Market Brief: The Telegraph of Overvaluation in AI Stocks and Its Echo in Crypto AI

MINIMAX dropped 9.2% on the Hong Kong exchange on July 22. Zhipu fell 3.1%. The market did not provide a catalyst. No technical bug, no regulatory shock, no earnings miss. Silence in the financial filings is the loudest warning sign. These two companies represent the frontier of China’s large language model race. MINIMAX, with its linear-attention architecture, raised over $600 million from investors like Alibaba. Zhipu, backed by Tsinghua University, claims its GLM-4 model rivals GPT-4 on Chinese benchmarks. Their stock prices, however, tell a different story. A single-day drop of 9% is not noise. It is a telegraph of structural repricing. Context: The bull market of 2023 rewarded any company with the label “AI.” Revenue was secondary. Narrative ruled. By mid-2024, the tide turned. Investors demanded proof of unit economics. The cost of a single inference query, the churn rate of API customers, the burn rate of GPU leases—these became the new metrics. MINIMAX and Zhipu, though private up to their IPOs, are now public and exposed to this scrutiny. Their price action suggests the market is recalibrating the value of AI models as commodities, not moats. Core Insight: The stock drop is a function of three variables. First, depreciation of the hype premium. When deepseek, a competing model, cut API prices by 80% in June, the entire race shifted from technology to price war. Second, burn rate visibility. MINIMAX’s IPO prospectus reported quarterly cash burn of $120 million. At that rate, its remaining $400 million cash pile gives it roughly 10 months of runway without additional revenue growth. Third, multiple compression. The P/S ratio for both companies hovered at 30x—unsustainable for pre-profit firms in a rising interest rate environment. The market is simply marking to reality. I have seen this pattern before. In 2020, I audited Curve Finance’s constant product formula. The math looked elegant, but the integer overflow edge case was hiding in plain sight. Similarly, these AI companies’ revenue models look solid until you stress-test the assumptions: What if API pricing falls to zero-margin levels? What if enterprise clients migrate to open-source alternatives? What if the next GPU generation makes training cheaper but also enables competitors to catch up? The mechanism autopsy reveals a fragile equilibrium. Trust is a variable. Verification is a constant. I verify by examining the tokenomics of crypto AI projects like Bittensor and Render Network. They face the same question: Does the token capture value from the underlying AI service, or is it just a speculative vehicle? For MINIMAX and Zhipu, the stock acts as the token. And the price action is screaming that the value capture mechanism is broken. Contrarian Angle: The bears are right about the short-term valuation, but they underestimate the long-term demand for foundational model infrastructure. The global AI compute market is projected to grow at 40% CAGR through 2030. MINIMAX and Zhipu control the model layer, which sits between chip supply and application demand. If they can commoditize the base model and monetize through proprietary fine-tuning or vertical-specific APIs, they may survive the shakeout. The contrarian bet is that the July 22 dip is a buying opportunity for those with a 24-month horizon. However, complexity is often a veil for incompetence. The narrative of “multi-modal supremacy” or “agentic future” should not mask the lack of current cash flows. Takeaway: The AI stock correction in Hong Kong is a microcosm of what awaits the crypto AI sector. Tokens like TAO, FET, and RNDR have rallied by over 200% in 2024 but are now trading at multiples that assume the network effect has already been achieved. The same forensic skepticism must apply. Ask: Where is the revenue? How much of the supply is unlocked? Can the protocol survive a 50% drop in token price? If the answer is “we trust the team,” then you have not done the due diligence. I wrote this not as a prediction of further decline, but as a call for verification. Check the financial statements. Benchmark the burn rate. Model the cascade scenarios. Code does not care about your roadmap. The chain remembers; the marketing team forgets. Based on my audit experience with Tezos in 2017, where formal verification missed runtime type-safety bugs, I learned that even the most rigorous theoretical framework can fail under practical stress. The same applies to business models. The market is now stress-testing AI valuations. The results are not yet final, but the early signal is clear: hype is expiring. Verification is the only constant.