The market is delivering a verdict. It’s not whispered in analyst notes or opined on Twitter. It’s written in the immutable ledger of the stock exchange. MiniMax and Zhipu AI, the two flagship pure-play AI model companies that IPO'd in Hong Kong this summer, are currently the subject of one of the most aggressive short-selling campaigns in the city's history. The data is stark: MiniMax's short interest sits at an eye-watering 20% of float. Their stock prices have been halved from post-IPO peaks, shedding billions in market capitalization in a matter of weeks.
Let’s follow the liquidity, not the narrative. The narrative on the streets is about overvaluation and the AI bubble. That's noise. The data is about capital flow mechanics, structural supply, and the monetization of hype. I have spent the last eighteen years tracking the collision between on-chain truth and market fiction, and while this isn't a crypto-native event, the forensic principles apply. When you see an anomaly like a 20% short interest, you don't just ask 'why now?' You trace the transaction. You look for the catalysts—the lock-up expirations, the earning reports, the fund flows. The recent price action shows a breakdown from a fragile range, but the real story is the mechanics of the attack.
We are witnessing a fundamental repricing. The market has shifted from a narrative-driven model, where the promise of AGI justified any price, to a forensic, earnings-driven model. The shorts are not just betting on a price drop; they are betting on a business model failure. They are looking at the balance sheet of the pure-play AI enterprise and finding it structurally flawed. The pattern of this bear raid is remarkably similar to the late-stage ICO sinkholes I audited in 2017: a project with a strong narrative, a massive token unlock schedule, and a fundamental inability to generate cash flow that matches its burn rate. Hashes don’t lie. Wallets do. And here, the wallets are showing the accumulation of a major short position, the timing of which is meticulous.
The Context: The Chinese AI Race and the Hong Kong Listing Route
The two companies at the center of this storm are MiniMax and Zhipu AI, two of the most prominent members of the 'AI Tiger Cub' group in China, or as they are often called, the 'Four Little Dragons' of AI. They are not legacy tech giants; they are pure-play AI model developers, which means their entire corporate valuation is hinged on the performance and monetization of their large language models (LLMs). For years, they have been funded by a combination of government-backed funds, sovereign wealth, and aggressive tech venture capital. Their primary assets are their engineering talent, their proprietary training data, and their models' capabilities, which they monetize through API access, cloud services, and licensing to downstream enterprises.
The journey to the Hong Kong Stock Exchange was seen as the culmination of a massive bull run in AI infrastructure. The IPO prices were not just based on earnings; they were based on a 'Total Addressable Market' (TAM) narrative that suggested every enterprise in the world would soon be paying for model inference. Zhipu AI, backed by a strong network of Chinese tech and government investors, and MiniMax, with its deep pocket Silicon Valley and Chinese backers, were the poster children for this shift. Their July listings were a major event for the Hong Kong market, seen as a vote of confidence for the region's ability to attract pure-play tech listings.
However, the market has a short memory. The very forces that propelled these stocks—the AI hype, the massive secondary market deals—are now the ones being used against them. The lock-up periods, typically 6 to 12 months from listing, are beginning to expire. This is the first major technical catalyst. The data on these unlocks is not hidden; it's in the IPO prospectus, a public ledger. But the market is ignoring the sum. For Zhipu, the July lock-up expiration released 25.68 million shares into the open market. For MiniMax, it was a staggering 150 million shares. At the time of writing, the combined value of these newly tradable shares is roughly $11.5 billion. That is not a 'float', that is a supply shock.
The Core: The Forensic Evidence of the Short Thesis
Now, let's trace the flow of this short thesis. The primary short thesis isn't just about the unlock. It's about the 'economic model' of the pure-play model company. In my 'Pre-Mortem' framework, I look for the structural reason why a company might fail. Here, the primary evidence is the breakdown of the 'narrative premium' that was baked into the IPO.
First, the technical edge is vanishing. The market no longer believes in a 'model architecture' moat. Jefferies, a major financial brokerage, released a research note on Zhipu AI's GLM-5.3 model. Their analysis concluded that it performed on par with the Kimi K3, a rival model, but with a 19% lower cost per task. In a vacuum, this is a positive. Zhipu is more efficient. But in the current market context, it was read as a warning. The market did not bid the stock up on this 'good news'. Instead, it seems to have triggered a 'sell the news' event. Why? Because it signals that the technology is not diverging. It's converging. When models are at parity, the only differentiators are price and brand. And price is a race to the bottom.
The data from Jefferies is a key clue, but you have to read the derivative. The market is not rewarding efficiency anymore. It's punishing the 'cost of capital' needed to achieve it. The recent price action on Zhipu AI, which fell roughly 24% in the days following the Kimi K3 release, while MiniMax dropped 18%, is a case study in negative information asymmetry. The market doesn't care that the model is good. The market cares that the model is a commodity, and commodity producers are price takers. This is the core of the bear thesis.
Second, the short-seller, Hedgeye, has a public track record of being early and right on some major tech shorts. They are not just analyzing a balance sheet; they are analyzing the specific 'fragmentation' of the AI value stack. Their contention is that Zhipu AI is under severe price pressure, limiting its ability to raise prices and improve margins. They argue that the pure-play model has no pricing power because the enterprise customer knows that switching costs are low—one API is very similar to another API. So the 'institutional flow' of the market is moving away from these pure plays and toward the infrastructure providers (the NVIDIAs of the world) and the application layers that are closer to the end-user and the revenue.
Third, let's dissect the on-chain data, so to speak. The 'smart money' flow is clear: The Southbound Stock Connect flow is the key signal. While it's often a narrative of 'Chinese retail is buying', the data shows that even with sustained buying from mainland investors (Zhipu is about 12%, MiniMax is 8.1% as of the latest data), the stock price is still falling. This tells me that the 'smart money' from the short side is not just hedging; they are using the liquidity provided by these retail buyers to exit and press their short position. The price is being supported by 'tender' demand, but the smart money is executing a classic "distribution" strategy. The Chinese buyers are seeing the 'low price' as a 'value buy' based on the long-term potential of AI, but the short-sellers are seeing the 'low price' as a 'starting point'.
The Contrarian Angle: The 'Value' Trap of Southbound Capital and the Misread of Efficiency
The contrarian angle here, the blind spot in the bear thesis, is the assumption that this is a zero-sum game where pure-play models are necessarily the losers. The 'narrative' is that they are 'stuck in the middle'—not as smart as the frontier labs, not as cheap as the 'open-source' upstarts. But the data might be pointing to a different conclusion.
The first blind spot is the cost of the alternative. The short thesis is based on a perfect market where enterprises will always choose the cheapest and the best model. But this is not the case in China. The regulatory environment and the need for data sovereignty mean that many Chinese enterprises, especially government and state-owned entities, are required to use domestically approved, compliant models. This provides a regulatory moat for companies like Zhipu and MiniMax that the market is pricing at zero. The bear thesis assumes the market is open and free, but the on-chain data of the Chinese market is that it is not. The market is 'regulated'. If the regulation demands a Chinese model, the 'price war' is not just about cost; it's about compliance.
The second blind spot is the institutional flow. We see the 20% short interest, but we have to ask: 'who is shorting?' Are they shorting the AI model, or are they shorting the 'AI bubble'? We saw the same pattern in the crypto market in 2021. The traders that shorted the DeFi tokens that had no revenue were correct, but the ones that shorted the 'ETH' which had a huge developer ecosystem and network effect, were severely burnt. The shorts might be correct that the model is a commodity, but they may be wrong about the time frame and the specific business models. Zhipu and MiniMax are not just API providers. They are building enterprise solutions, deploying models in vertical domains (finance, healthcare, government), and these are higher-margin services. The market is pricing them as pure API, but the actual revenue mix may be more diversified. The data I've seen from the latest half-year earnings is not yet public, but the pre-earnings estimate suggests that their enterprise business is growing faster than API.
The third and final blind spot is the the 'K' shaped outcome. The current short thesis assumes that the big players will just swallow the market. But the big players (the internet giants) have their own models. The demand for 'multiple models' is a real trend in the market. An enterprise does not want to be locked into one LLM. They want to have the ability to route different models for different tasks. In this scenario, the pure-players can survive as a 'best of breed' option for a specific cost or efficiency tier. The shorts are betting on a 'winner-takes-all' market. But the data from the infrastructure shows the 'Liquid' demand is for 'multi-model orchestration'.
The Takeaway: The Upcoming Earnings as the 'Denominator' Signal
The market is currently in a state of 'extreme fatigue' with the AI narrative. The one remaining catalyst to break this is the first half-year earnings report, due at the end of August. The recent price action is a reflection of a market that has already 'priced in' the worst-case scenario. The short interest is crowded, which means the risk of a 'short squeeze' is real if the earnings report contains even a marginal positive surprise.
For Zhipu AI, the earnings on August 31st will be the 'thesis' breaker. The market wants to see if the 19% cost advantage translates into a higher gross margin than the market assumes. The market expects a gross margin of maybe 10-15%. If the company posts a gross margin above 30% (which would be a signal that their efficiency is real), the short position is in trouble. For MiniMax, the earnings on August 26th will be more about growth. The market needs to see the revenue growth rate, not just the 'monetization' but the actual usage. If they show that the API usage is growing, not just through the price cuts but through new customers, the short thesis is weakened.
The market is asking a simple question: Can a pure-play AI model company be a profitable business in a commodity market? The short sellers have placed their bet on 'no'. But they are ignoring the potential that the market is just in a 'de-rating' phase, not a 'death knell'. I have seen this exact pattern in the crypto market with infrastructure projects. The 20% short interest is a record, but it is also a mechanical factor. It's a sign of a high-risk, high-reward trade. The 'takeaway' for the risk-averse investor is to stay out. The 'takeaway' for the long-term investor is that the 'unlock' might be the best buying opportunity, but only if the earnings are a beat. The 'sign' to look for is not in the share price, but in the 'institutional flow' of the OTC and the 'Southbound' balance. If the Southbound holding percentage increases while the price falls, it is a sign of 'bottom' formation. If it falls, the floor is giving way. The 'truth' is still in the data. The report is the next piece of the ledger. Watch the gas.