CATL earned a net profit of 21.2 billion yuan in a single quarter, yet fell nearly 15% in two days: what exactly is the AI rally killing?

A company with 21.2 billion yuan in net profit attributable to shareholders in a single quarter saw its stock fall nearly 15% over two trading days; a company that forecast profit growth of nearly sixfold saw its shares turn positive the day after hitting the limit down. In the same market, what determines a stock price is no longer how fast growth is, but the quality of earnings.

CATL's recently disclosed third-quarter report shows that in Q3 CATL's revenue was 165.221 billion yuan, up 58.58% year on year, a record high for a single quarter, with the first three quarters totaling 442.137 billion yuan.

The problem lies on the profit side: the consolidated gross margin declined for three consecutive quarters, from 24.82% in Q1 to 23.15% in Q2 and then to 20.69% in Q3, a new low since Q4 2022; single-quarter net profit attributable to shareholders was 21.246 billion yuan, up 14.54% year on year, below the lower bound of all mainstream institutional forecasts before the report was disclosed—Bernstein expected 22.9 billion yuan, Goldman Sachs 24 billion yuan, and CLSA 25 billion to 26 billion yuan.

In the earnings call, management attributed the profit pressure to three factors: long-term energy-storage contract orders signed during the low-price cycle of the previous two years were concentrated in Q3 deliveries; a 2% lithium-battery consumption tax took effect in September; and the battery export tax rebate rate was cut from 9% to 6% starting in April. The first was characterized as a one-off cost mismatch, while the latter two are long-term variables.

These explanations failed to stop the sell-off. On October 21, CATL's A-shares closed down 12.03% with 32.4 billion yuan in turnover, the second-largest single-day drop since listing, wiping about 140 billion yuan off its market value in one day; on October 22 it continued falling to 220.41 yuan, down nearly 15% over two trading days and nearly halved from its year-to-date high of 467.35 yuan on May 7.

On the same day, another class of assets moved in the opposite direction: Quectel hit the limit up for a second straight session, while Bestechnic and Rockchip rose more than 5%, with on-device AI the most concentrated destination for funds. The gains were driven by two factors: DeepSeek open-sourced its inference framework DeepInfer the previous day, leading the market to expect lower inference costs; and Counterpoint forecast that generative-AI phone penetration will reach 52% by 2027.

ChinaAMC STAR 50 ETF received another nearly 1.5 billion yuan in net inflows that day; even as the index fell, funds kept up net subscriptions.

After the close on October 12, optical-chip company Yuanjie Technology issued an earnings forecast, expecting Q3 net profit attributable to shareholders of 403 million to 453 million yuan, up 578.85% to 659.70% year on year, with single-quarter revenue growth above 264%. Its stock had previously hit the 20% limit down on October 8; on the first trading day after the forecast was disclosed it opened down 6.15% and closed up 3.64%. That same evening, Lianxun Instruments forecast first-three-quarter net profit of 1.9 billion to 2.1 billion yuan, up 1,866% to 2,073% year on year; it closed up 4.37% the next day, reclaiming the title of A-share king. Companies with high earnings growth still won fund recognition after disclosure.

The correction has lasted nearly a month. From September 22 to October 22, the Shanghai Composite fell 7.39% cumulatively and the ChiNext Index fell 24.20%, with the communications and electronics sectors both down more than 20%; in the week of October 13-16, the STAR 50 fell 6.77%, its fourth straight weekly decline, the ChiNext Index fell 8.90% for the week, the CSI Dividend Index fell 0.37% that week, and banks rose 1.52%.

In early October the style divergence was once very clear: as of October 12, the CSI Dividend Index was up 2.29% and banks up 2.28%, with several bank stocks hitting record highs; but this defensive rally failed to last more than two weeks. After peaking on October 14, the CSI Dividend Index gave back all its gains and turned to a 1.42% decline for October, and funds in the defensive direction also began to loosen.

Companies whose earnings greatly beat expectations were bought by funds, yet sector valuations kept moving lower. What this correction is killing determines how to assess the AI sector over the next quarter. The answer indicated by the divergence is broadly clear: what is being compressed is valuation; the earnings already delivered by the computing-power chain have not deteriorated; what has been shaken is the market's confidence in the sustainability of growth, and the logic by which the sector was previously valued on the AI narrative.

What has fallen is valuation

In the week of October 13-16, the Shanghai Composite Index fell 2.56%, the CSI 300 fell 3.40%, the Shenzhen Component Index fell 6.64%, the STAR 50 fell 6.77%, the STAR Composite Index fell 8.26%, and the ChiNext Index fell 8.90%. The size of the declines was positively correlated with growth attributes, and the style divergence was clear.

Figure 1: Performance of major A-share indices in the week of October 13-16. Growth names led the decline, and the size of the drops was highly correlated with duration characteristics. Source: public exchange market data

The first big bearish candle appeared on October 8, a week before the release of US September CPI. The trigger was a Morgan Stanley optical-module research note dated October 1, which reasoned that starting with the 3.2T generation, only if the US-origin value share in the bill of materials is no less than 65% could a product possibly obtain a US market exemption; the impact on current 800G and 1.6T products is limited. The scenario itself was relatively mild, yet optical-chip names fell sharply in response: Yuanjie Technology and Everbright Photonics hit the 20% limit down, and the STAR 50 fell nearly 5% in one day.

On October 13, the US FCC's 26-50 new rules took effect: equipment carrying logic-bearing hardware components produced by companies on the controlled list would be unable to obtain new equipment certification. Checking the original rule text against the list, Zhongji Innolight, Eoptolink and Yuanjie are not on it; the direct pressure falls on equipment in the Huawei and ZTE ecosystems, as well as routers, wearables, drones and other categories. What the market was pricing was the expectation of whether the list would expand. A week later, September CPI came in above expectations and rate-hike expectations heated up, becoming the second layer of pressure during the correction.

This is where the divergence arises: a rising discount rate is pressure on the denominator side, while what the research-note scenario and the expectation of an expanded FCC list shake is the sustainability of overseas revenue—that is a matter for the numerator side. At the company level, fundamentals have not deteriorated so far: Yuanjie Technology says shipments of its data-center CW products continue to grow and gross margins remain high, and management says demand visibility for the next two years is very high; the elasticity at Lianxun Instruments likewise comes from real orders. The same company is simultaneously under valuation pressure from overseas regulation and deepening its overseas ties—put the two directions together, and what they point to is valuation.

The divergence in valuation is also clear. As of October 16, the ChiNext Index's PE-TTM was about 32x, down 40% from about 54x in June; the STAR 50 was above 110x, down nearly 60% from its June high of about 264x—an absolute level that is not low. The correction has brought the AI sector's valuation down from its highs, but the valuation digestion of core computing-power assets is not yet complete. Among the core companies in the computing-power chain, no public case of a downward revision to earnings guidance has yet been seen.

Figure 2: PE-TTM coordinates for the ChiNext Index and STAR 50. ChiNext has returned to a historically low percentile, while the STAR 50, with a higher weighting in the computing-power chain, is still in its digestion phase. Source: Wind, as of October 16, 2026

The STAR 50's top ten weights total 59.66%, covering chip design, wafer manufacturing and semiconductor equipment. Cambricon weighs about 10%, and the only optical chip company, Yuanjie Technology, ranks sixth. It is essentially a shadow index of a basket of AI computing-power assets: a 6.77% drop equals dozens of computing-power stocks falling simultaneously on the same logic. Half of the market's worry comes from position rebalancing under a discount-rate re-rating, and half from the threat overseas regulation poses to revenue expectations.

Industry boom data are still pointing upward. SIA data show global semiconductor sales of $159.7 billion in August, up 144.3% year on year; under customs data, China's integrated circuit exports grew 108.1% in the first three quarters. But the increment in both sets of data comes mainly from memory price increases and re-exports, and institutional reviews also list concerns about the memory supply landscape as one of the triggers for this correction. Bulls and bears cite the same set of data; the disagreement is over future orders and the demand structure.

Taking all this evidence together, the main tone of this round is still valuation compression: earnings already delivered have not deteriorated, and the bottom of the correction is supported by earnings; when valuation recovers depends on the discount rate and risk appetite—that is, narrative and capital. And overseas regulation is now pushing the divergence from the denominator side toward the numerator side, which is the biggest variable going forward.

Trading money exits, allocation money takes over

Fund flows are often misread. Intuitively, falling share prices correspond to capital outflows, but the data give the opposite answer.

Wind data show that since September 28, equity ETFs have seen net inflows for eight consecutive trading days, totaling 94.753 billion yuan as of October 14, of which ChinaAMC STAR 50 ETF took in 15.463 billion yuan. By October 21, ChinaAMC STAR 50 ETF had reached 106.092 billion yuan in scale, with its shares up 35.25% over the past month—a record high—making it a newly minted hundred-billion-yuan ETF. Subscription intensity increased as the decline widened. The action on October 24 was even more direct: volume in several broad-based ETFs shifted from the usual late-session period to the morning, with E Fund ChiNext ETF trading 19.738 billion yuan for the full day, a new high since October 9, 2024, while Huatai-PineBridge CSI 300 ETF and ChinaAMC STAR 50 ETF traded 15.423 billion yuan and 15.197 billion yuan respectively.

Figure 3: ETF funds continued to flow in net during the correction. Prices fell while shares rose, with allocation-type funds as the takers. Source: Wind, Cailian Press, China Securities Journal

ETF shares continued to grow while the index fell, showing that the selling pressure and the buying came from different funds. Trading money is sensitive to the discount rate and crowding: it takes profits after gains accumulate and stops out when bad news appears; allocation money positions itself based on valuation percentiles and industry trends, adding more the lower valuations go. With the two types of funds trading in the same market, consecutive index declines and record-high shares can coexist.

The same phenomenon appears on ChiNext. As of October 15, ChiNext-related ETFs had seen net inflows of more than 15 billion yuan over the past week, and ChinaAMC ChiNext ETF had net inflows for 11 consecutive days. Galaxy Securities noted that demand for dividend and defensive allocations is heating up, and that state-owned capital increases, the return of passive funds and insurance capital positioning are supporting bank stock valuations. The bank side has passive funds as a backstop, while the growth side continues to attract allocations during the decline, and the support below is gradually strengthening.

The safe-haven side is also unstable: the dividend peak-and-pullback mentioned earlier occurred in these two weeks, and only the SSE 50 remained the most resilient weekly for four consecutive weeks. Capital has neither stayed in the AI sector nor remained in dividend assets; it is standing on the sidelines, waiting for a new main theme to form.

The two October 21 market moves at the beginning best illustrate how the market is currently pricing: companies whose earnings are questioned are sold off, while companies that deliver on orders are bought. That day's 1.43 trillion yuan turnover shows the market has ample liquidity; what it lacks is a consensus expectation.

Primary market: pricing power shifts to industrial capital

DeepSeek's second funding round has been confirmed complete. According to reports and business registration information, it raised about 100 billion yuan, with the target raised at least twice during negotiations—from an initial 50 billion yuan to 80 billion yuan, and finally to nearly 100 billion yuan—only four months after the first round of about 50 billion yuan in external financing. Tencent and CATL were the largest investors, the pre-money valuation remained unchanged at about 500 billion yuan, and new and old shareholders each contributed about half.

The change in the shareholder list sends a signal: old shareholders such as Tencent and IDG are mainly financial investors, while the new entrants Shixi Capital, SMIC Capital, CPE Yuanfeng, Legend Capital and Geely are mainly industrial players; among them, CATL's contribution rose from about 5 billion yuan in the first round to about 10 billion yuan. The weight of the end bound to the industrial chain has clearly risen, and the new funds are also concentrated in the industrial chain: in July it participated in the ChangXin Technology IPO subscription, in August it took a 141 million yuan strategic placement in Unitree Robotics, and in October it plans to lead a funding round for a robotics startup.

Industrial capital's valuation framework differs from P/E pricing; it places more weight on supply-chain positioning, computing-power synergy and order binding. The valuation logic shifts from financial return to industrial binding, and is therefore more stable than secondary-market sentiment. This also explains why the primary market can keep valuations unchanged while the secondary market plunges.

Zhipu has fallen more than 80% from its intraday high in June, but that 80% is the drawdown magnitude; it partly reflects that its free float at listing was less than 3% of total share capital, and the July lockup expiry and placement brought the scarcity premium to an end. It does not mean the secondary market is applying a uniform discount to AI. Nevertheless, a pre-money valuation of about 500 billion yuan in the primary market coexisting with a deep secondary-market drawdown means the price gap is already significant enough.

According to reports by Caijing and other media, DeepSeek has approached multiple investment banks since late August and has hired CITIC Securities as sponsor; it could file as early as the end of the year and list in 2027. To avoid the regulatory red line of 200 shareholders, some special-purpose funds have switched to blind-pool filing. In late September, Yan Wentao, formerly a partner at Hillhouse Venture Capital, became CFO, ending the company's three years without a CFO. Commercialization data are also starting to materialize: Liang Wenfeng disclosed at a recent investor meeting that the annualized revenue run rate has exceeded US$1 billion, more than double the less than US$500 million of a few months earlier, and customers have not churned after API price increases.

The IPO window and the deep drawdown appear at the same time; the window is narrowing, and those who list first gain pricing power while those who list later face harsher pricing.

A similar revaluation is happening at other leading companies. Moonshot AI's latest funding round was valued at about US$50 billion, though reports differ on pre- and post-money figures; its ARR rose from about US$300 million in June to about US$1 billion after August, corresponding to about 50x ARR. The pricing basis is the future revenue growth rate, not current revenue. On the very day this article was finalized, MiniMax fell below its HK$165 issue price, its first break below issue since listing—another example of the same logic.

Technology catalogue revision: open-source exemption, localized deployment registration

On October 16, the Ministry of Commerce and the Ministry of Science and Technology solicited public comments on a revision of the Catalogue of Technologies Prohibited or Restricted from Export in China, proposing to add 14 prohibited export technologies and 26 restricted export technologies, including 23 in the AI field, covering large models, agents, computing clusters, information processing and other links. On the same day, a State Council executive meeting heard a report on progress in the "AI+" initiative. Tightening and encouragement were issued on the same day; policy is not tightening unilaterally.

The exemption arrangements have a more direct impact than the control list. A Ministry of Commerce spokesperson made clear that technologies that have entered the public domain through open source, published academic papers and other means are not subject to export management; for AI technologies, open source, self-built and self-operated computing centers overseas, and overseas chip tape-out and manufacturing are eligible for facilitation measures such as automatic licensing, essentially "report and go"; enterprises using large models to provide external API calls, App/web interaction, SaaS applications and local on-device lightweight deployment are not within the scope of large-model technology export management. For large models included in the restricted category, overseas localized deployment falls within the scope of management; where deployment on overseas proprietary clouds and other risk-controllable circumstances applies, automatic licensing may be used.

At the industry level, if the provisions take effect as drafted, they clearly favor the open-source route represented by DeepSeek. Excluding open source from controls amounts to confirming that open source is an encouraged route; vendors relying on overseas deployment of closed-source models will see overseas delivery enter the scope of management and face new compliance costs. The announcement is still in the comment period; the official text prevails.

The last time the catalogue drew attention because of AI was in 2020, when personalized information push technology was added, directly changing the negotiating landscape of ByteDance's overseas transaction. This time, controls are being listed in the window when large-model capital operations are most intensive, with pre-export registration launched simultaneously; the rules precede the deals. The comment period itself is a window for companies to adjust their business models.

The other side of policy also provides a basis for differentiated pricing in the secondary market: vendors in the open-source ecosystem gain a layer of policy protection, while vendors pursuing closed-source overseas expansion must bear additional compliance costs. When industrial capital dominates pricing in the primary market and regulators draw boundaries at the policy level, valuation divergence among AI companies will no longer be determined by overall sector sentiment but by the companies' own differences.

Q3 reports are the nearest verification point

Several threads converge here: valuations have returned to a relatively reasonable range, allocation funds are buying against the trend, the primary market has provided steadier industrial pricing, and policy has clarified the boundaries of technology routes. Only two variables remain: when the discount rate will fall, and when a new narrative will coalesce.

A CITIC Securities research report dated October 9 judged that the next window for a tech rally is most likely in Q4, requiring verification of three conditions: a marginal reversal of rate-hike expectations, a new narrative coalescing into consensus, and a completed rebalancing of liquidity. Of the three, the first is an exogenous variable depending on the U.S. inflation path; the latter two will be tested first by Q3 reports.

Liquidity rebalancing is furthest along: ETF subscriptions against the trend have lasted more than three weeks, industrial capital is maintaining valuations in the primary market, and policy boundaries have been drawn.

What remains unsettled is the narrative. The previous consensus was an arms race in computing power: Microsoft cut the headline figure for its 2026 capital expenditure from about $190 billion to about $175 billion, Meta narrowed its guidance range, and overseas scrutiny of AI investment returns tightened—the old narrative is losing force. On-device AI scaling, falling inference costs, and a closed domestic computing-power loop have not yet coalesced enough to take over. On October 24, the humanoid robotics sector rose 7.98% for the full day, and Unitree Robotics recorded its first limit-up since listing—one of the new candidates, but a single day's move is not yet enough to constitute a narrative. A narrative vacuum, overseas regulatory shocks, and repeated swings in rate-hike expectations are the causes of this round of adjustment.

A-shares have entered the dense disclosure period for Q3 reports, and three things bear watching: whether order visibility in the computing-power chain can extend to 2027, whether gross margins rise or fall during the capacity ramp-up, and whether on-device AI scaling is a pulse or a trend.

The first question was answered on October 23–24: Zhongji Innolight announced that orders from major overseas customers are already booked through 2027, with some key customers giving clear guidance on 2028 demand; Eoptolink said its 1.6T optical modules have entered a phase of sustained scaling and announced a repurchase plan of 800 million to 1 billion yuan; TFC Optical Communication, Accelink Technologies, and Yuanjie Semiconductor simultaneously announced ample orders on hand and no change in industry conditions. The latter two questions await more earnings disclosures.

If all three are delivered, this round of adjustment is the starting point of a new narrative; if guidance is cut, it means the valuation selloff was merely pricing in doubts about 2027 ahead of time.

Finally, back to the contradiction at the outset: companies whose results greatly beat expectations are being bought with capital, yet sector valuations are still drifting lower; the primary market keeps valuations unchanged, while the secondary market repeatedly prices listed companies that are scarce in liquidity.

On October 24 there were signs that this contradiction was resolving. That morning, dragged down by a 2.6% overnight drop in the Nasdaq, the Shanghai Composite Index at one point fell to 3,574.27 points, breaking below 3,600; at midday, order announcements from five optical communication companies landed, and in the afternoon the market completed a V-shaped rally, with the ChiNext Index going from down more than 3% to closing up 4.16%, the STAR 50 up 3.04%, and full-day turnover of 1.96 trillion yuan, 427.6 billion yuan more than the previous trading day. On the same day, the banking sector fell 1.30%, with Bank of China down more than 4%, as capital withdrew from defensive positions and flowed back into growth.

The weekly line is still negative: the Shanghai Composite fell 0.69% this week, the ChiNext Index fell 3.18%, and the STAR 50 fell 0.97%; the single-day rebound failed to recover the previous month's losses. But from two consecutive weeks of net ETF subscriptions to those five midday order announcements, the direction of capital is already clear. (This article was first published on the TMTPost APP; author | AGI-Signal; editor | Jiao Yan)