OpenAI is not on the market, nor is Anthropic, so Wisdom has become a scarce entry point. The valuation of global AI assets has been rising. In the Hong Kong stock market, there are only two listed companies with large-scale base models as their core business, Zhipu and MiniMax. Zhipu is still the largest company, and hard technology funds looking for outlets are pouring in. At that time, Zhipu’s free-floating shares accounted for less than 3% of the total share capital. The chips were small and there were many people who wanted to buy. It was not surprising that the stock price was magnified several times.

Those who placed their bets first made a lot of money. According to Huxiu, Zhongke Chuangxing exclusively invested 40 million yuan in an angel round when Zhipu was founded in 2019. At that time, Zhipu’s post-investment valuation was only about 375 million yuan. Zhongke Chuangxing has transferred some old shares in 2023, but based on the original investment cost and rough calculation using the stock price in late August, the book return on this investment is still more than 100 times. The shares held by industrial capital companies such as Meituan and Ant are also in the order of tens of billions of Hong Kong dollars, and the shareholding value of relevant entities of Tsinghua University is also worth 16 to 7 billion Hong Kong dollars.

On July 8, Zhipu ushered in its first round of lifting of restrictions after listing. About 25.68 million shares in the hands of 11 cornerstone investors were lifted from the restrictions. This was its first stock price test since its listing. According to Xinhua Finance, not only did the stock price not fall, it closed up 13.35% on the day the ban was lifted, and then rose 11.34% the next day.

The wisdom spectrum at that time seemed to have passed the test. On July 9, it took advantage of the trend and announced the placement of new shares at HK$1,588 per share. It was completed a few days later, with a net raise of approximately HK$31.375 billion. Six institutions took over the shipment.

The chips changed from locked to tradable, and the circulation plate suddenly became larger. According to statistics from Jiemian News, there were only 11.74 million free-floating shares of Zhipu in the market before the ban was lifted. This time, about 25.68 million shares were released at once. Coupled with the nearly 20 million additional shares from the placement, the number of trading chips on the market has doubled.

Later, on the night of July 16, Dark Side of the Moon released Kimi K3, with a total parameter of 2.8 trillion. It is known as the open source model with the largest parameter scale in the world (the complete weight was released on July 27). It topped the Arena code list within a few hours of being online. User requests increased so much that, according to the Dark Side of the Moon announcement, new user subscriptions on the C-side had to be suspended 48 hours after it went online. Bloomberg compared this release to last year’s “DeepSeek moment,” while JPMorgan Chase said there were “DeepSeek 2.0-style concerns” in the market.

The next day is the “Black Friday” of the Hong Kong stock AI sector. Zhipu closed down 28.49% that day at HK$1,107, evaporating more than HK$200 billion in a single day; it fell another 19.56% on the next trading day, and its stock price fell below HK$1,000. The placement agency that just took over the order at HK$1,588 lost 30% in less than ten days. The fact that the stock price fell below the placement price was more damaging than the lifting of the ban itself.

On July 21, Zhipu announced the launch of a 1GW domestic computing power center and the completion of the acquisition of Zhongke Jiahe. The stock price once rebounded by more than 40% during the session. This kind of jumping up and down is not so much a change in fundamentals as it is a catharsis of emotions.

Tang Jie from Wisdom and Yang Zhilin from The Dark Side of the Moon had a teacher-student relationship, and the media simply called that week a “master-disciple showdown.” In fact, it has only been a month since K3 released open source GLM-5.2. Behind K3, there are also new models from Alibaba, DeepSeek and MiniMax.

After the release of K3, JPMorgan Chase re-evaluated the valuation framework of domestic large model companies. JPMorgan Chase believes that the iteration of cutting-edge models is getting faster and faster, and the market’s previous valuation premium for leading model companies needs to be re-examined. It then lowered the target price of Zhipu from HK$2,400 to HK$1,600 and MiniMax from HK$240 to HK$160. The shelf life of the leader has become shorter, and the premium given to the leader will naturally be discounted. In the view of some analysts, K3 is just the last straw. The increase in the supply of chips, the squeeze-out of valuation bubbles and the elimination of scarcity have already happened at the same time.

The market was originally waiting for Zhipu to respond with a big move. On August 14, GLM-5.3 was released with a parameter scale of about 750 billion. It shares the same base as GLM-5.2. According to the official statement of Zhipu, all the ability improvements come from post-training. This is not an iteration without highlights. Goldman Sachs called it another major leap in China’s AI model. It also raised Zhipu’s revenue forecast, but the rating is still “neutral”;

Daiwa Securities put it more bluntly, believing that this is just a regular iteration, not the flagship with larger parameters that the market expects. According to IT House, Qianwen Office, a subsidiary of Alibaba, also announced the launch of GLM-5.3 a few days later. Short-selling funds also became active at this time. On the day GLM-5.3 was released, Zhipu’s stock price closed down about 3.6%; according to Observer.com, on the eve of the crash on August 18, the short-selling indicators of Zhipu and MiniMax had reached new stage highs.

My income has nearly quadrupled, but I still don’t have enough money to spend.

On the evening of August 31, Zhipu released its first interim report after listing. On the same night, the quarterly adjustment of the MSCI China Index took effect, and Wisdom Spectrum was officially included. At the close of trading that day, the stock price was HK$1,195, and the total market capitalization was HK$556.4 billion, a decrease of nearly 60% from the high point. It only took more than two months to go from one trillion to more than 500 billion. Looking at growth alone, this interim report is actually not ugly, but the market’s attention is no longer on growth.

The interim report showed that revenue in the first half of the year was 954 million yuan, a year-on-year increase of nearly 400%, exceeding the full year of 2025 in just half a year. The revenue structure has also changed significantly. Open platform and API revenue reached 825 million, accounting for 86.5% of total revenue from 26.3% at the end of last year. However, localization deployment, which once supported 70% of revenue, fell by 20.5% in the first half of the year.

A company that originated as a government-enterprise project shifted its focus to pay-per-call cloud business within half a year. The average selling price of APIs has increased by approximately 101% compared to the beginning of the year, and the number of token calls has increased by more than 40 times. The gross profit margin of this business has increased from -0.4% in the same period last year to 24.6% now.

The business model has changed from the original “selling projects” to the current “selling consumption”. This is a qualitative change.

But the cost is also high. The adjusted net loss in the first half of 2026 was 1.964 billion. According to China Business News, it was 1.752 billion in the same period last year. The loss is still expanding; the comprehensive gross profit margin dropped from 50.0% in the same period last year to 26.4%, almost halved. Although the gross profit margin of the API business has improved from -0.4% to 24.6%, this level is far from healthy in the large model industry.

What’s more urgent is that Zhipu, which has just changed its business model, has been dragged into the price battlefield before it has entered the adaptation period.

On September 10, DeepSeek released V4.1 Flash and lowered the API price again. On that day, Zhipu and MiniMax closed down 10.34% and 8.98% respectively in Hong Kong stocks. It didn’t take long for leading manufacturers at home and abroad to follow suit. On September 21, Xiaomi open sourced MiMo-V2.6, and the cost of single-task inference was reduced to US$0.13; on September 22, OpenAI released GPT-6 Sol and GPT-6 Luna, and the API quotation was directly cut in half compared with the previous generation. On the same day, Anthropic released Claude Opus 5.5, and the cost under typical load was also about 40% lower than the previous generation.

When the world’s top models collectively cut their prices, the “high cost performance” label that Zhipu relied on to break through overnight changed from a differentiated advantage to an industry standard.

In addition, the ammunition inside is not sufficient. According to Zhipu’s placement announcement in July, of the 4.896 billion Hong Kong dollars in net proceeds received from the January IPO, 4.588 billion had been used by June 30, leaving only 308 million Hong Kong dollars unused. This amount of money will not last long. Based on the scale of losses in 2025, it will only be enough for one month. This also explains why the placement in July came so quickly.

Where did the money go? Zhipu gave the answer during a conference call on September 16, mainly about computing power.

According to China Fund News, management introduced at the meeting that after the release of GLM-5 in February, the demand for model calls increased by 10 times. Almost the same week as the release, the company’s computing power reserves were exhausted, and it had to stop selling its main product Coding Plan. Since then, Zhipu has purchased computing power at a high price, while relying on the All-in-Infra strategy and the acquisition of Zhongke Jiahe to increase the efficiency of computing power utilization by two or three times. As a result, its revenue has increased, but it has never really exploded.

It was not until the money arrived in July that the computing power was able to be expanded on a large scale. Coding Plan was re-opened and sales increased by more than 15 times. Tang Jie mentioned that when the team developed GLM-5.3-Flash, they did a lot of “stingy things” in order to save computing power. For large model companies, the ceiling of revenue is often the capacity of the computer room.

This is also the biggest difference between large models and traditional software. Selling more copies of the software hardly increases the cost. Every additional call received by the model consumes a real amount of computing power. The faster the income rises, the larger the gap in the computer room becomes. As mentioned earlier, Dark Side of the Moon suspended new C-end user subscriptions after the launch of K3. The reason behind this is also the same. This bottleneck is not unique to Zhipu.

On the evening of September 13, Zhipu announced another financing of approximately US$5 billion, of which US$2 billion came from the placement of new shares and US$3 billion was zero-interest convertible bonds, which the market calls “small stocks, big debts.” According to the Securities Times, the placement price was HK$714, a discount of approximately 9.96% from the closing price before the announcement, and less than half of the HK$1,588 round in July. Based on the stock price of around HK$610 around September 25, the six institutions that entered the market in July had lost about 60% of their books, and the group that had just entered in September had also lost more than 10%. The two groups of institutions entered the market at the same time, but the cost difference was more than double.

The conversion price of the convertible bonds is set at HK$892.5. According to calculations, if all are converted into shares, the share capital will be diluted by about 5%. According to Kuai Technology’s disclosure, the management made a calculation at the conference call: an investment of 30 billion yuan can build approximately 100,000P computing power, 40% will be reserved for training and research and development, and 60% will be used for business reasoning; according to the inference price-performance ratio of GLM-5.3, the gross profit of inference can theoretically reach 80%. If the inference computing power is used for one year, the revenue can reach 40 billion yuan.

This calculation is very beautiful, but there are many prerequisites. Any variable in computing power utilization, business discounts and fluctuations in computing power prices may discount the results.

At the same conference call, Zhipu raised its year-end ARR guidance from US$2.4 billion to US$3 billion. The full-service ARR has reached US$1.8 billion, and revenue sharing with domestic and overseas cloud vendors will be confirmed starting in October. According to Rui Finance, Co-work’s industry orders have exceeded 1 billion yuan, but this part needs to be confirmed in one to two years. The market’s reaction was quite confusing. It fell for two days in a row on September 14 and 15. The market value once fell to HK$316.6 billion. On the day of the call, it rose back to 5.88%.

Looking at the industry, this amount of money is actually not that much. Byte’s AI capital expenditure in 2026 has been raised to more than 200 billion yuan; Ali’s “three-year 380 billion” plan has accumulated a total investment of about 190 billion yuan as of the end of the June quarter, which is just halfway through. Management also stated that the actual investment will far exceed this amount.

Zhipu, which has raised a total of more than HK$75 billion in three fundraisings, stands next to these big companies, but its financial resources are still thin. Byte’s investment alone in one year is nearly three times the total amount raised since Zhipu’s listing.

Let the bullet fly a little longer

The one trillion that Zhipu lost was almost like the things that had been put on top of it layer by layer in the past five months, and then taken away layer by layer.

The highest pad is “scarcity”. It and MiniMax are the only companies in the Hong Kong stock market that are mainly engaged in large-scale pedestal models. The free-floating shares account for less than 3% of the total share capital. A small amount of capital can increase the flexibility of the stock price several times. But the lifting of the ban and two rounds of allotment suddenly increased the chips. In early September, news came out that Dark Side of the Moon secretly delivered watches. There were more and more available targets, and the scarcity premium was the first to collapse. Following the loosening of expectations for technological leadership, K3 and V4.1 Flash appeared one after another. The position of China’s strongest model will change every one or two months, and the market no longer dares to price smart spectrum based on long-term leadership.

What lies beneath is the business model itself. Revenue has indeed increased exponentially, but every 1 yuan of revenue still corresponds to an adjusted loss of approximately 2.1 yuan. The comprehensive gross profit margin has dropped from 50.0% in the same period last year to 26.4%. The world’s leading models have collectively cut prices at the same time. The more they sell, the more they lose, and there is no end in sight in the short term.

Computing power is eating up money at a speed far beyond what the outside world imagines. Zhipu could only raise funds at a discount twice in two months. The placement price dropped from 1,588 Hong Kong dollars to 714 Hong Kong dollars. The first institutions to enter the market suffered a floating loss of 60%. Subsequent funds will naturally push the price down.

The external anchors are also loosening at the same time. Altman told Fortune magazine that OpenAI will not be listed within the year; according to Reuters, Anthropic’s IPO timetable is also being pushed back, and it will not be listed until the end of October or early November at the earliest. Several forces overlapped, and the wisdom spectrum fell back from 1.33 trillion Hong Kong dollars to about 300 billion Hong Kong dollars. This is actually not surprising.

But just looking at these, it is easy to draw an overly pessimistic conclusion. Market capitalization has always been the result of the superposition of emotions and fundamentals. For the 1.33 trillion yuan in June, emotions accounted for the majority; for the current 300 billion yuan, there may be a lot of panic.

If you compare the intelligence scores of June and September together, you will see an interesting contrast. When it reached one trillion in June, the money raised from the IPO was almost spent, and the main product Coding Plan could not be sold because of insufficient computing power. By the end of September, two rounds of financing totaling more than 70 billion Hong Kong dollars had been received, and the computing power began to expand on a large scale. Coding Plan’s sales have increased more than 15 times since its resumption, and its end-of-year ARR guidance has been raised from US$2.4 billion to US$3 billion. Revenue sharing with domestic and overseas cloud vendors has been confirmed since October. Co-work still has more than 1 billion yuan in orders waiting to be completed. The ledger is much thicker than three months ago, but the market value is only about a quarter of its peak.

In terms of the attitude of institutions, Jefferies raised its revenue forecast from 2026 to 2029 by 37% to 119%, while lowering the valuation multiple of its cloud business from 50 times ARR to 30 times; Lyon raised its rating to “outperform” in mid-July; according to statistics from Zhitong Finance, Hong Kong Stock Connect funds have been adding positions in the past five trading days, and as of September 24, their shareholding ratio has risen to 14.37%.

These seemingly contradictory actions taken together show that the market has not denied Zhipu’s business, but is just unwilling to pay so much money in advance for its future. UBS made a similar judgment in late July. The decline in the market value of large model stocks does not mean that long-term fundamentals have shifted.

In the first half of the year, Zhipu approximately doubled the average API price while simultaneously increasing the number of calls. At least in high-value scenarios such as programming, customers recognize capabilities. Morgan Stanley judged that “flagship Pro plus lightweight Flash” will become the industry’s standard strategy. Each company is constrained by gross profit margin and is unlikely to fall into vicious competition without a bottom line. The low-end market is bustling, and it is the flagship end that really determines gross profit.

In the years since the Internet bubble burst, Amazon’s stock price fell by more than 90% from its high point, and Cisco’s stock price also fell by nearly 90%. The two companies later had completely different endings. Amazon relied on e-commerce and later cloud computing to continuously grow new businesses. From the bottom, its market value has increased thousands of times. Cisco’s network equipment is still easy to sell, but the stock price has not returned to that high point in the next two decades.

How much it falls actually doesn’t mean much. What matters is whether the business can continue to grow after the decline and whether the profits can be kept in one’s pocket.

Zhipu is now somewhat standing at the fork of these two roads. According to the calculation made by the management at the conference call, 100,000P computing power will be built with 30 billion yuan. In theory, the gross profit margin can reach up to 80%. The revenue after the inference computing power is fully used can reach 40 billion yuan a year. Even if it is only half realized, today’s 300 billion Hong Kong dollars is not outrageous; but if the computing power utilization rate does not increase and the price continues to fall, this investment may become a heavy depreciation on the report.

In the next few months, Zhipu still has several hurdles to overcome. According to the Economic Observer Network, on January 8, 2027, the original old shares, which account for more than 30% of the total share capital, will be lifted; in mid-September, ZCode was exposed to uploading user code warehouses in the background. After this crisis, developers’ trust will take time to repair. A-shares may be a variable. Zhipu’s A-share issuance plan was approved by the shareholders’ meeting on June 22 and plans to raise 15 billion yuan on the Science and Technology Innovation Board. If it goes well, it will create an additional pricing market, and A-shares have always been willing to give higher prices to scarce assets in hard technology.

Overall, Zhipu is not passively beaten in a battle that is bound to fail. It has model capabilities (first echelon on the open source list), has computing power autonomy (domestic clusters), and has a path to commercialization (API-led subscription system).

Related Reading

• Tax evasion? Luo Yonghao was reported under his real name! Open fire on all sides to be surrounded on all sides?2026-09-30
• Interface quick review2026-09-30
• Original: Why does Wufangzhai sell houses intensively?2026-09-29
• The west “breaks zero”, and the competition for the word “rice” in inland provincial capitals begins2026-09-29
• Original Luo Yonghao makes enemies everywhere, Zheng Gang resorts to “tax evasion” killer weapon2026-09-29