Hook
The market narrative around AI stocks has been stuck on interest rates. Yields tick up, Nasdaq bleeds. Yields flatten, tech rebounds. Simple, clean, wrong.
CITIC Securities just published a research report that flips the entire framework. Their core claim: the recent tech correction has nothing to do with Treasury yields. The real variable is internal to the AI industry itself—commercialization pace, compute conversion efficiency, and model gap evolution. And buried in their analysis is one phrase that should terrify every small AI lab and every investor betting on a competitive market: "anti-distillation."
Ledgers bleed, but code remembers the truth.
Context
CITIC's report identifies three verifiable pricing variables for AI stocks: the speed and scope of commercialization, whether compute advantages translate into market share and pricing power, and the trajectory of model capability gaps. They explicitly demote macro factors like US bond yields to secondary status. Their implied thesis: AI equities have entered a "validation period" where the market pays for execution, not imagination.
The report's hidden gem is the "anti-distillation" concept. Distillation—using a powerful model's outputs to train a smaller, cheaper model—has been the great equalizer in AI. It's how startups and open-source projects keep pace with frontier labs without spending billions on compute. If leading model providers can technically block this through output watermarking, API usage restrictions, or legal frameworks, the catch-up path for smaller players gets severed. The industry shifts from "many flowers blooming" to "oligopoly by design."
Core
Let's break down the commercialization variable first because CITIC correctly identifies it as the primary pricing anchor. OpenAI reportedly crossed $4 billion in annualized revenue, yet inference costs remain painfully high. Anthropic grows fast but margin-compressed. This is classic "revenue for market share" economics—unit economics unproven, customer acquisition masking weak depth monetization.
The market's patience window is narrowing. If the next two or three quarters don't deliver above-consensus commercialization data, CITIC implies a valuation system shift from PS multiples to PE logic. That's not a tweak; that's a repricing event. I've seen this movie before—in crypto, when projects move from "narrative phase" to "revenue phase" and 80% of them fail the transition. The LTV/CAC ratios that matter haven't been disclosed by any major AI lab. That silence is data.
On compute conversion efficiency, the report notes that 70%+ of capital expenditure at major AI firms now goes to compute-related infrastructure. But here's the nuance CITIC implies without stating: compute advantage alone doesn't create value. Google holds massive compute superiority yet trails OpenAI in commercialization. Why? Because compute is necessary but not sufficient. Productization, distribution channels, and enterprise sales cycles determine whether raw compute becomes revenue. The conversion efficiency gap between players is where alpha lives.
Now the anti-distillation variable. This is the report's most original contribution, and it's underdeveloped. Let me fill in the gaps from my own operational security background. Anti-distillation isn't just a technical problem—it's an information asymmetry weapon. If OpenAI or Anthropic can watermark outputs in ways that survive fine-tuning, they control the training data pipeline for all downstream competitors. That creates a positive feedback loop: compute → model → exclusive user interaction data → better model → more compute advantage. The gap becomes structural, not incremental.
Contrarian
The market consensus treats AI competition as a multi-player race. CITIC's report suggests otherwise—that we're one anti-distillation implementation away from a winner-take-most outcome. But here's what the report doesn't fully explore: anti-distillation cuts both ways. If leading labs successfully block distillation, they also slow the entire ecosystem's growth. Smaller companies stop building on their APIs, reducing the moats' network effects. Open-source alternatives gain relevance precisely because they're undistillable—the weights are public.
The K-shaped divergence convergence trade CITIC mentions is also worth questioning. The report suggests dollar weakness could trigger capital rebalancing from US AI leaders to other markets, including A-shares. That's a macro overlay contradicting their own "internal variables matter more" thesis. The rebalancing only persists if AI fundamentals support valuation convergence. Otherwise, it's a head fake.
Liquidity is just trust, quantified in gas.
Takeaway
Track three signals: quarterly commercial metrics from frontier labs (revenue growth, gross margin, retention), API term changes or watermarking implementations from OpenAI/Anthropic, and the performance gap between open-source and closed models. Anti-distillation is the variable that rewrites everything. If it succeeds, the model gap locks in and compute becomes a permanent moat. If it fails, the industry stays competitive and the current valuation dispersion corrects. We trade signals, not dreams, in the silence.
Yields vanish when the herd arrives at the gate.