There is a moment in every technology cycle when the market stops asking 'what if' and starts asking 'show me.' For artificial intelligence, that moment arrived quietly, buried inside a recent CITIC Securities research report that most retail investors will never read. The report does not make headlines with dramatic predictions. Instead, it performs something far more consequential: it redefines the very metrics by which AI companies will be valued. As someone who spent years auditing smart contracts for hidden centralization, I recognize this pattern. The market is conducting its own audit of the AI industry, and the findings are uncomfortable for those who bought the narrative without examining the code.
The report's central thesis is that AI stock pricing has shifted from macro factors—like US Treasury yields—to industry fundamentals. Three variables now dominate: the pace of commercialization, the efficiency of compute conversion, and the evolution of model gaps. But the most intriguing element is what the report identifies as the 'largest potential variable': anti-distillation. This is the practice where leading model developers use technical means—output watermarking, API usage restrictions—to prevent competitors from training new models on their outputs. It is, in essence, an attempt to build a moat around knowledge itself.
Let me be direct about what this means. The commercialization variable is the first test. The report notes that OpenAI's annualized revenue has surpassed $4 billion, yet inference costs remain stubbornly high. Anthropic's revenue grows quickly, but gross margins are under pressure. This is the classic 'revenue for market share' phase, where unit economics remain unproven. The market's patience window is narrowing. If the next two to three quarters do not deliver better-than-expected commercialization data, the valuation framework could shift from price-to-sales multiples to price-to-earnings logic. That shift would trigger a systematic de-rating across the sector.
The core insight here is that AI companies have not yet established pricing power. Current pricing models are cost-plus—per token, per seat. There is no mature value-based pricing that directly ties fees to customer outcomes. This is not a minor detail. In my years auditing ERC-20 standards, I learned that the absence of a fair mechanism is not neutrality; it is a bias toward whoever controls the infrastructure. Similarly, the absence of value-based pricing in AI means the companies cannot yet capture the value they create, and the market knows it.
The second variable—compute conversion—is where the report gets genuinely interesting. The claim is that compute advantage translates into market share through three channels: training scale, iteration speed, and inference cost. This is empirically sound. Google DeepMind's Gemini series and Anthropic's Claude series both validate the correlation between compute intensity and model performance. But the report also hints at a deeper concern: the model gap has narrowed from 'generational' to 'intra-generational.' The jump from GPT-3 to GPT-4 was massive; the jump from GPT-4 to GPT-4o is incremental. Yet inference cost gaps and long-context capability gaps are widening. This means even if model capabilities converge, cost and capability boundaries can sustain the incumbents' advantage.
The third variable—anti-distillation—is where the report's analysis becomes both prescient and incomplete. The report correctly identifies that if leading labs successfully implement anti-distillation, smaller AI companies lose the 'standing on giants' shoulders' path to catch up. They would be forced to train foundation models from scratch, dramatically raising entry barriers and accelerating market concentration. But the report does not adequately address whether anti-distillation is technically feasible at scale. Based on my experience with blockchain protocols, I can say this: any attempt to restrict the use of public outputs faces fundamental technical challenges. Watermarks can be stripped. API restrictions can be circumvented. The cat-and-mouse game between model providers and users is just beginning.
Here is where I must offer a contrarian perspective. The report's framework, while valuable, may be too focused on the Western AI giants. It implicitly acknowledges this by discussing 'K-shaped divergence' and the potential for capital to rotate from US AI leaders to other markets, including A-shares. But the deeper question is whether the compute-to-model-gap transmission chain is as deterministic as the report suggests. China's AI industry, operating under compute restrictions, is exploring alternative paths: algorithmic innovations like Mixture-of-Experts architectures, quantization techniques, and domestic chip substitutes. These are not perfect substitutes for high-end GPUs, but they may partially offset the compute disadvantage. The report's confidence level of B- (medium-high) is appropriate, but I would argue the uncertainty around anti-distillation is higher than the report admits.
The most significant insight from this report is the shift from 'paying for imagination' to 'paying for execution.' This is not merely a market cycle adjustment; it is a fundamental change in how technological progress will be valued. The era where a model demo could drive a stock price is over. What matters now is customer retention, gross margin improvement, and the boring, unglamorous work of enterprise deployment. The report's advice to 'avoid excessive grand narratives' is a warning against narrative inflation. The market has priced in AGI timelines and productivity revolutions that may not materialize on schedule.
I have seen this pattern before. In 2017, I audited token standards that promised decentralization but concentrated power in validator nodes. The code was elegant; the incentives were not. Today, AI companies promise intelligence, but the economics may concentrate power in those who control compute and data. The parallel is uncomfortable but instructive. Ethics is not a feature; it is the foundation. If the AI industry does not address the unit economics and the anti-distillation dilemma, it will repeat the mistakes of the crypto industry: building cathedrals of hype on foundations of sand.
Looking forward, the next 12 to 24 months will determine whether AI becomes a transformative technology or a cautionary tale. The signals to watch are not model benchmarks but quarterly reports: revenue growth, gross margins, customer retention rates. The 'killer app' that drives standardized deployment remains elusive. The compute gap may or may not become an irreversible model gap. And the global regulatory frameworks—the EU AI Act, China's model filing requirements—will shape the competitive landscape in ways that are difficult to predict.
The market is no longer paying for imagination; it is paying for proof. The question is not whether AI will change the world—it will. The question is whether the companies that dominate the narrative today will dominate the economics tomorrow. Based on my experience watching the blockchain industry's rise and fall, I would say the answer is uncertain. The technology is real, but the business models are not yet proven. The next few quarters will be a reckoning. And as always, the truth will emerge not from the headlines, but from the silence between the blocks—the quiet data points that reveal whether the emperor has clothes.
We are building libraries, not empires. The question is whether we will remember that before the next bubble bursts.