From Missing Out on AI to AI Overweight: The Pricing Lesson of STAR 50's 47% Drawdown in Four Months

(The author, Xue Hong, is Executive Vice President of Xingtu Financial Research Institute)

What changed was the price. The STAR 50 fell from 2,208 points at the end of June to 1,161 points at the end of October, a drawdown of about 47% in four months.

Why would an industry that appears to have no problems of its own be repriced once? The popular explanation is that "it had risen too much" and "sentiment has faded." Both are true, but neither answers the more important question: from what moment did the market stop paying for the same logic?

"The industry is not wrong" is true. But the mistakes that followed all attached themselves to it.

Second-quarter data were quite solid. The STAR 50 rose 75.7% in a single quarter, its largest quarterly gain since its launch; the ChiNext Index rose 36.4%, while the Shanghai Composite rose only 5.2% over the same period. Active equity funds' allocations to electronics and telecommunications rose to 43% and 17% respectively, 60% combined, both the highest since 2015.

These figures are not traces of "getting it wrong"; they are precisely traces of "getting it right." Once the industry trend was identified, capital priced that identification in through concentrated trading.

The macroeconomic environment at the time also made it hard to offer other choices. Consumption and investment data kept weakening, and few industries could still show growth in their financial statements; AI computing power was the one with the hardest fundamentals. Capital was not choosing among several stories; it picked the one with the clearest delivery amid overall dullness.

What made it worse was that the self-criticism in the quarterly reports itself pushed the crowded trade further along: those who thought in the second quarter that they had bought too little kept adding in the third quarter; most of those who lost money in the third quarter also did not dare to truly cut their total position, fearing they would miss the rally again. All they could adjust was the structure of holdings: cut the marginal names and move the money into the few they were most confident about. The total did not shrink; concentration rose instead. Self-criticism always ended up being about "which few stocks should we switch into" rather than "should we have this much position at all." No matter how often the list was changed, risk exposure did not decline, and crowding found no exit.

Once crowding reached this point, the "correctness" at the industry level had already been fully priced. The harder the industry logic, the earlier that day arrives: the more people believe it, the less money is left that can still come in. By this point, the question pricing had to answer had quietly changed: no longer "is it right," but "can it become even more right."

The trading structure of the A-share market compressed this process. Short-selling tools are limited, so those who are right can only express it by buying, and those who are wrong cannot hedge in the opposite direction; T+1 and price limits also make it hard for money that chased in on a given day to reverse course in time. Combined with annual ranking pressure, no one dared to be absent from the main line. Consensus was amplified by this structure, capital squeezed in the same direction, and a rally was completed within a very short window.

After buying power was exhausted, if prices were to keep rising, they could only rely on linear extrapolation. Extrapolation itself is not the problem; the capital market has always used current prosperity to infer future prosperity, and that is precisely how it prices. The watershed is: what will be used to deliver the line that has been extrapolated?

From 2025 to the first half of 2026, delivery looked nothing to worry about. The formula the market gave AI computing power was "capex revised upward—upstream orders rise—share prices rise." In this chain, capex sits on the numerator side: it means demand, orders, and insufficient capacity.

After mid-2026, the same capex was recorded on the denominator side. It needs to be supported by free cash flow and external financing; the longer it is supported, the longer the payback period and the higher the cost of capital tied up. This shift is clearly visible in the financial statements: capex first turns into construction in progress, then into fixed assets, and finally amortizes year by year into depreciation. The same money shows up as demand in the early stage and as amortization and interest in the later stage. The market will not wait until depreciation is fully provisioned before reacting; as soon as it sees capex growth still rising while operating cash flow is not rising in tandem, it will first shorten the discount period.

The same capex figure is an order in the early stage of expansion and a liability in the late stage of expansion.

Another explanation sounds more professional: the decline stemmed from rising U.S. Treasury yields. In October, long-end U.S. Treasury yields briefly broke above 5.3%, and the 30-year broke above 5.7%, both at highs of more than two decades, pushing the discount rate for long-duration assets further up. But the A-share decline had already begun before that. The problem in this round was not interest rates; the rise in rates only made an already fragile position harder to sustain.

There was no reversal at the industry level; the market switched to a different statement: from the income statement to the cash flow statement, from growth rate to duration.

The switch did not happen in an instant; it had an order. First, valuations were killed: once the discount rate changed, the present value of the most distant cash flows fell the most; second, earnings expectations were killed: once capex slowed, upstream orders no longer had incremental growth; only last did it reach the industry logic itself. That is the meaning of "pricing was wrong": the company is still the same company, but the price has been revalued once.

Therefore, the judgment that "the industry logic is wrong" often arrives last. Tops do not form when the public gets it wrong, but when the public gets it right and prices that "right" to the extreme. When people get it wrong, there is still disagreement in the price, and disagreement is a buffer; when people get it right and are one-sided, the buffer disappears, and any signal even slightly below expectations is enough to trigger a revaluation.

Third-quarter financial reports and share prices are the most direct evidence of this revaluation: the reports were still improving, while the share prices had already turned down. The difference is not in the company, but in the fact that the two are not using the same statement.

The most typical case is Cambricon. Revenue in the first three quarters was 9.526 billion yuan, up 106.76% year on year; net profit attributable to the parent was 3.715 billion yuan, up 131.49% year on year, with the single quarter still hitting a new high. In the third quarter its share price fell 36.8%, and Zhongji Innolight and Eoptolink also fell more than 35%.

The market is not refusing to recognize these results; it is just that these results are recorded on the income statement. Cambricon's profit is the receipt for downstream capital expenditure over the past four quarters, while the market is trading the next four quarters. When the growth rate of capital expenditure begins to be questioned, no matter how beautiful the receipt is, it is only an account of the past.

At the industry level, it is clearer: electronics and communications, which led the gains in the second quarter, became the top two decliners in the third quarter; petroleum and petrochemicals, coal, and banks, which fell in the second quarter, all turned to gains in the third quarter.

This is not "AI is failing," but rather funds at the same time re-choosing which kind of certainty to pay for. On one side is growth that is still expanding and whose returns take longer to arrive; on the other is dividend assets with modest growth but immediately available cash flow. After the discount rate rises, dividend assets are suppressed less, so their relative prices are higher.

Beyond price, for those holding these positions, there is another layer of mismatch, and many people are stumbling over it.

The realization of an industry trend is measured in five or ten years, while portfolio assessment is measured in one year or one quarter. An industry judgment can be correct and at the same time be completely ineffective during this period. The problem is not the judgment itself, but that the time scale of the judgment is inconsistent with the time scale of the account.

After a drawdown, the industry judgment does not need to be redone; what needs to be answered anew are two more specific questions.

The first question is: who ultimately pays for this logic. The ultimate payers for AI computing power are cloud vendors: domestically mainly Alibaba, Tencent, and ByteDance; overseas mainly Amazon, Microsoft, and Google. Their capital expenditure depends on their own cash flow and financing costs. Once the payers tighten, the high growth upstream is borrowed. Rather than watching upstream orders, it is better to watch whether downstream capital expenditure can continue.

The second question is: to what extent has this "correctness" already been priced in. Three positions can be examined: valuation percentile, institutional holdings, and share of trading volume. At the end of June, the valuation percentiles of electronics and communications were both above 90%, the electronics holdings percentile of active funds reached 100%, and TMT accounted for 44% of market-wide trading volume. All three were simultaneously near historical extremes, so subsequent room could only be opened by "more exceeding expectations," and "more exceeding expectations" is precisely the least sustainable.

Both questions have data that can be checked, but there is one signal that is more direct: leading companies' results beat expectations, yet their share prices fell. It means the market is no longer paying for "growth rate" but has begun paying for "visibility." In the third quarter of 2026, this signal appeared almost simultaneously across the computing power chain; this is not an accident for one company, but the replacement of the pricing benchmark for an entire chain.

Returning to the question at the beginning: for the same batch of stocks, up 75.7% in the second quarter and down about 47% four months later, the industry itself has not changed. What has really changed is not the industry, but the money the market is willing to pay for it, and the statement it looks at when paying.

If this round is broken down, it should be divided into at least three accounts: the industry account, the pricing account, and the account account. The industry judgment answers "whether it will happen"; the pricing judgment answers "how much has already been paid, and who else is willing to keep paying." The account account answers a more realistic question: given my assessment cycle, can I wait until it is realized. Only when the three accounts are mixed together does one mistake "getting the industry right" for "buying the price right." The industry can be very right, while the pricing can at the same time be very wrong.

For this reason, when the next round of narrative arrives, whether it is humanoid robots or other themes, one should not jump directly from "is the industry right" to "should I go heavily long." A more prudent order is: first see who ultimately pays for this logic, then see to what extent this "correctness" has already been priced in, and finally see whether one's own account time scale can wait until it is realized.

The most expensive lesson of this round is not just chasing highs, but mistaking an industry judgment for a pricing judgment. There are many people who get the industry right; there are few who, while the industry is still correct, can clearly explain to what extent this "correctness" has been priced in. Next round, it is worth asking before the money has been paid out: who ultimately pays, to what extent this "correctness" has been priced in, and whether my account can wait until it is realized. Only by answering this question does the industry judgment have somewhere to land.

Exclusive first release on Yicai's Yicai Account; this article represents only the author's views and does not constitute investment advice.