On August 24, Hong Kong-listed AI concept stocks experienced a sharp correction. Zhipu fell over 11%, MiniMax dropped more than 10%. The data comes from Bitget market feeds. Headlines frame this as a sector-wide panic. I frame it differently: this is a repricing event, and the market is finally asking the question that code audits have been asking for two years — where is the revenue?
History verifies what speculation cannot. In 2018, I spent three months auditing ICO refund contracts while the market bled. The lesson was simple: when narratives collapse, only structural integrity survives. The same principle applies to AI valuations today. The market is not punishing Zhipu or MiniMax for technical failures. It is punishing them for the absence of verifiable commercial metrics.
Context: The Valuation Gap
Zhipu, backed by its GLM architecture, closed a funding round in early 2024 at a valuation exceeding RMB 20 billion. MiniMax, with its MoE-based abab series, crossed the $1 billion mark. Both are considered part of China's "AI Dragon Four." Neither has disclosed meaningful API revenue. The P/S ratios are not just high — they are unverifiable. In a bear market for risk assets, unverifiable multiples become liabilities.
The Hong Kong market is particularly unforgiving. Unlike US markets, where AI names trade on narrative momentum, Hong Kong investors demand milestones. They want customer counts, retention rates, and gross margins. Zhipu and MiniMax have not provided these. Consequently, the market is doing what markets do: it is marking down uncertainty.
Core: The Technical Reality Behind the Selloff
Let me be precise about what this decline is not. It is not a reaction to a model failure. There is no evidence of a GLM or abab regression. No security breach. No regulatory violation. The selloff is a sentiment event, but sentiment is not random. It is a lagging indicator of structural pressure.
First, the price war. Since mid-2024, major players — ByteDance, Alibaba, Baidu — have slashed API prices by over 90%. This is not competition; it is attrition. For startups like Zhipu and MiniMax, whose unit economics were already thin, this compression is existential. The market understands this. It is pricing in a future where gross margins approach zero for undifferentiated models.
Second, the competitive stack. Baidu's Ernie, Alibaba's Tongyi, and ByteDance's Doubao are not just models; they are integrated with cloud ecosystems. They offer compute, storage, and model inference as a bundle. Zhipu and MiniMax are trying to sell models à la carte. In a price war, bundled offerings win. This is not a technical gap — it is a distribution gap. And distribution gaps are harder to close than model quality gaps.
Third, the differentiation problem. DeepSeek has captured the open-source community with its V3/R1 series. Moonshot AI owns the long-context narrative with Kimi. Zhipu and MiniMax are caught in the middle — not the cheapest, not the most open, not the most specialized. In a market that rewards extremes, the middle is a dangerous place to sit.
Contrarian: The Blind Spot in the Panic
Here is the counter-intuitive angle. This selloff may be more informative than any rally. It reveals that the market has shifted from a "story-driven" to a "data-driven" regime. That is a healthy correction. But it also exposes a blind spot: the market is treating all AI startups as interchangeable. They are not.
Zhipu has a credible path to vertical solutions — government and enterprise contracts that require private deployment. MiniMax has a consumer-facing angle with its entertainment and social products. These are different business models with different risk profiles. The blanket selloff ignores this granularity. Pressure reveals the cracks in logic, and the logic here is lazy.
Another blind spot: the compute constraint. Both companies rely on domestic chips or downgraded GPUs due to US export controls. This is a cost disadvantage, but it is also a moat. Companies that can optimize inference on Huawei Ascend or A800-class hardware will have a structural cost advantage over those that cannot. The market is not pricing this. It is only looking at the top-line narrative, not the bottom-line engineering.
Takeaway: What to Watch
Silence is the strongest proof of truth. The next three months will be defined by what Zhipu and MiniMax do, not what they say. Watch for three signals. First, any disclosure of API call volumes or paying customer counts. Second, a pivot to vertical industry solutions with named enterprise clients. Third, any announcement of compute optimization that reduces inference costs. If these companies can demonstrate unit economics, the current valuation becomes a discount. If they cannot, the decline is not a correction — it is a preview.
Evidence does not negotiate. The market has issued its verdict on narrative. Now it awaits the evidence of execution. Patience is a technical requirement, and the patient will be rewarded with clarity. The rest will be left holding a story without a balance sheet.