Hong Kong's AI Giants Are Bleeding. The Shorts Smell Blood.

The trap isn't a falling stock price. The trap is the belief that a falling stock price is just a buying opportunity. Over the past month, Hong Kong's pure-play AI model companies have been hit by a wave of short selling that signals something far more systemic than a simple sector pullback. MiniMax's short interest has hit a staggering 20% of float. Zhipu AI sits at around 6%. This isn't fear. This is forensics. For over two decades, I have watched capital markets digest technological narratives. From the ICO madness of 2017 to the DeFi liquidity traps of 2020, the one constant is that the market eventually separates the signal from the noise. The current situation in Hong Kong's AI sector is a textbook case of narrative rejection. The market is not just doubting a few companies. It is doubting the entire business model of the 'pure large model' entity. The catalyst for this latest bloodletting was the release of Kimi K3 by Moonshot AI in July. The immediate market reaction was brutal. ZhipuAI dropped 24%. MiniMax fell 18%. These are not minor corrections. They are valuation resets triggered by a single product release. In my experience auditing tokenomics and tech valuations, a price reaction of this magnitude suggests the market views Kimi K3 not as an incremental improvement, but as a generation leap. This brings us to the core issue. The era of competing purely on model capability is over. We have entered a dual-vector war: capability versus cost efficiency. ZhipuAI's response to Kimi K3 was not to claim superior intelligence. Instead, they highlighted that their GLM-5.3 model offers similar performance at a 19% lower cost per task. This is a defensive move, a classic 'follower' strategy. It admits that on pure capability, they are trailing the front-runner. They are attempting to buy a competitive moat through engineering efficiency rather than foundational research. The problem? Cost optimization is an engineering problem, not a scientific breakthrough. It involves techniques like quantization, speculative sampling, and batch optimization. These are the low-hanging fruit of inference costs. Once Moonshot AI deploys the same optimizations, that 19% advantage evaporates. It is not a durable source of alpha. MiniMax's position is even more precarious. The analysts at Hedgeye summed it up with a brutality that resonates with my own 2020 research on DeFi yield traps: they are neither the smartest nor the cheapest. This is the dead zone. In a market where the technology is rapidly commoditizing, you must have a clear differentiation. If you cannot command a premium for superior intelligence, and you cannot win on price, you are stuck in the middle. You have no pricing power, and you have no margin protection. Chaos is just data that hasn't been processed yet. Let's process the data points that tell the real story. First, the shorts are not acting on rumor. They are positioning ahead of the interim earnings reports. MiniMax reports on August 26th. ZhipuAI reports on August 31st. The short sellers are placing their bets on the expectation that the financials will expose the broken unit economics. The market expects these companies to show massive operating losses, with revenue growth failing to outpace the cost of compute and R&D. Second, the 'southbound' capital is trying to catch a falling knife. Mainland investors have been buying the dip via the Stock Connect. ZhipuAI's holding is around 12%, MiniMax's around 8%. But this buying has not stabilized the price. This tells me that the institutional selling pressure is overwhelming the retail flow. The mainland money is acting as a 'reverse indicator'—a signal that the correction may have further to run. Third, the lockup expiration. Both companies had significant share unlock in July, coinciding with the Kimi K3 panic. ZhipuAI has 25.68 million shares and MiniMax has 150 million shares unlocked. This is a massive supply overhang. Early investors are sitting on enormous gains—ZhipuAI is still up 800% from its IPO price, despite being 50% off its highs. The incentive to sell and lock in profits is huge. The market is forcing us to confront a question that has been avoided for years. Can a pure AI model company ever be a good business? This is not about technology. It is about structural economics. The input costs are skyrocketing, driven by GPU scarcity and electricity. The output prices are collapsing due to the price war. As a macro analyst, I see this as a margin squeeze that no amount of volume can fix. The shorts are providing the evidence. The market is liquidating the 'AI narrative premium'. We are seeing a repricing from 'hope' to 'earnings'. This is the 'Real World' accounting. But I see a blind spot in the bearish thesis. The market is treating the cost advantage as permanent and the capability gap as insurmountable. It is not. The 20% short interest on MiniMax is an extreme position. If the earnings report shows any sign of traction, even a hint that the burn rate is slowing, we could see a vicious squeeze. The shorts are betting on a total collapse. The risk is that they are only betting on a minor disappointment. My takeaway is not about predicting the direction of these stocks. The trap isn't a falling market. The trap is the illusion of infinite growth. It is the assumption that revenue growth will eventually translate into profit. In the AI race, this assumption is now being stress-tested. The next few weeks will not just tell us the fate of these two companies. They will tell us the fate of an entire business model. Watch the reaction to the earnings. Watch the short interest changes. The market is not just pricing in a bad quarter; it is pricing in a paradigm shift.