Record short interest in Hong Kong's AI sector is not noise. It is a systematic audit of a business model that has yet to prove its unit economics. Over the past week, data from S&P Global confirmed that MiniMax's short ratio has surged to 20%, while Zhipu AI sits at a comparably elevated 6%. This is not a random bout of bearishness. It is a concentrated bet, placed ahead of interim earnings, that the market has fundamentally mispriced the profitability of China's pure-play large language model companies.
From my seat managing a digital asset fund, this pattern is familiar. It mirrors the DeFi liquidity stress tests I ran in 2020, where the market's faith in a protocol's revenue model was the last thing to break before the price did. The mechanics differ, but the underlying principle is identical: when a sector's valuation is predicated on narrative rather than cash flow, the market eventually demands a margin statement. Short sellers are simply the first to demand it.
Context: The Liquidity Map of AI Valuations
To understand the current positioning, one must map the global liquidity flows. Hong Kong's equity market is a conduit for mainland capital via the Stock Connect, but it is also a venue for global funds that require a different risk premium than their US counterparts. The AI narrative that drove these IPOs was built on a global liquidity glut and a scarcity of high-growth assets. That tide is ebbing. With US rates remaining restrictive and risk appetite rotating toward tangible earnings, the market is now applying a discount rate that these pre-profit AI firms cannot withstand.
The trigger for the recent sell-off was the release of Kimi K3 by Moonshot AI. The market interpreted this not as an incremental update, but as a generational leap. Zhipu AI's stock dropped 24% and MiniMax fell 18% in the aftermath. These are not minor price adjustments; they are repricing events. The market is signaling that model capability is the primary determinant of competitive position, and a gap in that capability is a direct threat to future revenue.
Core: Auditing the Unit Economics of Pure-Play AI
The core issue is not the technology; it is the cost structure. A pure-play LLM company is a capital-intensive operation with a cost of goods sold dominated by compute. In a price war, where API costs are being driven down by competitors like ByteDance and Alibaba, the ability to generate margin is severely constrained.
Zhipu AI's response to Kimi K3's release was to position its GLM-5.3 model as offering "similar performance at a 19% lower cost." Based on my engineering background, a 19% cost advantage is likely the result of inference optimization—techniques like quantization, speculative sampling, and batch processing—rather than a fundamental architectural efficiency. This is an engineering problem, not a scientific one. It is a moat that can be crossed. My experience auditing over 400 smart contracts in 2017 taught me that competitive advantages built on process optimization are ephemeral; they are standardized and replicated within a single cycle.
MiniMax faces a more dire situation. Hedgeye's assessment that MiniMax is "neither the smartest nor the cheapest" is a classic stuck-in-the-middle position. In a market where differentiation is the sole basis for pricing power, MiniMax has neither a technological premium to justify higher prices nor a cost structure to win on volume. This is not a temporary setback; it is a structural flaw in its go-to-market strategy.
The data tells the story of a supply overhang. Both companies are trading more than 50% below their peaks. Zhipu AI's stock is still 800% above its IPO price, which indicates the froth in the initial valuation. More critically, the lock-up expiry in July unleashed a wave of supply. Zhipu AI had 25.68 million shares and MiniMax had 150 million shares released, totaling approximately $11.5 billion at then-prices. This is a wall of selling that will suppress any rallies.
Contrarian: The Decoupling Thesis
The prevailing narrative is that short sellers are correct because these companies cannot make money. My contrarian view is that the short thesis is correct on the business model but may be mistimed on the price action. A short ratio of 20% is extreme. It creates a mechanical risk: the squeeze. If either company reports earnings that are merely "less bad" than the consensus fear, we will see a violent short-covering rally. The shorts are positioned for a binary event. The market is a discounting mechanism, and a 50% drawdown already prices in a significant amount of failure.
Furthermore, the market is ignoring the possibility of a sectoral decoupling from the US. While US AI names trade on narrative, Chinese AI names are increasingly being priced on domestic capital costs and policy support. The southbound flows—mainland investors holding ~12% of Zhipu AI and ~8.1% of MiniMax—are not simply value investors. They are often strategic buyers with a longer time horizon than the average global macro fund. This is not a floor, but it is a source of bid that is indifferent to short-term volatility. In my 2020 stress test, the market exited UST 48 hours before the crash because the data on depeg risk was clear. Here, the data on profitability is not clear; it is speculative. The shorts are betting on a conclusion that has not yet been written.
Takeaway: Positioning for the Post-Earnings Cycle
The upcoming interim reports on August 26 (MiniMax) and August 31 (Zhipu AI) are the catalysts. The market is not waiting for a profit; it is waiting for a path to one. I want to see if Zhipu AI's "19% cost advantage" translates into a gross margin figure that is defensible. I want to see if MiniMax can articulate a vertical strategy that escapes the middle.
The 2017 ICO audit taught me to look for structural vulnerabilities before the market does. The structural vulnerability here is the absence of a differentiated cost curve. The shorts have found this vulnerability. The question is whether the fundamentals have hit an inflection point where the downside is priced. We do not predict the wave; we engineer the hull. The hull of these companies is being stress-tested now. The data suggests the engineering is incomplete, but the market may have already discounted the failure. In the short term, the squeeze potential is real. In the medium term, only a clear path to margin expansion will prevent the slide from continuing. This is a trade on the report, not a thesis on the technology.
