Here's a number that should unsettle any investor still clinging to the narrative that 'AI is the new electricity': the short interest in Hong Kong-listed MiniMax has hit 20%. For context, anything above 10% is generally considered a crowded, high-conviction short. To see a figure double that on a company that was, until recently, a darling of the AI boom, is not just a trade; it's a referendum on a business model.
This isn't a story about one bad company. It's a story about a category. The market has begun to ask a very uncomfortable question about Chinese 'pure-play' large language model companies: if you only sell the model, and the model is becoming a commodity, what exactly is the equity worth?
The Kimi K3 Shockwave
To understand the current short interest, you have to rewind to a specific catalyst: the release of Kimi K3 by Moonshot AI in July. The market's reaction was violent and immediate. Zhipu AI fell roughly 24%, and MiniMax dropped around 18% in the wake of the release. You don't see moves like that from incremental feature updates. The market priced Kimi K3 as a generational leap, not a step forward.
This is the central tension of the AI capex supercycle. While billions pour into Nvidia GPUs and datacenter infrastructure, the end product—the model itself—is subject to brutal deflationary pricing pressure. A new version of a leading model is the one thing that can temporarily pause that deflation, and the market reacts to that as a zero-sum game. When one player jumps ahead, the others don't just lose market share; they lose the narrative of 'being the smartest'—a narrative that is often the only thing holding up their valuation multiples.
The 'Stuck in the Middle' Trap
In this sector, the market punishes ambiguity. According to Jefferies, Zhipu's counter-move is a GLM-5.3 model with 'performance comparable to Kimi K3, at 19% lower cost.' This is the classic follower's strategy: accept the technical deficit, and pivot to a cost-efficiency play. On the surface, that is rational. But it reveals a structural weakness—the 'cost advantage' is likely born of engineering optimization—quantization, speculative sampling, batching—rather than architectural novelty. Those are real, but they are also replicable. Moonshot, with its superior infrastructure efficiency, can adopt the same optimizations tomorrow, eroding Zhipu's margin of differentiation.
Then there is MiniMax, which finds itself in the most dangerous position of all. Hedgeye's assessment is particularly cruel and precise: the company is 'neither the smartest nor the cheapest.' In a market where you must either command a premium for intelligence or a volume for price, being in the middle is not a place of safety. It's a dead zone. The shorts are betting that you can't get out of it.
The Sell-Side Signals and the Buy-Side Denial
The short sellers aren't just looking at the tech. They are looking at the calendar. The positioning has been building ahead of the interim earnings reports (August 26th for MiniMax, August 31st for Zhipu). That is a clear institutional wager that the numbers will reveal the weakness of the unit economics.
Meanwhile, Southbound capital has been a consistent buyer, with holdings rising to roughly 12% for Zhipu and 8.1% for MiniMax. The stock price has done nothing but fall. That is the definition of a buyer absorbing a falling knife. The classic 'buy the dip' narrative is failing because the seller is not a retail panic; it's the lock-up expiry. In July, the lock-up ended, freeing up roughly 25.68 million Zhipu shares and 150 million MiniMax shares—a supply worth nearly $11.5 billion at the time. These are early-stage holders with astronomically low cost bases. Even with the share price halved, they are sitting on enormous profits. They have every incentive to exit, and their exit is creating a gravity well that retail value-buyers cannot escape.
The Contrarian Angle: The Squeeze and the Base Rate
Let me play devil's advocate. A 20% short interest is a double-edged sword. If the H1 numbers show a 'beat'—not even necessarily a profitable quarter, just a miss that's smaller than feared—the short squeeze could be violent. We saw in a previous cycle how shorting the wrong AI name can ruin a fund, not just a position. The shorts are also correct at a massive level. The industry is in a price war, and pricing power is nonexistent for 'similar' models.
However, there is a subtlety here. The market is pricing a pure 'commodity' scenario. But the Chinese AI market is not a pure free market. It is heavily influenced by the 'national team' capital. The south-bound buying, which looks foolish to a western quant, might be strategic positioning by entities looking to secure a foothold in the national champion of AI. This is not a purely financial trade.
The Takeaway: The Death of the 'Pure-Play'
The de-rating of Zhipu and MiniMax is not just a warning about these two firms; it is a market rule about the 'pure-play' AI model company in China. The market is telling us that the standalone model provider is a structurally sub-scale business without a distribution moat or a proprietary data feedback loop. It is a deflationary asset with no differentiated buyer.
If the earnings in August don't show a path to margin expansion, or if the companies fail to pivot to an application-layer strategy that locks in a user, the shorts will be proven right. But if one of them has an ace up their sleeve—a massive enterprise contract that the analysts don't model—the squeeze will be a sight to behold. The clock is ticking. Which side are you on?