The Short-Seller's Ledger: Why Hong Kong's AI Darling Stocks Are Bleeding

Let us examine the balance sheet of a narrative. The market has spoken, and it is not a whisper. It is a record. In Hong Kong, the AI darlings MiniMax and Zhipu AI are under siege. Short interest in MiniMax has hit a staggering 20%, a level not seen for tech IPOs of this magnitude. The data indicates that the market is not just cautious; it is aggressively betting on a decline. Risk is not optional. It is a variable. And for these two companies, the variable has been calculated. Shares have halved from their peaks. Lock-up expirations have unleashed a flood of paper—approximately $11.5 billion in combined stock. This is not a market correction. This is a repricing. The context is not merely a bearish call from a single analyst. It is a structural shift. The market is no longer paying for a story. It is demanding an equation. The narrative of 'pure-play' large language model companies is being audited. The bullish thesis was simple: these are the architects of the next technological era. The bearish reality is more complex: they are burning capital in a price war, without a clear path to profitability. The data indicates a fundamental shift. The market has moved from a "story-driven" to a "data-driven" evaluation. This is a direct consequence of the broader macro environment. With interest rates no longer at zero, investors are demanding return on capital, not just revenue growth. The "at any cost" growth phase is over. The new mandate is: "Show me the unit economics." The catalyst is clear. July's release of Kimi K3, a competing model from another lab, should have been a bullish signal for the sector. It wasn't. Instead, it triggered a sharp decline in both Zhipu and MiniMax—a drop of 24% and 18%, respectively. This is the market's verdict. The technology is not the differentiator; the cost of keeping up is the liability. "Volatility is the tax on uncertainty," and the uncertainty here is the AI profit curve. The core issue is a "middle-zone" trap. Zhipu's GLM-5.3 is competitive on performance, with a 19% cost advantage per task compared to Kimi K3. Yet, this technical edge has not translated into stock price support. The market is not rewarding efficiency; it is punishing the perception of a price war. This is the core of the bearish thesis. A 20% short interest is not a position. It is a consensus. It signals that sophisticated capital is convinced the share price has yet to reflect reality. They are betting on a de-rating. The price, despite being down 50%, still holds a significant premium over the IPO price. Zhipu is still up 800% from its issue price; MiniMax up 80%. The bears are not merely betting on a fall; they are betting on a convergence to a fundamental value, which they believe is far below current levels. The immediate supply shock is the lock-up expiry. The insiders are now free to sell. The market assumes they will. This is the principle of "Trust the contract, doubt the community." The contract states the shares are unlocked. The community expects the founders to hold. The smart money prices in the former, not the latter. "Liquidity vanishes; principles remain." But the principle of early investors is usually to cash out. The market's reaction to the K3 release is the key insight. It is not a verdict on K3's quality. It is a verdict on the cost of competition. For MiniMax and Zhipu, competing means spending more on R&D to keep up with the "Alibaba, ByteDance, and DeepSeek" of the world. This spending is a negative for their income statements. The market is not rewarding "performance parity"; it is punishing the capital expenditure required to get there. This is where the contrarian angle emerges. The most dangerous position is not to be a bear or a bull, but to be a "value buyer" in a falling knife. The Southbound capital flows (mainland Chinese investors buying HK-listed stocks) have been buying the dip. Zhipu's southbound ownership is 12%; MiniMax is 8.1%. They are providing "support," but the price has not stabilized. This is a classic "value trap" scenario. The narrative is "the long-term AI thesis is intact." The reality is that the quarterly numbers will not be able to prove that thesis. The logic is simple. The southbound flow is large, but it is not infinite. They are buying a thesis. The short sellers are buying a fact. The thesis is "AI is the future." The fact is "These companies are burning cash with no end in sight." The conflict is the market. When the facts are presented (the next earnings report), the thesis may be broken. The half-year results are the pivotal event. They are the "truth serum." MiniMax reports on August 26; Zhipu on August 31. This is the moment when the "data-driven" market gets its data. The analysts will be looking for three things: Revenue growth, gross margin, and the burn rate. A revenue beat will not be enough if the margins are shrinking. The market wants proof of a path to breakeven, not just a story of growth. My own framework from backtesting the 2024 ETF arbitrage taught me that the market is a discounting machine. It prices in the next two quarters. The current price of these stocks is a wager that the earnings report will be a disaster. If the report is merely "bad" (not catastrophic), we will see a relief rally. If it is "good" (with a path to profitability), the short squeeze will be violent. The bears are betting on "disaster." The buyers are betting on "bad but survivable." The fundamental issue is the lack of a moat. Zhipu has the cost advantage, but it is not a "cost advantage" in a commodity market. It is a cost advantage in a race to the bottom. When you have a parity in intelligence, price is the only differentiator. This is the "middle-zone" problem. You are not a leader, and you are not the cheapest. You are a follower. The market's verdict is clear: "Audit the code, not the hype." The hype was the Chinese AI sector. The code is the financial statement. The shorts are making a legal claim that the code is broken. The true risk is not the "company" but the "narrative." The risk that the entire "pure AI" business model is flawed. The risk is not just that these companies will fail. It is that the market will price them as "media" companies with a tech premium, not as "infrastructure" companies with a utility return. This is a derating. The takeaway is not a target price. The takeaway is a framework. The price action is the information. The market is telling you that the risk is not an AI regulation, but a P&L statement. The earnings report is the next black swan event. Be prepared. The market owes you nothing. The earnings report is the only ledger that matters. And that ledger is coming due.