Title: Hong Kong's AI Giants Are the Short-Seller's Mirror
Article:
The number is stark. Twenty percent of MiniMax's free float is now held as a short position. That is not a hedge. That is a thesis. In my years auditing protocol architectures and stress-testing value flows, I have learned to treat extreme positioning as a signal. It is the market's way of saying the current price is a liability, not an asset. When a stock's short ratio hits levels that usually precede delisting or insolvency, we are no longer analyzing a company. We are analyzing the structural integrity of an entire business model.
Over the past seven days, the narrative in Hong Kong has shifted from AI narrative premium to forensic accounting. Zhipu AI has seen its share price fall roughly 24% from a July peak. MiniMax has dropped 18%. Both have declined more than 50% from their all-time highs. This is not a correction. This is a de-rating. The market is re-pricing "pure play" large language model companies from a story of future dominance to a current liability.
The catalyst was the release of Kimi K3 by Moonshot AI. The market did not treat this as an incremental improvement. It treated it as a generational gap. Zero knowledge is a liability, not a virtue. The price action following Kimi K3 demonstrates that the market believes Moonshot has achieved a structural leap in model capability. This has rendered the offerings of Zhipu and MiniMax, if not obsolete, then at least in the second tier of a two-tier market.
As a core protocol developer, I have seen this pattern before. In 2017, I spent six weeks auditing the Golem Network's smart contract release. I found an integer overflow in the task distribution logic. The team missed it because they were deploying fast. Speed without structural integrity is just deferred debt.
The Kimi K3 release is a structural shock in the AI model market. When a competitor's launch causes a 20% plus drawdown in the stock price of its peers, it signals that the market is mapping technical capability directly to competitive position. It is not just a product. It is a verdict. In the same way that I documented a vulnerability in the interest rate adjustment function of Aave V1, the market has identified a vulnerability in the business model of the followers.
The data points are clear. Zhipu AI's GLM-5.3 has been positioned by Jefferies as offering "similar performance" at a 19% lower cost per task. That is not a competitive advantage. That is a survivorship strategy. The company is admitting that it cannot win on capability, so it will try to win on price. This is a "follower" strategy. It is a rational response to a competitive disadvantage, but it is also an admission that the technical leadership is not there.
MiniMax is in a worse position. Hedgeye analyst Nathan J. said the company is "neither the smartest nor the cheapest." That is a fatal position in any market. It is the "stuck in the middle" position that business strategists have warned about for decades. MiniMax cannot command a premium because it is not the smartest. It cannot win price-sensitive customers because it is not the cheapest. The short interest of 20% is the market's way of saying that the company has no structural integrity.
The Profitability Question
The market is asking a question that no amount of user growth can answer: Can a pure play LLM company be profitable? This is not a question about current earnings. It is a question about structural viability.
The short sellers are not just betting on bad earnings. They are betting on the model. They are betting that in a market where API prices are dropping and models are becoming commoditized, companies without a clear differentiation will be squeezed. The cost structure of a large language model company is dominated by compute. Training and inference costs account for more than 50% of operating expenses. When the price per API call drops, the model must be either much better or much cheaper to maintain margins. If you are neither, you are in the trap.
Zhipu AI's "cost advantage" is a temporary engineering optimization. It is not a structural advantage. I have seen this pattern in DeFi. In the summer of 2020, I spent 400 hours simulating flash loan attacks on Aave V1. I found a reentrancy edge case in the interest rate adjustment function. The flaw was in the assumptions about how value flows between pools. Zhipu's cost advantage is similar. It is an optimization, not a defense. It can be replicated by competitors. The 19% cost advantage will disappear as soon as Moonshot or others optimize their own inference stacks. The "performance similar" claim is also suspicious. Which benchmarks? Under which conditions? The selectivity of the metrics is a red flag. As I always say, the bug is always in the assumption.
The shorts are making a more significant bet. The financial data will be released soon. MiniMax's interim results are due August 26. Zhipu's are due August 31. The shorts are increasing their positions before the earnings date. That is not a hedge. That is a conviction. They believe the earnings will expose the weakness in the unit economics. They believe the market will finally see that revenue growth does not equal profit.
The Unlock and the Southbound Trap
The technical picture is made worse by the liquidity dynamics. The IPO lock-up period ended in July. The unlock is massive. Zhipu AI has 25.68 million shares unlocked. MiniMax has 150 million shares unlocked. At current prices, this represents a supply overhang of approximately $11.5 billion. This is an enormous overhang for any company, but especially for companies with a short ratio of 20%.
The supply pressure is the structural issue. Early investors are sitting on huge gains. Zhipu AI's stock is still 800% above its IPO price, despite the drawdown. This means that even after the 50% decline, the early investors can sell. They will sell. The "sell on the news" pressure is not a possibility; it is a mathematical certainty. The unlock is a supply event. The short is a supply event. They are not correlated.
The Southbound capital flows are an interesting counter-signal. Mainland investors have been buying through the Stock Connect. Zhipu's holdings are about 12% of the float. MiniMax is about 8.1%. This is often described as a "supportive" flow. I see it as a risk. It is a "value investing" or "buy the dip" approach that has not stopped the decline. The market is telling us that the seller is stronger than the buyer. The fundamentals are stronger than the flow. The Southbound capital is not a support. It is a sponge. It is absorbing the selling, but the seller has infinite supply.
The short ratio of 20% also creates a very specific risk: the short squeeze. If the earnings report is better than expected, the shorts will be forced to cover. This could lead to a sharp short-term rally. This is the symmetry of the short. But in my experience, in a market where the technical advantage is widening, the expectation of a positive surprise is low. The shorts are not positioned to get squeezed. They are positioned to be right.
The Narrative Gap
The key issue is the gap between the narrative and the numbers. The narrative of AI is one of infinite potential. The numbers of AI are the reality of high costs and low margins. The market is starting to price the numbers, not the story. This is a healthy adjustment. But it is painful for the companies involved.
The article from Beating AI has a low conflict of interest, but the information selection is biased towards the negative. The short data comes from S&P Global, which is a neutral source. But the selection of the short ratio as the headline, combined with the declining stock prices and the negative analyst comments, creates a bearish narrative. This is not necessarily wrong. It is a structural view. But it is not a balanced view.
As someone who has been in the technology sector for over 20 years, I have seen this cycle. I saw the dot-com crash. I saw the Terra/Luna collapse. The pattern is the same. A narrative that is not supported by unit economics will eventually be corrected by the market. In May 2022, I wrote a 15,000-word analysis of the TerraUSD anchor program. I proved that the incentive structure was mathematically unsustainable. The market did not care. The market cared about the crash. The market cares about the price.
The Infrastructure Level
There is a layer that is not discussed enough in the short-selling analysis: the infrastructure. The cost of computing power is the core variable in the profitability of the LLM companies. The "cost advantage" of Zhipu AI is likely to be the result of inference optimization, such as quantization, speculative decoding, and batch processing. These are engineering. They are not scientific breakthroughs. They can be copied.
The second issue is the US export controls. The Chinese AI companies are facing restrictions on the purchase of high-end GPUs. This is a structural constraint. It affects training efficiency and cost. It is a hidden risk factor. The shorts have not yet priced in the geopolitical risk. This is a potential downside.
The pricing of the AI model is a commodity. The model is becoming a commodity. The infrastructure is the moat. The companies that own the infrastructure, or that have the most efficient infrastructure, will survive. The companies that rely on the API of others will be squeezed. The pure play LLM companies are in the most difficult position. They have no distribution. They have no existing customer base. They have to build everything from scratch. The costs are high. The price is dropping. The short sellers are not just betting on the current earnings. They are betting on the lack of a structural path to profitability.
The Final Observation
Logic does not care about your narrative. The market is now in a phase of "profitability reality check." The valuations of the AI giants in Hong Kong are being tested. The test is not about the technology. The test is about the business.
Trust is a variable, not a constant. The market's trust in the "pure play" model is declining. The high short ratio is a warning. The price drop is a confirmation. The upcoming earnings are the verdict. Will the earnings show a path to profitability? Or will they show the structural weakness?
The shorts are betting on the latter. The supply is betting on the latter. The technical lead of Kimi K3 is betting on the latter. The only ones who are not betting on the latter are the retail investors and the Southbound funds. They are the counter-party. They are the exit liquidity.
As a protocol developer, I am looking for the bug in the assumption. The assumption is that a pure play LLM company can be a profitable public company. The bug is that in a commoditized market, the only path to profitability is either a structural cost advantage or a structural capability advantage. If you have neither, you are in a position. The short seller is betting that MiniMax and Zhipu AI will not find that path before the cash runs out.
The shorts are not the cause. They are the effect. The cause is the business model. The cause is the lack of differentiation. The cause is the reality of the market. The shorts are just the mirror. And the mirror is showing a picture that is not flattering. The question is not whether the shorts are right. The question is whether the company can prove them wrong. The clock is ticking. The earnings are coming. The answer is coming. The price will tell the truth.