The financial market is a brutal ledger. It does not care about model benchmarks, nor does it reward the most eloquent technical whitepaper. It cares about the balance sheet. And right now, the balance sheet of China's most prominent AI startups is screaming a warning that the technology press is too enamored with the "magic" of large language models to hear.
According to a report from Crypto Briefing, short bets against Zhipu AI and MiniMax have hit record levels. The stated reason is simple: investor anxiety over a destructive AI price war. But the code whispers what the pitch deck screams. This is not just about a pricing skirmish. This is the market's forensic audit of a business model that has spent years prioritizing scale over sustainability.
Let's dissect the anatomy of this bet. It is not a random act of speculative aggression. It is a calculated conclusion drawn from observable data points. In my years auditing cryptographic protocols and DeFi architectures, I have learned that the market's most sophisticated players do not act on hype; they act on the underlying logic of the system. Here, the system is the AI industry's unit economics. The short sellers are essentially reading the bytecode of the financial statements, and they see an integer overflow in the profit model.
The Context is not a vacuum. Zhipu AI and MiniMax are not fringe players. They are the champions of China's "AI National Team" narrative, having raised billions in funding at staggering valuations. They are positioned as the counterweights to Western giants like OpenAI and Anthropic. Yet, they operate in a hyper-competitive landscape dominated by Big Tech (Baidu, Alibaba, Tencent, ByteDance) and aggressive newcomers (DeepSeek, Moonshot). The pressure to capture market share in the enterprise and consumer sectors has led to a race to the bottom on API pricing. The market is now asking a fundamental question: Can these companies translate technical leadership into financial leadership? The short interest suggests the answer is a resounding "no."
The Core of my analysis—the systematic teardown—focuses on the economic architecture of these companies. The problem is not the model quality; it is the cost of goods sold (COGS). Training a frontier model requires tens of thousands of high-end GPUs (H100s or A100s), a capital expenditure that is effectively a sunk cost. The variable cost, however, is the inference cost—the electricity and compute required to answer each query. In a price war, this variable cost is the battleground. If you cut your API price by 50%, you need to quadruple your volume to maintain the same gross profit, assuming you haven't optimized your inference engine. This is a brutal mathematical reality.
The market is betting that Zhipu and MiniMax cannot achieve the operational efficiency required to make their current pricing sustainable. They are betting that the "blob" of data flowing through these models will not translate into a "blob" of revenue high enough to cover the infrastructure costs. This is not a novel critique. It mirrors the narrative we saw in the early days of cloud computing, where growth was prioritized over profit, leading to eventual consolidation and a "winner takes all" dynamic. However, the AI landscape is more fragmented, and the cost of experimentation is higher. The short sellers are essentially saying that the "toll bridge" these companies are building is too expensive to cross, and the traffic (users) is not paying enough.
The core insight is the "per-token economics" are structurally flawed. The industry has been obsessed with parameter count and benchmark scores. But the real metric is cost-per-thousand-tokens. If your cost-per-token is higher than your price-per-token, you are losing money on every interaction. To win the price war, you need a technological moat in inference optimization. This includes quantization (running models on lower-precision arithmetic), speculative decoding (guessing tokens to generate them faster), and leveraging cheaper, specialized hardware. If Zhipu and MiniMax rely on standard NVIDIA deployments without advanced optimization, their margins will be crushed. The short sellers are likely betting that these startups, despite their research pedigree, lack the infrastructure engineering talent of a Baidu or Alibaba, who have spent years optimizing their cloud operations.
Furthermore, the short bet signals a deeper anxiety about the "value capture" problem. In the crypto world, we talk about "fat protocol" vs. "thin application" theories. In AI, the reverse might be true. The model itself is becoming a commodity. As open-source models like Llama and Qwen approach parity with proprietary models, the "magic" is evaporating. The value shifts to the distribution layer—the interface, the workflow integration, and the proprietary data used for fine-tuning. If Zhipu and MiniMax cannot differentiate their offerings with unique datasets or vertical-specific solutions, they are just selling standardized "neurons" at a loss.
My Contrarian angle is to defend the bulls for a moment. Beauty is the most sophisticated rug pull. The market might be wrong. The short interest could be a reaction to macro-level fears (e.g., US export controls on chips) rather than company-specific weaknesses. Zhipu AI, with its strong academic roots and government support, might have access to domestic chip alternatives (Huawei Ascend) that lower its cost structure in the long run. MiniMax, with its focus on consumer-facing products like "Glow" and its success in the short-drama and gaming sectors, might be building a data flywheel that reduces customer acquisition costs. The bulls argue that the AI adoption curve is still in its infancy, and the current revenue is not indicative of future potential. They see the price war as a necessary "land grab," similar to Amazon's strategy of prioritizing market share over short-term profits.
However, the Contrarian view fails to address the "capital intensity" required to survive. Amazon could sustain losses because it had a monopoly on cloud infrastructure to subsidize its retail division. In China, Zhipu and MiniMax do not have a cash cow to fund their burn rate. They rely on external capital. A record short interest makes future fundraising significantly harder. It triggers a negative feedback loop: lower valuation → harder to attract top talent (who want equity) → less competitive product → lower revenue → further shorting. This is the "death spiral" that the short sellers are banking on.
Truth hides in the assembly, not the press release. The assembly here is the cost structure of the compute. Let's look at the infrastructure layer. The short sellers are likely aware that the cost of training models is not the primary expense; it is the inference. In a price war, you are essentially subsidizing your user's intelligence. If you charge $0.50 per million tokens but it costs you $0.75 to generate them, you are paying users to use your product. This is not a business; it is a charity. To fix this, you need algorithmic breakthroughs, not just hardware. The market is betting that Zhipu and MiniMax have exhausted their algorithmic edge and will resort to financial engineering to stay afloat. This is a dangerous assumption, but the risk is asymmetrical.
The Takeaway is not to panic or to blindly follow the short sellers. It is to recognize that the AI industry is entering a phase of "selective survival." The initial euphoria, the "hype cycle" of 2023-2024, is over. We are now in the "trough of disillusionment" where the financial community is holding AI companies to the same standards as traditional SaaS companies: they must demonstrate a path to profitability. Based on my audit experience, I find that most of these "high-flying" startups fail not because of technical incompetence, but because of financial mismanagement. They burn through their "seed" money on vanity metrics instead of building a solid foundation.
Silence is the only honest consensus mechanism. The silence from Zhipu and MiniMax regarding their gross margins is deafening. In the absence of transparent financial data, the market will default to its worst assumptions. The short sellers are filling the information vacuum with their own bearish thesis. This is the "legal" manipulation of sentiment through leverage. The only way to counter this is with data. These companies need to release case studies showing cost-per-token improvements and enterprise adoption rates that justify the pricing.
As a security auditor, I look for the flaw in the logic. The logic here is simple: if the unit economics are bad, the project fails. The short sellers have found the vulnerability in the business model. The question is whether the developers can issue a "patch" before the market crashes. This is not a question of code, but of economics. The cost of intelligence is dropping, but the price is dropping faster. The "rug pull" here is not a malicious developer draining a liquidity pool; it is a market correction that evaporates the value of over-leveraged, under-profitable enterprises.
The AI industry needs a "Cold Dissector." It needs to stop romanticizing the technology and start auditing the economics. The record short bets against Zhipu and MiniMax are not a tragedy; they are a necessary correction. They are the market's way of forcing accountability. Innovation without integrity is just theft—in this case, stealing capital from investors with the promise of future riches that may never materialize.
The forward-looking thought is this: watch the "blob" metrics. In the next two years, we will see a clear bifurcation. AI companies will either pivot to high-margin verticals (finance, healthcare, legal) where they can charge premium prices, or they will become commoditized utilities with razor-thin margins, surviving only on the backs of massive infrastructure partnerships. The short sellers are betting that Zhipu and MiniMax cannot make that pivot fast enough. They might be right. The market is a harsh auditor, and it never accepts a "good idea" as a substitute for "good revenue." The code of the financial markets is unforgiving. It has no sympathy for unprofitable elegance.