On July 22, 2024, the Hong Kong AI stock sector bled. MINIMAX-W dropped over 9% in a single session. Zhipu AI lost more than 3%. The Hang Seng Index barely budged. This was not a market-wide panic. It was a focused repricing of two names that had ridden the AI narrative wave for months.
The data is clean. Prices collapsed on above-average volume. But the question every battle-tested trader must ask: Was this a signal or noise?
I have spent 21 years in markets, the last eight auditing smart contracts and deploying DeFi strategies. I have seen narratives inflate valuations beyond reason only to snap back when the code fails to deliver. The AI narrative is no different. The code does not lie, only the audits do.
Context: The AI Premium and Its Fragile Foundation
MINIMAX and Zhipu are poster children of China's large language model race. MINIMAX, backed by Alibaba, boasts a proprietary linear-attention architecture. Zhipu, spun out of Tsinghua University, pushes the GLM series. Both command billion-dollar valuations despite generating negligible revenue relative to their market cap.
Their stock prices reflect hope, not earnings. In a high-interest-rate environment, hope is a liability. The July 22 selloff was not driven by news of a model failure or a competitor breakthrough. It was a slow realization that the market had overpriced future cash flows that may never materialize.
As a DeFi yield strategist, I see a direct parallel to algorithmic stablecoins. Terra's LUNA collapsed when the market finally understood that its yield was not sustainable. AI stocks are facing the same reckoning. The only difference is the time horizon.
Core: The Order Flow Analysis — Who Sold and Why?
On-chain data for stocks is less transparent than crypto, but volume and order books tell a story. The sell-off was concentrated in the first hour of trading. Large block trades hit the tape, not retail-sized orders. This is classic distribution: smart money exiting while retail holds.
Let me connect this to my 2020 DeFi Summer experience. I deployed a Python script to automate yield farming across Uniswap V2 and Curve Finance. I learned that when liquidity providers start pulling capital en masse, it is never a single catalyst. It is the accumulation of small cracks in the narrative. The same principle applies here.
What cracks? Three, based on my forensic risk exposure mapping:
- Capital Burn Rate: Both companies spend millions monthly on GPU compute. Their revenue is a fraction of that. In my 2017 ICO audits, I saw projects with similar burn rates that never launched a mainnet. The code did not lie then; the cash flows do not lie now.
- Commoditization Pressure: Baidu, Alibaba, and ByteDance have slashed API prices by 90%+ in 2024. MINIMAX and Zhipu cannot compete on cost without eroding margins. This is the same dynamic I observed in DeFi lending protocols: when Aave drops fees, Compound must follow or lose TVL.
- Low Switching Costs: Users can replace one LLM with another in days. There is no lock-in, no smart contract immutability enforcing stickiness. In crypto, liquidity is sticky because of impermanent loss. In AI, users leave for the next free tier.
The market is beginning to price these risks. The July 22 drop is just the first threshold breach.
Contrarian: The Blind Spots of the AI Narrative
Many analysts will call this a buying opportunity. They will point to the technical achievements: GLM-4's strong Chinese benchmarks, MINIMAX's efficient architecture. They will say the sell-off is overdone.
They are wrong.
My contrarian angle comes from the 2022 Terra collapse. I spent three weeks tracing on-chain transactions, watching the death spiral in real time. I learned that circular liquidity is an illusion. AI stocks are caught in a similar loop: high valuation attracts talent, talent builds models, models require compute, compute burns cash, cash demands revenue, revenue is elusive. The loop breaks when funding dries up.
Retail investors see the model demos. Smart money sees the balance sheet. The code does not lie, only the audits do. And no audit can fix a broken business model.
But there is a nuance. Not all AI plays are equal. Companies with diversified revenue streams — like cloud providers or hardware makers — will weather this correction better. The pure-play AI startups are the most exposed. This is analogous to Layer-1 blockchains vs. DeFi protocols in a bear market. L1s have more resilient valuation floors.
Smart contracts execute logic, not intentions. AI stocks execute on revenue, not on whitepapers. The market is finally enforcing that logic.
Takeaway: Actionable Levels and the Long Game
For traders, the technical picture is bearish. MINIMAX-W broke below its 50-day moving average on July 22. The next support is at the 200-day MA, approximately 20% lower. Zhipu AI is clinging to a trend line that dates back to March. If it breaks, expect a 15% slide to the next consolidation zone.
Fundamentally, this is a signal to rotate out of AI narrative plays and into infrastructure or application layers with proven unit economics. In crypto terms, it is like moving from speculative memecoins to liquid staking derivatives — lower upside but higher data integrity.
My personal rule, hardened by five market cycles: unless you can verify the code (or the financials) yourself, do not catch the falling knife. The July 22 selloff is not a dip to buy. It is a warning to short or wait.
Auditability is the only yield. And until these AI companies show auditable revenue that covers their burn rate, their stock price will remain a derivative of hope. Hope is not a trading strategy.
I will revisit this thesis when Q2 earnings drop. Until then, I am watching the volume profile. The code does not lie, only the audits do. This time is not different.
