Produced by | Bullet Finance

Art Design | Qianqian

Reviewed by | Songwen

On January 8, 2026, Zhipu made its debut on the Hong Kong Stock Exchange with an issue price of HK$116.2 per share, earning the distinction of being the “world’s first listed large model company.” This pioneering status bestowed upon Zhipu a rare and highly coveted prestige.

Following its listing, Zhipu's stock price skyrocketed, reaching a peak of HK$2,980 per share—an increase of over 24 times from its issue price—and its market capitalization briefly surpassed HK$1 trillion.

However, this meteoric rise was soon followed by a sharp reversal.

As the lock-up period expiration loomed, Zhipu's stock price entered a sustained downward spiral. On July 17, the stock plummeted by more than 28% in a single day; on July 20, it fell an additional 19.56%. After a brief rebound, the stock collapsed by 13.3% on August 18, with intraday losses at one point exceeding 16%.

By the close on August 25, Zhipu was trading at HK$1,059 per share, down over 64% from its all-time high, with a market capitalization of HK$493.1 billion. From its peak, the company’s market value has diminished by more than HK$500 billion.

From being a highly sought-after and rare asset to experiencing a stock price “halving” as players flocked into the large model sector, can the scarcity narrative that once underpinned Zhipu's lofty valuation hold water?

Founded in 2019 and spun off from Tsinghua University's Knowledge Engineering Group (KEG), Zhipu was co-founded by Professor Tang Jie, positioning it as one of China’s earliest large model companies.

In August 2022, prior to the launch of ChatGPT, Zhipu open-sourced GLM-130B, the world’s first open-source Chinese-English bilingual model boasting 100 billion parameters.

Subsequently, Zhipu integrated “open-source” into its core philosophy. In 2025, it fully open-sourced its flagship model under the highly permissive MIT License, enabling developers to use it commercially free of charge.

This strategy successfully attracted a global developer community to the GLM series. By the end of 2025, Zhipu’s products had reached 218 countries and regions, serving over 12,000 institutional clients and more than 242,000 paid developers.

For a company with a strong enterprise focus, this represented a solid foundation.

In the past year, Zhipu’s GLM-5.2 model gained traction in programming scenarios due to its “post-training” capabilities. Post-training involves keeping the model’s foundational parameters (pre-trained weights) unchanged while feeding it high-intensity, high-density quality data in specific domains to enhance performance—akin to a student raising their grades from passing to near-perfect through relentless practice.

Leveraging this approach, GLM-5.2 performed on par with or even surpassed some overseas closed-source models in programming benchmarks. Thus, when Kimi K3 debuted in July with 2.8 trillion parameters, the market eagerly awaited Zhipu’s response.

On August 14, Zhipu released GLM-5.3, featuring over 750 billion parameters. Built on the same foundation as GLM-5.2, all improvements stemmed from post-training. Specifically, GLM-5.3 achieved notable advancements in code generation, multi-round reasoning, and long-context understanding.

Objectively, this was far from a lackluster upgrade. Goldman Sachs hailed it as “another major leap in China’s AI model progress” and raised its revenue forecast for Zhipu. Yet, intriguingly, Goldman maintained a “neutral” rating.

Daiwa Securities was more blunt, arguing that GLM-5.3 was merely a routine iteration, not the larger flagship model the market anticipated, making it unlikely to catalyze unexpected growth in the short term.

Even more telling, on the evening of August 14, Qianwen Office proudly announced the integration of GLM-5.3 and DeepSeekV4Pro as its first external models. However, just three days later, both models vanished from the product interface, and related WeChat articles were set to private.

While Alibaba offered no explanation, this wavering attitude indirectly shook investor confidence in Zhipu’s “external ecosystem integration” business. After all, if the stability of B-end distribution channels is questionable, the credibility of the company’s ecosystem narrative naturally diminishes.

(Figure: Qianwen interface no longer shows external models)

At dawn on August 19, Zhipu officially launched the API (application programming interface) for GLM-5.3, allowing developers to pay for access to the new model. Yet, the stock price remained weak, closing down 2.97%.

Amid this delicate timing, Zhipu faced its first wave of post-listing lock-up expirations.

On July 8, Zhipu’s first batch of restricted shares became eligible for sale, and the stock gapped down at the open, hitting a low of HK$1,450 per share. However, it quickly rebounded, closing up 13.35%; the next day (July 9), it rose another 11.34%.

However, extending the timeframe reveals that after July 9, Zhipu’s stock price clearly entered a downtrend. Daily short-selling data from the Hong Kong Stock Exchange showed a marked increase in short-selling activity following this lock-up expiration.

(Figure: East Money)

On August 14, the day GLM-5.3 was released, daily short-selling turnover surged to HK$1.16 billion, with short-selling accounting for 7.74% of total trading volume and 885,300 shares shorted—both hitting new highs since July 9.

(Figure: East Money)

Meanwhile, Zhipu proceeded with a share placement. On July 9, the company announced it would place new H-shares under general mandate at HK$1,588 per share—a 12.99% discount to the July 8 closing price of HK$1,825. Up to 19.78 million new H-shares could be subscribed by six professional, institutional, and other investors.

Just four days later (July 13), the placement was completed, raising a total of approximately HK$31.41 billion (about RMB 27.2 billion). After deducting commissions and estimated expenses, the net proceeds were about HK$31.375 billion.

Soon after, the stock price plummeted.

By August 25, the stock traded at HK$1,059 per share. Based on the placement price of HK$1,588, participating institutions faced losses exceeding 30%. In less than a month, the HK$30+ billion placement shifted from “a steal” to “being left holding the bag.”

Yet compared to these latecomers, Zhipu’s early investors had already reaped 100-fold returns.

The prospectus revealed that since its inception, Zhipu had completed eight funding rounds, raising over RMB 8.3 billion from more than 50 institutional investors, including industrial capital like Meituan, Ant Group, Alibaba, Tencent, Xiaomi, and Kingsoft; top venture firms such as Legend Capital, Sequoia Capital, Hillhouse Capital, and Qiming Venture Partners; and local state-owned entities from Beijing, Shanghai, Chengdu, Tianjin, and Hangzhou.

Among them, Zhongke Chuangxing invested approximately RMB 40 million in 2019, when Zhipu’s post-money valuation was just RMB 375 million. After subsequent funding rounds and equity dilution, Zhongke Chuangxing still holds nearly 1.47% of the shares. Even at the current depressed stock price, this stake is worth about HK$7.2 billion—a roughly 100-fold return on paper.

Internet giants also fared well: Meituan invested RMB 300 million in Series B2 and holds 17.2173 million shares (about 3.9% post-listing, now diluted to ~3.70%), valued at ~HK$18.2 billion. Ant Group invested ~RMB 600 million and holds 16.0847 million shares (~3.61%, now ~3.45%), worth ~HK$17 billion.

Additionally, Tsinghua University’s Huakong Technology Transfer Co., Ltd., a core founder, holds 15.5344 million shares (~3.48%, now ~3.34%), valued at over HK$16.4 billion. The Tsinghua University Education Foundation’s JinYi Capital, as a cornerstone investor, subscribed to ~HK$54 million in shares. Combined, Tsinghua-related entities hold Zhipu shares worth ~HK$16.9 billion.

While institutions that participated in the placement face losses exceeding 30% after just a month, early shareholders still hold paper gains of several times post-lockup. Not all investors share the same fate.

As AI adoption accelerated, Zhipu’s revenue soared. In 2025, revenue reached RMB 724 million, up 131.85% year-over-year.

Enterprise-grade general large models (private on-premises deployments) generated RMB 366 million, accounting for 50.4% of total revenue, primarily serving central SOEs and financial institutions.

Enterprise-grade agents contributed RMB 166 million (22.9%), while open platforms and APIs (cloud-based) added RMB 190 million (26.3%). Thus, Zhipu predominantly serves government, enterprises, and financial institutions.

(Figure: Zhipu’s 2025 annual report)

However, unlike its revenue growth, Zhipu’s net loss attributable to shareholders widened to RMB 4.698 billion in 2025, up 58.91% year-over-year.

From 2022 to 2025, Zhipu incurred losses of RMB 143 million, RMB 788 million, RMB 2.956 billion, and RMB 4.698 billion, respectively, with cumulative losses nearing RMB 8.6 billion over four years.

Behind these persistent losses lies alarming cash burn. After its January 2026 IPO and overallotment exercise, Zhipu raised a net total of ~HK$4.896 billion. By June 30, 2026, it had spent ~HK$4.588 billion—over 90% of the proceeds within six months of listing.

This explains why the company urgently proceeded with a HK$30+ billion placement in July. Such frequent fundraising is rare in the Hong Kong market.

For Zhipu, it seems a never-ending race: to build stronger models, it needs more computing power; to afford more computing power, it must raise more capital. High growth, high investment, and heavy losses are the shared fate of large model companies today.

However, Zhipu’s stock price volatility reflects a broader shift in the valuation logic of the large model sector. The “scarcity premium” that once supported a trillion-dollar market cap is rapidly eroding.

At its listing, Zhipu commanded such a high valuation largely due to its “first-mover” advantage. With no other pure-play large model companies listed in Hong Kong at the time, its status as a “rare domestic large model asset” drove frenzied capital inflows and inflated valuations.

But as AI large models rapidly evolve, many companies are rushing toward capital markets. Overseas, Anthropic filed for a confidential IPO with the U.S. SEC on June 1, valuing it at ~US$965 billion post-investment, with a potential listing by Q4 2026; OpenAI followed on June 8, targeting a ~US$852 billion valuation.

Domestically, multiple media outlets report that Kimi (Yuezhi’s Dark Side) has sent listing proposals to investors, preparing for a Hong Kong IPO with a pre-money valuation raised to US$50 billion (~RMB 338.5 billion); DeepSeek is also planning an A-share STAR Market listing.

As global large model leaders flock to capital markets, the “golden label” of “Hong Kong’s only pure-play large model stock” loses its luster.

More intriguingly, Zhipu has announced plans for an A-share listing, meaning its “A+H” structure will diversify financing channels beyond Hong Kong.

As scarcity fades, technological moats are also eroding at a “weekly update” pace. Since July, models like Kimi K3 (2.8 trillion parameters), Grok4.6, DeepSeekV4Pro, and Qwen3.8 have launched in rapid succession, with top players jostling for leadership every week.

Scarcity is being replaced by “oversupply,” forcing capital to shift from concentrated bets on a few assets to diversified wagers.

To be clear, Zhipu is not alone in its stock price decline. MiniMax, another Hong Kong-listed large model company, has tumbled from a high of HK$1,330 per share to around HK$300.

Over seven months since its listing, Zhipu’s stock price gyrations epitomize a broader realignment in the large model sector’s valuation logic. As more players enter and technology iterates faster, stories built solely on “scarcity” will eventually run their course.

Whether the company can prove its worth through other means and warrant revaluation remains to be seen. The large model narrative is far from over, but the rules of storytelling have changed.

*Featured image from: Shutterstock, under VRF protocol.