kimi
HECTOR RETAMAL/Getty Images

On the evening of September 2, LatePost reported that Moonshot AI (Kimi) had confidentially filed an A1 document with the HKEX this week, formally starting the process for a Hong Kong IPO. An A1 is one of the formal application forms a company submits when applying to list on the HKEX. Kimi declined to comment, saying there was no information to disclose for now. Gelonghui's September 3 morning briefing included the same item with the phrase "reportedly."

The same report carried another number: the company is advancing a new funding round at a pre-money valuation of $50 billion, which LatePost judged is likely Kimi's final round before the IPO.

Put the two numbers side by side and the problem emerges. In mid-June, the STAR Market Daily reported that Kimi's annualized revenue (ARR) had surpassed $300 million; at the September 2 close, the total market capitalization of Zhipu, already listed in Hong Kong, was about HK$520 billion, and MiniMax's was about HK$120 billion. A company founded three years earlier and employing more than 300 people is preparing to use financials that have never been made public to command a valuation benchmark above both listed peers.

Which Calculation Produces the $50 Billion

Putting the three companies on the same scale makes the contrast most intuitive. Converting at roughly HK$7.8 to $1, Zhipu's HK$520 billion is about $66.7 billion, implying roughly 42 times its $1.6 billion ARR disclosed in August; MiniMax's about HK$120 billion is roughly $15.4 billion, implying about 19 times its $800 million ARR; Kimi's $50 billion pre-money valuation, against $300 million—the last publicly verifiable annualized-revenue figure—implies about 167 times. In other words, Zhipu's publicly disclosed annualized revenue is more than five times Kimi's, while its market cap is only about 1.3 times Kimi's target valuation.

The multiple cannot be treated as a conclusion on its own, because the denominator itself is not airtight. LatePost specifically cautioned in the report that ARR has two meanings: one is Annual Recurring Revenue, which mainly measures software subscriptions; the other is Annual Run Rate, a measure of annualized revenue scale that extrapolates a full year from recent revenue and can include subscription fees, API calls and even custom-project revenue. Neither Zhipu nor MiniMax has said which measure its ARR is closer to; the data Anthropic released this year uses the latter.

The numerator is also changing. On July 18, the day after K3 was released, Kimi President Yutong Zhang said that after the release, enterprise ARR had grown several-fold and that the company had recorded its largest single-day increase in history. If that account holds, the 167-times multiple will quickly become smaller, but outsiders cannot obtain new figures and can only accept one fact: Kimi has so far not disclosed any verifiable revenue, gross-margin or loss figures.

The primary-market pricing trajectory better shows where the support comes from. According to LatePost: at the end of 2025, the valuation was about $4.3 billion; in May 2026, it exceeded $20 billion; after K3 was released in July, another round was completed with a post-money valuation of $35 billion, rising in a little over half a year to about eight times the previous level; it is now advancing a pre-money valuation of about $50 billion. 36Kr, citing an evaluation by the third-party leaderboard Artificial Analysis, said K3 ranked third overall, behind only top models from Anthropic and OpenAI, and at lower cost. The product cadence has been K2.5 in January, K2.6 in April and K3 in July, almost once every three months. Zhenxin Huang, head of enterprise business, told Jiemian News in June that overseas paying users had grown fourfold, API revenue had risen 400%, and the product had entered more than 200 countries and regions, while the company remained a lean team of more than 300 people.

Technology rankings can be quantified by leaderboards; a $50 billion valuation cannot. Its only anchor at present is the degree to which investors believe that, after K3, call volume will continue to double.

Why Now, Not Next Year

The first reason is compute. LatePost put it bluntly: model companies, unlike traditional software companies, have high marginal costs for incremental revenue; stronger models bring more calls and rapidly consume more compute. After K3 was released, Moonshot AI briefly suspended new-user subscriptions, prioritizing existing paying users for compute. 36Kr added details: on July 19, it suspended new consumer-side subscriptions after request volume over 48 hours exceeded the cluster's limit; the suspension has since been lifted. That pause reads like a product strategy, but it is in fact a statement of funding needs—it sells, but there is no compute available to fulfill delivery. K3.1 is also due to be released soon; in late July, posts on social media claimed it is a minor iterative version of K3, mainly optimizing inference speed and Agent capabilities, but as of September 2 it had not been officially released.

The second reason is that the Hong Kong market has just validated for peers that listing equals a financing line. Zhipu completed a placing in July 2026, raising about HK$31.4 billion, with new shares accounting for about 4.25% of the enlarged total share capital; MiniMax completed a placing of 35.6 million shares plus HK$6.5 billion of convertible bonds. In the same period, Zhipu disclosed that, as of the end of June, it had used more than 93% of the net proceeds from its Hong Kong IPO. For a company whose revenue growth is tied to compute procurement, secondary-market placings and convertible-bond channels are closer to readily available than a single private round in the primary market.

The third reason is the window. At the end of 2025, founder Zhilin Yang's public stance was that there was no rush to IPO in the short term; 36Kr, citing an internal letter, said cash reserves exceeded RMB 10 billion after the Series C round. Between not rushing and confidentially filing the application lay exactly a jump in valuation from $4.3 billion to $35 billion.

What the Confidential Filing Withholds

The word "confidential" is easily misread. On September 2, Anji Microelectronics Technology, an A-share STAR Market company, announced that it had already submitted application materials for issuing H shares and listing with the HKEX in confidential form, while cautioning that it still needed to obtain filings, approvals or consents from the China Securities Regulatory Commission, Hong Kong's Securities and Futures Commission and the HKEX, among others, and that whether those would be obtained and when remained uncertain. This shows that confidential filing is a normal step permitted by Hong Kong's current system, not a backroom operation. As of September 3, no public filing for Moonshot AI could be found on HKEXnews, consistent with the nature of a confidential submission.

But for investors, the confidentiality period means a real information vacuum: what revenue recognition basis is used, the gross margin, the size of losses, how long cash can last, whether compute is self-built or rented, whether the company will use Chapter 18C for specialist technology companies, the sponsor and the intended amount to be raised—all must wait until public disclosure. The last time Kimi's financials were publicly discussed was a single line: ARR had surpassed $300 million.

This vacuum has also changed the cost of rumors. On August 3, market reports said Moonshot AI planned to file an application with the HKEX as early as that month and raise about $3 billion; the company responded that the information was untrue. On August 8, other media said it might file before September 30. The confidentiality mechanism removes the cycle of leaks and denials from public view, at the cost that outsiders find even the denials themselves difficult to verify.

On the revenue side, there is also a path that has yet to materialize. Cailian Press, citing sources, reported that Moonshot AI is negotiating revenue sharing with Microsoft, Google and Amazon over K3, initially seeking to take up to 30% of revenue from K3-related services, while industry expectations put the upper limit at about 20%. If an agreement is reached, it would be the first large-model revenue-sharing agreement between a Chinese AI company and U.S. cloud giants; all parties declined public comment.

What to Watch Next

The revenue basis in the first public document will matter more than the revenue figure itself. Zhipu's revenue in the first half of 2026 was RMB 954 million, up nearly 400% year on year; among that, open-platform and API revenue was RMB 825 million and its share rose from 15.2% to 86.5%, but adjusted net loss still reached RMB 1.964 billion. MiniMax's first-half revenue was $117 million, and its adjusted net loss was $293 million. The two earnings reports have already put the high-growth-plus-high-loss paradigm on the table; Kimi's prospectus will be only the third sample of it.

Second is how much of the use of proceeds goes to compute; this determines whether the market reads the raise as buying growth or sustaining delivery. Third is the share-price reactions of Zhipu and MiniMax: if the $50 billion pre-money valuation holds, the valuation anchors of the two listed companies will need to be recalculated; if it does not, what exits is a full round of primary-market pricing expectations. Fourth are comparables in the same track: LatePost noted that the market widely expects DeepSeek may also list in the first half of next year, while The Information said DeepSeek's actual revenue in the first seven months was about RMB 475 million, about 10 times its full-year 2025 figure.

What this deal really needs to answer is not whether $50 billion is expensive, but whether Chinese AI startups can form a second set of pricing rules in the secondary market. Zhipu and MiniMax have already shown that the market is willing to pay a premium of dozens of times revenue for the scaling of intelligence; but the condition for the premium to be honored has never been model rankings—it is the compute cost behind every yuan of revenue, the cost that once caused subscriptions to be suspended.

Reference Data Sources

  1. HKEXnews official homepage · Search for Moonshot AI public application documents found no record · 2026-09-03 · https://www.hkexnews.hk/index_s.htm
  2. Wall Street CN reprint of a LatePost exclusive · Moonshot AI confidentially files with HKEX to formally launch IPO · 2026-09-02 · https://wallstreetcn.com/articles/3780927
  3. Jiemian News · After interim earnings, are Zhipu and MiniMax's share prices reasonable? · 2026-09-02 · https://www.jiemian.com/article/15046549.html