The Short-Seller's Verdict on China's AI Duopoly
Record short interest against Zhipu AI and MiniMax isn't just a market signal. It's a technical audit of two companies caught between narrative and unit economics. The ledger remembers what the market forgets.
The numbers paint a stark picture. Short bets on these two Chinese AI heavyweights have hit unprecedented levels, driven by what analysts politely call 'price war anxiety.' Less politely: investors are betting these companies can't make the math work. This isn't a sentiment play. It's a structural assessment of a market where API tokens are becoming commodities before the infrastructure is even amortized.
Context matters here. Zhipu AI and MiniMax represent two distinct paths in China's AI race. Zhipu, with its GLM series and government-enterprise focus, positions itself as the sober, institutional player. MiniMax, riding the consumer wave with short-form content and interactive applications, plays the growth game. Both now share one unfortunate metric: their stock loan utilization rates suggest the market views them as overvalued stories in a sector that's rapidly transitioning from technological moonshots to profit-and-loss statements.
Let's examine the order flow. The short thesis isn't complicated, but it's compelling. These companies are selling access to large language models at prices that are collapsing faster than model training costs. The market has seen this pattern before—in cloud computing, in ride-sharing, in every sector where venture capital subsidizes usage to buy market share. What's different here is the capital intensity. Training frontier models requires billions in compute, and serving them at scale requires even more. When your marginal revenue per token falls while your marginal cost per token stays flat or rises, you have a structural problem, not a pricing problem.
Structure survives where sentiment collapses. Based on my experience auditing smart contracts and building hedging strategies in volatile markets, I see a familiar pattern: the market is pricing in a scenario where these companies' differentiation claims fail to materialize into pricing power. The shorts are betting that GLM-4's long-context capabilities and MiniMax's multimodal strengths won't translate into customer lock-in. They're betting that Chinese enterprises, notoriously price-sensitive, will treat AI models as interchangeable utilities.
Now, the contrarian angle. The market might be missing something. Short interest this concentrated often represents a crowded trade, and crowded trades have a tendency to reverse violently. More importantly, the shorts are betting against the one thing Chinese AI companies have proven they can do: optimize costs through vertical integration and government support. The narrative ignores that Zhipu and MiniMax aren't just competing against each other. They're competing against Baidu, Alibaba, and ByteDance—companies with deeper pockets and existing distribution. The price war isn't a choice; it's survival.
But here's what the short thesis gets wrong. It assumes this price war ends in mutual destruction. In reality, price wars in technology markets historically end with the strongest player emerging with dominant share and improved margins. The question isn't whether Zhipu and MiniMax survive the price war. It's whether they emerge as the consolidators or the consolidated. The shorts are betting on the latter without fully accounting for the potential of government intervention to support national champions in strategic technology sectors.
The infrastructure reality adds another layer. Both companies face the same compute constraints as every Chinese AI firm, magnified by export controls on advanced semiconductors. This is where I see the real vulnerability. Not in the pricing model, but in the supply chain. If you can't get the chips, you can't train better models, and you can't reduce serving costs. The shorts might be positioning for a scenario where these companies hit a compute wall, forcing them into a perpetual catch-up game against better-resourced competitors.
We do not predict the wave; we engineer the board. The market is sending a clear message: the era of paying for AI promise is ending. The next funding rounds for Zhipu and MiniMax will be priced on metrics like gross margin, customer retention, and path to profitability—not on model benchmarks. This is the transition every technology sector eventually faces, but few experience it with this velocity.
Liquidity dries up; logic remains solvent. The short sellers are doing what the market should always do: forcing discipline. They're asking the hard questions about unit economics, about the sustainability of subsidized pricing, about the difference between usage growth and value creation. Whether they're right about the timing or the magnitude, they're right about the underlying tension. AI companies must eventually prove they can convert technological capability into financial returns.
The data points I'd watch aren't the stock prices. They're the API pricing sheets, the customer churn rates, and the compute utilization metrics. Time decays options; patience decays noise. In six months, we'll know whether this short thesis was prescient or premature. In the meantime, the market has delivered its verdict on China's AI duopoly: show me the margins, or I'll show you the door.
The lesson here extends beyond Zhipu and MiniMax. It's a warning to every AI company globally operating on subsidized economics. The market's patience for narrative-driven valuations has a limit, and that limit is measured in quarters, not years. The companies that survive this reckoning won't be the ones with the best models. They'll be the ones with the best cost structures and the clearest path to profitability.
The real question the market should be asking isn't whether Zhipu and MiniMax are overvalued. It's whether the entire AI sector is facing a moment of truth where technological leadership must finally translate into economic leadership. The shorts have made their bet. The ledger will reveal the outcome.