Within three months, Zhipu's market value has evaporated by 1 trillion.

When future generations record the development history of large models, June 22 will likely be an unavoidable date.

That day was the first trading day after the Dragon Boat Festival holiday. Less than an hour after the Hong Kong stock market opened, Zhipu AI's share price saw a sharp surge, peaking at HK$2,980 during the intraday session. Calculated based on the intraday high, Zhipu AI's market capitalization once rose to HK$1.33 trillion. It had been less than half a year since it rang the listing bell on the Hong Kong Stock Exchange. When it went public on January 8, Zhipu AI's market capitalization was approximately HK$52.8 billion. It transformed from the original HK$52.8 billion to HK$1.33 trillion, completing this shift in just five months. Why is this regarded as a highly significant milestone in the history of large models? Because Zhipu AI is the first large model company in China with a market capitalization exceeding HK$1 trillion.

During that period, there were only two listed companies on the Hong Kong Stock Exchange focusing on foundational large models: Zhipu AI and MiniMax. Most of the capital looking to buy the "Chinese version of Anthropic" flooded into Zhipu AI. A company that generated RMB 724 million in revenue and posted a loss of RMB 4.7 billion in 2025 was pushed to a market size nearly half that of Alibaba. According to estimates, Zhipu AI is roughly equivalent to three Meituan, and also outpaces established technology companies such as Xiaomi and NetEase. Based on its 2025 revenue, its price-to-sales ratio exceeds 1,000 times.

Three months later, the landscape changed. On September 25, Zhipu AI's share price once fell to around HK$610.5 during the intraday session, equivalent to roughly 20% of its June 22 high, and its total market capitalization also dropped to around HK$300 billion. Calculated from the intraday high, HK$1 trillion had vanished, representing a nearly 80% decline.

Zhipu AI was far from idle during this period. It successively launched new models, recorded a nearly 4-fold increase in interim report revenue, and secured two large sums of capital from the market one after another. Including its IPO, it raised more than HK$75 billion in total within nine months. As it raised more capital, its share price kept moving lower.

Over five months, Zhipu AI's market value surged by HK$1 trillion; over the next three months, it lost that same HK$1 trillion. This dramatic reversal happened extremely quickly. On Xueqiu, there was widespread criticism of Zhipu AI, but as a pioneer in the large model sector, Zhipu AI's experience in this cycle serves as an excellent sample, offering lessons for other large model companies that are preparing for listing.

The HK$1 trillion valuation was propped up by "scarcity"

Brokerages later named that round of surge for Zhipu AI the "scarcity premium".

At that time, Zhipu AI happened to have almost all the elements that make an AI company most likely to secure a high valuation: foundational models, Coding, MaaS, developer communities, plus the extremely scarce identity of a "publicly listed large model company" in China's capital market.

OpenAI was not listed, and neither was Anthropic, making Zhipu AI a scarce access point. As the valuation of global AI assets kept rising, there were only two listed companies on the Hong Kong Stock Exchange with foundational large models as their core business: Zhipu AI and MiniMax, and Zhipu AI was the larger of the two. Hard tech capital searching for investment outlets flooded in all at once. At that time, Zhipu AI's free-floating shares accounted for less than 3% of its total share capital, with a small number of chips and a large number of buyers, so it was not surprising that the share price was amplified several times.

Early investors reaped huge profits. According to reports from Huxiu, CAS Star made an exclusive RMB 40 million angel investment when Zhipu AI was founded in 2019, when Zhipu AI's post-investment valuation was only about RMB 375 million. CAS Star transferred some of its old shares in 2023, but based on the original investment cost and the share price in late August, the paper return on this investment still exceeds 100 times. The shares held by industrial capital such as Meituan and Ant Group are also worth tens of billions of Hong Kong dollars, and the market value of shares held by Tsinghua University-related entities is also around HK$16-17 billion.

On July 8, Zhipu AI welcomed its first round of share unlock after listing, with approximately 25.68 million shares held by 11 cornerstone investors released from lock-up, which was its first share price test since listing. According to reports from Xinhua Finance, not only did the share price not fall, it closed up 13.35% on the unlock day, and rose another 11.34% the next day.

At that time, Zhipu AI seemed to have passed the test. On July 9, it took the opportunity to announce the placement of new shares at HK$1588 per share, which was completed a few days later, raising approximately HK$31.375 billion in net proceeds, with six institutions taking up these shares.

As locked shares became tradable, the floating share volume expanded dramatically. According to statistics from Jiemian News, before the unlock, Zhipu AI's free-floating shares in the market were only 11.74 million shares. This time, about 25.68 million shares were released at once, plus nearly 20 million new shares added through the placement, the number of tradable chips in the market increased several times.

Later on the evening of July 16, Moonshot AI released Kimi K3 with a total parameter count of 2.8 trillion, claiming to be the world's largest open-source model in terms of parameter scale (full weights were released on July 27), and it topped the code ranking of the Arena platform within hours after launch. User requests surged so rapidly that according to Moonshot AI's announcement, it had to suspend new C-end user subscriptions 48 hours after launch. Bloomberg compared this release to last year's "DeepSeek Moment", while JPMorgan stated that the market had seen "DeepSeek 2.0-style concerns".

The next day was the "Black Friday" for the Hong Kong stock AI sector. Zhipu AI closed down 28.49% at HK$1107 that day, with its market value evaporating by more than HK$200 billion in a single day; it fell another 19.56% in the next trading day, with its share price dropping below HK$1000. The placement institutions that took up the shares at HK$1588 had a floating loss of 30% in less than 10 days. The share price falling below the placement price was more damaging than the unlock itself.

On July 21, Zhipu AI announced that its 1GW domestic computing power center had been launched and that it had completed the acquisition of Zhongke Jiahe, and its share price once rebounded by more than 40% during the intraday session. This sharp volatility was more a vent of market sentiment than a change in fundamentals.

Tang Jie from Zhipu AI and Yang Zhilin from Moonshot AI have a teacher-student relationship, and the media simply called that week the "master vs. apprentice showdown". In fact, K3 was released only one month after Zhipu AI open-sourced GLM-5.2, and new models from Alibaba, DeepSeek and MiniMax followed right after K3.

After K3's release, JPMorgan re-evaluated the valuation framework for domestic large model companies. JPMorgan believes that the iteration of frontier models is getting faster and faster, and the valuation premium that the market previously gave to leading model companies needs to be re-examined. It immediately lowered Zhipu AI's target price from HK$2400 to HK$1600, and MiniMax's target price from HK$240 to HK$160. The "shelf life" of technological leadership has shortened, so the premium for leaders naturally has to be discounted. In the view of some analysts, K3 was just the last straw: the increase in chip supply, the squeeze of valuation bubbles and the elimination of scarcity were all happening simultaneously.

The market originally expected Zhipu AI to come up with a major countermeasure. On August 14, GLM-5.3 was released, with a parameter scale of approximately 750 billion, sharing the same base model as GLM-5.2. According to Zhipu AI's official statement, all capability improvements came from post-training. This was not an iteration without highlights. Goldman Sachs called it another major leap for Chinese AI models, and also raised Zhipu AI's revenue forecast, but its rating remained "Neutral".

Daiwa Securities put it more bluntly, believing that this was just a regular iteration, not the larger-parameter flagship model the market expected. According to reports from IT Home, Alibaba's Qianwen Office announced the first launch of GLM-5.3 a few days later. Short selling capital also became active at this time. On the day GLM-5.3 was released, Zhipu AI's share price closed down about 3.6%; according to reports from Guancha, by the eve of the sharp decline on August 18, both Zhipu AI and MiniMax's short selling indicators had hit new phase highs.

Revenue nearly quadrupled, but capital is still insufficient

On the evening of August 31, Zhipu AI released its first interim report after listing. On the same night, the quarterly adjustment of the MSCI China Index took effect, and Zhipu AI was officially included. At the close of trading that day, its share price was HK$1195, with a total market capitalization of HK$556.4 billion, nearly 60% less than its peak. It took only more than two months for its market value to drop from HK$1 trillion to more than HK$500 billion. Judging solely from growth, this interim report is actually not bad, but the market's attention is no longer fully focused on growth.

The interim report shows that revenue in the first half of the year was RMB 954 million, a year-on-year increase of nearly 400%, exceeding the full-year revenue of 2025 in just half a year. There have also been major changes in the revenue structure: open platform and API revenue reached RMB 825 million, accounting for 86.5% of total revenue, up from 26.3% at the end of last year, while localized deployment, which once supported 70% of revenue, fell by 20.5% in the first half of the year.

A company that originated from government and enterprise projects shifted its focus to the usage-based cloud business within half a year. The average API selling price rose by about 101% compared to the beginning of the year, the growth in Token usage exceeded 40 times, and the gross margin of this business changed from -0.4% in the same period last year to 24.6% now.

The business model has transformed from the original "selling projects" to the current "selling consumption", which is a qualitative change.

But the cost is also very high. The adjusted net loss for the first half of 2026 was RMB 1.964 billion, which was RMB 1.752 billion in the same period last year according to reports from Yicai, and the loss is still expanding; the overall gross margin dropped from 50.0% in the same period last year to 26.4%, almost halved. Although the gross margin of the API business turned positive from -0.4% to 24.6%, this level is far from healthy in the large model industry.

More urgently, Zhipu AI, which had just transformed its business model, was not given an adaptation period before being dragged into a brutal price war.

On September 10, DeepSeek released V4.1 Flash and further lowered its API prices. On that day, Zhipu AI and MiniMax closed down 10.34% and 8.98% respectively on the Hong Kong Stock Exchange. Shortly after, leading domestic and international manufacturers followed suit. On September 21, Xiaomi open-sourced MiMo-V2.6, pushing the single-task inference cost to US$0.13; on September 22, OpenAI released GPT-6 Sol and GPT-6 Luna, with API prices directly cut in half compared to the previous generation. On the same day, Anthropic released Claude Opus 5.5, whose cost under typical loads is also about 40% lower than the previous generation.

When the world's top models collectively cut prices, the "high cost-performance" label that Zhipu AI previously relied on to break through overnight changed from a differentiated advantage to an industry standard.

In addition, its internal ammunition is no longer sufficient. According to Zhipu AI's placement announcement in July, the net proceeds of HK$4.896 billion raised from its IPO in January had been used up by HK$4.588 billion by June 30, leaving only HK$308 million unused. This amount of capital cannot last long. Based on the loss scale in 2025, it is only enough for about one month, which also explains why the July placement came so quickly.

Where did the capital go? Zhipu AI gave the answer at its September 16 investor call, which was mainly used for computing power.

According to reports from China Fund News, the management introduced at the meeting that after the release of GLM-5 in February, model invocation demand surged by 10 times. Almost in the week of release, the company's computing power reserves were exhausted, and it had to stop selling its main product Coding Plan. Since then, Zhipu AI has purchased computing power at high prices, while using the All-in-Infra strategy and the acquisition of Zhongke Jiahe to increase computing power efficiency by two to three times. Revenue has grown accordingly, but it has never truly exploded.

It was not until the proceeds from the July placement arrived that computing power could be expanded on a large scale, and Coding Plan was relaunched for sale, with sales increasing by more than 15 times. Tang Jie mentioned that when the team developed GLM-5.3-Flash, they did a lot of "stingy" things 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 one more copy of software hardly increases costs, but every additional invocation for a model consumes a real unit of computing power. The faster revenue grows, the larger the gap in computer room capacity. As mentioned earlier, Moonshot AI suspended new C-end user subscriptions after K3's launch for the same reason, and this bottleneck is not unique to Zhipu AI.

On the evening of September 13, Zhipu AI announced another financing of approximately US$5 billion, of which US$2 billion came from new share placement and US$3 billion was zero-coupon convertible bonds, a structure the market calls "small equity, large debt". According to reports from 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$15888 placement price in July. Calculated based on the share price of around HK$610 around September 25, the six institutions that entered in July had a paper loss of about 60%, and the new batch of institutions that entered in September also had a loss of more than 10%. The two batches of institutions entered one after another, but their costs differed by more than double.

The conversion price of the convertible bonds is set at HK$892.5. According to estimates, if all converted into shares, the share capital will be diluted by about 5% further. According to disclosure from MyDrivers, the management did the math at the investor call: investing RMB 300 billion can build about 100,000P of computing power, 40% of which is reserved for training and R&D, and 60% for business inference; based on the inference cost-performance of GLM-5.3, the maximum theoretical gross margin of inference can reach 80%, and if the inference computing power is fully utilized for one year, revenue can reach RMB 40 billion.

This calculation looks very attractive, but it has many prerequisites. Any variable such as computing power utilization, business discounts and fluctuations in computing power prices may reduce the actual results.

At the same investor call, Zhipu AI raised its full-year ARR guidance from US$2.4 billion to US$3 billion. The full-business caliber ARR has reached US$1.8 billion, and revenue sharing with domestic and overseas cloud vendors will be recognized starting from October. According to reports from Rui Finance, industry orders under the Co-work model have also exceeded RMB 1 billion, but this part will be recognized over one to two years. The market reaction was rather tangled: Zhipu AI's share price fell for two consecutive days on September 14 and 15, with its market capitalization once dropping to HK$316.6 billion, before rebounding by 5.88% on the day of the investor call.

In the context of the entire industry, these funds are not that much. ByteDance's 2026 AI