The record short interest against Zhipu AI and MiniMax is not a market anomaly. It is a technical verdict. Investors are not betting on a price drop. They are betting on a fundamental failure of unit economics in China's large language model sector. As someone who spent the 2022 bear market compiling ZK-SNARK circuits on local hardware, I have a certain respect for verification. Short sellers are performing a form of adversarial verification on business models. Their thesis is simple: the cost of inference will outpace the revenue from API calls. The market is pricing in a mathematical certainty that these companies cannot escape the commodity trap. Let me walk you through the mechanics of why their logic holds, and where it might break.
Context: The AI Gold Rush is Over, The Audit Has Begun
The narrative surrounding Chinese AI companies has shifted. It is no longer about parameter counts or benchmark leaderboards. The conversation has moved to burn rates, customer acquisition costs, and gross margins. Zhipu AI and MiniMax represent two distinct approaches to the market. Zhipu leans into the enterprise and government sector, leveraging its GLM series for vertical deployments. MiniMax pushes consumer-facing applications, specifically in the short-form video and interactive entertainment space. Both require massive upfront capital expenditure for compute. Both face a brutal pricing environment, driven by the aggressive strategies of Baidu, Alibaba, and ByteDance. The record short bets are a signal that investors believe the technical moat of these two companies is insufficient to justify their valuations. It is a vote of no confidence in their ability to convert model quality into pricing power.
Core: The Mechanics of the Cost Trap
The core issue is not intelligence. It is the cost of delivery. The short thesis hinges on the relationship between token generation cost and API pricing. In a competitive market, the price of an API call trends toward the marginal cost of compute. If Zhipu and MiniMax cannot differentiate their output, they are forced to match the price of the giants. The giants, with their cloud infrastructure and scale, have a structural advantage. They can subsidize their model divisions with profits from other business units. A standalone AI company does not have this luxury. The AMM model hides its truth in the invariant, and similarly, the AI business model hides its truth in the cost-per-token curve.
Let us consider the technical variables. Inference cost is driven by memory bandwidth and compute utilization. During the 2021 Axie Infinity forensics, I spent weeks reverse-engineering smart contracts to find edge cases in breeding fees. The logic here is similar. We are looking for the edge case where the cost structure collapses. For Zhipu and MiniMax, the edge case is the lack of proprietary hardware. They rely on NVIDIA GPUs, or at best, adapted versions of domestic chips like the Huawei Ascend. The major cloud providers are investing heavily in custom silicon, like Google's TPU or Amazon's Trainium. These custom ASICs dramatically reduce the cost per token for the giants. A standalone company, purchasing GPUs at market rates, is structurally disadvantaged. They are fighting a war with borrowed weapons.
The short sellers are not just betting on a price war. They are betting on a capital expenditure race that Zhipu and MiniMax cannot win. The cost of a single training run for a frontier model is astronomical. The cost of serving that model to millions of users is even higher. If the market forces prices down, the margin evaporates. My analysis of the 2020 Uniswap V2 liquidity mechanism showed how the constant product formula could create arbitrage opportunities for high-frequency traders. The same principle applies here. The short sellers are the high-frequency traders. They are exploiting the structural arbitrage between the high cost of standalone inference and the low cost of subsidized cloud inference. They are front-running the inevitable decline in pricing power.
Zero knowledge isn't magic; it's math you can verify. The same applies to profitability. The market is trying to verify the profitability equation, and the math is not working out. The fundamental issue is that model quality is becoming a commodity. As open-source models improve, the gap between proprietary and open-source narrows. If a developer can achieve 90% of the performance of Zhipu's GLM-4 by using a free, open-source alternative, why would they pay for the API? This is the commoditization threat. The short thesis relies on the assumption that these companies cannot build a sufficient moat through data flywheels or distribution advantages. The record short interest suggests the market believes the moat is shallow.
I don't see the moat in the public technical documentation. The competitive advantage is not in the architecture; it is in the data. The question is whether Zhipu's enterprise deployments and MiniMax's consumer engagement generate enough proprietary data to improve their models faster than the competition. This is a data flywheel effect. If it works, it creates a barrier. If it does not, they are stuck in a commodity market. The short sellers are betting on the latter. They are betting that the user-generated content from MiniMax's apps is not high-quality enough to train a significantly better model. They are betting that Zhipu's enterprise contracts do not provide the diverse, high-volume data needed to push the frontier. This is the crux of the bearish case.
Contrarian: The Blind Spot in the Short Thesis
The contrarian angle is the assumption that cost is the only variable. The short thesis focuses on the price war and the inability to compete with cloud giants on cost. But it ignores the possibility that these companies can win on distribution and integration. Zhipu has a strong position in the Chinese government and state-owned enterprise sector. This is a market that values data sovereignty and compliance over price. If a government entity requires on-premise deployment, they are not comparing Zhipu's API price to Alibaba's API price. They are comparing the cost of a full-scale deployment. This is a different sales cycle, with different economics. The short thesis fails to account for the switching costs in this sector. Once a model is integrated into a government workflow, the cost of switching to a competitor is high. It is not just a matter of API pricing.
Furthermore, the short thesis ignores the potential for regulatory protection. The Chinese government has an interest in maintaining a competitive domestic AI ecosystem. They do not want a monopoly controlled by one or two cloud giants. There is a political incentive to ensure that independent AI companies survive. This could manifest in subsidies, preferential access to compute, or regulatory requirements for diversity. The 2024 ETH ETF technical due diligence taught me to look at the centralization risks in proposed models. The same lens applies here. The government may see a centralized AI market as a systemic risk. They may intervene to support Zhipu and MiniMax as a counterweight. The short sellers are pricing in a pure market outcome, but the AI sector in China is not a pure market. It is a strategic sector.
The safety argument also cuts both ways. The analysis suggests that price pressure might force companies to cut corners on safety alignment. This is a valid concern. But it is also a potential barrier to entry for the giants. If the regulatory environment punishes unsafe models, the cost of compliance is higher for a massive, general-purpose model. A smaller, specialized model, like those from Zhipu or MiniMax, might be easier to align and certify. This could be a cost advantage, not a disadvantage. The short thesis assumes a race to the bottom in safety. But the Chinese regulator has a strong incentive to avoid a scandal. They might impose standards that favor the nimbler, more focused companies. This is a potential counterweight to the cost disadvantage.
Takeaway: The Verification is Still in Progress
The market is performing a security audit on the Chinese AI sector, and the findings so far are critical. The record short bets are a red flag. They are a warning that the financial foundation of these companies is not as solid as the technology suggests. The next twelve months will be a period of verification. We will see if the cost curves bend in their favor, if the data flywheels start to spin, and if the regulatory environment provides a buffer. The market is looking for a specific invariant: the point at which revenue growth outpaces the cost of compute. Until that point is proven, the short thesis will hold. The AI war is not about who has the best model; it is about who can afford to serve it. The market is betting that the standalone players cannot. The math is on their side, for now. But in my experience, the math is only as good as the assumptions. And the assumptions about government intervention and market dynamics are far from settled. The verification is still in progress.
