The AI spending bull arrives at the lockup expiration

Sun, 21 Jun 2026 21:10:00 +0800

The awkward part of this round of AI stocks is that, while some people are shouting “capex bull,” many A-shares and Hong Kong stocks have not risen at all, and some have even dropped significantly. It looks like a contradiction, but it’s actually two sides of the same coin: money is still flooding into AI infrastructure and a handful of application gateways, but the secondary market is starting to ask who will ultimately be able to turn this spending into cash flow.

Let me directly answer your questions first.

This article can be read along three threads: North American tech giants aren’t without growth drivers—they’re betting on the next curve with heavy capital expenditure; coding is currently the easiest AI application to monetize; the key question for Zhipu and MiniMax isn’t whether they have a story, but whether their valuations and the supply unlocked by lock-up expirations can be absorbed by the market.

After the marginal efficiency of the old growth drivers at North American tech giants declines, AI has become one of the few directions that can still sustain a narrative of hundreds of billions of dollars in capital expenditure. Cloud, advertising, search, social media, and e-commerce are all still profitable, but if market caps are to continue being priced at high multiples, there must be a next curve large enough. What makes AI special is that it can be pitched both as a new product and as a reinvestment cycle for cloud computing, chips, data centers, electricity, and enterprise software.

“Capex Bull” makes sense. What is currently priced most richly is not necessarily the companies that have already earned AI net profits, but the companies standing at the front end of hyperscaler capital expenditure, collecting the money: GPUs, ASICs, HBM, optical modules, servers, switches, data centers, power equipment, cloud resources. The problem also lies here: if capex growth outpaces the realization of final revenue, stocks will first rise into a very narrow bull market, and later use pullbacks to filter out who has real cash flow.

Is the big tech industry running out of money? My take is: it’s not that they’ve become poor, but that AI investment is so massive that even companies with strong cash flow need to optimize their balance sheets. Goldman Sachs, JPMorgan, and other institutions, in their recent reports and media coverage, have estimated hyperscaler capex for 2026 at over $700 billion; Axios reported in mid-June that Nvidia has also entered a window of large-scale bond financing, driven by the growth of debt financing for AI data centers and supply chains. Having plenty of cash and continuing to issue debt are not mutually exclusive. It’s more of a signal: this cycle is not an asset-light internet bull market, but a heavy-asset balance sheet expansion.

Are investors anxious to cash out? In part, yes. IPOs, A+H listings, lock-up expirations on the secondary market, and M&A are all paths that turn paper valuations in the primary market into liquidity. This is especially true for listings like Zhipu and MiniMax, where the post-IPO float is extremely thin and stock prices have been pushed up by index inclusions and scarcity premiums. Before and after the lock-up expiration, the core issue is not simply “is the company good or not,” but rather that an originally very small tradable supply suddenly becomes much larger, and the market must reprice the liquidity premium.

Why has the coding track become white-hot? The reason is straightforward: it offers a faster path to paid willingness than “general AI changing the world.” Developers are willing to pay monthly fees for efficiency gains, while enterprises are willing to purchase seats for code generation, testing, migration, and security review. Anthropic’s earlier official fundraising materials revealed that Claude Code’s run-rate revenue exceeds $2.5 billion, and OpenAI has also turned Codex into an enterprise-grade coding agent, emphasizing in the 2026 Gartner Magic Quadrant for Enterprise AI Coding Agents that more than 4 million people use Codex weekly. Going forward, competition will not only be about model benchmark scores, but also about repository context, permission control, CI/CD integration, enterprise compliance, inference cost, and stability.

Is Zhipu’s Hong Kong stock market capitalization reasonable? Based on the closing price of 2513.HK at HK$2,094 returned by Yahoo Finance on June 18, 2026 and a total share capital of 440,230,190 shares after listing, Zhipu’s total market capitalization is approximately HK$922 billion.¹ The company’s revenue in 2025 was RMB 724.3 million. Even without precise currency conversion, this represents a price-to-sales ratio on the order of one thousand times. It has gone beyond the realm of “somewhat expensive” and can only be explained by very distant future revenue and profits. Unless you believe Zhipu will transform from a model company into China’s AI infrastructure platform, with future revenue increasing by two orders of magnitude, the current market capitalization is difficult to support based on present financial statements.

First, compress the financing data into a narrow table:

Entity	Capital Action	Amount / Valuation
OpenAI	Official new funding round	$122B committed capital; $852B post-money
Anthropic	Series H	$65B; $965B post-money
SpaceX / Cursor	Reported stock acquisition	~$60B transaction scale
Nvidia / AI Bond Financing	Reported large bond issuance	~$20B–$25B
Zhipu (智谱)	Hong Kong IPO, advancing STAR Market proposal	IPO raised ~HK$4.3B; market cap ~HK$922B
MiniMax	Hong Kong IPO, initiated A-share tutoring	IPO net proceeds ~HK$4.596B; market cap ~HK$154B

My take is set apart from the table. OpenAI and Anthropic are no longer ordinary SaaS financings, but rather infrastructure and gateway financings; Cursor has been pushed to a $60 billion deal benchmark, indicating that programming agents have shifted from a tool competition to a strategic gateway competition; core supply-chain companies like Nvidia are also starting to extend their capital duration, suggesting that AI infrastructure financing is increasingly resembling a capital markets engineering project.

Zhipu and MiniMax represent a different problem: the secondary market has assigned them a high forward platform premium, but they still have to contend with losses, secondary financing, expansion of the float, and absorbing the overhang from lock-up expirations. The meaning of this table is not “all AI companies are safe,” but rather: financing has become part of the business model. The larger the capital, the more the market will ask two questions: whether more money can still come in, and when revenue will prove that these expenditures are not sunk costs.

Programming First Becomes a Cash Flow Entry Point

Before, everyone wanted to build general-purpose AI, because the general-purpose AI story was the biggest. Then they realized: the biggest story does not equal the fastest return. C-end chat products have traffic, but paid conversion rates, retention, computing costs, and competitive subsidies are all hard to look at; enterprise general assistants can talk about efficiency, but the delivery cycles and internal compliance are heavy; programming is different — it stands directly on the software production chain.

A developer paying an extra $20, $100, or $200 per month is easy to convince, as long as it saves them a few hours. The same applies to enterprises. What truly delivers real-world value is not the slogan of “replacing programmers,” but rather fixing bugs, migrating frameworks, writing tests, reviewing code, doing refactoring, integrating internal knowledge bases, and opening PRs. These involve clear objects of work, clear before-and-after comparisons, and a clear paying audience.

So the competition between OpenAI and Anthropic will increasingly resemble a three-layered battlefield.

The first layer is model capability. Whoever can more stably understand large repositories, long contexts, and multi-file changes will have the advantage.

The second layer is the engineering shell. Tools like Codex, Claude Code, and Cursor are essentially not just model invocations—they also have to handle permissions, terminals, sandboxes, Git, testing, logs, rollbacks, and human review. If the model is strong but the shell is poor, developers will quickly get frustrated.

The third layer is enterprise distribution. The real money lies in team seats, private code security, audit logs, IDE and CI/CD integration, and admin consoles. In the future, many companies won’t just buy a single chatbot; they’ll procure AI coding agents as R&D infrastructure.

Things are very likely heading toward consolidation along this line. Standalone tools that lack their own models, compute, or deep distribution channels will easily get squeezed from both sides: upstream model providers raising prices or throttling access, and downstream big tech players bundling features directly into IDEs, cloud platforms, and office suites. The rumor of a SpaceX deal pushing Cursor’s valuation to around $60 billion is a clear sign that coding agents are now viewed as a strategic entry point, not just small utilities.

Where Zhipu’s Market Value Comes From

Zhipu’s problem is not “whether it has an AI story.” Of course it does. Its prospectus and annual results have shown that it has models, enterprise customers, R&D investment, and a capital market identity. The issue is that the secondary market has priced it too far ahead.

By the broadest measure:

\[ \text{Crude P/S Ratio} \approx \frac{\text{Total Market Cap}}{\text{Annual Revenue}} \]

On June 18, the closing price of HK$2,094 multiplied by 440,230,190 shares gives a total market capitalization of approximately HK$922 billion. Revenue in 2025 was RMB 724.3 million. Even without getting bogged down in the HK dollar and RMB conversion, this ratio has already exceeded the explanatory range that ordinary software stocks, cloud companies, and most growth stocks can bear.

To argue that it is reasonable, one must simultaneously believe several things: Zhipu’s future revenue will not be in the billions, but in the tens or hundreds of billions; the gross margin can be squeezed out from model inference and delivery costs; the ratio of R&D investment to revenue will drop significantly; and domestic government and enterprise clients, internet companies, end users, and the developer ecosystem will be willing to entrust their budgets to it over the long term.

These things are not necessarily impossible, but this is no longer “looking at the 2025 annual report to buy the company”—it is: pricing today based on a platform position that might emerge after 2030. When the float is very small, stocks like these can rise dramatically, but once lock-up restrictions are lifted, the market will ask again: are those willing to take the shares betting on future cash flow, or are they simply buying scarce liquidity?

MiniMax’s valuation is similarly not cheap. Based on a rough calculation using the closing price of 0100.HK at HK$497.6 on June 18 and total shares outstanding of 309,255,668, the total market capitalization is approximately HK$154 billion. The company’s 2025 revenue was US$79.038 million, which converts to roughly HK$600 million. The price-to-sales ratio is likewise in the hundreds range. The only difference is that Zhipu’s latest market cap is even more exaggerated, while MiniMax’s supply shock in July was greater.

What Really Changed After the July Lift

The unlock table needs to be viewed from two perspectives: the total share capital basis and the tradable share basis. The former shows potential dilution and the supply from existing shareholders, while the latter shows how many shares the secondary market can absorb on that day. To avoid crowding a wide table together, only the changes in the circulating shares are listed here:

Company	Float Before Unlock	Float After July
Zhipu `2513.HK`	~11.74 million shares	~37.42 million shares
MiniMax `0100.HK`	~12.69 million shares	Up to ~166 million shares²

Zhipu’s total share capital after listing is 440,230,190 shares. On July 7, the lock-up period on 25,681,600 shares held by cornerstone investors will expire, accounting for approximately 5.83% of the total share capital. This will increase the theoretical tradable supply from approximately 11.74 million shares to approximately 37.42 million shares, about 3.2 times the original amount.

After its IPO, MiniMax has a total share capital of 309,255,668 shares. On July 8, 16,504,040 shares held by cornerstone investors become unlocked; the upper limit of shares released from lock-up by existing shareholders is approximately 136,997,480 shares. Together, they account for about 49.63% of the total share capital, and under the upper-limit calculation, the tradable supply jumps from approximately 12.69 million shares to around 166 million shares.

The most easily misread entry in this table is MiniMax. In its allotment results announcement, the lock-up commitments shown for multiple tranches of Class A ordinary shares held by pre-IPO shareholders are set to expire on July 8, 2026; summing them item by item from the announcement gives roughly 137 million shares, consistent with the market-reported figure of “approximately 44.29% of total shares.” Combined with the 16,504,000 shares held by cornerstone investors, this yields a theoretical supply ceiling of nearly half of the total share capital.

But the key point here is: the lifting of commitments does not mean everything can be sold that same day. A footnote in the MiniMax announcement also mentions that some Pathfinder SII shares are subject to longer lock-up periods under Rule 18C; another batch of existing shareholder shares is locked up until the earlier of “the 20th trading day after inclusion in Stock Connect” and “nine months after listing.” In other words, the actual selling pressure depends on whether multiple lock-ups overlap, the nature of the shareholders, the progress of Stock Connect inclusion, and the secondary market price. The article uses the upper-bound figure to give readers a sense of the magnitude of the supply shock — it does not mean all of these shares will be sold on the same day.

Zhipu’s July unlock is much smaller. The cornerstone investors’ 25.6816 million shares will be unlocked, expanding the already thin float, but the truly large-scale lock-up period for existing shareholders falls on January 7, 2027. Zhipu’s allotment results announcement states this clearly: under Rule 18C, there is a six-month requirement, but all existing shareholders are also subject to the applicable Chinese laws prohibiting the disposal of their holdings within 12 months after listing. In other words, July is the first float expansion for Zhipu, while next January is the bigger supply test.

List the large-scale time points separately:

Zhipu, July 7, 2026: Cornerstone investors’ lock-up expires on 25,681,600 shares, approximately 5.83% of total share capital, expanding the free float from extremely low levels.
Zhipu, January 7, 2027: Existing shareholders’ 12-month lock-up expires; this includes approximately 178 million previously listed H-shares, with additional unlisted domestic shares subject to conversion and full circulation arrangements. This represents a larger potential supply window.
MiniMax, July 8, 2026: Cornerstone investors hold 16,504,040 shares, plus existing shareholders’ release commitment cap of approximately 136,997,480 shares. July is the main impact window, but overlapping lock-ups must be deducted.
MiniMax, the earlier of the 20th trading day after inclusion in Stock Connect or October 8, 2026: Another batch of existing shareholders’ shares of approximately 56.95 million shares. Whether this occurs earlier than October depends on Stock Connect trading availability.
MiniMax, January 8, 2028: Controlling shareholders’ 24-month lock-up expires, approximately 79 million shares. This corresponds to founder and control-related holdings, with a more distant timeline.

So, if the question is who is under more pressure in July, the answer is straightforward: MiniMax. Zhipu will also expand its circulation in July, but its real test comes later. MiniMax’s problem is that the theoretical supply in July is too large. If the stock price is still in a high valuation range, early shareholders don’t need to sell all of their holdings — the market will first price in the risk that they “might sell.”

Why Have Many A-Shares and Hong Kong Stocks Not Risen

This stock market rally is not a broad-based bull market. The situation is particularly evident in A-shares and Hong Kong stocks, because many companies have not directly benefited from AI capex, and may even be weighed down by high interest rates, weak consumption, the real estate chain, local government finances, export expectations, and domestic demand expectations.

The funding path for the AI capex bull market is narrow: first buy the American large caps and the chip chain, then expand into electricity, data centers, optical communications, storage, and a few application entry points. Among A-shares and Hong Kong stocks, there are many companies with AI in their names, but there aren’t that many that can truly turn orders into revenue, revenue into gross profit, and gross profit into cash flow. Once the market shifts from “having an AI concept” to “whether it can deliver,” most stocks will be left behind.

Hong Kong stocks have an additional issue: when many new-economy companies go public, the tradable float offered is quite small, and short-term price gains can be pushed significantly higher by scarcity and index-related buying. Once lockups expire, the market finally confronts the real supply. Zhipu and MiniMax are a microcosm of this logic: inclusion in the Hang Seng Tech Index, expectations for Stock Connect eligibility, and their status as scarce AI targets all attract buying; but lockup expirations, losses, secondary financings, and elevated price-to-sales ratios cause that buying to start hesitating.

My overall assessment is: AI is not without growth drivers, nor is AI purely a bubble; however, if this bull run is primarily driven by spending, it will naturally lead to divergence. Companies that sell the picks and shovels benefit first, products that can turn AI into a high-frequency paid workflow benefit next, and companies that only talk about long-term general-purpose AI without visible cash flow will find it increasingly harder to justify their valuations.

The programming track will continue to see fierce competition going forward, because it was the earliest direction that proved “AI can make money by seat, by usage, and by team budget.” OpenAI, Anthropic, Cursor, as well as the major players’ IDE/cloud platforms will all be crowding into this space. It won’t stop at “writing a few lines of code,” but will dig into the software delivery process: requirements, code, testing, review, deployment, security, and operations can all become billable points for agents.

But the stock market won’t wait for all the answers to become clear. It will first give high valuations to the companies that most resemble infrastructure, and then gradually test them through lock-up expirations, financing, debt, quarterly reports, and pullbacks. The question now is not “does AI still have a future,” but rather: how much future has already been priced into these valuations.

This article only provides a compilation of publicly available materials and personal analysis, and does not constitute any investment advice. Market conditions, market capitalization, and lock-up release figures may change with announcements and market developments. Specific trading decisions should still be based on officially disclosed documents and your own risk constraints.

References

OpenAI: OpenAI raises $122 billion to accelerate the next phase of AI
Anthropic: Anthropic raises $65B in Series H funding at $965B post-money valuation
Anthropic: confidentially submits draft S-1 to the SEC
OpenAI: named a Leader in enterprise coding agents by Gartner
OpenAI: Codex for every role, tool, and workflow
Axios: SpaceX will buy Cursor for $60 billion
Axios: AI debt boom ramps up with Nvidia bond sale
Business Insider: Goldman Sachs on AI capex supercycle
MiniMax Group Inc. Prospectus, HKEX filing
MiniMax Group Inc. Allotment Results, HKEX filing
MiniMax Group Inc. 2025 Annual Results Announcement, HKEX filing
Beijing Zhipu Huazhang Technology Co., Ltd. Prospectus, HKEX filing
Beijing Zhipu Huazhang Technology Co., Ltd. Allotment Results, HKEX filing
Beijing Zhipu Huazhang Technology Co., Ltd. 2025 Annual Results Announcement, HKEX filing
Yahoo Finance: Historical quotes for 0100.HK, 2513.HK, ^HSI, retrieved on 2026-06-21; 0100.HK only returned one record on 2026-06-18 this time, so no complete historical series was expanded based on this.

写作附记

Original Prompt

$blog-writer I actually haven’t quite figured it out—have the major North American tech companies all run out of new growth drivers? They’re all rushing into AI, and some say this bull market is a “capex bull,” driven by investments from the big players. Analyze the funding data and the latest funding news, organize them into a table, and share your perspective on this funding data. Are the big companies themselves short on cash? Are investors also anxious to cash out? Earlier, everyone wanted to build general-purpose AI, but later they realized it still needs to make money. The coding track is profitable—OpenAI and Claude are competing, and it’s already turning white-hot. Where will things go from here? Is the market cap of the Hong Kong-listed Zhipu reasonable? In July, both MiniMax and Zhipu had a wave of lock-up expirations. Analyze the changes in floating share data before and after the lock-up expirations—when is the large-scale lock-up expiration, and what are the corresponding data changes. In this stock market cycle, many stocks haven’t risen, and some have even dropped significantly—this is especially evident in A-shares and Hong Kong stocks. As a professional financial analyst, please answer my questions first.

Writing Approach Summary

This piece retains the core questions from the original prompt: Kai-fu Lee’s expenses, funding news, the programming track, Zhipu valuation, MiniMax and Zhipu lock-up expirations, and the A/H stock divergence. What was suppressed are unverifiable absolute statements, such as “the major companies can’t find any growth points” and “investors must be anxious to cash out.”

The body of the text divides the materials into three layers: the company’s official financing and performance are first-hand facts; capex, bonds, and the Cursor transaction use recent media reports as the latest background; market cap, free float, and the unlocking schedule are rough estimates based on HKEX disclosure documents and the market data as of 2026-06-18. MiniMax’s July unlocking is specifically marked with “ceiling caliber” and “overlapping lock-up” to avoid directly writing theoretical supply as same-day selling pressure.

The Hong Kong stock price and rough market capitalization in this article are calculated using the closing price on 2026-06-18; 0100.HK only returned one quote for 2026-06-18 from Yahoo Finance this time, so the full historical price trend of MiniMax is not expanded accordingly. ↩︎
This refers to the theoretical upper-bound calculation. Certain MiniMax shareholders, including Pathfinder SII and others, may still be subject to Rule 18C, Stock Connect trading hours, or other lock-up arrangements. The lifting of the lock-up commitment does not mean that all shares will be sold on the same day. ↩︎

Zhipu on Uncle Xiang's Notebook