ZhiPu's Hong Kong stock price has surpassed HK$2,380, with its total market capitalization exceeding HK$1 trillion, achieving a gain of over 1,900% since its listing less than six months ago.Article author and source: Tencent Technology
Three and a half years after the explosion of generative AI, the market has reached a new point of divergence: optimism continues to accelerate, while skepticism is mounting. Determining whether a bubble has arrived is insufficient to explain today’s complexity. The series “AI Faith vs. Bubble” will examine key variables from multiple perspectives—market, technology, industry, and companies. This is the sixth article in the series.
At 600 billion, almost everyone thought Zhipu was too expensive, but it rose to a trillion.
Soros once said: "Prices in financial markets are always wrong, but the errors can be self-reinforcing for a considerable period, even ultimately turning themselves into right."
On the morning of June 22, the first trading day after the Dragon Boat Festival holiday, Zhipu's Hong Kong-listed shares opened more than 13% higher, surpassing HK$2,380 per share, with its total market capitalization officially reaching HK$1 trillion. Within less than six months of listing, the stock has surged over 1,900%.
Go back four days, and a conversation began to spread on X.
Someone asked Musk: "When is China's large model expected to reach Fable level? GLM-5.2 has undoubtedly narrowed the gap." Musk replied with two words: "Probably Q1."
Professor Tang Jie from Tsinghua University and founder of Zhipu directly replied below with four words: "Won't take that long."
Elon Musk added that while you might be right if you only look at benchmark scores, creating real, practical value that generates profit is an entirely different matter.
Tang Jie replied again: “Focus is all we need.” This retort is not translated—it may carry rich meaning.
a16z founder Marc Andreessen shared this conversation with the word: “Interesting.”
Three days later, when the Hong Kong stock market opened, Zhipu surpassed one trillion.
01. Everyone thinks Zhipu's valuation is outrageous.
Reviewing Zhipu's first annual report, the company reported total revenue of RMB 724 million for the full year of 2025, a net loss of RMB 4.718 billion, and R&D expenses of RMB 3.18 billion.
For intuitive reference: Alibaba Cloud's revenue for fiscal year 2026 was RMB 158.1 billion, with AI-related revenue growing at triple-digit rates for seven consecutive quarters. Zhipu's annual revenue is approximately 1/218 of Alibaba Cloud's, yet the secondary market has assigned it an independent valuation close to one trillion HKD. During the same period, Alibaba's market capitalization stood at just RMB 1.96 trillion HKD.
Everyone noticed this numerical discrepancy. Some in the investor community said it was "seriously inflated," while others directly asked, "Can we short it?"
The question of whether it’s reasonable or not isn’t a real question—it will likely take three years before a final judgment can be made. This article attempts to address another question: why is the price still rising when everyone feels it’s too expensive?
Trillions are not all air—the underlying force driving the rally is indeed fundamental.
The newly released flagship model, GLM-5.2, features a total of 744 billion parameters and a MoE architecture with 40 billion activated parameters, supporting lossless context up to 1 million tokens. In the blind test rankings on Arena.ai's Code Arena, GLM-5.2 Max ranked among the top two with a score of 1593.
In the Design Arena web design evaluation, the June 20 update shows GLM-5.2 topping the leaderboard with an Elo score of 1360, surpassing Claude Fable 5. On the Terminal-Bench 2.1 evaluation, GLM-5.2 scored 81.0, up from 63.5 for its predecessor GLM-5.1, representing a 17.5-point improvement over two months. In the FrontierSWE real-world software engineering tasks evaluation, GLM-5.2 trails only slightly behind Claude Opus 4.8 by approximately 1 percentage point and surpasses GPT-5.5.
Chinese open-source large models are now competing on par with Anthropic’s and OpenAI’s flagship models in programming, a high-commercial-value capability.
But it’s also honest to add that Musk’s comments have merit—being close in a single benchmark is not the same as having similar model productivity.
Some developers have created websites to monitor model usage; under similar task success rates, GLM-5.2 Max consumes several times the duration and tokens of comparable models, indicating that the model is still inefficient at figuring out the correct approach and requires more trial and error.
Although this is a personal developer’s website and its objectivity is yet to be verified, it nonetheless suggests that GLM-5.2 still has some gaps in efficiency and cost when handling complex engineering tasks requiring multi-step reasoning and long-range agent operations.
Another notable signal is that multiple developers have reported that Zhipu’s Coding Plan is difficult to obtain. According to the official documentation, the Coding Plan supports GLM-5.2, but it imposes a 5-hour quota, a weekly quota, and dynamic concurrency limits. During peak hours, GLM-5.2 usage is deducted at three times the normal rate.
On January 23, 2026, Zhipu issued an official announcement: due to a surge in users following the launch of GLM-4.7, which caused severe compute resource constraints, the daily new subscriptions for the GLM Coding Plan were reduced by 80% and changed to a limited daily release at 10:00. Subsequently, online tutorials titled "How to Secure the Coding Plan" emerged, yet even users setting alarms to refresh precisely at 10:00 still found slots sold out instantly. In February, Zhipu issued an apology for drastically cutting quotas and opened refunds for existing users.
However, Coding Plan functions more as an entry-point product for Zhipu’s entry into the developer ecosystem rather than a primary revenue source. It leverages subscriptions to integrate workflows like Claude Code, Cline, and OpenCode, rapidly expanding GLM-5.2’s presence in coding scenarios. However, the drawbacks of Coding Plan are clear: coding agents consume substantial context and multiple rounds of calls, making it difficult for subscription pricing to cover the actual inference costs.
Therefore, after the release of GLM-5.2, the difficulty in securing Coding Plan slots, strict usage limits, and doubled deductions during peak hours reflect not just a sales strategy, but likely indicate Zhipu’s efforts to manage inference resource allocation and mitigate subscription model losses. In contrast, API billing by token is better suited to generate sustainable revenue and is more appropriate for long-term enterprise use.
02. "Scarcity": The Fundamental Logic Behind Large Model Valuations in 2026
If we examine the valuation rankings of global large model companies in 2026, we will find an counterintuitive pattern: this market cycle rewards companies with supply-demand imbalances, not those with the most users.
The most prominent example is Anthropic surpassing OpenAI. In May 2026, Anthropic completed its Series H funding round of $65 billion, reaching a valuation of $965 billion and surpassing OpenAI’s $852 billion for the first time.
From a product perspective, Claude has always been harder to use than GPT. Claude Pro/Max has consistently had much stricter rate limits than ChatGPT Plus, causing heavy users to frequently hit speed limits mid-conversation, and waiting in line for Claude Code during peak hours is normal. Claude’s API pricing has also consistently been higher than GPT’s equivalent models.
Users have developed the perception that "Claude = scarce = premium," much like a restaurant that requires waiting in line—psychologically, they assume it’s better than the one next door that’s empty.
The same applies to the capital side. Both companies are not yet publicly listed, and their valuations are derived from private financing and secondary share transactions. In October 2025, OpenAI conducted a large-scale tender offer, in which more than 600 current and former employees collectively cashed out $6.6 billion, with approximately 75 individuals selling the maximum allowed amount of $30 million each. Buyers interested in acquiring OpenAI equity have relatively ample supply to choose from.
The founding team of Anthropic holds a concentrated stake, and early strategic investors, such as Google and Amazon, are unlikely to sell before the IPO. There are very few sellers willing to part with their shares, while buyers are competing fiercely for the limited supply.
Product-side scarcity plus capital-side scarcity created a dual resonance, pushing Anthropic’s valuation beyond that of OpenAI.
Zhipu exhibits a similar scarcity effect: on the product side, the Coding Plan is limited to daily sales; on the capital side, the current float in the Hong Kong market is under 4%. Whether it’s its token or its stock, both are in a state of supply falling short of demand.
03. Why do people still buy when they know it's expensive?
However, from a valuation perspective, these fundamentals are entirely insufficient to support a trillion-dollar market cap; it is widely accepted by the market that Zhipu is overvalued.
And even when the market was at $500 billion or $600 billion, people said it was expensive—so why is it still rising?
In 1936, Keynes proposed in The General Theory that the stock market is not a game of judging who is the most beautiful, but a game of predicting what the majority will consider the most beautiful.
A large portion of buyers of Zhipu don’t need to believe it’s worth trillions. As long as the float remains extremely small, passive demand continues, and the narrative of leading coding capabilities stays strong, the existence of a next buyer remains a high-probability event.
In stocks with low free float, the inclusion in an index, passive fund allocation, and limited supply can create a short-term amplification effect. Zhipu was announced for inclusion in the Hang Seng Tech Index by Hang Seng Indexes Company in May 2026, with official effectiveness beginning on June 8, and an initial weighting of approximately 0.53%.
The special aspect of this inclusion is that, although Zhipu has a high market capitalization, the free float truly unrestricted by lock-up restrictions at the time of listing is limited. According to the prospectus, approximately 11.7379 million H-shares were not subject to sale restrictions at listing, accounting for about 2.67% of the total post-listing share capital. This means that passive funds rebalancing to the Hangke Index may encounter insufficient tradable shares in the short term, thereby amplifying price volatility.
Thus, each participant acts rationally within their own constraints: passive funds are bound by rules to buy; short-term traders engage in strategies with clearly defined time windows; and long-term institutional investors see the configuration logic in the scarcity of pure-play Chinese AI assets. Potential competitors like DeepSeek are not publicly listed, and Moonshot AI has not yet gone public—only Zhipu is a publicly traded company with the strongest AI foundation model capabilities on the open market.
No one is insane. It’s just that the collective rational actions of everyone have produced a price that appears insane.
However, this wild price will be a double-edged sword.
For Zhipu, a high valuation is a good thing. Similar stories have happened in history: Amazon’s market capitalization peaked at around $30 billion during the 2000 dot-com bubble, then dropped 90% in 2001, but today its valuation is dozens of times higher than it was at the peak of the bubble. It ultimately turned the window of opportunity provided by capital markets into long-term infrastructure capabilities, paving the way for later growth drivers like AWS.
If Zhipu can truly leverage the resources gained from this round of ultra-high valuation—funding, talent, customers, and computing power—to establish itself as China’s leading AI infrastructure platform within two to three years, then today’s trillion-dollar valuation will not be a bubble.
In Q1 2026, Zhipu's MaaS API generated an annualized revenue of approximately $250 million, with API pricing rising 83% quarter-over-quarter and paid tokens increasing fourfold compared to the previous quarter. If this pace is sustained, full-year 2027 revenue could reach $6 to $8 billion, reducing the price-to-sales ratio from over 1,200x to just over 100x. While still high, this would bring it into the acceptable range for high-growth technology companies.
According to Soros' concept of reflexivity, market prices do not merely reflect fundamentals—they can also反过来 alter them. As long as this cycle continues to function smoothly, simply shouting "PS is too high" cannot burst the bubble, because the denominator is catching up to the numerator at a triple-digit rate.
04. Whose risk?
On the other side of the coin, for secondary market investors, the current risk-reward ratio is extremely asymmetric. Company benefits and investor benefits are two different things.
In the short term, approximately 25.68 million restricted shares will be unlocked on July 8, accounting for about 5.76% of the total shares outstanding. While this is not the largest unlock event, it could significantly alter the short-term supply dynamics given the currently very low actual float. A larger unlock window is scheduled for January 8, 2027.
Over a six-month to one-year horizon, quarterly revenue growth rate is a key metric to monitor. The current triple-digit growth, driven by an 83% increase in API pricing and a fourfold rise in token volume, underpins the market’s willingness to assign a very high price-to-sales (P/S) ratio.
Once quarterly growth drops from triple digits to double digits—even from 150% to 70%—the market will rapidly reprice. At a price-to-sales ratio of 1,200x, it implies that the timeline for the future to catch up to the numerator has been significantly extended, causing the valuation model to collapse under revision.
Conversely, if the company maintains triple-digit growth for two or three consecutive quarters, the market may conclude that the reflexive cycle is indeed in motion, and prices could rise again after the lock-up release shock has been absorbed.
In the long term, over the next two to three years, the key will be whether Zhipu can sustain its competitive advantages.
For example, if DeepSeek releases a model in the coming months with coding capabilities on par with GLM-5.2 and offers it at an extremely low price or even for free, Zhipu’s Coding Plan advantage of “high demand and limited supply” will disappear.
Once alternatives emerge and scarcity disappears, pricing power will no longer belong to Zhipu.
Zhipu’s trillion-dollar market cap has an underlying logic; there’s no denying that Zhipu is a strong company, and the market cannot be simply dismissed as irrational. However, “having logic” and “being a good investment” are two different things. At this price, perhaps only 50% reflects “real growth,” while the other 50% stems from “structural amplification + narrative premium.”
The bottom 50% must endure the ongoing test of time.