On July 17th, shares of Z.AI (02513.HK) plummeted 28.49%, wiping out over 200 billion Hong Kong dollars in market capitalization in a single day. Two days later, on July 20th, the stock fell another 19.56%. In just two trading sessions, more than 300 billion Hong Kong dollars in value vanished.

On the surface, this was a sudden assault by Kimi K3. This open-source model with 2.8 trillion parameters, released on the early morning of July 16th, scored 1679 points on Code Arena, surpassing GLM 5.2 and directly challenging Z.AI’s most prized coding territory. However, K3 was merely the spark. The true detonation was the valuation landmine buried since Z.AI’s IPO: its trillion-dollar market cap was never built on earnings, but on the unique status of being the “first major model stock.” When that scarcity evaporated, the market cap began its inevitable descent. This is not a simple "black swan event," but a long-overdue valuation correction.

The Genesis of a Trillion-Dollar Market Cap: Not Performance, but a Lack of Alternatives

On January 8, 2026, Z.AI listed on the Hong Kong Stock Exchange at an IPO price of 116.1 Hong Kong dollars, raising net proceeds of approximately 48.96 billion Hong Kong dollars. Six months later, its share price hit a high of 2980 Hong Kong dollars, pushing its market cap above 1 trillion Hong Kong dollars—a gain of over 25 times. But a look at Z.AI’s annual report reveals a stark financial reality. For the full year of 2025, the company posted revenue of just 724 million yuan, a net loss of 4.718 billion yuan, and R&D spending of 3.18 billion yuan. How could a company with less than 800 million yuan in revenue and losses exceeding 4.7 billion yuan be worth 1 trillion Hong Kong dollars? Even within the context of the 2025 AI investment frenzy in Chinese stocks, this number was fantastical. The answer is that capital had no other choice.

In the first half of 2026, the pool of publicly tradeable “pure-play” large model stocks in China was extremely scarce. ByteDance was not listed, Moonshot AI (Kimi) was not listed, and MiniMax, despite being public, had limited liquidity. Z.AI’s early move to list on the Hong Kong Stock Exchange gave it the identity of the “first major model stock,” making it the only channel for global capital to bet on China’s large model narrative. To participate in this story without being able to buy ByteDance or Alibaba’s Tongyi Qianwen, investors were forced to buy Z.AI. This scarcity premium was inflated to a trillion-dollar level. This was not a rally driven by fundamentals, but by a capital market with "no other choice."

The core argument supporting this valuation was the MaaS platform’s ARR (Annual Recurring Revenue), which had reached 1.7 billion yuan (about 250 million US dollars) by March 2026, a 60-fold increase over the past year. While that growth rate sounds impressive, consider this: a 1.7 billion yuan ARR against a 1 trillion Hong Kong dollar market cap gives a price-to-sales ratio of over 400 times. Even for the most aggressive SaaS companies in the US, this multiple is a fairy tale. Furthermore, this ARR figure is worth questioning. It is not an audited financial number, but an operational metric disclosed by the company. What was the base for that 60-fold growth? If the ARR was less than 30 million yuan a year ago, how much water is there in this "growth narrative"? The same script has already been played out by MiniMax. After its lock-up period expired in early July, MiniMax’s stock collapsed, and its market cap quickly fell below 100 billion Hong Kong dollars. The company reported revenue of 79.038 million US dollars for 2025 and a net loss of 1.872 billion US dollars. The only difference between Z.AI and MiniMax is that Z.AI listed earlier and its scarcity premium lasted longer. The ending of the script is unlikely to be different. When the market has only one option, capital is willing to pay a premium. But when supply increases and scarcity fades, the premium will correct itself as quickly as possible.

The Shattering of the Illusion: Open-Source Models Eliminate Scarcity

On the early morning of July 16th, Kimi K3 was officially released. It is the world's largest open-source model with 2.8 trillion parameters. It scored 1679 points on Code Arena, surpassing GLM 5.2 to rank first globally. It features a 100 million Token ultra-long context and has an API price on par with Sonnet 5. Looking at these parameters together, it’s clear why the market reacted so violently against Z.AI: K3’s positioning overlaps heavily with Z.AI’s GLM 5.2—both target the coding track, have ultra-long context capabilities, and focus on the agent direction. However, K3 is more powerful and free and open-source. That is the fatal blow. What does an open-source model mean for a large model company’s business model? It means customers can migrate to a better model at zero cost. Z.AI’s MaaS platform ARR grew 60-fold in the past year, but on what basis was this growth built? It was built on the fact that customers had no better choice. Now that K3 is open-source, free, and more powerful, why would customers continue to pay for GLM 5.2’s API? The coding track was Z.AI’s deepest moat and the core pillar of its valuation narrative. GLM 5.2’s ranking on Code Arena was its trump card—one of the world’s most powerful coding models, a breakthrough for a Chinese company in this niche. But with K3 surpassing GLM 5.2 at 1679 points, this trump card has been directly crushed.

The deeper issue is that in the large model race, the title of “world number one” has never been stable. GPT-4 was followed by Claude, Claude by Gemini, Gemini by GLM, and now GLM has been followed by K3. Every few months, the rankings are reshuffled. What does this mean? It means that any large model company’s claim of “technological leadership” is temporary and cannot form a lasting competitive barrier. Open source further compresses this temporary advantage to zero. Closed-source models can still bind customers through their APIs, but open-source models leave customers with no reason to be locked in. On July 17th, Z.AI shares fell 28.49%, closing at 1107 Hong Kong dollars. On July 20th, they fell another 19.56%, closing at about 890 Hong Kong dollars. The cumulative drop over two trading days was over 44%. The Hong Kong stock AI sector also crashed, a day dubbed “Black Friday,” as related stocks like MiniMax and 4Paradigm collectively slumped. This was not just a panic sell-off. It was a re-evaluation of the valuation model. When the scarcity of the “first major model stock” is diluted by open-source models, capital begins to recalculate: If customers can migrate for free to a better open-source model at any time, is Z.AI’s ARR still worth 1.7 billion yuan? If the top coding position can be lost at any moment, what is Z.AI’s differentiation? If the “world’s best” title changes every few months, is a 400x price-to-sales ratio reasonable? From its peak of 2980 Hong Kong dollars, Z.AI’s share price has already fallen over 70%. Its market cap has dropped from 1 trillion Hong Kong dollars to about 470 billion Hong Kong dollars. And this may not be the end. If the next earnings season confirms a slowdown in ARR growth, the valuation could have further room to compress.

The Reality of the Cash Burn: Financing Speed Can Never Keep Up

Z.AI’s cash burn rate would make any traditional tech company envious. The net proceeds from its IPO were about 48.96 billion Hong Kong dollars. By June 30, 2026, 93% of this money had already been spent, with 45.88 billion Hong Kong dollars used and only 3.08 billion Hong Kong dollars remaining on the books. Less than six months after listing, the IPO funds were nearly depleted. For a company with annual R&D spending of 3.18 billion yuan and a net loss of 4.718 billion yuan, 3.08 billion Hong Kong dollars would not even last two months at the 2025 cash burn rate. What does this burn efficiency mean? For every 1 Hong Kong dollar Z.AI raises, it burns through it in less than two months. This led to a top-up placement announcement on July 9th, with up to 19.78 million shares issued at 1588 Hong Kong dollars per share. The placement was completed on July 13th, raising net proceeds of about 313.75 billion Hong Kong dollars. The company expects this money to be completely used up by the end of 2027—in just 18 months. The cycle of "burn cash, raise funds, burn cash again" is now in its second round. The first round was the IPO, which was spent in six months. The second round is the placement, expected to last 18 months. Where will the third round come from? At what price? Will the market still accept it?

While 313.75 billion Hong Kong dollars seems like a massive sum, in the context of the large model race, it may be less durable than you think. The construction of a 1GW domestic data center is underway, which is a multi-billion dollar capital expenditure. The acquisition of Zhongke Jiahe to lay out heterogeneous computing software requires capital. The evaluation of cooperating with domestic chip companies to develop custom AI chips demands even more. The 3.18 billion yuan R&D figure is from 2025, and it will only be higher in 2026. For a rough but meaningful comparison, Anthropic had an ARR exceeding 47 billion US dollars in June 2026 and a valuation of 965 billion US dollars. Z.AI’s ARR is 1.7 billion yuan (about 250 million US dollars), with a peak market cap of 1 trillion Hong Kong dollars (about 128 billion US dollars). Anthropic’s valuation-to-ARR ratio is about 20 times, while Z.AI’s was over 500 times. This comparison is, of course, not entirely fair—Anthropic operates in the US market with mature SaaS payment habits, while Z.AI is in China educating customers from scratch. But even factoring in all the "China premium" and "first major model stock scarcity," the gap between 500 times and 20 times still points to one conclusion: either Anthropic is severely undervalued, or Z.AI is severely overvalued.

Now look at Kimi. Kimi’s ARR is about 300 million US dollars, with a valuation of about 31.5 billion US dollars, giving it a valuation-to-ARR ratio of about 100 times. Before the K3 release, 100 times was already high. But after the K3 release, Kimi’s narrative logic has actually strengthened: open-source, free, and globally top-performing, which is backed by a different business logic: using open source to build an ecosystem and using the ecosystem to secure the future. This logic is at least harder to disrupt than the "coding track differentiation." What is Z.AI’s business logic? It sells APIs, sells MaaS services, and charges based on coding capabilities. This logic made sense before K3, but after K3, why would customers pay for a capability they can get for free? Z.AI now faces two paths: either cut prices to retain market share or watch customers leave. Neither path is appealing. The more troublesome issue is the signal sent by the placement price. The placement price of 1588 Hong Kong dollars was already a discount to the market price in early July. The institutional investors who participated in the placement are now facing a floating loss of over 40%. If the share price continues to fall, the price for a third round of financing will only be lower, further diluting existing shareholders.

Tang Jie's 'Touch High' Plan: Redefining the Narrative Before the Collapse

On July 11th, five days before the K3 release, Z.AI CEO Tang Jie published an internal letter titled "The Great Wave Has Arrived." In it, he announced a new strategy called the "Touch High" plan. This timing was no coincidence. The K3 release was on July 16th, and Tang’s letter was on July 11th. Major model releases usually have a warm-up period, and the news of K3’s upcoming launch was likely an open secret in the industry. Tang sent the letter five days before the valuation collapse, and the timing itself speaks volumes: he knew change was coming and needed to redefine Z.AI’s narrative before the market started asking, “What do you have after coding is surpassed?” The core operation of this redefinition is very clear: coding was almost invisible in the letter. Remember, coding was always Z.AI’s most core and differentiated label. The key selling point of GLM 5.2 was its coding ability, and the Code Arena ranking was its PR trump card. But in the "Great Wave Has Arrived" internal letter, coding is no longer the protagonist. It has been replaced by four new directions: Long Horizon Task, Autonomous Agent System, Fully Self Training, and Security Governance (mechanistic interpretability).

The narrative shift Tang Jie wants is to upgrade Z.AI from a "coding company" to an "AGI company." There is an iron law in the capital market: once a story is realized, it is no longer the future. Z.AI had already told the coding story to its fullest. The Code Arena ranking of GLM 5.2 being overtaken by K3 means the coding narrative has been disrupted. The capital market will not continue to pay a premium for a "number one that has already been overtaken." Today you are the coding leader; tomorrow K3 is; the day after tomorrow, another model might be. This fluidity in rankings severely devalues the investment appeal of the coding narrative. Tang Jie needs a bigger story, one that is harder to verify, more distant, and more vague. AGI is just such a story. Long Horizon Tasks—AI that can plan and execute complex goals over extremely long time spans—sounds grand, but when will it be realized? No one knows. Autonomous Agent Systems—AI that can complete the entire process from planning to execution independently—is also a long-term vision. Fully Self Training—models that can self-evolve without human annotation—is technically appealing but far from commercial reality. Security Governance and mechanistic interpretability are the most distant directions, with no near-term commercial potential. The common thread of these four directions is that they are grand enough, distant enough, and vague enough. Vagueness means room for imagination, and imagination means room for valuation. You tell the story while the market is still willing to listen. However, the July 17th crash has already provided an initial answer. On the first trading day after the "Touch High" plan was announced, the stock price did not rise but instead crashed 28.49%. This suggests the market did not buy it—at least not fully. Investors can see through this maneuver: the coding story is played out, so quickly switch to a story that is harder to disprove. But changing the story does not change the fundamentals. Tang Jie’s dilemma is that he must tell a story, otherwise the valuation will collapse even faster. But the new story he tells may not be believed by the market. Between belief and disbelief, Z.AI’s share price will experience more volatility.

How Long Will 313 Billion Last?

At the close on July 20th, Z.AI’s market cap was about 470 billion Hong Kong dollars. This is more than half of its peak of 1 trillion Hong Kong dollars. It is still several times higher than the IPO price of 116.1 Hong Kong dollars, but for investors who participated in the placement at 1588 Hong Kong dollars, the floating loss is already over 40%. How will these institutional investors respond? Will they cut their losses and leave, or will they grit their teeth and hold on, waiting for a rebound? Their choice will largely determine the short-term direction of Z.AI’s stock price. The 313.75 billion Hong Kong dollars from the placement is Z.AI’s lifeline. The company expects it to be used up by the end of 2027, in 18 months. In those 18 months, what does Z.AI need to prove? It needs to prove that the story beyond coding is credible—that grand narratives like Long Horizon Tasks, Autonomous Agents, and Self Training are not just concepts on a PowerPoint but require technical progress, product deployment, and customer adoption. It needs to prove that the GLM series models still have paying customers despite the K3 shock, and that ARR will not drop off a cliff. It needs to prove that the "Touch High" plan is not an emergency rebranding, but that real technological breakthroughs are happening. But the reality is harsh. The 1GW domestic data center is under construction, the acquisition of Zhongke Jiahe for heterogeneous computing software, and the evaluation of partnerships with domestic chip companies for custom AI chips—these investments are all burning cash with no short-term returns. Z.AI is on a "full-stack self-research" path, building chips, software, and models in-house. This path is tough and expensive.

Z.AI’s real enemy is not Kimi. Kimi is just the first alarm bell. The real enemy is time: if 18 months pass and the 313 billion is gone while the story remains unfulfilled, where can Z.AI go for its next round of financing? At what valuation? Will the market still believe in the grand AGI narrative then? The beauty of the scarcity illusion is that when it exists, everyone believes it will last forever. When it disappears, no one remembers why they believed in it in the first place. Z.AI’s trillion-dollar market cap was ultimately just a capital market reaction to early-stage supply scarcity in the large model track. As open-source models become more powerful, as the "world number one" title changes owners every few months, and as customers can migrate between models at zero cost—how much is the identity of the "first major model stock" worth? The answer is being rewritten by the market. And the rewriting process is far from over.

Data Source Note: Financial data (revenue 724 million yuan, R&D spending 3.18 billion yuan, net loss 4.718 billion yuan, IPO net proceeds 48.96 billion Hong Kong dollars, used IPO proceeds 45.88 billion Hong Kong dollars, remaining cash 3.08 billion Hong Kong dollars, placement net proceeds 313.75 billion Hong Kong dollars, MaaS platform ARR 1.7 billion yuan) are sourced from Z.AI’s annual report, prospectus, and company announcements. Share price data (IPO price 116.1 Hong Kong dollars, intraday high 2980 Hong Kong dollars, July 17 closing price 1107 Hong Kong dollars, July 20 closing price about 890 Hong Kong dollars) are sourced from Hong Kong Stock Exchange market data. Kimi K3 parameters (2.8 trillion parameters, Code Arena score 1679) are sourced from the official release by Moonshot AI. Anthropic ARR of 47 billion US dollars and valuation of 965 billion US dollars are from public reports in June 2026. MiniMax financial data (revenue 79.038 million US dollars, net loss 1.872 billion US dollars) are sourced from MiniMax’s annual report. The four directions of Tang Jie's "Touch High" plan are sourced from the internal letter "The Great Wave Has Arrived."