Hook
On July 22, 2024, two of China's most-hyped AI unicorns—MiniMax and Zhipu—saw their Hong Kong-listed shares tumble 9% and 3% respectively in a single session. The news wires called it a 'sector-wide correction.' I call it the first crack in a valuation facade that has been built on code that doesn't ship, burn rates that don't blink, and a product market fit that's still being invented. In my years auditing Layer2 protocols—where 'decentralization' was marketed but never delivered—I've learned to watch for the moment when the market stops believing the roadmap and starts checking the math.
Context
MiniMax (ticker 00100.HK) and Zhipu AI (02513.HK) are two of the most prominent large language model (LLM) companies in China. MiniMax is backed by Alibaba and known for its 'linear attention' architecture (the 'Big Model' approach). Zhipu is the 'Tsinghua faction,' running the GLM series that scores top marks on Chinese benchmarks. Both went public via SPAC or direct listing during the 2023–2024 AI frenzy, raising billions at valuations that assumed they'd dominate the Chinese generative AI market. But the core business metrics—revenue, active users, API call volume—remain opaque. The only publicly verifiable data point is the stock price, and on July 22, that price screamed.
The dip wasn't isolated. The entire Hong Kong AI basket—including Baidu's ERNIE-linked stocks—fell 2–5%. The narrative was 'sector rotation' or 'macro headwinds,' but those are excuses, not root causes. When I see a 9% single-day drop in a company that has no earnings, no clear path to profitability, and a cost structure dominated by GPU rental, I see a market that finally decided to read the footnotes.
Core Analysis
Let me break this down the way I'd audit a zk-Rollup circuit: line by line, cost by cost, invariant by invariant.
1. The Revenue Illusion
MiniMax and Zhipu both generate revenue primarily through API calls and custom model fine-tuning for enterprise clients. But here's the arithmetic that no pitch deck includes: the marginal cost of a single inference call on an NVIDIA H100 cluster is roughly $0.003–$0.005 for a typical 7B parameter model. For larger models—130B+—that cost jumps to $0.02–$0.05. Competitors like DeepSeek, Baidu, and Alibaba have slashed prices by 80–90% since Q1 2024. MiniMax and Zhipu have matched those cuts. So what's the unit economics?
During my 2020 audit of a zk-Rollup's circuit constraints, I discovered that the fraud proof window duration was artificially inflated to mask a computational bottleneck. The same trick is happening here: companies are extending 'training cost' amortization schedules to disguise the fact that inference margins are negative. If you do a simple back-of-the-envelope calculation using publicly available cluster sizes (e.g., Zhipu claimed 10,000 GPUs in 2023), and assume 30% utilization, the annual inference cost alone exceeds $200M. None of these companies have disclosed API revenue anywhere near that figure.
2. The Valuation Disconnect
At their respective IPO valuations—MiniMax at ~$2.5B, Zhipu at ~$3B—these companies were priced at 50–80 times trailing revenue (if we generously estimate revenue at $50M each). In a high-interest-rate environment where the risk-free rate is 5%, that's a 400-year payback period. Compare that to the Layer2 space: in 2024 I analyzed sequencer centralization metrics for three major rollups. Two of them had a single sequencer processing >90% of transactions—a single point of failure. The market ignored that because the narrative was 'scaling Ethereum.' The same mental accounting is happening here: investors are buying the 'AI disruptor' story without verifying the underlying infrastructure.
3. The Technical Debt
MiniMax's linear attention architecture is genuinely novel—it sidesteps the quadratic complexity of Transformer self-attention. But novelty doesn't mean practicality. When I manually reconstructed the proof constraints for an early Optimistic Rollup in 2020, I found that the 'efficiency' gains came with a hidden assumption: the fraud proof window had to be extended to cover the additional computation. The same principle applies to linear attention: it reduces training FLOPs but introduces accuracy trade-offs in long-context tasks. The market hasn't priced that risk because the benchmarks are cherry-picked. Check the math, not the roadmap.
4. The Capital Expenditure Trap
MiniMax and Zhipu are burning cash at a rate of $50–$100M per quarter just on GPU rent and data center costs. Their cash reserves—last disclosed in their prospectus—are roughly $500M each. That gives them a runway of 12–18 months. If the stock price stays depressed, they lose the ability to raise follow-on equity at favorable terms. The same dynamic killed several Layer2 projects in 2023: they had great tech, but the treasury was denominated in their own tokens, which collapsed when the market turned. These AI companies are denominated in fiat, but the burn rate is just as unforgiving.
Contrarian Angle
Now for the counter-intuitive take: the July 22 selloff might be overdone for the wrong reasons. Market observers are blaming 'AI bubble fears,' but the real risk isn't a bubble bursting—it's a slow suffocation from capital misallocation. The companies themselves are not bad; they have world-class researchers and genuinely impressive models. Zhipu's GLM-4 holds its own against GPT-4 on Chinese language tasks. MiniMax's linear attention could be a breakthrough if applied correctly. The problem is the market valued them as if the revenue would materialize instantly, ignoring the 10-year compressed timeline.
In the Layer2 world, we saw the same pattern: Arbitrum and Optimism traded at billions in fully diluted valuation before they had any meaningful fee revenue. But at least those Layer2s have a clear value capture mechanism (sequencer fees). AI models don't; the marginal cost of inference trends to zero as hardware improves, while Layer2 fees are bounded by L1 congestion. The contrarian question: is there a scenario where these AI companies become the 'Ethereum' of AI—the settlement layer for autonomous agents? If so, the current dip is a buying opportunity. But that requires a fundamental shift in their business model, not just a better model.
Takeaway
The Hong Kong dip is a preview of the reckoning coming for every tech company that sold a vision before it built the infrastructure. I've spent the last six years auditing protocols where the code didn't match the whitepaper. What I see in AI is the same pattern: audited only by marketing teams, never by engineers who can trace the data flow. The market will eventually demand proof—just like it did for Optimistic Rollups when users realized withdrawals took seven days. Complexity is the enemy of security, and valuation complexity is the enemy of investment sanity. Check the math, not the roadmap. Audits are snapshots, not guarantees. And code does not care about your vision.