On July 22, MINIMAX crashed 9%; Zhipu fell 3%. The Hong Kong AI concept stocks bled collectively. The headlines scream "sell-off." But what does the code say? Nothing—because there is no code to audit. These are not smart contracts; they are black boxes promising intelligence. As a zero-knowledge researcher, I see a familiar pattern: the gap between promised innovation and implemented engineering has finally triggered a market recalibration. The structure of the trust is breaking down.
Context: MINIMAX and Zhipu are two Chinese AI startups riding the LLM wave. MINIMAX's "big form" architecture (based on linear attention) once drew hype; Zhipu's GLM-4 inherited Tsinghua pedigree. But by mid-2024, the narrative has shifted from model capabilities to revenue reality. Both are burning cash on GPU compute, slashing API prices to compete with Baidu and DeepSeek, and facing regulatory compliance costs. The market is no longer buying promises—it is demanding proof.
Core: Let's dive into the structural vulnerabilities that the price action reveals.
1. Technical Black Box Problem No one can verify the actual performance of these models on an arbitrary test set. Unlike a blockchain where every state transition is transparent, AI inference is opaque. The market relies on self-reported benchmarks, which are prone to data contamination. Math doesn't lie, but marketing does. In my audit of the 0x protocol v2 in 2018, I found seven critical edge cases because the code was open. Here, we have no code. The risk premium is infinite.
2. Commercialization Math Let's assume inference cost per query is $0.01 and the average API price is $0.002. The burn rate per query is $0.008. Multiply by millions of queries per day. That is a running negative sum. Math doesn't. Unit economics alone justify a major haircut. Compare to a crypto protocol: if gas cost exceeds transaction fee, the chain becomes unsustainable. Same here.
3. Competitive Game Theory Every Chinese LLM player is in a prisoner's dilemma. They all cut prices to capture market share, but the collective result is margin erosion. No one can deviate. The equilibrium outcome is a race to the bottom. This mirrors the L2 war in Ethereum: everyone claims decentralization, but only the ones with real liquidity survive. Zhipu and MINIMAX are fighting for the same enterprise customers. The winner takes little.
4. Valuation Recalibration In 2021, AI startups were trade at 100x revenue (if any). Now, investors are looking at burn multiples. The discount rate is up, yet these companies have negative earnings. This is a classic "growth stock" squeeze—identical to what happened to high-flying DeFi tokens after the Terra collapse. In my 2022 retreat studying algorithmic stablecoins, I concluded that any system with unclear liabilities will eventually face a sudden repricing. AI stocks are no different.
5. Regulatory Overhead China's cyberspace administration requires model censorship compliance. This adds engineering cost and limits use cases. Privacy is a protocol, not a policy. These companies are forced to implement content filters that degrade performance. The regulatory drag is a silent deadweight loss. Trust is a vulnerability, not a virtue; centralized content moderation introduces a single point of failure.
6. Infrastructure Dependency Both companies rely on NVIDIA GPUs, which are scarce and expensive. If revenue doesn't cover GPU rental, they must raise more capital at lower valuations. This is exactly the issue with ZK-rollups: proving costs dominate, and if revenue per transaction fails to cover proof generation, the chain spirals. I co-authored a ZK-rollup standardization proposal in 2024, and the number one lesson was: optimize proof cost before anything else. AI companies are ignoring this.
7. Market Sentiment Cascading A 9% drop triggers stop-loss algorithms and margin calls. The dip becomes a self-fulfilling prophecy. This is pure game theory of liquidity: once the sell order book thins, slippage dominates. In crypto, we see this every week. The structural lack of buy-side conviction in AI stocks is a replay of the 2022 crypto bear.
Contrarian: But this overcorrection may present an opportunity. The fundamental technology—large language models, multimodal understanding—is real. The blind spot is that the market currently lacks a mechanism to verify AI model claims. Here is where blockchain can intervene. Imagine a trustless AI inference network where each output is accompanied by a zero-knowledge proof of correctness, proving that the model ran exactly as advertised. If Zhipu or MINIMAX could integrate such a verifiable inference layer, they would restore trust instantly. Privacy is a protocol, not a policy; verifiability is also a protocol. I have seen how Zcash's shielded pool managed to prove transaction correctness without revealing data. The same principle can apply to AI. The contrarian read: the dip is the market pricing in an absence of protocol-level trust. Build that trust, and the valuation resets upward.
Takeaway: The unwinding of AI hype is not a crash; it is a transition from speculation to verification. Over the next 12 months, companies that can demonstrate code-level verifiability (open weights, auditable inference, transparent revenue models) will survive. Those that remain black boxes will bleed. Math doesn't care about your narrative. The proof is in the protocol—or the lack thereof.
