The Ledger Remembers What the Market Forgets.
Over the past 12 months, I have watched institutional capital rotate through AI equities with the same pattern I observed in crypto during the 2021 bull run: narrative leading, fundamentals trailing, and the gap between the two eventually closing with force. CITIC Securities' latest research note on the AI sector adjustment is notable not because it predicts a crash, but because it reframes the cause. The report explicitly shifts the attribution of tech stock corrections from external macro factorsโspecifically US Treasury yields โ to internal industry variables. This is a subtle but significant change in the consensus ledger.
As a macro strategy analyst who has spent the last five years tracking liquidity flows across crypto and traditional markets, I recognize this shift. It is the same transition crypto underwent in 2022 when narrative-driven valuations collapsed and only protocols with verifiable revenue survived. The market is now asking AI companies the same questions the market asked DeFi protocols: what is your revenue, what is your gross margin, and who is actually paying you.
The CITIC report identifies three verifiable pricing variables: the pace of commercialization, the efficiency of computing power conversion, and the evolution of model gaps. It identifies "distillation countermeasures" โ technical means to prevent competitors from training models on your output โ as the largest potential variable. The report's core conclusion: AI has entered an expectation verification period where valuation will depend on verifiable industrial progress rather than macro liquidity.
This is the framework I have been applying to digital assets for years. And it is worth examining closely.
Context: The AI Industrial Ledger
Let me establish the baseline. The AI industry is currently operating under a specific set of constraints that map directly onto the liquidity frameworks I have used in crypto. The current situation resembles the DeFi summer of 2020, but with more institutional participation and higher capital requirements.
The global AI infrastructure market is currently characterized by GPU shortages, energy constraints, and export controls. The supply chain tension in this sector โ GPU supply, export controls, energy constraints โ is reminiscent of the mining hardware cycle of 2021, except that the stakes are higher. The capital expenditure of top AI firms on computing power now exceeds 70% of their total budget. This is not infrastructure. This is the core production factor.
The market has transitioned from the "technology breakthrough expectations" of 2023 โ which anchored valuations on GPT-4 release and multimodal progress โ to the "commercialization realization" anchor of 2024. The market tolerance for AI companies with only technical narratives has dropped significantly. My data shows this shift is real and measurable: the rate of revenue growth in top AI companies still relies on new customer acquisition rather than deep monetization of existing customers.
Let's take the publicly available numbers. OpenAI's annualized revenue has surpassed $4 billion, but inference costs remain high. Anthropic's revenue is growing rapidly, but gross margins are under pressure. The unit economics have not yet been verified. The industry is still in the "revenue for market" phase. This is where I start to see the structural problem.
The corporate client adoption metrics tell a similar story. Microsoft's Copilot penetration rates have been disputed. Salesforce's Einstein GPT actual adoption rates remain unclear. Enterprise AI budgets are growing, but the deployment speed is slower than early optimistic expectations. The gap between "pilot" and "production deployment" is wider than the market narrative suggests. This is the same pattern I observed in DeFi when "total value locked" was growing, but the actual user activity remained concentrated in a small number of protocols.
Core: The Computing Power and Commercialization Pipeline
The CITIC report makes a valuable point: the computational advantage to market share chain has become the core competitive logic in the AI industry. Compute is the barrier. The barrier is the pricing power. This logic is remaking the value chain.
I have seen this dynamic in Bitcoin mining. Compute โ or hash power โ directly determines market share. The same applies to AI. Companies with the most computing power iterate models faster, offer services at lower costs, and respond to customers more flexibly. Together, these three factors convert into market share. Google DeepMind's Gemini series and Anthropic's Claude series both demonstrate this logic: the intensity of computational investment correlates positively with model performance in the market.
But there is a critical nuance that the report captures. The model gap has narrowed from "generational difference" to "intra-generational difference" โ the jump from GPT-4 to GPT-4o is less significant than the jump from GPT-3 to GPT-4. However, the inference cost gap and the long-context capability gap are still widening. Even if model capabilities converge, the differences in cost and capability boundaries are enough to maintain the competitive advantages of top companies.
This is where "distillation" comes in. The CITIC report identifies this as the largest potential variable. The core concept: if leading model vendors use technical means โ such as output watermarking, API terms restrictions โ to prevent competitors from using their outputs to train new models, the "catch-up path" for smaller AI companies will be cut off. The industry could accelerate from "all flowers blooming" to "oligopoly".
The deeper implication is this: if model gaps are solidified through anti-distillation, the speed of AI innovation diffusion will significantly slow down. This has a particularly profound impact on the Chinese AI industry, which relies on a "open source + distillation" path to catch up. In the context of computing power controls, the question is whether China's AI industry can find an alternative path around the "computing power โ model gap" transmission chain.
I have studied this dynamic in the context of crypto mining. When China banned Bitcoin mining in 2021, the computing power rebalanced globally. But the model here is different. In AI, the data is the key. If the leading model vendors successfully build a "data moat" at the model level, their computing power advantages will be expressed not only in model training but also in exclusive access to high-quality training data โ user interaction data. This creates a "computing power โ model โ data โ computing power" positive feedback loop. The barrier to entry becomes structural.
The report also touches on a question that I consider critical: whether computing advantages can be converted into market share and pricing power. The computing power itself does not directly create value. Only through productization, channels, and service systems can it be converted into commercial value. This explains why Google has top-tier computing power but its AI commercialization progress is not as advanced as OpenAI's. Computing power is a necessary condition, not a sufficient condition.
Contrarian: The Decoupling Thesis
Here is where I have to take a contrarian stance on the market's current pricing. The market is still treating AI stocks as if they were a single asset class โ all rising and falling together. That is the wrong framework. The CITIC Securities report is moving in the right direction by differentiating the three variables, but I think it does not go far enough.
The valuation anchor has shifted. AI stocks are entering an "expectation verification period". The market is shifting from paying for imagination to paying for execution. This is exactly what happened in the crypto market in 2022 when the FTX contagion and Terra/Luna collapse forced the market to differentiate between protocols with real usage and protocols with only narratives. The differentiation was violent. I expect the same in AI.
The market's K-shaped divergence will continue. In the crypto market, we saw the K-shaped divergence between Bitcoin and altcoins in 2023. Bitcoin, with its institutional adoption narrative and ETF flows, recovered to new highs. Most altcoins never returned to their previous highs. The same pattern is emerging in AI. Companies with verifiable commercial traction and clear paths to profitability will continue to grow. Companies with only technical narratives will face valuation compression.
The second contrarian angle is about the "distillation" variable. The market has not yet priced in this risk. If distillation measures become industry standards, the competitive landscape changes fundamentally. The open-source ecosystem's ability to catch up will be severely constrained. This means the gap between open-source and closed-source models may widen, not narrow. And that will lead to a structural consolidation of the AI industry that the market has not yet fully priced in.
The third angle: the CITIC Securities report suggests that the US bond yield is not the fundamental cause of the tech stock correction. My take is that this is partially correct but potentially misleading. Even if the rate environment improves, AI stocks without commercial validation will struggle to achieve valuation recovery. But it's also true that the macro environment affects the willingness to pay for long-duration assets. The correction in AI stocks is not caused by the rate environment alone, but the rate environment determines the risk appetite that is applied to the "narrative premium" in valuations.
Takeaway: Positioning for the Verification Phase
The market is now in the verification phase. We do not build on hype; we build on consensus.
For the next 6-18 months, I will be watching the following signals: the Q3 earnings reports of major AI companies (OpenAI, Anthropic, Microsoft, Google) โ revenue growth, gross margins, customer retention rates; the actual execution of enterprise AI budgets; and any changes in API terms or technical implementations that indicate distillation efforts.
The takeaway is this: the market will increasingly separate the "narrative premium" from "fundamental support". The AI stocks that can demonstrate clear commercial paths and verifiable revenue growth will get premium valuations. The rest will face compression. This is the same pattern I have seen in crypto โ from the ICO boom to the DeFi summer to the NFT cycle. The market always returns to the ledger.
The final question is the one the CITIC report implicitly raises but does not answer: how much of current AI stock valuations is "narrative premium" rather than "fundamental support"? My estimate is that for the average AI stock, the narrative premium is still significant. This means the downside risk is also significant if the narrative does not convert into tangible commercial results. The ledger remembers what the market forgets. In the next 12-24 months, we will see who has built the fundamental base and who is just running on narrative โ and the ledger will not lie.