Record short interest. Lock-up expiry. A 50% drawdown from peak. Over the past month, two of China's most prominent AI model companies have gone from IPO darlings to the market's favorite short targets. MiniMax and Zhipu AI are getting squeezed in Hong Kong, and the on-chain data is telling a story that narrative-driven retail investors are ignoring. Gas spike detected. Run.
The setup was always fragile. Both companies listed in Hong Kong in 2025, riding a wave of AI enthusiasm that priced in Chinese AI dominance. Then the music stopped. Short interest on MiniMax hit 20%—a record. Not for the sector. Not for the region. A record for Hong Kong large-cap tech. Let that sink in. Zhipu isn't far behind. But the numbers are only part of the story. The real signal is the convergence of technical exhaustion, dilution mechanics, and a fundamental shift in how the market values AI pure plays.
The Unlock Window
Here's the cold math. The 7-month IPO lockup expired in July for both companies. That unlocked 25.68 million shares of Zhipu and 150 million shares of MiniMax. Combined, that's roughly $11.5 billion in share supply hitting the tape at once. This isn't just a liquidity event. It's a price discovery mechanism. When early VCs and employees dump shares, they're not betting on the company. They're betting that the current valuation is peak and they're cashing out. And they know more than the market does.
But the short interest data is what I'd call the real smoking gun. Twenty percent short interest means one in five shares is being borrowed and sold. That's not a hedge. That's a conviction trade. It means institutional players have looked at the fundamentals—the burn rate, the pricing pressure, the lack of a moat—and concluded this isn't a dip. It's a repricing.
The price action confirms it. Zhipu is down over 50% from its peak, even though it's still 800% above its IPO price. MiniMax is down over 50% as well, and only 80% above its IPO price. The divergence between those two numbers—the IPO pop and the current decay—tells you everything. The market initially paid a premium for the AI story. Now it's paying a discount for the business reality.
The Same Model, Lower Price
Let me get technical. This is where the story gets interesting for those of us who actually read the code and run the benchmarks.
In July, Moonshot released Kimi K3, the flagship model. It was good. Probably the best open-source model in China at the time. Zhipu responded with GLM-5.3, claiming performance on par with Kimi K3. But with a crucial difference: Zhipu's cost per task was 19% lower. That's a real number. That's a genuine engineering advantage.
But look at what happened to the stock price. It dropped 24%.
Think about that. A company that is achieving the same performance at lower cost—the kind of efficiency that should generate fatter margins—is being rewarded with a sell-off. Why? Because the market is no longer valuing model performance. It's valuing the business model. And the business model is under attack.
Hedgeye, one of the short sellers, called it: Zhipu is under pricing pressure that limits its ability to raise prices or expand margins. And MiniMax is "neither the smartest nor the cheapest"—the most dangerous position in a commodity market. That's a killer line because it's accurate.
The technical reality is that the Chinese LLM market has shifted from a seller's market to a buyer's market. The consumer has options. DeepSeek, Alibaba's Qwen, ByteDance's Doubao—all offering comparable performance at aggressive price points. The differentiation window has closed. When models converge in capability, the price becomes the only differentiator, and then there's a race to the bottom.
The Math of the Trap
Let me break down the valuation problem from my perspective. I've been running the numbers on this sector since the LUNA collapse taught me to check every transaction log.
Zhipu's stock is down 50% from peak, but still 800% above IPO. MiniMax is down 50%, but still 80% above IPO. That's a massive gap. The market is pricing in a future where these companies are either dominant or dead.
Here's the problem: the short interest tells you which one the market expects.
When you see 20% short interest, you're seeing a market that has made its judgment. It's not a question of whether these companies are good. It's a question of whether they can be profitable. And the market is betting they can't.
The catch is the Southbound capital. Chinese mainland investors have been buying the dip. Zhipu has 12% Southbound holdings. MiniMax has 8.1%. They're treating this as a buying opportunity, hoping to catch the AI wave. But here's the dangerous part: if the fundamentals don't improve, these investors are just providing liquidity for the shorts to cover into. This is a classic value trap setup. And value traps in growth sectors are the most dangerous kinds.
The Core Number: The Valuation Gap
Let's look at the actual earnings. The core number is the valuation. These companies are not yet profitable. They are burning cash. And the bearish case is that they're in a competitive environment that makes profitability impossible.
The key point: the market has shifted from story-driven to data-driven. The old model of AI stocks was, "Invest because AI is the future." The new model is, "Invest because this company can make money."
This is a fundamental shift. It means the market is starting to value AI companies on the basis of their business model and their cash flow, not just their technology. And for companies that are not yet profitable, this is a dangerous place to be.
It's not a coincidence that the shorts are targeting the two companies. They are targeting the ones with the weakest fundamentals. The ones with the least differentiation. The ones most vulnerable to price wars.
The companies are in a bind. They need to spend on R&D to stay competitive. But they need to be profitable to satisfy investors. And they're caught in the middle. They don't have the scale of the giants, and they don't have the niche focus of the smaller players. They're stuck in the middle.
This is where the model wars end. This is the tipping point.
The Contrarian Angle
Here's the part the narrative misses. The market's obsession with profit is right. But the timeline might be wrong.
The market is pricing these companies as if they will never be profitable. That might be the case. But there's a counter-narrative: the AI market is growing so fast that even small players can be profitable if they can find a niche.
The contrarian angle is that the market is also ignoring the possibility of a turnaround. If the market gets a signal that these companies are focusing on enterprise solutions or specialized applications, the stocks could rebound violently.
The other angle is the take-private and acquisition opportunity. At these valuations, these companies could be targets for larger players looking to enter the AI market or consolidate.
But here's the more uncomfortable truth. The market is also ignoring the possibility of the positive case for AI. The entire AI sector has been oversold, and these companies might be a good long-term investment.
I'm not saying that's the case. I'm just saying that the market is pricing in a certain future, and that future is uncertain.
The shorts have a thesis. It's a good one. But it's not a guarantee. The AI market is still evolving, and these companies are still evolving with it.
The Future
The shorts are right that the pure-play model companies are in trouble. But the data doesn't tell you which ones are going to survive. It only tells you which ones are at risk.
The market is forcing these companies to face a reality check. They need to show the market they can make money. If they can, the shorts will be forced to cover. If they can't, the stock will continue to bleed.
I'm watching the July earnings reports. That's the next test. If they show strong revenue growth, the shorts are going to get squeezed. If they show weakness, the shorts are going to be rewarded.
The bottom line: The market has moved from the story to the numbers. The days of unlimited capital for AI labs are over. The AI models that survive will be the ones that can show a path to profit. And for those that can't, the market is making it clear that it doesn't care about your benchmark scores.
The shorts are in control. But the data is the judge. And the verdict is coming in July.
I've seen this before. The ICO era was filled with great tech and bad business models. The ones that survived were the ones that focused on the fundamentals. The ones that failed were the ones that kept telling the story. The same thing is happening now. The era of storytelling is over. The era of numbers has begun.
Watch the earnings. Watch the burn. Watch the margins. Because the data is the only truth in this market. And right now, the data is saying something that many people don't want to hear.
Gas spike detected. Run.