipu landed on the Hong Kong Stock Exchange first, with an issue price of HK$116.2 and an initial market capitalization of about HK$52.8 billion, and its Hong Kong public offering received about 1,159 times over-subscription. One day later, MiniMax priced its issuance at the upper limit of HK$165, surged 109% on the first day of trading to close at HK$345, with its market capitalization once exceeding HK$105 billion. Its public offering was over-subscribed by more than 1,837 times, attracting about 420,000 participants.
In the early stage after listing, MiniMax's market capitalization was nearly twice that of Zhipu.
This is not surprising, as the market was more willing to pay for C-end products at that time. MiniMax owned products such as Talkie and Hailuo AI, with relatively visible user scale and growth curves. According to statistical standards, the revenue contribution of AI-native products (C-end) in 2025 was about RMB 53.075 million, accounting for 67.2% of total revenue and achieving a rapid growth of 143.4%, making it an absolute pillar. At that time, MiniMax had 236 million individual users worldwide. In the industrial context back then, for investors, a C-end story that could quickly show user growth and activity was easier to understand than a B-end-focused API narrative.
But Zhipu followed a different logic at that time: model capabilities, enterprise services, and API calls. It did not have a strong "consumer hit product" aura, and its story did not sound as attractive as "AI companionship" or "AI video".
However, the capital market has a cruel side: visible growth does not necessarily equal sustainable moats. C-end AI products can gain users quickly, but they may also be replaced quickly. The user migration cost is low, and the sense of freshness of products fades fast. The opposite is true for B-end APIs: the development is slow in the early stage, but once embedded in customers' engineering processes and business systems, the replacement cost will rise rapidly.
This difference later became the seed for the valuation divergence between the two companies.
02. "China's Anthropic" VS "Global Multimodal Platform"
After listing, Zhipu showed a very steep upward curve.
On June 8, Zhipu was included in the Hang Seng Tech Index and the Stock Connect eligible list; on June 17, it open-sourced GLM-5.2, and against the backdrop of widespread price cuts across the industry, it raised API prices by no less than 30% against the trend. On June 22, Zhipu's intraday price surged to HK$2,980, with its market capitalization once hitting HK$1.33 trillion, making it the first domestic AI company with a market value exceeding HK$1 trillion.
MiniMax's trend was completely the opposite. After touching its all-time high of HK$1,330 on March 18, it began to weaken continuously. By June, Zhipu's market capitalization was about 7 times that of MiniMax.
The divergence at this stage appears to be a difference in share price strength on the surface, but essentially it is the capital market's re-pricing of the two different development strategies.
Let's start with Zhipu. Before and in 2025, Zhipu's revenue was dominated by localized deployment income, and product implementation and delivery required a series of processes from integration, customization to environment deployment. In layman's terms, it means that Zhipu's AI large model system is directly installed and run on the servers or computing devices owned by customers, instead of calling Zhipu's cloud services through the Internet, which is a typical "government and enterprise project company".
Under the logic evolution of "China's Anthropic", Zhipu has shifted to cloud calls in the Coding stage since 2025, and Zhipu has chosen a focused path: Coding and long-horizon Agent, with its capability narrative unfolding along the ladder of Coding → Agent → Co-work. Code generation, debugging, repair, and engineering task execution have a natural advantage — the results can be verified. Whether the code can run, whether the vulnerabilities are fixed, and whether the tasks are completed, customers can relatively clearly measure how much the model is worth. For B-end customers, this capability can be directly converted into engineer time, project cycle and labor cost.
In the first half of 2026, Zhipu has completed the core shift from a "government and enterprise project company" to a "cloud API company", with its core products shifting from "localized deployment projects for government and enterprises" to the MaaS open platform + GLM Coding Plan subscription. Financial reports show that the proportion of Zhipu's localized revenue has dropped from 84.8% in H1 2025 to 13.5% in H1 2026. At the same time, its half-year API revenue (RMB 825 million, up more than 27 times year on year) has exceeded its total revenue for the whole year of 2025 (RMB 724 million), with the proportion rising from 15.2% to 86.5%. For reference, about 85% of Anthropic's revenue comes from enterprise APIs and developers, and Zhipu's revenue structure has now shifted to the same form. At the 2026 interim results conference call, Xiao Lei, the Secretary of the Board of Directors, even said bluntly: "Compared with revenue growth, the shift in revenue composition is more worthy of attention."
MiniMax, on the other hand, has adopted a two-line battle for multimodal, targeting a "global multimodal platform". One line is the language model, which aims to hold the Coding and Agent market; the other line is the video generation model, trying to carve out a space in the market dominated by closed-source giants. Both lines have imagination, but it also means that resources need to be dispersed, and the company has to tell two stories at the same time.
In the first half of 2026, MiniMax's revenue from open platform and enterprise services surged from USD 9.2 million in the same period of the previous year to USD 73.9 million, up 703.1% year on year, and its proportion rose from 30.3% to 63.4%, becoming the largest revenue source. Its AI-native product revenue was USD 42.64 million, up 100.9% year on year, but the proportion dropped from about 70% to 36.6%. This means that MiniMax has also completed a drastic business model shift within half a year.
MiniMax disclosed at the results meeting that its ARR in August exceeded USD 800 million, of which the To B business accounted for about 80% and the To C business accounted for about 20%. Compared with the same period of last year, when the To B proportion was about 30% and the To C proportion was about 70%, the model calling demand from enterprises and developers is becoming MiniMax's new main revenue driver. The Token consumption in July reached 20 times that of January, and the number of enterprise customers and developers has exceeded 2 million, which is 10 times that at the end of 2025.
From the perspective of revenue growth, this is a good thing for MiniMax, but the shift itself brings new pressure: the growth of B-end revenue is highly dependent on the continuous leading of model capabilities. However, this does not seem easy for MiniMax.
On the language model line, MiniMax released its flagship model M3 in June 2026. According to part of the evaluation data, M3 scored 59.0 on SWE-Bench Pro and 83.5 on BrowseComp web autonomous Agent evaluation. A JPMorgan research report clearly pointed out that since M2.7, MiniMax has not launched a new domestic SOTA model, and its pure model capability is still in the catching-up stage, with the newly released M3 "lacking sufficient competitiveness". Data from third-party benchmark testing firm Artificial Analysis also shows that M3 ranks 53rd among 230 evaluated models worldwide. Greater pressure comes from the strong dependence of B-end revenue on model capabilities. In July, Token consumption was 20 times that of January, ARR in August exceeded USD 800 million, and the B-end contributed more than 80% of revenue. This growth relies on enterprise customers accessing the model into real business scenarios, but the migration cost of enterprise customers is not high. Once the model capability is surpassed, customers may quickly turn to competitors.
Then let's talk about the other line of the two-line battlefield: the video generation line. The competition pattern of the video generation track is far more severe than that of the language model. ByteDance's Seedance has occupied more than 80% of the market share, and Kuaishou's Keling accounts for about 14%. This is a market dominated by closed-source giants. Yan Junjie disclosed at the results meeting that the training resource investment of the text model is about 4 times that of the video model. This means that the video generation line is in a relatively secondary position in resource allocation, but it has to face head-on competition from giants such as ByteDance and Kuaishou. MiniMax's narrative of entering the market with "open source + cost performance" is foreseeably full of challenges and fierce battles.
The capital market is not afraid of companies doing difficult things, but it is afraid of companies doing two difficult things at the same time with unclear relationships between them. Once the resource allocation is not transparent, it will be difficult for investors to judge: which line is the core of the company? Which line deserves valuation? If neither line has achieved absolute advantages, the market tends to discount first. As a result, Zhipu's "focus" is interpreted as certainty, and MiniMax's "multi-line strategy" is interpreted as uncertainty, thus the valuation gap is widened.
03. The Unlocking Test: The Same Exam, Two Different Shareholding Structures
On July 8, Zhipu unlocked 25.6816 million shares, accounting for only 5.76% of the total share capital. On the unlocking day, its share price rose 13.35% to close at HK$1,825. The next day, the company announced that it would place 19.78 million new H shares at HK$1,588 per share, raising about HK$31.4 billion, and the share price continued to rise by 11.34%.
One day later, MiniMax unlocked about 153 million shares, accounting for 48.9% of the total share capital. On the unlocking day, its share price fell 17.98% to close at HK$297.4. The next day, the company announced that it would place 35.6 million new shares and issue HK$6.5 billion zero-coupon convertible bonds, raising a total of about HK$16 billion. The share price fell another 9.68% to close at HK$268.6.
Both experienced unlocking, with one rising and the other falling. This is not a coincidence, but a concentrated exposure of two different shareholding structures in the same time window.
A considerable part of Zhipu's cornerstone investors are long-term funds with state-owned capital background. After the unlocking, these shareholders sent signals that they would continue to hold the shares. Their assessment cycle and exit logic are different from those of financial investors, and they do not need to cash in returns on the unlocking day, nor do they necessarily pursue short-term book returns. For the market, this forms an expectation of "lock-up".
MiniMax's unlocked shares, on the other hand, mostly come from early financial investors. These institutions have low entry costs and high book returns, so their willingness to exit is naturally stronger. Their business model determines that they have a clear exit cycle and profit-taking demands. After unlocking, the selling pressure is not an emotional issue, but a structural one.
Unlocking itself does not create value, nor does it destroy value. It just lays the company's real shareholding structure out in the sun, allowing the market to re-answer a question: Who is holding this company? Why are they holding it? How long will they hold it? Zhipu's unlocking let the market see the presence of long-term capital; MiniMax's unlocking let the market see the profit-taking impulse of financial capital. The same incident gets completely different interpretations, and the share price trend is the result of these interpretations.
04. 4x Valuation Gap: Similar Financial Performance, Divergent Narratives
Zhipu's total market capitalization is HK$381.35 billion, while MiniMax's is about HK$101.98 billion, the former being about 4 times the latter.
This gap has narrowed from the extreme 7 times at its peak in July, but it is still huge. More importantly, the revenue and loss scales of the two companies in the first half of 2026 are highly close, which shows that the difference in the market's pricing of the two companies does not reflect their current financial performance, but the judgment on the certainty of their business paths and the sustainability of their capabilities.
Gross margin is a key indicator to observe the commercialization quality of large models, and it is also the most intuitive reflection of the difference between the two companies.
Zhipu's overall gross margin in the first half of the year was 26.4%, a significant decline from 50.0% in the same period of last year. This change seems negative, but it is actually an inevitable result of the revenue structure shift: the proportion of high-gross-margin localized deployment (gross margin 59.1%) shrinks, while the low-gross-margin API business (gross margin 24.6%) expands rapidly. It is worth noting that the gross margin of the API business has turned positive from negative, indicating that the scale effect of cloud services is emerging.
MiniMax's overall gross margin in the first half of the year was 17.9%, which is lower than Zhipu's but has improved from 12.1% in the same period of last year. The gross margin of its C-end AI-native products is only 4.7%, which is the main factor pulling down the overall gross margin. However, with the increase in the proportion of B-end revenue, the overall gross margin is expected to continue to improve.
The cost structures of the two companies are highly similar: 70%-80% of R&D expenditure is used for computing power. Zhipu's R&D investment in the first half of the year was RMB 2.131 billion, of which the computing power service fee was about RMB 1.1 billion, accounting for 71.8%; the cloud computing service expenditure related to training in MiniMax's R&D expenditure accounted