Record Short Selling Hits Hong Kong's AI Giants: MiniMax Short Ratio Reaches 20%

The numbers are stark. MiniMax's short interest has hit 20% — a level that typically signals extreme bearish conviction. Zhipu AI sits at approximately 6%, still elevated by any standard. Both companies have fallen more than 50% from their peaks. Yet Zhipu AI's stock still trades 800% above its IPO price. This disconnect — between the narrative of AI supremacy and the reality of unit economics — is the story the market is finally pricing.

Over the past seven days, I have watched the data flow in from S&P Global and third-party analysts. The picture is consistent: shorts are positioning aggressively ahead of the interim earnings reports scheduled for late August. They are not betting on a single bad quarter. They are betting on a structural flaw in the business model of pure-play large language model companies.

The Kimi K3 Shockwave

The trigger for the recent selloff was the release of Kimi K3 by Moonshot AI in July. The market reaction was immediate and brutal: Zhipu AI fell roughly 24%, MiniMax dropped 18%. These are not marginal moves. They are repricing events.

As someone who has spent years auditing model architectures and deployment pipelines, I read this as a clear signal: Kimi K3 is not an incremental improvement. It represents a generational leap in capability. When a competitor releases a model that the market perceives as categorically superior, the entire valuation framework shifts. You are no longer competing on features. You are competing on survival.

Jefferies' assessment of Zhipu AI's GLM-5.3 — "similar performance at 19% lower cost" — is revealing. This is not a strategy of frontal competition on capability. It is a flanking maneuver on price. But here is the uncomfortable truth: cost advantages in inference are engineering problems, not scientific breakthroughs. Model quantization, speculative sampling, batch optimization — these are replicable techniques. If Moonshot AI decides to optimize its own inference costs, Zhipu AI's 19% advantage evaporates within two quarters.

Hedgeye's commentary on MiniMax is even more damning: "neither the smartest nor the cheapest." This is the classic "stuck in the middle" trap. Without a defensible differentiation on either capability or cost, the company's pricing power is squeezed from both directions. In a market where API prices are in freefall, that is a fatal position.

The Unit Economics of Pure-Play LLMs

Let me be direct: the market is questioning whether pure-play LLM companies can ever achieve sustainable profitability. This is not a cyclical concern. It is structural.

Consider the cost structure. Compute costs — both training and inference — typically account for more than 50% of operating expenses for these companies. In a price war where API costs decline quarterly, gross margins are perpetually under pressure. The shorts understand this. They are not trading on sentiment; they are trading on arithmetic.

The fact that Zhipu AI's stock did not rebound after GLM-5.3's release confirms that the market no longer believes the "technology improvement → commercial monetization" transmission mechanism. Investors have been burned by this narrative before. They are now demanding proof of margin expansion, not promises of capability gains.

Southbound capital has been buying the dip — Zhipu AI's Southbound ownership is around 12%, MiniMax around 8.1%. But the stock continues to fall. When selling pressure overwhelms persistent buying, it tells you something important: fundamental concerns outweigh capital flows. The Southbound money is catching a falling knife, not providing a floor.

The unlock schedule adds another layer of pressure. In July, after the lock-up period expired, Zhipu AI and MiniMax released 25.68 million shares and 150 million shares respectively — approximately $11.5 billion in total at the time. Early investors hold massive unrealized gains. Even after a 50% drawdown, they are sitting on enormous profits. The incentive to sell is overwhelming.

The Contrarian Angle: What the Market Is Missing

Here is where I diverge from the consensus bearish narrative. A 20% short interest ratio is a double-edged sword. If the interim earnings reports surprise to the upside — even modestly — the short squeeze potential is significant. Shorts will be forced to cover, driving prices sharply higher in a compressed timeframe. This is a symmetric risk that many retail investors overlook.

More importantly, the market may be underpricing regulatory risk. In China, large model filing requirements, content safety compliance, and data governance are prerequisites for commercialization. If any of these companies faces a compliance setback, the downside is not just financial — it is existential. The shorts have not yet incorporated this into their models. It remains a latent risk factor.

There is also the question of technical convergence. The market assumes Kimi K3's lead is structural. But in my experience auditing model architectures, leads can erode quickly. If Zhipu AI or MiniMax achieves a breakthrough in architecture — not just engineering optimization — the competitive landscape shifts again. The market is pricing in permanent inferiority. That is rarely a safe assumption in this industry.

A Final Word on the Business Model

The core question remains unanswered: can a pure-play LLM company achieve profitability without a differentiated moat? Based on my audit experience across DeFi protocols and AI systems, I have learned that sustainable competitive advantage comes from structural barriers, not incremental improvements. Moonshot AI may have built that barrier with Kimi K3. Zhipu AI is betting on cost efficiency. MiniMax is, for now, stuck in between.

The interim earnings reports on August 26 and August 31 will provide the first real test. Revenue growth, gross margins, customer retention — these numbers will tell us more than any analyst commentary. Trust no one, verify the proof, sign the block. The market is about to verify a lot of claims.

If these companies cannot demonstrate a credible path to profitability, the short thesis is validated. If they can, the squeeze will be spectacular. Either way, the next two weeks will define the trajectory of China's AI sector for the next 12 months. The data is clear. The conclusion is not.

The chain remembers everything. So does the market.