Technology determines whether a company can enter the game, business model determines how far it can go, and profit determines its final outcome. This is a lesson Z.AI (HKEX: 02513) taught investors once again at a cost of over HK$1 trillion.
On January 8, 2026, China's prominent large model company Z.AI (Beijing Zhipu Huazhang Technology Co., Ltd.) listed on the Hong Kong Stock Exchange at an issue price of HK$116.2, with a market value of approximately HK$52.8 billion. Less than half a year later, on June 22, its share price surged to HK$2,980 intraday, with market value briefly touching HK$1.33 trillion, making it the first domestic large model company to reach a trillion-Hong Kong-dollar market value. As of the most recent trading day's close, Z.AI's share price stood at HK$501.5, with a total market value of HK$244.525 billion. In five months, it gained approximately HK$1.28 trillion; in four months, it lost over HK$1.1 trillion. What exactly happened in between?
Z.AI did not actually stop moving forward. GLM-5.2 and GLM-5.3 were released in succession, its first interim report after listing showed revenue growth of nearly four times year-on-year, and it completed three rounds of financing in nine months, raising over HK$75 billion cumulatively. The more money it raised, the lower its share price fell. This is precisely what makes this case worthy of serious study. A company whose model capabilities have not regressed and whose revenue is still growing rapidly saw its market value shrink by 80%. As judged in a Shanghai Securities News report, the model capability did not suddenly disappear — what disappeared was the valuation premium the market had once been willing to grant it.
Scarcity premium
Brokerages gave the June rally a name: the scarcity premium. At that time, Z.AI happened to possess nearly all the elements that most easily earn an AI company a high valuation — foundation models, coding, MaaS platform, developer ecosystem, and the extremely scarce identity in China's capital markets of being a "listed large model company." OpenAI was not listed, nor was Anthropic, making Z.AI a rare entry point. Domestically, only Z.AI and MiniMax were listed companies primarily engaged in foundation large models, and Z.AI was the larger of the two, so hard-tech funds searching for an exit poured in.
More critically, the shareholding structure: Z.AI's free float in the early days after listing was less than 3% of total share capital. According to Jiemian News statistics, before the July lock-up expiry, only 11.74 million shares were freely tradable on the market, meaning a small amount of capital could push the share price up by multiples. The supply gate opened in July. On July 8, approximately 25.68 million shares held by 11 cornerstone investors were released from lock-up, the first test after Z.AI's listing. Public information shows that on the day of the lock-up expiry, the share price not only did not fall but closed up 13.35%, and rose another 11.34% the next day, appearing to be a smoothly passed stress test. On July 9, Z.AI announced a placement of new shares at HK$1,588 per share, completed a few days later with net proceeds of approximately HK$31.375 billion, with six institutions taking up the batch. The chips thus shifted from locked status to tradable status. The lock-up release freed 25.68 million shares, and the placement added nearly 20 million more, multiplying tradable chips several times over. In early September, news emerged that Moonshot AI had secretly filed, and the number of selectable targets kept growing. The scarcity premium was the first layer to collapse in this downturn, and the most thoroughly collapsed layer.
It reveals an often-forgotten truth: scarcity is a valuation, not a moat. The definition of a moat is that even when competitors flood in, it remains effective; the definition of scarcity is precisely that once competitors flood in, it ceases to exist. Pricing scarcity as if it were a moat is essentially paying for time, and time never stands on the side of scarcity.
The shelf life of technological leadership
On the evening of July 16, Moonshot AI released Kimi K3, with total parameters of 2.8 trillion, the largest open-source model by parameter scale at the time, topping the Arena code leaderboard within hours of launch. According to Moonshot AI's announcement, due to user requests growing too rapidly, it had to suspend new C-end user subscriptions 48 hours after launch. Bloomberg compared this release to the earlier "DeepSeek moment." The next day was called "Black Friday" for the Hong Kong AI sector by media. Z.AI closed down 28.49% that day at HK$1,107, with market value evaporating over HK$200 billion in a single day; the next trading day it fell another 19.56%, with the share price dropping below HK$1,000. The placement institutions that had just taken up shares at HK$1,588 were sitting on a paper loss of 30% in less than ten days. The share price falling below the placement price hurt market sentiment far more than the lock-up expiry itself.
Afterward, JPMorgan reassessed the valuation framework for domestic large model companies, cutting Z.AI's target price from HK$2,400 to HK$1,600 and MiniMax's from HK$240 to HK$160. According to another report, JPMorgan lowered Z.AI's 2030 expected P/E ratio from 30 times to 20 times, on the grounds that the leading position in frontier models is rotating rapidly, and the certainty of a single vendor maintaining long-term technological leadership is lower than previously expected. The meaning behind this statement is that Z.AI's own model did not regress, but a competitor releasing a stronger model was enough to prompt investment banks to reassess its long-term value.
The market had been waiting for Z.AI to deliver a heavyweight response. On August 14, GLM-5.3 was released, with a parameter scale of approximately 750 billion, sharing the same base as GLM-5.2, and according to Z.AI's official statement, all capability improvements came from post-training. Goldman Sachs called it another major leap for Chinese AI models and raised its revenue forecast but maintained a "neutral" rating; Daiwa Securities was more direct, considering it merely a routine iteration rather than the larger-parameter flagship the market had expected. On the day of the release, Z.AI's share price closed down about 3.6%.
Here lies a counterintuitive valuation fact of the AI era: technological leadership is an asset that automatically depreciates over time. A bestselling car model can contribute two to three years of sales for an automaker, but the leading window for a state-of-the-art model may be only months or even weeks. Large models are typical relative-performance assets; GLM ranking in the global top three today does not mean it will still be there three months later. In financial terms, this technological premium has obvious Theta decay characteristics — even if the underlying price remains unchanged, time value erodes as the next release approaches. What is more complex is that large model companies' valuations are simultaneously constrained by two product cycles: one is their own model iteration rhythm, and the other is the iteration rhythm of all competitors. The latter is completely outside their control.
This scene played out in the new energy vehicle industry. A few years ago, NIO, XPeng, and Li Auto had market values differing by multiples; today the gap among the three has narrowed significantly. Hit products rotate, technology catches up, and once seemingly insurmountable gaps often narrow again after a few rounds of product iteration. The market is learning not to treat the leadership of a particular product generation directly as a valuation advantage that can be enjoyed long-term.
Unit economics model
On August 31, Z.AI released its first interim report after listing. Looking at growth alone, the report was not unattractive. First-half revenue was RMB 954 million, up about 400% year-on-year, exceeding the full-year 2025 revenue of RMB 724 million in just half a year. The change in revenue structure was even more dramatic: open platform and API revenue reached RMB 825 million, with its share rising from 26.3% at the end of last year to 86.5%, a year-on-year increase of over 27 times; the localized deployment business, which once contributed 70% of revenue, declined 20.5% in the first half. A company that started with government and enterprise localized deployment switched its focus to call-based cloud billing within half a year. The average API selling price rose about 101% compared to the beginning of the year, Token call volume grew over 40 times, and the segment's gross margin turned from negative 0.4% in the same period last year to 24.6%. From selling projects to selling consumption, this was a qualitative shift, and the direction was correct.
The cost was equally visible in the report. First-half consolidated gross margin fell from 50.0% in the same period last year to 26.4%, nearly halved; adjusted net loss was RMB 1.964 billion, versus RMB 1.752 billion in the same period last year, with losses still widening; net loss for the period was RMB 2.072 billion, and R&D investment was RMB 2.131 billion, 2.2 times current revenue. Roughly converted, every RMB 1 of revenue still corresponded to about RMB 2.1 of adjusted loss.
The root of the problem lies in the fundamental difference between this business and the software business. Traditional software selling an extra copy adds almost no cost, with marginal cost approaching zero; but every additional API call a model handles genuinely consumes compute. The faster revenue grows, the bigger the data center shortfall becomes. Z.AI's experience is evidence: according to management on a September 16 conference call, after the February GLM-5 release, model call demand surged 10 times, and almost in the same week as the release, the company's compute reserves were exhausted, forcing its flagship Coding Plan product to be suspended from sale. Afterward, while purchasing compute at high prices and relying on the All-in-Infra strategy and the acquisition of Zhongke Jiahe to improve compute utilization efficiency by two to three times, revenue grew but never truly exploded. It was not until the July financing arrived that compute could be expanded on a large scale, Coding Plan went back on sale, and sales grew more than 15 times.
For large model companies, the ceiling on revenue is often the capacity of the data center. Just as the business model switch was completed and before there was time to adapt, the price war arrived. On September 10, DeepSeek released V4.1 Flash and cut API prices again, and that day Z.AI and MiniMax closed down 10.34% and 8.98% respectively. Subsequently, leading domestic and international players followed suit: on September 21, Xiaomi open-sourced MiMo-V2.6, pushing single-task inference cost down to US$0.13; on September 22, OpenAI released GPT-6 Sol and GPT-6 Luna, with API pricing halved compared to the previous generation, and on the same day Anthropic released Claude Opus 5.5, with costs under typical workloads also about 40% lower than the previous generation. When the world's top models collectively cut prices, the label of cost-effectiveness shifted from a differentiated advantage to an industry entry qualification. Z.AI was forced to raise prices for its Coding package against the trend, preserving revenue in the short term while exposing the fact that its pricing power in the general API field is weak.
The speed at which ammunition was consumed also exceeded outside expectations. According to Z.AI's July placement announcement, of the HK$4.896 billion net proceeds raised in the January IPO, HK$4.588 billion had been used by June 30, leaving only HK$308 million. Based on the 2025 loss scale, that amount would only sustain about one month. This also explains why the July placement came so urgently. Thus came the second discounted financing within two months. On September 13, Z.AI announced a financing of approximately US$5 billion, of which US$2 billion was a new share placement and US$3 billion was zero-coupon convertible bonds, which the market called "small equity, large debt." According to Securities Times, the placement price was HK$714, a discount of about 9.96% to the closing price before the announcement, less than half of the HK$1,588 round in July. Based on a share price of around HK$610 in late September, the six institutions that entered in July were sitting on paper losses of around 60%. Two batches of institutions entered back-to-back, yet their costs differed by more than double.
Placing Z.AI in the context of the industry, this money is not much. ByteDance's 2026 AI capital expenditure has been raised to over RMB 200 billion; Alibaba's "RMB 380 billion over three years" plan had already invested about RMB 190 billion cumulatively as of the end of June, just past the halfway mark, and management said actual investment would far exceed that number. Z.AI, having raised over HK$75 billion cumulatively across three financings, still looks thinly capitalized next to these giants. ByteDance's annual investment alone is close to three times Z.AI's total financing since listing. The capital market's focus on large model companies has shifted from whether revenue can double to how much revenue each yuan of compute can generate, whether gross margin can continue to improve, and how much cash rapid growth can ultimately leave behind. This is an irreversible shift.
Strategic positioning and trust
At the very bottom lies the question the business model itself has yet to answer clearly: is Z.AI a cloud-native company that can scale and charge by usage, or an IT services provider dressed in AI clothing? This question directly determines the unit economics model and therefore the valuation center. Although private deployment is heavy, gross margin once approached 60%; APIs appear scalable but are squeezed from both sides by compute costs and peer price cuts. Z.AI has tried to extend into Agents and enterprise workflows, but first-half enterprise-level intelligent agent revenue was only RMB 55.56 million, a very small share, and it is highly delivery-oriented, not yet showing recurring revenue characteristics.
In September, an incident added new uncertainty to this question. On September 18, developers discovered that Z.AI's AI coding tool ZCode, while users were logged in, silently packaged and encrypted the entire workspace and uploaded it to the cloud, including complete Git history and operation records, with the encryption private key stored only in the cloud and undecryptable locally by users. Z.AI apologized the same day, attributing the cause to the "codebase indexing" feature being enabled by default, and promised that data would be destroyed immediately. Later, Taiyuan Chengming Technology sent a letter seeking accountability, pointing out that upload records still existed in the early hours of the apology day; according to China Newsweek, nearly 400 developers formed a rights protection group. On September 21, Z.AI announced that ZCode would be officially open-sourced, invited the China Academy of Information and Communications Technology and NSFOCUS to conduct security audits, and launched a "no data content retention" feature, promising no retention and that relevant code had never been used for model training. The share price still responded with a significant decline that day. For a company whose API revenue already accounts for 86.5%, the trust of developers and enterprise customers is not a public relations issue but the foundation of revenue. Data boundary disputes will directly affect renewal decisions.
A lesson for investors
Z.AI's journey from a trillion to just over two hundred billion is the first valuation correction in the AI era fully completed by the public market. The lessons it offers are equally important for entrepreneurs and investors.
First, scarcity is the most fragile moat. Scarcity is a function of time and will be automatically erased by supply. The premium a company gains from scarcity begins to erode the moment competitors appear and chips increase. To judge whether an advantage is a moat, one need only ask: if three equally powerful rivals appeared tomorrow, would it still exist?
Second, distinguish technological leadership from the ability to lead continuously. Topping a model leaderboard once is a result; continuously producing leading models is a capability. The latter is composed of organization, engineering systems, data flywheels, compute autonomy, and commercialization feedback. The market is learning to pay long-term premiums only for the latter. JPMorgan's reason for cutting Z.AI's long-term valuation expectation is worth remembering for all technology companies: the shelf life of leadership has shortened, so the premium for leaders must naturally be discounted.
Third, revenue growth does not equal value growth. When gross margin is negative or very thin, growth itself may be accelerating capital consumption. The question that truly needs answering has shifted from whether revenue can double to whether gross margin can hold, how high the renewal rate is, how large customer migration costs are, and how much revenue each yuan of compute can generate. Parameters and benchmark scores cannot be directly converted into profit; Tokens must be converted into profit, and narrative must be converted into renewal rates and gross margins.
Fourth, the pricing split between the primary and secondary markets is itself a signal. According to reports, the pre-IPO financing of the still-unlisted Moonshot AI is being pursued at a valuation of about US$50 billion, equivalent to about 50 times ARR; DeepSeek's new financing round targets a valuation of RMB 500 billion, which at current annualized revenue of about US$1 billion is about 75 times ARR. Meanwhile, publicly traded Z.AI is having its expectations corrected quarter by quarter through financial reports. When two pricing systems appear in an industry, it usually means the pricing logic is at a turning point.
Fifth, the size of the decline itself does not explain much; what matters is whether the business can continue to grow after the fall. During the years of the internet bubble burst, Amazon's share price fell more than 90% from its high, and Cisco also fell nearly 90%. The two companies later took completely different paths: Amazon continuously opened new curves through e-commerce and the cloud computing business that later grew, and its market value rose thousands of times from the low; Cisco's network equipment still sold well, but its share price never returned to that high over the following two decades.
Where to go from here?
Comparing Z.AI in June and September reveals an intriguing contrast. When it surged to a trillion in June, the money raised in the IPO was almost spent, and the flagship Coding Plan could not sell well due to insufficient compute; by the end of September, two rounds of financing totaling over HK$70 billion had arrived, compute began to expand on a large scale, Coding Plan's sales grew more than 15 times after resuming sales, the year-end ARR guidance was raised from US$2.4 billion to US$3 billion, all-business ARR had reached US$1.8 billion, revenue sharing with domestic and overseas cloud vendors began to be recognized from October, and Co-work industry orders exceeded RMB 1 billion.
The books were much thicker than three months earlier, yet the market value was only about a quarter of its peak. Institutional attitudes also showed this split. Jefferies raised its 2026-2029 revenue forecasts by 37% to 119% while compressing the cloud business valuation multiple from 50 times ARR to 30 times; CLSA upgraded its rating to "outperform" in mid-July; according to Zhitong Finance statistics, as of September 24, the Stock Connect shareholding ratio had risen to 14.37%. These seemingly contradictory actions together indicate that the market has not rejected this business but is simply unwilling to pay so much in advance for its future.
There are still several hurdles ahead. According to Economic Observer, on January 8, 2027, original old shares accounting for more than 30% of total share capital will face lock-up expiry; after the ZCode incident, developer trust needs time to repair. Z.AI's A-share issuance plan was approved by the shareholders' meeting in June, proposing to raise no more than RMB 15 billion on the STAR Market, and if it proceeds smoothly, it will gain an additional pricing market. Z.AI still possesses three things: model capabilities in the first tier of open-source leaderboards, compute autonomy based on domestic clusters, and an API-led subscription-based commercialization path. It has the ability to surge to a trillion in five months, and it may not lack the chance to return there. But for all observers, the value of these four months goes far beyond one company's share price curve. It confirmed, in a rather intense way, an underlying rule of the AI industry: over the past three years, the most important question in the large model industry was whose model was stronger; in the next few years, this question will gradually become who can sell intelligence at a higher price while pushing compute costs lower. Technology determines whether a company can enter the game, business model determines how far it can go, and profit determines its final outcome. This is a lesson Z.AI taught investors once again at a cost of over HK$1 trillion.
Comments