There is a particular silence that falls over a trading floor when the narrative shifts. It is not the loud panic of a crash, but the quiet, collective intake of breath when everyone realizes the story they were telling themselves is no longer sufficient. I felt it in 2022, watching the corpses of overvalued protocols wash ashore, and I felt it again last week, reading a research report from CITIC Securities that dared to suggest the recent tech sell-off was not about bond yields, but about something far more intimate: the slow, painful process of AI growing up.
For months, we have been told the correction was a macro story. A spike in the 10-year Treasury yield, a hawkish whisper from the Fed, a rotation out of growth. It was a clean, comfortable narrative that absolved the industry of any internal sin. But this report, buried in the middle of a broader market analysis, performs a kind of intellectual surgery. It cuts away the macro fat to reveal the bony structure of the problem: AI stocks are no longer being priced for their potential, but for their receipts. The market has stopped asking 'what if?' and started asking 'show me.'
This is the story of that transition. It is a story about the death of the narrative premium, the rise of the 'anti-distillation' moat, and the uncomfortable truth that in a bear market for ideas, execution is the only currency that still spends.
The End of the Imagination Premium
To understand where we are, we must first understand how we got here. The 2023 AI rally was a classic narrative-driven market. The release of GPT-4 was not a product launch; it was a cultural event. It was a Rorschach test for the future, and investors saw in it the shape of AGI, the promise of a productivity revolution, and the obsolescence of entire job categories. In that environment, valuation was a function of imagination. You were not paying for revenue; you were paying for the probability of a paradigm shift. The discount rate was irrelevant because the terminal value was infinite.
CITIC's report, which I have parsed with the care of a curator examining a fragile manuscript, argues that this era is over. The first and most critical variable they identify for AI stock pricing is no longer model capability, but the 'pace and scope of commercialization.' This is a profound shift. It means the market is now demanding that the steep, upward-sloping curve of capital expenditure be met by a corresponding, verifiable curve of revenue. The time lag between these two curves—the 'technical investment curve' and the 'revenue realization curve'—is the central tension of the current market. And the market is repricing that lag with brutal efficiency.
I have seen this movie before. In the DeFi summer of 2020, I led a governance working group for MakerDAO. We analyzed over 500 voting proposals, and we saw the same pattern. Projects with beautiful narratives and no users were valued at billions. The market was paying for the dream of decentralized finance, not the reality of its clunky user interfaces and impermanent losses. When the music stopped, the projects with real revenue—the ones with actual borrowers and lenders—survived. The rest became footnotes. The same Darwinian logic is now being applied to AI. The market is no longer asking if AI is the future; it is asking which specific companies have a viable business model in the present.
The evidence is in the numbers. OpenAI's annualized revenue has reportedly crossed the $4 billion mark, a staggering figure by any historical standard. Yet, its inference costs remain stubbornly high. Anthropic's revenue is growing, but its gross margins are under pressure. This is the classic 'revenue for market share' phase, where unit economics are sacrificed for growth. The market, which once rewarded this land-grab strategy, is now growing impatient. The report hints at a 'narrowing window of patience,' suggesting that if the top players cannot deliver better-than-expected commercialization data in the next two to three quarters, the valuation framework will shift from a price-to-sales (PS) multiple to a price-to-earnings (PE) logic. That shift, when it comes, will not be a gentle correction. It will be a systemic de-rating.
The Tyranny of the Compute Moat
The second variable the report identifies is the 'conversion of compute advantage into market share and pricing power.' This is the cold, hard physics of the AI industry. Compute is no longer just an IT infrastructure line item; it is the core factor of production, the strategic asset that determines who can train the largest models, iterate the fastest, and serve customers at the lowest cost. The report's analysis of the transmission chain—'compute advantage leads to market share, which leads to model gap'—is a stark and accurate description of the current competitive landscape.
I have spent years in the governance trenches, and I have learned that power is rarely distributed evenly. In the AI world, that power is concentrated in the hands of those who control the silicon. The capital expenditure of top AI companies is now over 70% compute-related. This includes GPU procurement, cloud service fees, and data center construction. This is not a technology budget; it is a national defense budget for the digital age. The report correctly notes that this creates a 'generational advantage' for incumbents. New entrants cannot simply buy their way into the game; the supply of high-end GPUs is constrained, the export controls are tightening, and the energy requirements are staggering.
But the report's most intriguing contribution is its focus on 'anti-distillation.' This is the concept that leading model vendors will use technical means—such as output watermarking or API usage restrictions—to prevent competitors from using their outputs to train new models. This is the ultimate moat. It is not just about having more compute; it is about using that compute to create a proprietary data flywheel that is closed to outsiders. If successful, anti-distillation would sever the 'standing on the shoulders of giants' path for smaller AI companies. They would be forced to start from scratch, training base models from zero, which would dramatically increase the barriers to entry and accelerate the industry's march toward oligopoly.
This is where my own experience with curation and authenticity comes into play. In 2021, I curated a small, invite-only DAO called 'The Ethereal Archive.' We rejected the mainstream hype of the NFT frenzy, focusing instead on on-chain provenance as a form of digital storytelling. We manually verified the artistic intent behind 300 unique digital pieces. When the market crashed in 2022, our archive's value remained stable because it was built on genuine cultural connection, not speculation. The same principle applies here. Anti-distillation is an attempt to protect the authenticity of a model's output, to ensure that its 'soul' cannot be cloned. It is a defensive mechanism against the derivative clones that threaten to dilute the value of the original. In a world of derivative clones, curating the soul of your data is the only sustainable strategy.
The K-Shaped Divergence and the Search for Signal
The report's third variable is the 'evolution of the model gap.' It argues that the capability gap between models has narrowed from a 'generational difference' to an 'intra-generational difference.' The jump from GPT-3 to GPT-4 was monumental; the jump from GPT-4 to GPT-4o is incremental. However, the report astutely points out that the gap in inference cost and long-context capability is still widening. This means that even if models converge in raw intelligence, the cost and capability boundaries will still maintain the competitive advantage of the top players. This is a nuanced and important insight. It suggests that the race is no longer about who has the smartest model, but who can serve the most capable model at the lowest price.
This leads to the report's implicit warning about 'K-shaped divergence.' The term refers to a recovery where different segments of the economy grow at vastly different rates. In the context of AI, it suggests that the gap between the 'haves' (those with compute, data, and distribution) and the 'have-nots' (those with only a good idea) will continue to widen. The report hints that a weaker dollar and reduced rate hike expectations could trigger a rebalancing of capital flows from US AI leaders to other markets, including A-shares in China. But it cautions that this rebalancing will only be sustainable if the underlying fundamentals support the valuation convergence. In other words, you cannot arbitrage a narrative gap; you can only arbitrage a fundamental gap.
This is where the report's advice to 'avoid excessive grand narratives' becomes crucial. The market's expectations for AI are now laden with 'grand narrative' components—the imminent arrival of AGI, the promise of a productivity revolution. These narratives have a powerful gravitational pull on valuations. But if they fail to materialize as concrete business results, the risk of a valuation correction will significantly amplify. I have seen this dynamic play out in the crypto markets, where the 'narrative premium' on projects like 'Web3 social' or 'metaverse' evaporated overnight when it became clear that user adoption was not following the hype. The same fate awaits AI companies that are long on vision and short on execution.
The Contrarian View: Compute is Not Destiny
For all its analytical rigor, the CITIC report falls into a subtle trap: it treats compute as a near-deterministic factor. The implicit assumption is that more compute equals a better model, which equals more market share. But my experience in the trenches of governance and protocol design tells me that this is a dangerous oversimplification. Compute is a necessary condition for success, but it is not a sufficient one. The report itself hints at this when it notes that Google, despite having top-tier compute, has not achieved AI commercialization success commensurate with its computational advantage. Why? Because compute is just a raw material. It must be productized, distributed, and supported by a go-to-market strategy. It must be wrapped in a user experience that solves a real problem.
I recall a conversation I had with a builder during the depths of the 2022 bear market. He was working on a decentralized compute network, and he was frustrated that his superior technology was being ignored in favor of less efficient, but more user-friendly, centralized alternatives. He asked me, 'Why don't they see the truth?' I told him that the truth is not enough. You need to translate the truth into a story that people can understand and a product they can use. The same applies to AI. A company with a massive compute cluster but a poor product will lose to a company with a smaller cluster and a brilliant product. The report's focus on 'conversion efficiency' is the key. It is not about who has the most compute; it is about who can convert that compute into market share and pricing power most effectively.
This is the contrarian angle that the report misses. The 'anti-distillation' moat, while powerful, is a defensive strategy. It protects an existing advantage, but it does not create new value. The real opportunity lies in the 'conversion layer'—the ability to take raw compute and turn it into a service that customers are willing to pay for. This is where the 'K-shaped divergence' becomes an opportunity, not just a risk. The companies that can navigate this conversion layer, that can build the bridges between the raw power of the model and the mundane needs of the enterprise, will be the ones that thrive. They will be the curators of the AI age, selecting the right tools for the right jobs and presenting them in a way that is both powerful and accessible.
The Takeaway: From Imagination to Execution
The CITIC report is a valuable document because it marks a turning point in the market's collective consciousness. It is an admission that the era of paying for imagination is over. The market is now in a 'expectation verification period,' where valuations will be determined by verifiable industrial progress, not by macro liquidity. This is a healthy correction. It forces a level of discipline and accountability that was sorely missing during the hype cycle. It separates the wheat from the chaff, the builders from the storytellers.
But we must be careful not to overcorrect. The pendulum of market sentiment always swings too far in one direction before settling in the middle. The report's focus on 'commercialization' could lead to a dangerous short-termism, where companies are punished for making long-term investments in research and development. The 'anti-distillation' moat, if it becomes a standard industry practice, could stifle innovation and lead to a less dynamic ecosystem. We must find a balance between the need for financial discipline and the need for visionary exploration.
As I look at the AI landscape, I am reminded of the early days of the internet. There was a similar bubble, a similar crash, and then a similar period of consolidation. The companies that survived were not necessarily the ones with the most impressive technology or the most compelling narrative. They were the ones that built sustainable business models on top of the new infrastructure. They were the ones that understood that the real value was not in the network itself, but in the services and experiences that the network enabled.
The same will be true for AI. The models are the new infrastructure. The value will be created by those who can build the applications, the services, and the experiences that leverage this infrastructure to solve real problems. The market is now demanding proof of this value creation. It is demanding receipts. And that is a good thing. It is the beginning of the end of the derivative clones and the start of a new era of authentic, sustainable value creation. The question is not whether AI will change the world; it is whether we have the wisdom to build a world that is worthy of the change. The answer, as always, lies not in the code, but in the choices we make as we write it.
