We assume that a falling stock price is merely a reflection of financial performance, a lagging indicator that catches up to the numbers after the fact. But the recent trading action around Hong Kong's AI darlings suggests something more unsettling: the market is pricing in a failure of narrative before the financials have even been published. When short interest on MiniMax hit a staggering 20% of free float, and Zhipu AI saw its bearish bets climb to 6%, the signal was not just about projected losses. It was a collective judgment that the fundamental value proposition of the "pure AI model company" is broken. Truth is not what is seen, but what is trusted—and the trust, for now, is gone.
The context is a peculiar confluence of events that has turned a textbook bull market narrative on its head. In July, the release of Kimi K3 by Moonshot AI triggered a seismic shockwave through the sector, wiping out roughly 24% of Zhipu AI's market value and 18% of MiniMax's in a matter of days. This was not a subtle recalibration. It was a violent repricing based on a perceived generational leap in model capability. As a protocol PM who has spent years watching value accrue to networks based on their technical edge, I found this reaction oddly validating: the market still treats model intelligence as the primary driver of competitive moats. But the reaction also revealed a deeper, more uncomfortable truth that the short sellers have now capitalized on. The era of "capability at any cost" has ended, replaced by a brutal, unforgiving "capability-to-cost" duopoly.
The core of the bearish thesis is not just that these companies are unprofitable—that was priced in long ago. The problem is that they have no defensible position in the value chain. Jefferies' assessment of Zhipu's GLM-5.3 as offering "similar performance at 19% lower cost" sounds like a strategic win, but it is a confession. It admits that on a pure capability basis, Zhipu cannot compete head-to-head with the frontier models. The strategy pivots to efficiency, to engineering optimization—quantization, speculative sampling, and batch inference tricks that my own audit work tells me are replicable within a single development cycle. It is a software moat, not a structural one. It buys time, not a future. I have seen this movie before in the Layer 2 wars, where teams claimed "cheaper transactions" as a feature, only to be undercut by a competitor with a superior architecture that made the optimization irrelevant.
MiniMax's situation is even more precarious, and Hedgeye's brutal epithet—"neither the smartest nor the cheapest"—captures the essence of its strategic trap. It is a "stuck in the middle" scenario, a position I have long considered to be the most dangerous in any competitive landscape. You cannot command a premium for a product that is not best-in-class, and you cannot win a race to the bottom when your cost structure is not the lowest. The market has correctly identified that this "middle ground" is not a plateau; it is a slide. The short sellers are not gambling on a quarterly miss; they are betting that these companies cannot find a reason to exist in their current form.
This is where the contrarian angle emerges, and it is not one that is simply bullish. The narrative spun by the short sellers is that these companies are value traps with massive overhangs—the 1.5 billion shares of MiniMax and 25.68 million of Zhipu that came off lockup in July represent roughly $11.5 billion in potential supply. The stock has already fallen 50% from peaks, yet Zhipu still trades 800% above its IPO price. On the surface, this is a mathematical nightmare. But it ignores a critical dynamic: the symmetric risk of a high short float. A short ratio above 15% is not just a bearish indicator; it is a powder keg. If the upcoming interim earnings reports—scheduled for August 26 and August 31—show any signs of a rational path to profitability, or even a stabilization of gross margins, the resulting short squeeze could be violent. The bears are comfortable in their conviction, but they have forgotten that markets are not a one-way street. The "sure thing" has a way of reversing with brutal efficiency.
Furthermore, the assumption that the $115 billion lockup expiry creates a linear selling pressure is naive. It assumes that early investors have a diversified, rational strategy. But many of these investors are strategic, not financial. They are aligned with the long-term mission of the company, not a quick exit. The 12% stake held by Southbound capital in Zhipu is a reflection of strategic national interest in AI sovereignty, not a hedge fund rotation. These are not weak hands. They are long-term holders who understand that the current price levels are a reflection of sentiment, not utility. The institutions are learning to speak in hash rates, and their commitment is not to a quarterly P&L, but to a decade-long infrastructure build-out.
The technical analysis of the broader industry tells a similar story of premature judgment. The market is treating Kimi K3's success as a terminal condition for its competitors, extrapolating a single model release into a permanent generational gap. But based on my experience auditing failed protocols during the 2022 collapse, I know that a temporary technical lead is often a liability, not an asset. It creates complacency. The real question is not who is ahead today, but who has the architecture and the cost structure to iterate fastest when the next paradigm shift—be it in test-time compute or agentic workflows—renders current advantages obsolete. Moonshot's lead may be real, but it is not structurally unassailable. The shorts are betting that MiniMax and Zhipu cannot catch up, but they are ignoring the fact that the front-runner is also carrying the heaviest weight in terms of maintaining its lead.
The ultimate takeaway from this short-selling wave is not a prediction of bankruptcy, but a demand for a new definition of value. The market has moved from a phase of "narrative valuation" to "fundamental scrutiny." The next twelve months will be a Darwinian filter, not just for these two companies, but for the entire class of pure-play model providers. The ones who survive will not be those with the best benchmarks, but those who can translate intelligence into a defensible, high-margin, mission-critical service. The ones who fail will be those who cannot answer the question: What is the trust that your token—or your model—actually represents? For the shorts, the answer is "nothing." The price action will soon tell us if they are right, or if they have underestimated the stubborn resilience of those who are building the next constitution, one layer at a time.