The math whispers what the network shouts. In the AI industry, the network is shouting about valuation corrections, but the math—the actual compute flows, model distillation pathways, and unit economics—tells a quieter, more structural story. A recent CITIC Securities research report attempts to reframe the ongoing tech stock selloff, shifting the blame from macro interest rates to internal industrial variables. It identifies three pricing variables: commercialization pace, compute conversion efficiency, and model gap evolution. And it flags "anti-distillation" as the largest potential variable. This is a useful framework, but based on my years auditing protocol logic and tracing value flows through decentralized systems, the report stops short of the real bottleneck. It treats compute as a monolithic asset, when in fact the binding constraint is not raw FLOPs, but the efficiency of conversion—from compute to model capability, and from model capability to revenue. The market is not just pricing in a slowdown; it is pricing in a fundamental uncertainty about whether the current AI stack can ever achieve the unit economics of traditional software. Trust is not given; it is computed and verified. And right now, the market is trying to verify a thesis that the industry itself has not yet proven.
The CITIC report's core contribution is its analytical pivot. It argues that AI stock valuations have moved from a "technology breakthrough expectation" anchor (the GPT-4 era) to a "commercialization realization" anchor. This is correct. The market's patience window is narrowing. The report correctly notes that OpenAI's annualized revenue surpassing $4 billion is impressive, but inference costs remain high. Anthropic's revenue is growing, but gross margins are under pressure. This is the classic "revenue for market share" phase, where unit economics are unvalidated. The report's hidden implication is that if top players fail to deliver exceptional commercialization data in the next 2-3 quarters, the valuation system could shift from PS multiples to PE logic, triggering a systemic de-rating. This is a sharp observation. However, the report's analysis of the "commercialization pace and scope" variable lacks a crucial distinction: it does not differentiate between vertical deep-diving (achieving excellence in a few scenarios) and horizontal expansion (rapid multi-scenario rollout). In a high-interest-rate environment, horizontal expansion requires massive capital expenditure, making it less viable. The market likely prefers the former, but the report leaves this ambiguous.
The report's most intriguing element is the introduction of "anti-distillation" as a potential game-changer. The logic is that if leading model vendors use technical means—output watermarking, API usage restrictions—to prevent competitors from training on their outputs, the "catch-up path" for smaller AI firms is severed. This would accelerate the industry's shift from a "hundred flowers bloom" to an "oligopoly." From my perspective as someone who has spent years analyzing zero-knowledge proofs and data sovereignty, this is where the report's analysis is both insightful and incomplete. The report treats anti-distillation as a future possibility, but the technical reality is more nuanced. Distillation is not just about copying outputs; it's about extracting the latent knowledge embedded in a model's responses. Watermarking can be evaded. API restrictions can be circumvented. The real moat is not technical restriction, but data exclusivity—the ability to generate proprietary, high-quality interaction data that competitors cannot access. This is the "compute → model → data → compute" positive feedback loop the report hints at but does not fully explore. The report's confidence level of B- for this variable is appropriate, but it underestimates the technical difficulty of implementing effective anti-distillation at scale.
Here is the contrarian angle the report misses: the compute advantage itself is becoming commoditized. The report correctly states that compute is a core production factor, with capital expenditure on compute exceeding 70% of total CapEx for top AI firms. But it fails to acknowledge that the marginal value of additional compute is diminishing. The model capability gap has narrowed from a "generational difference" to an "intra-generational difference." The jump from GPT-4 to GPT-4o is smaller than the jump from GPT-3 to GPT-4. This means that while compute is necessary, it is no longer sufficient for competitive advantage. The report's own evidence supports this: Google has top-tier compute (TPU v5p), yet its AI commercialization lags OpenAI. Compute is a necessary condition, not a sufficient one. The real differentiator is the efficiency of the conversion loop—how quickly a firm can turn compute into a product that users pay for. This is a product and distribution problem, not just a compute problem. The report's framework, while shifting from macro to micro, still over-indexes on the supply side (compute) and under-indexes on the demand side (product-market fit and distribution channels).
Proving truth without revealing the secret itself. The secret the market is trying to uncover is whether AI can achieve the gross margins of SaaS (70-80%) or will remain stuck in the lower-margin world of infrastructure (30-40%). The CITIC report does not answer this. It provides a framework, but not the quantitative signals needed to validate it. Based on my experience auditing DeFi protocols during the 2020 summer, I learned that the most dangerous risks are not the ones you can model, but the ones you cannot see. The same applies here. The report's top three risks—commercialization shortfall, anti-distillation entrenchment, and compute supply chain issues—are all visible. The invisible risk is the narrative premium. How much of current AI valuations is "story premium" rather than fundamental support? The report warns against "excessive grand narratives," but it does not quantify this premium. This is the critical gap. The market is not just pricing in a slowdown; it is pricing in a fundamental uncertainty about whether the current AI stack can ever achieve the unit economics of traditional software. The math whispers what the network shouts. The network is shouting about rate cuts and K-shaped convergence. The math is whispering a more uncomfortable truth: the unit economics of AI are still unproven, and the window to prove them is closing. The next 6-18 months will determine whether AI is a new software layer or just an expensive feature. The market is not just pricing in a slowdown; it is pricing in a fundamental uncertainty about whether the current AI stack can ever achieve the unit economics of traditional software. Trust is not given; it is computed and verified. The market is now demanding the proof.
