Over the past 90 days, a subtle but seismic shift has occurred in how institutional capital prices AI equities. The correlation between AI stock performance and 10-year Treasury yields has weakened by an order of magnitude, while the correlation with quarterly revenue growth guidance has strengthened proportionally. This is not a market anomaly. It is a regime change. The market has stopped paying for imagination and started paying for execution. Based on my experience auditing protocol economics during the DeFi summer of 2020, I can tell you this pattern is familiar: when the narrative matures, the unit economics take over. The question is whether the current AI cohort can survive the transition.
CITIC Securities recently published a research note that attempts to formalize this shift. The report's core thesis is that AI stock pricing has moved from a macro-driven model to a fundamentals-driven model, anchored on three verifiable variables: commercialization pace, compute conversion efficiency, and model gap evolution. The report also identifies "distillation" as the largest potential variable—a term that deserves far more technical scrutiny than it has received. This analysis framework is directionally correct, but it lacks the granularity required for actual position sizing. Let me break down what the report gets right, what it misses, and where the real risks lie.
The Commercialization Reality Check
The report correctly identifies commercialization as the primary pricing variable. This is not a controversial claim; it is an observable fact. OpenAI's annualized revenue has crossed the $4 billion threshold, but inference costs remain stubbornly high. Anthropic's revenue is growing rapidly, yet gross margins are under pressure. These are classic signs of a market-share acquisition phase, not a unit-economics validation phase. The market is now asking a question it did not ask in 2023: what is the customer lifetime value to customer acquisition cost ratio for an enterprise AI deployment? The answer, based on my analysis of available data, is that we do not yet have a clear answer. This uncertainty is the root cause of the valuation compression we are witnessing.
The report's framing of "commercialization pace and scope" as a single variable obscures a critical distinction. There are two paths to commercialization: vertical depth (dominating a few high-value use cases) and horizontal expansion (spreading across many scenarios quickly). The market is currently signaling a preference for the former. Horizontal expansion requires massive capital expenditure, and in a high-interest-rate environment, that capital is expensive. Vertical depth, by contrast, can be achieved with more modest resources and generates higher margins per unit of revenue. The report does not explicitly state which path it believes the market will reward, but the implication is clear: companies that can demonstrate deep, defensible adoption in specific verticals will command higher multiples than those pursuing broad but shallow deployment.
The Compute Conversion Problem
The second variable—compute conversion efficiency—is where the report's analysis aligns most closely with my own technical background. The claim that compute advantage translates into market share and pricing power is empirically supported. Google DeepMind's Gemini series and Anthropic's Claude series both demonstrate a positive correlation between compute investment intensity and model market performance. But the report misses a crucial nuance: compute is a necessary condition, not a sufficient one. Google has arguably the best compute infrastructure in the world, yet its AI commercialization lags OpenAI. This is not a compute problem; it is a productization and distribution problem. Compute only creates value when it is converted into products that customers will pay for. The conversion efficiency varies significantly across companies, and this variance is not captured in the report's framework.
There is also a temporal dimension to compute advantage that the report does not address. The current model capability gap has narrowed from a generational difference to an intra-generational difference. The jump from GPT-3 to GPT-4 was transformative; the jump from GPT-4 to GPT-4o is incremental. However, the inference cost gap and long-context capability gap are widening. This means that even if model capabilities converge, cost and capability boundary differences are sufficient to maintain competitive advantages for incumbents. The report touches on this but does not fully explore the implications for new entrants. In a market where the top players have locked in multi-year compute supply agreements, the barrier to entry is not just technical skill; it is capital access.
The Distillation Blind Spot
The report's identification of "distillation" as the largest potential variable is the most intellectually interesting part of the analysis. Distillation, in this context, refers to the practice of using a large model's outputs to train a smaller or competing model. This is a common technique in the AI industry, and it has been a primary pathway for smaller players to catch up with the frontier labs. The report suggests that if frontier labs successfully implement technical measures to prevent distillation—such as output watermarking or API terms of service restrictions—the catch-up path for smaller AI companies would be severed. This would accelerate the industry's movement from a fragmented landscape toward oligopoly.
From a technical perspective, the feasibility of anti-distillation measures is questionable. Output watermarking can be detected and stripped. API restrictions can be circumvented through careful prompt engineering. The cat-and-mouse game between distillation and anti-distillation is likely to be a perpetual arms race, not a decisive victory for either side. However, the report's deeper point stands: even imperfect anti-distillation measures raise the cost of entry for smaller players. If a small AI company must spend 10x more on data acquisition and training to achieve the same result, the competitive dynamics shift significantly. This is a real risk, and it is underappreciated in the broader market narrative.
The report also hints at a geopolitical dimension to the distillation question. In the context of export controls on high-end GPUs, Chinese AI companies have relied heavily on distillation from Western models as a catch-up pathway. If this pathway is closed, the gap between Chinese and Western AI capabilities could widen significantly. The report does not explicitly state this, but the implication is clear. This is a risk that institutional investors should be tracking, particularly those with exposure to Chinese AI equities.
The Valuation Framework Shift
The report's most valuable contribution is its reframing of the valuation debate. By shifting the attribution of tech stock adjustments from external macro factors (Treasury yields) to internal industry variables (commercialization, compute conversion, model gap), the report provides a more actionable framework for investment decisions. The implicit investment logic is that AI stocks have entered a "expectation validation period" where valuations will depend more on verifiable industry progress than on macro liquidity. This means investment strategy needs to shift from beta-driven sector allocation to alpha-driven stock selection.
The report's dismissal of Treasury yields as the root cause of tech stock adjustments is a bold claim that deserves scrutiny. My analysis suggests that interest rates still matter, but their influence has been overstated. The 2023 AI rally was primarily driven by narrative and liquidity; the 2024 correction is primarily driven by fundamentals. This is a healthy development. It means that companies with real revenue growth, improving margins, and strong customer retention will be rewarded, while companies with only technical narratives will be punished. The market is becoming more discriminating, and this is a positive sign for long-term investors.
The K-Shaped Divergence
The report's mention of "K-shaped divergence convergence" is a subtle but important signal. The K-shaped pattern refers to the divergence between AI leaders and laggards, which has been a defining feature of the current market cycle. The report suggests that a weaker dollar and reduced rate hike expectations could trigger a rebalancing of capital from US AI leaders to other markets, including A-shares. This is a plausible scenario, but its sustainability depends on whether AI industry fundamentals support valuation convergence. If US AI leaders continue to deliver superior commercialization results, the rebalancing will be temporary. If their results disappoint, the rebalancing could be structural.
The Unanswered Questions
Despite its analytical rigor, the report leaves several critical questions unanswered. First, what is the current unit economics of AI commercialization? The report does not provide specific metrics on LTV/CAC, gross margins, or customer lifetime value. Second, what is the conversion rate from pilot to full deployment for enterprise AI customers? The report mentions the Copilot penetration controversy and Salesforce Einstein GPT adoption rates but does not provide a quantitative framework for tracking these metrics. Third, does AI commercialization exhibit winner-take-all characteristics, or will it form a multipolar landscape across vertical scenarios? The report does not take a definitive position on this question.
The Contrarian Angle
The contrarian angle here is that the market's focus on commercialization may be premature. The AI industry is still in its early innings, and the most valuable applications may not have been invented yet. The internet went through a similar phase in the late 1990s, where companies with questionable business models were punished, only for the real winners to emerge years later. The current focus on near-term revenue may be causing the market to undervalue companies with long-term strategic positions but delayed monetization. This is a risk for investors who are too focused on quarterly results.
There is also a contrarian angle on the distillation question. If anti-distillation measures are successfully implemented, it could actually benefit the open-source ecosystem. Open-source models like Llama and Qwen have already achieved remarkable performance levels, and they are not subject to the same distillation restrictions as closed-source models. If the closed-source frontier labs close their outputs, the open-source community may become the primary vehicle for AI innovation diffusion. This would be a positive development for the industry as a whole, even if it is negative for the closed-source labs' near-term competitive position.
The Takeaway
The CITIC Securities report provides a useful framework for understanding the current AI valuation regime, but it is not a complete investment playbook. The three variables it identifies—commercialization pace, compute conversion efficiency, and model gap evolution—are all important, but they need to be supplemented with more granular data and scenario analysis. The report's identification of distillation as the largest potential variable is intellectually stimulating, but its analysis of this topic is too shallow to be actionable.
My recommendation is to focus on the signals that are actually measurable. Track the quarterly reports of OpenAI, Anthropic, Microsoft, and Google for revenue growth, gross margins, and customer retention. Monitor the API terms of service changes and technical measures implemented by frontier labs. Watch the performance gap between open-source and closed-source models. These are the data points that will determine the next major move in AI equities. The market has shifted from paying for imagination to paying for execution. The companies that can demonstrate execution will be rewarded. The companies that cannot will be left behind. This is the new reality, and it is not going to change anytime soon.