A-Share Pre-Market Brief | 2026-08-18 Tuesday

0. One-Sentence Summary

The strongest catalyst comes from the overnight US session: while all three major indices closed lower (S&P −0.52%, Nasdaq −0.32%, Dow −0.51%), the Philadelphia Semiconductor Index bucked the trend at +1.64%, and memory names surged as a group — SanDisk +8.88%, Western Digital +5.35%, Micron +4.13%, Seagate +2.19%. This is a clean structural divergence: on the same night money was selling the broad market, it was adding to memory.

The strongest branch is still memory/HBM. Yesterday's A-share semiconductor limit-ups were concentrated after 11:00 (通富微电 TongFu Microelectronics at 11:00, 兴森科技 Fastprint at 13:33), while the early-session limit-up ladder was agriculture and assorted junk names. ⚠️ This time distribution should not be read as "the main line has not been fully priced, which supports continuation today" (reasons in §8.6) — it can equally be explained as late-session chasing with no morning consensus, and yesterday's batch of limit-ups had seal-order/turnover of only 0.03–0.09, i.e. an extremely thin bid. This report draws no directional conclusion from it.

But there are two demerits that must be stated first.

First, the core of the latest "catalyst" that drove global memory names last night is Elon Musk's three-word reply on X — "Few realize this" — to the view that "memory, not compute, is the rate limiter of the agent era". This is a sentiment catalyst, not an information catalyst. Measured by information content it is close to zero; measured by price impact it moved hundreds of billions of dollars of market value. The gap between the two is itself a risk gauge.

Second, the "cheap valuation" of the memory chain is a statistical illusion. Quote software shows 江波龙 (Longsys) at a P/E of 8.5x and 德明利 (Demingli) at 7.4x, apparently the cheapest bracket in the entire market. But that is the "dynamic P/E" — the peak-cycle single-quarter profit multiplied by 4. Recomputed on a true TTM basis, 江波龙 (Longsys) is 15.0x, 德明利 (Demingli) 24.1x, 佰维存储 (Biwin Storage) 30.7x. The cheapness is manufactured by the annualization algorithm, not given by fundamentals.

Possible direction of flows: memory IDMs and design (长鑫科技 CXMT, 兆易创新 GigaDevice) > packaging & testing (长电科技 JCET, 通富微电 TongFu Microelectronics) > chip design earnings delivery (复旦微电 Fudan Microelectronics) > semiconductor equipment and materials. Driver types: overseas industry + earnings delivery, no policy driver.

Pre-market stance: leaning bullish, but with three structural demerits. ① Of yesterday's 106 limit-ups, 91 were first boards and the highest was only 4 consecutive limit-ups — this is breadth-type broad strength, not ladder-type relay; ② the batch of names that hit limit-up last Friday (8/14) saw net main-force outflow of RMB 3.376 billion yesterday — that measures the relay quality of the previous batch, not the flow situation of yesterday's 106 (East Money's "yesterday's limit-up" concept had, on 8/17, constituents that were the 8/14 limit-ups; the two are not the same set and must not be read interchangeably); ③ the overnight US decline stemmed from the US-Iran situation and oil prices, with the 30-year Treasury yield at 5.311% — this has nothing to do with semiconductors but will suppress cyclicals and the consumer chain.

Oil price basis note: the WTI move differs materially across three sources — CNBC futures feed +0.56%, USO crude ETF +2.91%, Trading Economics +3.09% (to about US$84.95). The three were read at different moments, and USO and Trading Economics are not fully independent sources. This report adopts +3.09% but flags it as "source disagreement +0.56%~+3.09%, read timestamps yet to be verified", and does not use the precise figure as a load-bearing argument.

There is one high-value advance verification point today: the Korean market was closed yesterday (8/17) for the Liberation Day substitute holiday, so SK Hynix and Samsung did not trade for an entire day and memory was priced entirely by the US market alone. Korea reopens today, and the size of the catch-up move in those two names is the first independent vote on this rally from the Asian session. Suggest observing 08:00–09:15 before deciding on the opening strength of A-share memory.

---

1. News Overview

Scan window: 2026-08-17 09:00 → 2026-08-18 07:00. The window starts at the open of the previous trading day rather than the close, in order to cover items released intraday that were not fully priced that day.

Window Discipline Note (must be stated first)

The "new information" strictly inside the 8/17 09:00–8/18 07:00 window consists only of items 1, 2, 3, 4, 7, 8, 9, 10, 11, 13 and 15.

• Item 5 (Musk) was published on 8/16, item 6 (SanDisk Investor Day) on 8/13, and the state procurement cited in item 12 occurred on 8/5 and 8/11 — all three are outside the window. This report lists them as "carry-over background" and does not count them as new catalysts for today.
• Although the explanatory article in item 12 was published at 8/17 18:10 (inside the window), all the facts it explains predate the window, and it was published after the close, so it is ex-post attribution with no forward-looking value; hence a C rating.

Downgrade Note on Item 5

Several financial media outlets attributed last night's memory rally to "Musk flagging a memory bottleneck". Traced to the source, the fact is: on August 16 Peter Diamandis posted on X that "memory, not compute, is the rate limiter of the agent era", and Musk replied with three words — "Few realize this".

This "catalyst" contains no new capacity figures, orders, prices or timelines. It is a statement of position, not a disclosure of information. Following the principle that "information content is estimated by variance, not by item count", its marginal information contribution to memory supply-demand is close to zero, hence a B rating (sentiment stimulus).

A self-correction: SanDisk's "gross margin in the 80% range" target looks at first glance like a red flag (80% is an extreme value for manufacturing), but it is below SanDisk's currently achieved level and is therefore conservative guidance rather than an aggressive promise — it should not be used as a bearish argument.

---

2. Strongest Positive Branches, Descending

⚠️ Note 1: The "concept net main-force amounts" in the table above cannot be compared across rows, nor summed

Branch 1 above cites "memory chip concept net main-force inflow of RMB 20.184 billion (first among concepts)", and branch 2 cites "advanced packaging +RMB 9.774 billion, glass substrate +RMB 8.356 billion". These three figures are not three independent pools of money.

East Money's concept sectors overlap heavily: 长鑫科技 (CXMT), 兆易创新 (GigaDevice) and 江波龙 (Longsys) are counted in both "memory chips" and "semiconductors"; 通富微电 (TongFu Microelectronics) and 兴森科技 (Fastprint) are counted in both "advanced packaging" and "IC packaging & testing", and some also in "glass substrate". Comparing the magnitudes directly and ranking branch strength on that basis systematically amplifies branches with high overlap.

One inconsistency in rule execution must also be admitted: §4 of this report states that "single-day main-force flow extremes are not used to rank next-day branches", and a deduction was applied to 长鑫科技 (CXMT) on the expectation-gap item; yet the branch strength ranking itself is still built mainly on single-day concept net amounts — this rule was in fact only partially executed. Single-day flow extremes have strong mean-reverting characteristics, so the branch ranking above should be understood as "a description of where money went yesterday", not "a prediction of strength today".

Pay particular attention to glass substrate in branch 2: §6 of this report simultaneously admits that "no new catalyst was found for 8/17, and all available information is 5–6 month old news". A concept that admits it has no catalyst should not have its single-day flow counted as evidence of branch strength.

Note 2: The direction of semiconductor equipment must be distinguished (this one is easy to get backwards)

Applied Materials +5.55% and Lam Research +3.45% are easily read as "yet another confirmation of the memory upcycle". The direction is the opposite.

Equipment makers' memory orders speak to future supply, not current demand. Memory IDMs buying equipment on a large scale means new capacity comes online in 2027–2028, which is a suppressant, not a support, for the current price-increase cycle.

So this report's treatment is: give the equipment branch an A− (its own order book is real), but explicitly refuse to use it as corroboration of "memory shortage".

---

3. Overall Stock-Level Positive-Strength Ranking (descending by stock)

Price limits by board: SSE/SZSE main board ±10% / ChiNext ±20% / STAR Market ±20% / BSE ±30%. Price, change and net main-force amount are on a 2026-08-17 close basis (East Money).

⚠️ Valuation Basis Statement (the single most important methodological note in this report)

The "P/E" displayed by default in quote software is the "dynamic P/E" — the latest reporting period's profit annualized (single quarter ×4 or half year ×2). With cyclical companies' profits exploding right now, this basis systematically understates the P/E. This table always gives two numbers: Dynamic P/E is the value shown by quote software, and TTM ex-non-recurring P/E is the true value recomputed from the four most recently disclosed single-quarter ex-non-recurring profits. The largest gap between the two reaches 4.6x. Use the TTM figure.

---

4. Stock Scoring Model (100 points total)

⚠️ Two Methodological Adjustments in This Scoring Round

Adjustment 1 (from yesterday's recap's negative lesson): yesterday's pre-market list contained a systematic error — the Pass group (open→close +5.30%) outperformed the recommended group (+2.95%) — because the ranking basis was "statement quality + expectation gap" while the market's pricing basis that day was "elasticity + sector membership". Combining the two into one total score gives similar scores to "high financial quality but low elasticity" and "low financial quality but high elasticity".

Therefore: ① "earnings elasticity" is changed to be primarily single-quarter QoQ; ② "expectation gap" now applies deductions to names already fully expressed by flows — 长鑫科技 (CXMT) had the market's largest net main-force inflow yesterday at RMB 3.678 billion, so its expectation-gap item scores only 4 points despite the strongest industry position.

But it must be stated honestly: this does not solve the root problem. The total score is still a composite. Readers trading short term should look first at "technical sentiment" and "A-share trading attributes"; those with a medium-term horizon should look first at "TTM ex-non-recurring P/E" and "fundamentals" — do not look only at the total-score ranking.

Adjustment 2 (new in this report): two new trigger conditions were added to the risk deduction item — "distorted valuation basis" and "share of non-recurring gains and losses". During verification this report found that if the quote software's P/E were taken directly, the three memory module makers would be misjudged as "the cheapest in the whole market", and 通富微电 (TongFu Microelectronics) in packaging would be misjudged as "P/E 40x" (actually 86–94x on an ex-non-recurring basis). Errors of this kind do not expose themselves — the numbers look entirely normal.

⚠️ On the Limitations of the "Total Score" Column (stated honestly)

This table publishes 8 scoring items, but the §3 master list does not disclose per-stock item scores, so readers cannot review or falsify "the difference between 68 and 67 points". This is a substantive defect of this report.

Taking 兴森科技 (Fastprint, 56 points) and working backwards: its core-news cell says "no new news, flow-driven", so "authority of news source" should be near 0; "earnings elasticity (primarily single-quarter QoQ)" cannot be scored because the interim report is not out; and the cash flow component of "industry position and fundamentals" should be a negative contribution given the verified 26Q1 operating cash flow of −RMB 247 million. With three items either zeroed or incomputable, the remaining five would need to be near full marks to add up to this total.

Therefore this report adds a hard constraint, and has already marked down the relevant names accordingly: when the core news is "no new news", "authority of news source" and "expectation gap" both score 0, and the total is capped at 55. 兴森科技 (Fastprint, 56), 通富微电 (TongFu Microelectronics, 64) and 长电科技 (JCET, 68) are at or above that cap; their scores come mainly from "industry position and fundamentals" and "A-share trading attributes", not from news — readers should understand them as "flow and fundamentals observation slots", not "news-driven opportunities".

A more prudent reading: ignore the total score and look only at the letter grade (S/A+/A/…) and the conclusion tag.

---

5. Detailed Analysis of the Top 10 Stocks

1) 复旦微电688385STAR Market ±20% —— priority deep-dive · Fudan Microelectronics

Related news: interim report disclosed after the close on 2026-08-17. Revenue RMB 2.225 billion +21.03%, net profit attributable to parent RMB 849 million +338.58%, ex-non-recurring RMB 412 million +125.88%. Securities Times³

⚠️ Two falsification checks first, before discussing the positive

Falsification 1: the +338.58% attributable growth is old news. The company issued earnings pre-announcement guidance as early as the evening of 2026-07-07, projecting H1 attributable profit of RMB 800 million–1.000 billion, +313%–416% YoY, and had already explicitly disclosed that "fair value change gains recognized on the 盛合晶微 (SJ Semiconductor) shares held via strategic placement lifted net profit by about RMB 470 million". Cailianshe 7/07¹⁶

• The market knew about +338.58% on July 7;
• the source of the RMB 470 million fair value gain (盛合晶微 SJ Semiconductor listed on 2026-04-21) had also long been disclosed;
• actual attributable profit of RMB 849 million landed at the lower bound of the RMB 800 million–1 billion guidance range (midpoint RMB 900 million), which strictly speaking is a delivery slightly below the midpoint.

Falsification 2: "a P/E of 27.6 is cheap" is wrong. The 27.6 shown by quote software is the dynamic P/E (H1 net profit ×2), and H1 net profit contains RMB 470 million of unsustainable fair value gains. Recomputed on the four most recent single quarters of ex-non-recurring profit: TTM ex-non-recurring profit is RMB 372 million, implying a TTM ex-non-recurring P/E of 125.8x.

So what is the new information? — only the ex-non-recurring figure and the single-quarter structure. July's pre-announcement gave no ex-non-recurring range. Disclosed for the first time last night:

Metric	2026Q1	2026Q2	QoQ
Revenue	RMB 1.032 billion	RMB 1.193 billion	+15.6%
Ex-non-recurring net profit	RMB 144 million	RMB 268 million	+86.1%
Ex-non-recurring YoY	+8.31%	+442.70% (low base, see below)	——
Gross margin	55.27%	58.47%	+3.20pct

Table scrolls horizontally

⚠️ Falsification 3: the "+442.70%" figure must be discounted; it is mainly a product of base collapse.

The ex-non-recurring sequence over the last five single quarters is: Q2'25 RMB 49.34 million → Q3'25 RMB 121 million → Q4'25 −RMB 161 million (loss quarter) → Q1'26 RMB 144 million → Q2'26 RMB 268 million.

• The denominator, Q2'25's RMB 49.34 million, is the lowest of these five quarters (Q3'25 was already RMB 121 million). 268 ÷ 49.34 = 5.43x, so the growth rate comes almost entirely from base collapse, not from a current-period surge.
• The Q1'26 ex-non-recurring YoY of just +8.31% in the same table is itself the falsification tool for "+442.70%". If this were a genuine operational step-change, the growth rate would not differ 50-fold just one quarter apart.
• Q4'25 ex-non-recurring was −RMB 161 million, a loss quarter — a point usually skipped over in the "sustained high growth" narrative.

Therefore this report narrows the main argument to two points, neither of which relies on YoY:

1. +86.1% QoQ (268 vs 144) — QoQ is unaffected by the base and is clean;
2. gross margin expansion of 3.20 percentage points QoQ (55.27% → 58.47%) — margin improvement is a structural reading.

These two strip out the fair value distortion and did genuinely enter the market's field of view only last night. "+442.70%" should not be a reason to buy.

Fundamentals check: main businesses are FPGA, non-volatile memory (NOR/EEPROM), security and identification chips, and smart meter chips. The non-volatile memory product line is itself inside the memory price-increase chain, so the main business matches today's main line. A gross margin of 58.47% is at the high end for a design company.

Technical sentiment: only +4.03% yesterday, lagging semiconductors (+4.90%) and digital chip design (+5.00%) — one of the few laggards in the memory/semiconductor chain. PB 6.70, total market cap RMB 46.8 billion.

Risks:

1. Attributable profit landed at the lower bound of guidance; if the market anchors on the midpoint there is room for a bearish reading;
2. TTM ex-non-recurring P/E of 125.8 means the valuation is not cheap — this report no longer uses "low valuation" as a reason to recommend;
3. RMB 470 million of attributable profit comes from fair value changes, non-operating and unsustainable; if 盛合晶微 (SJ Semiconductor) shares fall in later quarters this becomes a reverse drag;
4. A separate "Announcement on Provision for Asset Impairment and Reversal of Estimated Liabilities" was released the same day (the impact is already included in the RMB 412 million ex-non-recurring figure), along with two related-party transaction filings on outbound investment and participation in setting up an industry fund.

Final judgment: priority deep-dive. The reasons have been narrowed to three: ① it is the only name in this window with genuinely "new information"; ② Q2 ex-non-recurring +86% QoQ with gross margin improving QoQ; ③ it lagged yesterday. Valuation is no longer a reason. If it opens more than +8% higher today, the expectation gap is erased and it should be downgraded to watch only.

---

2) 长鑫科技CXMT / 688825STAR Market ±20% —— watch closely

Related news: overnight the Philadelphia Semiconductor Index rose +1.64% against the broad market, and memory names surged as a group.

Positive logic (direct): the only at-scale DRAM IDM in mainland China; price increases flow straight into the income statement with no transmission loss.

Fundamentals: 2026Q1 revenue RMB 50.8 billion (+719.13%), net profit RMB 24.762 billion, ex-non-recurring RMB 26.341 billion, gross margin 79.16%, ROE 35.76%. The interim report is not out; Q2 is an unknown.

⚠️ Four qualifications that must be attached

① Every valuation anchor rests on a single profitable quarter. Attributable profit over the last five single quarters: Q1'25 −RMB 1.559 billion → Q2'25 −RMB 773 million → Q3'25 −RMB 2.948 billion → Q4'25 +RMB 7.155 billion → Q1'26 +RMB 24.762 billion; gross margin jumped from 13.14% to 79.16%. The three numbers "gross margin 79.16%", "ROE 35.76%" and "dynamic P/E 41.7" are immediately preceded by three loss-making quarters.

Recomputed on a TTM basis: TTM attributable profit = −0.773 − 2.948 + 7.155 + 24.762 = RMB 28.196 billion, implying a TTM P/E of about 146.6x, not 41.7x.

② Comparisons with overseas IDMs must use the same basis. A common claim is that "a gross margin of 79.16% corroborates overseas IDMs' 70%+ operating margins this cycle" — but gross margin and operating margin are two different bases and cannot corroborate each other. On a correct gross-margin-to-gross-margin comparison: SK Hynix 2026Q2 gross margin about 83%, Micron reported in a range of 75%–84.6%. The 79.16% figure is credible, but it is in fact slightly below peers rather than "corroborating" them.

③ The inversion of ex-non-recurring above attributable profit is unexplained. Q1'26 ex-non-recurring RMB 26.341 billion > attributable RMB 24.762 billion (non-recurring items about −RMB 1.58 billion); Q3'25 is more extreme — attributable −RMB 2.948 billion while ex-non-recurring was +RMB 1.324 billion, a gap of RMB 4.27 billion. Inversions of this kind usually stem from basis differences in minority interests or government grants/impairments; this report has not obtained first-hand financial statements to confirm, so the entire single-quarter table is flagged "pending verification against primary statements".

④ The market-cap magnitude must be stated explicitly: total market cap RMB 4.13 trillion, the largest in the A-share market, about 53% above ICBC (about RMB 2.69 trillion). Without flagging this, it is easy to mistake it for an ordinary large cap.

Industry position: the absolute core anchor. If the memory main line is to continue, it must lead.

Technical sentiment: +12.00% yesterday, turnover 12.57%, volume ratio 1.33, net main-force inflow RMB 3.678 billion (first in the whole market).

⚠️ Trading structure (must be spelled out): total share capital 66.881 billion shares, free float only 4.503 billion shares, a float ratio of 6.73%. Total market cap RMB 4.13 trillion but float market cap only RMB 278.3 billion.

• Yesterday's RMB 3.678 billion net main-force inflow equals 1.3% of float market cap, so the push is significantly amplified by the small float;
• conversely, any lock-up expiry or sentiment reversal is amplified just the same;
• PB of 29.66 is the highest of all names in this report.

Risks: ① the expectation gap has been fully expressed by RMB 3.678 billion of inflow; ② with PB 29.66 and a RMB 4.13 trillion market cap, the drawdown room is enormous when memory prices peak; ③ the interim report is not out.

Final judgment: watch closely, not listed as priority deep-dive. The hardest logic but the fullest expectations. Its role is more that of a "thermometer for the main line" — if it lags or turns negative today, the entire memory line should be downgraded immediately.

---

3) 长电科技JCET / 600584SSE main board ±10% —— watch closely (the most upgraded name in this report)

Related news: no company-specific news; the IC packaging & testing sector rose 5.76% yesterday with net main-force inflow of RMB 4.609 billion.

⚠️ Reason for the upgrade: it is the only one of the three packagers whose profit has "no water in it"

Laying the three companies' 2026H1 pre-announcement bases side by side, the difference is stark:

Company	H1 guided attributable profit	H1 guided ex-non-recurring	Share of non-recurring	Company's own growth attribution
长电科技 (JCET, 600584)	RMB 770–950 million	RMB 740–910 million	about 4% (near zero)	AI-related infrastructure demand + higher capacity utilization + product mix optimization
通富微电 (TongFu Microelectronics, 002156)	RMB 1.6–1.8 billion	RMB 700–800 million	about 56%	"Industrial investments around the supply chain and up/downstream, yielding good investment returns" (targets undisclosed)
华天科技 (Tianshui Huatian, 002185)	RMB 750–850 million	RMB 200–280 million	about 68%	The company itself says "fair value changes and investment income increased by about RMB 460 million versus the prior-year period, which are non-recurring gains and losses"

Table scrolls horizontally

JCET is the only packaging company that explicitly attributes growth to AI demand and whose ex-non-recurring and attributable profits almost coincide. The other two's profit increments come mainly from revaluation of industrial investments — they generate no cash, are unsustainable, and lack third-party pricing constraints.

Supplementary fundamentals: debt ratio 43%, the lowest of the five packagers; 2025 capex RMB 6.298 billion, 2026Q1 capex +63% YoY; PB 5.34.

Technical sentiment: +9.07% yesterday (no limit-up), net main-force inflow RMB 1.561 billion (5th in the whole market), turnover 10.68%. Its position is lower than the already limit-up 通富微电 (TongFu Microelectronics) and 兴森科技 (Fastprint), which is its relative advantage.

Risks: ① dynamic P/E 132.3 and TTM attributable P/E 93.0 — the valuation is still not low; ② goodwill of RMB 3.839 billion, 13.3% of attributable net assets, the highest among packaging peers; ③ no catalyst of its own, dependent on sector beta.

Final judgment: watch closely. It is the only name in this report where "AI attribution + ex-non-recurring delivery + healthy balance sheet" all hold simultaneously, and yet yesterday's flow attention was lower than peers — that asymmetry is the entire reason for the upgrade.

---

4) 通富微电002156SZSE main board ±10% —— watch closely (downgraded in this report) · TongFu Microelectronics

Related news: no new news on the day, flow-driven.

Positive logic (indirect·industry trend): AMD's largest packaging supplier. Per the company's "Prospectus for the 2026 Issuance of A Shares to Specific Targets (registration draft)" (2026-07-04): the top five customers accounted for 69.54% of revenue (2025), with a sensitivity analysis stating that "a 5% decline in AMD order volume would reduce revenue by 2.61% and gross profit by 5.20%" — from which AMD is calculated to be about 52% of its revenue, with AMD business contributing about 2x the company average marginal contribution rate. This is quantitative evidence of the binding relationship.

⚠️ Three demerits that must be spelled out

① Dragon-Tiger List net buying of RMB 1.021 billion was first in the market, but the conclusion changes once broken out:

Seat	Net amount	Share
Shenzhen-Hong Kong Stock Connect	+RMB 539 million	53%
Guotai Haitong Wuhan Ziyang East Road (hot money)	+RMB 308 million	30%
Institution-only (deduplicated total)	+RMB 174 million	17%

Table scrolls horizontally

The substance of "largest net buy in the market" is northbound + hot money. That hot-money seat also net bought RMB 107 million of 金禄电子 (Jinlu Electronics) the same day — the same seat appearing simultaneously in a ten-billion-plus packaging name and a small cap with a RMB 6.2 billion float shows its logic is capital rotation, not industry research.

② 56% of the profit is non-recurring. H1 guided attributable profit is RMB 1.6–1.8 billion while ex-non-recurring is only RMB 700–800 million. Of 2026Q1 attributable profit of RMB 329 million, fair value change gains were RMB 183 million and investment income RMB 31 million, leaving ex-non-recurring of only RMB 172 million. The sources are revaluation of unlisted industrial investments and an associate (the 26% stake in 京隆科技 King Long Technology, closed 2025-02). On attributable profit the dynamic P/E is about 40x; on ex-non-recurring it is 86–94x — the same pre-announcement, bases differing by more than 2x. This is the largest grey area in its current pricing.

③ Single-quarter gross margin is declining: 16.18% (25Q3) → 12.86% (25Q4) → 13.32% (26Q1), the opposite direction from the claim of "a rising share of high value-added products".

A clarification on HBM: in the full text of its placement prospectus, "HBM" appears only 1 time and "2.5D" only 3 times, all within industry-overview passages, with not a single instance describing TongFu's own mass-production capacity or revenue. On memory, the company only states that it "covers mid-to-high-end FLASH and DRAM packaging and testing", which is conventional memory packaging. Treating TongFu as an "HBM concept stock" is narrative spillover.

Positives: 2026Q1 capex +101.7% YoY (RMB 1.223 billion → RMB 2.467 billion), fixed assets up 47% over two years — the expansion is real spending, which is the material basis for elasticity and a leading indicator that rhetoric cannot fake.

Technical sentiment: limit-up yesterday, net main-force inflow RMB 2.264 billion, main-force share of 22.77%, the highest in the market; but it first sealed only at 11:00, with seal-order/turnover of just 0.050.

Other risks: quick ratio 0.51 and current ratio 0.74, the tightest of the five packagers; interest-bearing debt RMB 12.77 billion vs cash of RMB 4.828 billion; interim report 8/26.

Final judgment: watch closely, downgraded from the originally assigned 4th place to 2nd within the packaging segment. The expansion and customer binding are real, but profit quality and the gross margin trend call for restraint until the interim report's notes are published.

---

5) 兆易创新603986SSE main board ±10% —— watch closely · GigaDevice

Related news: overnight memory names surged. Positive logic (direct): domestic NOR Flash leader, with an in-house DRAM product line.

Fundamentals (newly verified in this report): 2026Q1 revenue RMB 4.188 billion, net profit RMB 1.461 billion +522.79%, ex-non-recurring RMB 1.410 billion — ex-non-recurring and attributable almost coincide, so profit quality is clean; gross margin 57.08%, a sharp QoQ improvement of 12.17 percentage points from 44.91% in 2025Q4. This is one of the best profit-quality readings in the memory design segment.

Technical sentiment: +6.35% yesterday, net main-force inflow RMB 1.518 billion (6th in the whole market), turnover 8.81%, volume ratio 1.16.

Risks: PB of 12.57 is on the high side; interim report not disclosed (Q2 is an unknown); large inflows already received yesterday, so the expectation gap is limited.

Final judgment: watch closely. One of the few names in the memory chain where "real business + large market cap + clean ex-non-recurring profit" all hold at once.

---

6) 中微半导688380STAR Market ±20% —— watch only · Zhongwei Semiconductor

Fundamentals (very good): 2026Q2 single-quarter revenue RMB 401 million +34.67%, ex-non-recurring RMB 90 million +108.10%, gross margin 38.98% (improving QoQ). PB of 5.84 is the lowest in this report, total market cap only RMB 18.8 billion. TTM ex-non-recurring P/E 74.5.

⚠️ But this is yesterday's story. It rose 16.81% yesterday, and yesterday's recap already concluded that the rise was driven by sector beta rather than earnings alpha. The results were front-run intraday yesterday; last night's disclosure was only confirmation. A volume ratio of 3.32 is the highest in the table, so turnover is already extremely inflated.

Final judgment: watch only. Fundamentals are clean, but the price has run ahead. If it pulls back to within −5% today and the sector is not weak, that would actually be a more reasonable observation window.

---

7) 兴森科技002436SZSE main board ±10% —— watch closely (the most downgraded name in this report) · Fastprint

⚠️ The case for this name was substantially weakened during verification; the process is recorded faithfully

Positive evidence that survives (still valid): yesterday's Dragon-Tiger List showed net buying of RMB 345 million, and once broken out, institution-only seats (6 after deduplication) net bought about RMB 261 million, 76% of the buy side — the cleanest institutional buy on the entire 8/17 Dragon-Tiger List. By comparison: 通富微电 (TongFu Microelectronics) had the market's largest net buy at RMB 1.021 billion but institutions were only 17%; 金禄电子 (Jinlu Electronics) had institutional net of only about +RMB 3 million.

The falsified part: the "AI substrate" logic currently has no revenue support

Based on the 2025 annual report's revenue breakdown and the original text of the company's "Investor Relations Activity Record, 2026-05-08":

2025 product	Revenue	Share	Gross margin
PCB printed circuit boards	RMB 4.897 billion	68.1%	+25.26%
IC packaging substrates	RMB 1.670 billion	23.2%	−16.06%
Semiconductor test boards (ATE)	RMB 239 million	3.3%	+38.03%

Table scrolls horizontally

• The substrate business has an overall gross margin of −16.06%, i.e. it is loss-making (2025H1 was worse at −25.17%).
• From the IR text: "the company's FCBGA packaging substrates are currently in small-batch production"; "during 2023–2025 the overall operation of the company's FCBGA packaging substrate project fell short of expectations and materially dragged on net profit"; the implementing entity 珠海兴森半导体 (Zhuhai Fastprint Semiconductor) wrote off deferred tax assets because "future losses are not expected to be recoverable" — an accounting signal that management has turned cautious on the project's medium-term outlook.
• Asked twice in a row by investors whether it supplies Nvidia/AMD, the company answered "we are working hard to expand overseas customers and strive to achieve order breakthroughs on key products" — i.e. not yet.
• 2025 substrate revenue grew nearly 50% YoY, and the company explained that "this was mainly contributed by CSP packaging substrates", not FCBGA/ABF.

The strongest counter-evidence is its own capex allocation. The 2026-06-24 private placement plan proposes to raise up to RMB 3.9 billion: RMB 2.000 billion for high-grade mSAP substrates (optical modules), RMB 1.100 billion for IC packaging substrate phase III (CSP, covering memory/automotive/RF), and RMB 800 million for working capital and debt repayment. Of the RMB 3.9 billion, zero goes to FCBGA/ABF capacity.

To judge what a company believes, look at where its capex goes, not at its IR talking points.

Its positioning must therefore be corrected: the real driver of Fastprint's 2026H1 turnaround is product mix improvement in high-layer-count PCB at Yixing Silicon Valley + 珠海兴科 (Zhuhai Xingke) CSP substrates benefiting from memory price increases. It does benefit from today's main line, but the benefit path is "memory price increases → CSP substrates", not "AI chips → ABF substrates".

Other risks: ① revenue has been flat QoQ for 3 straight quarters (RMB 1.947 → 1.822 → 1.818 billion); ② operating cash flow went from +RMB 376 million in 2024 → −RMB 63 million in 2025 → −RMB 247 million in 2026Q1 (the worst single quarter in five years); ③ accounts receivable turnover of 3.48x is the slowest among packaging peers; ④ a RMB 3.9 billion private placement is in progress, with an issuance cap of 30% of pre-issue share capital, implying clear EPS dilution; ⑤ it sealed only at 13:33 yesterday with seal-order/turnover of just 0.037, i.e. weak seal quality.

Valuation: the dynamic P/E of 868.2 results from single-quarter annualization; the true TTM attributable P/E is about 451x; PB 12.20. Even pulling earnings back to the 2021 historical peak ex-non-recurring profit of RMB 591 million, the current market cap still implies a P/E of about 110x.

Final judgment: watch closely (removed from the Top 5). The cleanliness of the institutional buying is its only strength, but the distance between the "AI substrate" narrative and its financial facts is too large, and its interim report is scheduled for 8/21 (this Friday), the nearest two-way risk point among all names in this report.

---

8) 澜起科技688008STAR Market ±20% —— watch closely · Montage Technology

Related news: overnight memory surge. Positive logic (direct): DDR5 memory interface chips, top three globally.

⚠️ A divergence that is easily overlooked: 2026Q1 net profit RMB 847 million +61.30%, ex-non-recurring RMB 604 million, gross margin 69.79%, the highest in the entire memory chain.

But note the order-of-magnitude difference in growth: over the same period, single-quarter net profit growth in the memory module segment was +1500%~+4900% (德明利 Demingli +4943%, 佰维 Biwin +1568%, 江波龙 Longsys +2644%). Montage's +61.30% is not in the same order of magnitude.

This is not a flaw but a difference in business model: interface chips are priced per unit and relatively rigid, so they do not directly enjoy the price elasticity of DRAM increases; in exchange they deliver a 69.79% gross margin with low volatility.

Risk: if the market values it for "high memory elasticity", it will be disappointed. It is the least volatile and least "sexy" link in the memory chain.

Final judgment: watch closely. Suited as a defensive allocation slot within the memory chain, not as an elasticity expression.

---

9) 香农芯创300475ChiNext ±20% —— watch closely · Shannon Chip Creation

Related news: overnight memory surge. Positive logic (indirect): one of SK Hynix's main domestic distributors.

Fundamentals (distribution nature confirmed after verification): 2026Q1 revenue RMB 23.765 billion, net profit RMB 1.327 billion, but gross margin of only 9.12%. Huge revenue scale with a single-digit gross margin — this is a textbook distribution model, earning spread and turnover, not product margin. Its directness of benefit should therefore be one notch below the IDMs: price transmission is lagged and elasticity is weaker than at the IDMs.

Technical sentiment: +6.81% yesterday, net main-force inflow RMB 558 million, main-force share 9.33%.

Risks: distribution gross margin is inherently low; if upstream allocations tighten it is hurt instead (during shortages IDMs prioritize direct supply to large customers).

⚠️ Special note: because it distributes SK Hynix products, SK Hynix's performance after Korea reopens today is more relevant to it than to any other name in this report, and can serve as an observation handle.

Final judgment: watch closely.

---

10) 江波龙301308ChiNext ±20% —— watch only (the most important falsification in this report) · Longsys

⚠️ A contradiction that must be explained: apparently the cheapest in the market + single-quarter results still accelerating + yet lagging and net sold on a day when the sector surged

First, the results — they have not deteriorated; they are accelerating:

Metric	2026Q1	2026Q2	QoQ
Revenue	RMB 9.909 billion	RMB 14.180 billion	+43.1%
Ex-non-recurring net profit	RMB 3.943 billion	RMB 6.105 billion	+54.8%
Gross margin	55.53%	61.26%	+5.73pct

Table scrolls horizontally

H1 totals: revenue RMB 24.088 billion +136.26%, attributable profit RMB 10.577 billion, ex-non-recurring RMB 10.047 billion, EPS RMB 25.20. The statements are extremely strong; there is no direct evidence of "earnings peaking".

Now how the market prices it — the exact opposite: yesterday the memory chip concept rose 4.81% with net main-force inflow of RMB 20.184 billion (first among concepts), while 江波龙 (Longsys) rose only 3.15%, saw net main-force outflow of RMB 109 million, and its volume ratio of 0.97 is the only reading below 1 in this report.

Four explanations, all pointing to the same conclusion:

① The "lowest P/E in the market" is an illusion created by the annualization algorithm.

Stock	Dynamic P/E (shown by quote software)	TTM ex-non-recurring P/E (recomputed)	Multiple gap	PB
江波龙 (Longsys, 301308)	8.5	15.9	1.9×	8.10
德明利 (Demingli, 001309)	7.4	24.2	3.3×	14.91
佰维存储 (Biwin Storage, 688525)	10.4	31.7	3.0×	14.23
(reference) 复旦微电 (Fudan Microelectronics)	27.6	125.8	4.6×	6.70
(reference) 中微半导 (Zhongwei Semiconductor)	54.9	74.5	1.4×	5.84

Table scrolls horizontally

The dynamic P/E treats peak-cycle single-quarter (or half-year) profit ×4 (or ×2) as a full year. Demingli's dynamic 7.4x versus a true 24.2x is a 3.3x gap — and its "best single quarter" happens to be 2026Q1, with margins falling 15.3 percentage points the very next quarter (see ④).

② The divergence between P/E percentile and PB percentile is 3x that of the design companies.

Stock	P/E (TTM)	P/E historical percentile	PB	PB historical percentile	Percentile gap
江波龙 (Longsys, 301308)	15.0	0.9%	9.69	79.5%	+78.6pct
德明利 (Demingli, 001309)	24.1	11.2%	14.91	79.5%	+68.3pct
佰维存储 (Biwin Storage, 688525)	30.7	7.9%	14.23	74.4%	+66.5pct
(reference) 普冉股份 (Puya Semiconductor)	157.4	72.2%	25.72	97.9%	+25.7pct
(reference) 兆易创新 (GigaDevice)	108.4	57.5%	12.33	81.4%	+23.9pct

Table scrolls horizontally

The three module makers show percentile gaps of 66–79 points, while the two design companies show only 24–26 points. The difference is not about cheap versus expensive but about the sustainability of the denominator E: the module makers' P/E percentiles sit at historical lows of 1%–11% while their PB percentiles are at 74%–80% — the market is pricing them on PB, which amounts to a verdict that current earnings cannot be capitalized.

Basis qualification (must be stated): none of the three has been listed for a full 5 years (江波龙 Longsys listed 2022-08), so there is no "5-year percentile"; the table uses full comparable-history percentiles, with a sample covering only one complete cycle, statistically weaker than for mature names.

③ The profit has not turned into cash, and it was stockpiled with borrowed money.

Longsys's H1 attributable profit was RMB 10.577 billion, while net operating cash flow was −RMB 3.151 billion (per the interim report). Broken out, the problem is not on the collection side:

• cash received from sales of goods RMB 23.352 billion vs revenue RMB 24.088 billion → a cash collection ratio of 96.9%, customers are paying normally;
• cash paid for goods RMB 24.469 billion > sales cash receipts RMB 23.352 billion → 100% of the cash gap comes from the procurement side;
• borrowings obtained in the same period RMB 9.559 billion, with long-term borrowings rising from RMB 4.377 billion at the start of the year to RMB 10.493 billion (+139.7%).

That is: roughly RMB 9.6 billion was borrowed to buy more wafers. This is a leveraged long inventory position betting that prices keep rising.

Inventory structure (from the original table in the interim report's "Inventory" note): total inventory RMB 25.875 billion, of which raw materials (memory wafers) RMB 14.370 billion, up 187% in half a year; inventory is 60.12% of total assets and 1.39x net assets; yet only RMB 97.4 million of write-down provisions have been made, a ratio of 0.38% — almost no cushion.

④ The accounting source of the gross margin: it is selling goods stockpiled at low prices, not an improvement in operating efficiency.

From the original product breakdown in Longsys's H1 report: revenue +136.13% while cost of sales rose only +11.41%, with gross margin expanding 45.80 percentage points YoY to 59.08%.

Revenue doubling while costs barely moved = matching low-priced inventory against high selling prices, i.e. time-lag arbitrage. And the RMB 14.370 billion of raw materials now sitting on the balance sheet were bought at high prices, so the cost base of the next shipments must rise sharply.

Reverse operating leverage estimate (holding Q2 revenue at RMB 14.179 billion and the expense ratio at 11.0%):

Assumed gross margin	Quarterly operating profit
61.3% (26Q2 actual)	+RMB 7.13 billion
30%	+RMB 2.69 billion
19.4% (full-year 2025)	+RMB 1.19 billion
11.0%	0 (breakeven line)
8.2% (2023 actual)	−RMB 400 million

Table scrolls horizontally

Actual gross margins in the last down-cycle: 江波龙 (Longsys) 8.2% in 2023 (a loss of RMB 828 million), 佰维 (Biwin) 1.7% in 2023 (a loss of RMB 624 million). The historical gross margin range of this business is 2%–61%; there is no profit floor.

Final judgment: watch only. The statements are the most beautiful in the whole report, but the beauty of the statements is precisely the risk itself.

⚠️ A boundary of inference that must be written down: Longsys's P/E percentile (0.9%) is more extreme than Demingli's (11.2%), with inventory at 1.39x net assets and cash flow of −RMB 3.151 billion. It is telling the same story as Demingli, only Demingli has already turned the page first (see §6). Treating Longsys as a "safe low P/E" and treating Demingli as "the cheapest low P/E" are two stages of the same error.

⚠️ Counter-side note (required by yesterday's recap lesson): yesterday's reconciliation showed that during periods of euphoria, fundamental negatives are diluted by a broad rally, and the pricing of a falsification is postponed rather than cancelled. If the memory main line keeps strengthening today, Longsys could well catch up or even hit limit-up. Judging that its "quality is questionable" and predicting that it "will fall today" are two different things; this report asserts only the former. Yesterday the Pass group outperformed the recommended group by 2.35 percentage points, which is the price of this note.

⚠️ Another counter-indication: on 8/11, the day it disclosed its interim report (attributable profit up several hundred-fold YoY), it rose only 1.21%, then −0.77% on 8/12 and −4.18% on 8/13. Near-zero reaction to such an extreme set of statements shows the positive was long since priced in — which both supports "no need to chase" and means part of its downside has likewise been absorbed.

⚠️ Counter-side note (required by yesterday's recap lesson): yesterday's reconciliation showed that during periods of euphoria, fundamental negatives are diluted by a broad rally, and the pricing of a falsification is postponed rather than cancelled. If the memory main line keeps strengthening today, Longsys could well catch up or even hit limit-up. Judging that its "quality is questionable" and predicting that it "will fall today" are two different things; this report asserts only the former. Yesterday the Pass group outperformed the recommended group by 2.35 percentage points, which is the price of this note.

---

6. Pass List

⚠️ Dedicated Note on 太辰光 (T&S Communications) (recording a methodological failure faithfully)

The common reason for passing on T&S is "H1 ex-non-recurring −2.55% YoY, revenue up but profit down". That constitutes selective citation and must be corrected:

The two bases point in completely opposite directions, and the H1 decline comes entirely from Q1 (RMB 61.61 million vs RMB 78.21 million in the prior-year period, −21%); Q2 has already reversed.

The more serious problem is this: §4 of this report has just laid down the rule that "earnings elasticity is changed to be primarily single-quarter QoQ, because cumulative YoY double-counts already-disclosed old information". And yet on this one name I used the very cumulative basis I had downgraded to pass it, while ignoring that on my own designated primary metric it is one of the strongest QoQ-elasticity names in the report. The rule was executed only where it was unfavourable to a particular stock — the same class of error criticized in yesterday's recap.

The Pass rationale for T&S has therefore been narrowed to three independent grounds unrelated to earnings: ① TTM ex-non-recurring P/E 158.8 and PB 26.39 are both the highest in the table; ② operating cash flow per share −0.320; ③ the Dragon-Tiger List was hot-money dominated (Huaxin Shanghai Lujiazui +RMB 108 million) while institutions were only RMB 61 million. "Weak earnings" is no longer a reason to Pass.

---

7. Intra-Branch Rankings

Branch 1: Memory / DRAM / NAND (strength S)

• Hardest single name: 长鑫科技 (CXMT, 688825) — the only IDM, price increases go straight into the statements.
• Best profit quality: 兆易创新 (GigaDevice, 603986). See "the cash divide" below.
• Highest trading visibility: 长鑫科技 (CXMT, 688825).
• Downgraded to watch only / Pass: the three module makers — not because results are weak, but because of the combination of "dynamic P/E illusion + PB percentile of 74–80% + negative operating cash flow"; among them Demingli has already shown single-quarter margin compression and is Passed outright.

⚠️ The cleanest dividing line in this branch: did the profit become cash, or inventory

All three module makers show "sharply higher net profit + negative operating cash flow"; both design companies show positive conversion. This dividing line is cleaner than any valuation metric — within the same upcycle, one group's profits turned into cash and the other's turned into inventory.

⚠️ A conceptual illusion that must be dispelled: none of the three module makers produces HBM

"HBM" appears 41 times in 佰维存储 (Biwin Storage)'s 2025 annual report, but after checking every instance in context — all of them sit in the "industry overview" chapter, describing things like HBM market size as industry background, with not one referring to Biwin's own products. Of the 5 HBM mentions in 江波龙 (Longsys)'s interim report, 3 compare the specifications of its in-house AIDIMM with the HBM2/HBM2E figures on SK Hynix's website (i.e. "I am not HBM, but the bandwidth is close").

HBM is the business of Samsung / SK Hynix / Micron. Keyword frequency will make you read "industry background" as "company business".

Verifiability of each company's AI-related revenue: 江波龙 (Longsys) discloses enterprise-grade storage H1 revenue of RMB 2.140 billion (+208.80%), but that is only 8.9% of total revenue — the only one of the three to disclose an amount; 佰维 (Biwin) had 2025 AI glasses revenue of RMB 960 million (real, but consumer edge-side); 德明利 (Demingli) offers only qualitative description with no disclosed amount, and its product mix is 100% consumer-grade.

The main cause of all three companies' profit jumps is price increases and inventory appreciation in consumer-grade SSD / eMMC / memory modules, not a structural pull from AI demand.

Branch 2: Packaging & testing / advanced packaging / substrates (strength A+)

• Hardest single name: 长电科技 (JCET, 600584) — the only one where "AI attribution + ex-non-recurring delivery + healthy balance sheet" all hold at once.
• Highest trading visibility: 通富微电 (TongFu Microelectronics, 002156).
• Highest institutional content: 兴森科技 (Fastprint, 002436, 76%), but its fundamental support is the weakest — that divergence is itself the most worthwhile thing to watch in this branch.
• Who is Passed: 甬矽电子 (Forehope Electronic, 688362), 华天科技 (Tianshui Huatian, 002185).

Branch 3: Semiconductor equipment / materials (strength A−)

• What makes this branch special: yesterday A-share equipment names (+3.46%+4.48%) clearly lagged memory (+12%+16.81%), so money has not rotated from memory into equipment. That is both an opportunity (not yet started) and a risk (it may never rotate).
• Direction reminder: equipment orders point to new supply in 2027–2028; do not treat rising equipment names as corroboration of a memory shortage.

---

8. Opening Verification Signals for Today

8.1 Before the open (08:00–09:15) —— a lead signal unique to today

The Korean market was closed yesterday (Liberation Day substitute holiday), so SK Hynix and Samsung did not trade for an entire session. Korea reopens today, giving this memory rally its first independent pricing outside the US market.

⚠️ Read timestamps (extremely easy to be off by an hour)

The Korea Exchange opens at 09:00 Korea time (KST). Korea is 1 hour ahead of Beijing, i.e. = 08:00 Beijing time. Not 09:00. That is 75 minutes before the A-share call auction (09:15) and 90 minutes before the A-share open (09:30). Suggested read timestamps: 08:05 Beijing time (opening reaction) and 08:30 (whether it holds), twice.

This is today's highest-information-value lead observation point, and the reading is available 90 minutes before the A-share open.

8.2 Auction signals (09:15–09:25)

• 长鑫科技 (CXMT, 688825)'s opening gap is the master gate for the main line: a gap of +3%~+8% is healthy; a gap above +12% erases the expectation gap and makes chasing very poor value; flat or lower open = the main line was a one-day affair.
• 复旦微电 (Fudan Microelectronics, 688385): a gap of 0%~+8% is the best observation zone; a gap above +8% should trigger a downgrade to watch only; if it opens lower, beware of the market reading "attributable profit at the lower bound of guidance + impairment provision" as a negative.
• 长电科技 (JCET, 600584): yesterday's lack of a limit-up is its relative advantage. If today's opening gap exceeds TongFu Microelectronics', the market is starting to re-rank the three packagers by "profit quality" rather than "flow heat" — the inference this report most wants to verify.
• 兴森科技 (Fastprint, 002436): yesterday's seal-order ratio was only 0.037. If it gaps up and then quickly falls back, yesterday's afternoon seal was a sentiment pulse; if it opens flat and strengthens, the institutional buying is confirmed.
• Unexecutable one-line board note: 中石科技 (Jones Tech, 300684) seal-order/turnover 2.928, 桂发祥 (Guifaxiang, 002820) 2.378, 天洋新材 (Tianyang New Materials, 603330) 1.405 — most likely still impossible to buy today, and should not be counted in any return expectation.

8.3 Sector signals

• Healthy: 3 or more memory-chain names hitting limit-up quickly + CXMT rising on volume + a second/third-board ladder emerging.
• Warning: of yesterday's 106 limit-ups, 91 were first boards and the highest was only 4 boards — the ladder is extremely thin. If the number of first boards keeps expanding today while the consecutive-board height does not rise, that is money spreading out rather than a theme relay.
• Liquidity-siphon note: 频准激光 (Pinzhun Laser, 688826) lists on the STAR Market today, with an IPO price of RMB 186.88, the most expensive new issue of the year. Per the known pattern, a newly listed stock significantly pollutes the flow data of its sector (laser/semiconductor equipment) on listing day — it must be excluded when tallying sector net main-force amounts after the close today, otherwise the magnitude will be distorted.

8.4 Stock-level signals

• Top-5 absorption: watch the intraday absorption and large-order net inflow of 长鑫科技 (CXMT, 688825), 长电科技 (JCET, 600584) and 通富微电 (TongFu Microelectronics, 002156).
• Seal quality: yesterday the seal-order/turnover of semiconductor limit-ups was generally 0.03–0.09 (TongFu 0.050, Fastprint 0.037, CETC Chip 0.050), extremely weak. Only if semiconductor limit-ups with seal-order/turnover above 0.3 appear today can ladder quality be said to have substantively improved.
• Blocked by a stronger name: if 复旦微电 (Fudan Microelectronics, 688385) lags while 中微半导 (Zhongwei Semiconductor, 688380) keeps strengthening, the market is recognizing only one name on the "earnings delivery" line.

8.5 Risk signals (downgrade on appearance)

1. SK Hynix / Samsung opening flat or lower after Korea reopens — the highest-priority falsification signal.
2. 长鑫科技 (CXMT, 688825) gapping up and fading — if the core anchor of the main line does not follow, the whole line is void.
3. Yesterday's limit-ups falling back as a group: yesterday the "yesterday's limit-up" concept saw net main-force outflow of RMB 3.376 billion, prices up but money out; if outflows continue today, there is no relay money.
4. US-Iran tensions / oil prices continuing higher: WTI +3.09% yesterday to about US$84.95, 30-year Treasury 5.311%. This has little impact on semiconductors but suppresses consumption, property and cyclicals; if the index is dragged down, the wealth effect in tech names will be diluted.
5. Two-way risk during the dense interim reporting period: 兴森科技 (Fastprint) 8/21, 通富微电 (TongFu Microelectronics) 8/26, 德明利 (Demingli) 8/28 — any miss would ripple through the whole branch.
6. High-level junk names receding: the 4-board names 天洋新材 (Tianyang New Materials, 603330), 金螳螂 (Gold Mantis, 002081) and 澳洋健康 (Aoyang Health, 002172) are all unrelated to the main line, and Gold Mantis had 9 failed seals yesterday. If this group weakens together, it will cool overall sentiment.

8.6 ⚠️ Retracting an inference: "afternoon sealing = not fully priced" does not hold

"Semiconductor limit-ups clustered in afternoon seals" is easily taken as evidence that "the main line has not been fully priced → supports continuation today". This inference should be retracted, for three reasons, each sufficient on its own to negate it:

① This report assigns opposite signs to the same fact. In the summary, "afternoon sealing" is bullish; but §5 writes of 兴森科技 (Fastprint) that it was "driven by afternoon money rather than morning accumulation… weak seal quality", and §8.2 goes further, saying it "shows yesterday's afternoon seal was a sentiment pulse". The same timestamp is bullish in the summary and bearish at the stock level, depending on which conclusion needed support at the time.

② The evidence set itself is contaminated. Of the 6 names that sealed in the afternoon: 中瓷电子 (CETC Chip Technology, 13:21) and 金禄电子 (Jinlu Electronics, 13:38) are Passed by this very report, on the grounds of "pure hot money, no catalyst"; 奥来德 (Olead, 14:42) is an OLED materials company and does not belong to the semiconductor main line, and appears nowhere else in the text. Using a 6-name sample in which 2 are self-judged hot money and 1 is outside the industry to argue that "the main line is unpriced" makes the sample itself a counter-argument.

③ It does not hold logically. A limit-up is the day's price ceiling, so regardless of the sealing time, the day's information is already priced into the board; "what time it sealed" measures the speed at which the bid formed, not the completeness of pricing. And yesterday's semiconductor seal-order/turnover of just 0.037–0.050 (which §8.4 itself calls "extremely weak") points precisely to a thin queued bid, the opposite direction from "not fully priced". This report also provides no base rate (how afternoon-sealed names have historically performed the next day).

It is therefore rewritten as a neutral description, with no directional judgment drawn from it:

Yesterday's semiconductor limit-ups were all formed after 11:00, and seal-order/turnover was generally 0.03–0.09. Both characteristics are present. A bull can read it as intraday capital migration, a bear as chasing without morning consensus — this report draws no directional conclusion from it.

Observation point available today to adjudicate: if semiconductor names seal quickly in the morning session (09:30–10:30) with seal-order/turnover rising above 0.3, bid quality has improved; if action again comes only in the afternoon, or names gap up and fade, yesterday was a late-session pulse.

---

9. Final Recommendation Conclusions

① The 5 stocks most worth watching today

Note: of the 5, only 复旦微电 (Fudan Microelectronics, 688385) is tagged "priority deep-dive"; the rest are "watch closely". This distribution is deliberate — this report has only one catalyst that qualifies as in-window "new information" (Fudan's ex-non-recurring structure); everything else is a continuation of yesterday's action or transmission of overseas sentiment.

Three ranking points that need explaining (stated honestly): ① 兴森科技 (Fastprint) is not in this table, because its "AI substrate" logic lacks revenue support upon verification (see §5-7); ② 长电科技 (JCET) is ranked 3rd, because it is the only packager with no water in its profits; ③ 通富微电 (TongFu Microelectronics) is ranked 5th rather than higher, because non-recurring gains and losses account for 56%.

② The 3 strongest branches today

③ Directions not recommended for chasing today

1. The three memory module makers (江波龙 Longsys 301308, 德明利 Demingli 001309, 佰维存储 Biwin Storage 688525). "Cheapest in the market" is an annualization illusion of the dynamic P/E: Demingli's dynamic 7.4x versus a true TTM 24.2x; Biwin's dynamic 10.4x versus a true 31.7x. P/E historical percentiles of 0.9%–11.2% against PB historical percentiles of 74%–80%, a divergence of 66–79 points, 3x that of the design companies (24–26 points). Add Longsys's H1 operating cash flow of −RMB 3.151 billion (inventory 60.1% of total assets, write-down provisions only 0.38%) and Demingli's Q2 single-quarter attributable profit already −17.7% QoQ — this is the shape of a cycle top, not a valuation trough.
2. Buying semiconductor equipment names as "corroboration of a memory shortage." The direction is the opposite: equipment orders point to new supply in 2027–2028.
3. Optical comms / CPO. Overnight Coherent +7.79% and Lumentum +4.62% are tempting, but the three A-share leaders were distributed as a group yesterday (亨通光电 Hengtong Optic-Electric −RMB 919 million, 中天科技 ZTT −RMB 358 million, 剑桥科技 CIG −RMB 403 million). This is the only branch in this report where overseas and A-share flows point in opposite directions.
4. The two packaging names with an excessive share of non-recurring gains and losses: 华天科技 (Tianshui Huatian, 002185, about 68%), 甬矽电子 (Forehope Electronic, 688362, TTM P/E about 400).
5. 中瓷电子 (CETC Chip Technology, 003031) and high-level pure hot-money names. CETC Chip has had no filing for a third straight day and its attribution piece is AI template reuse; 蓝盾光电 (Landun Photoelectron, 300862) had a single hot-money seat buy RMB 334 million and sell 0; 金禄电子 (Jinlu Electronics, 301282) had institutional net of only RMB 3 million.
6. 太辰光 (T&S Communications, 300570). +20% limit-up yesterday, TTM ex-non-recurring P/E 158.8 and PB 26.39 both the highest in the table, operating cash flow per share −0.320, Dragon-Tiger List hot-money dominated with institutions at only RMB 61 million. ⚠️ But it must be stated: its Q2 single-quarter ex-non-recurring was +67.2% QoQ, so earnings are not weak — using "deteriorating earnings" as the Pass reason is selective citation (see the dedicated note in §6). The basis for this item is valuation and flow structure, not earnings.
7. One-line boards with a seal-order/turnover ratio > 1. 中石科技 (Jones Tech, 300684) 2.928, 桂发祥 (Guifaxiang, 002820) 2.378, 天洋新材 (Tianyang New Materials, 603330) 1.405 — a limit-up you cannot buy produces no return.

④ Final one-sentence judgment

The memory main line received a second independent confirmation in the overnight US session (Philadelphia Semiconductor Index +1.64% against all three major indices in the red), which is the hardest piece of evidence within this report's field of observation; but the latest "catalyst" driving this bout of sentiment is merely a three-word social media reply, and the module segment's "lowest P/E in the market" turns out on verification to be an annualization illusion of the dynamic basis — the true TTM is 2–3.3x that.

This report's observation framework can be summed up in one sentence: within the same upcycle, one group of companies' profits turned into cash (兆易创新 GigaDevice operating cash flow/net profit +122%, 普冉股份 Puya Semiconductor +97%) and another group's turned into inventory (江波龙 Longsys −30%, 佰维存储 Biwin Storage −93%). This dividing line is cleaner than any valuation metric, and it is this report's sole basis for distinguishing "profits that survive the ex-non-recurring and cash flow tests" from "profits propped up by a price-increase narrative".

This report gives no position, trading or stop-loss advice. Everything listed above consists of objects of observation and verification points; every verification signal in §8 also states its falsification direction — if SK Hynix opens flat or lower after Korea reopens, this report's strength judgment on the memory branch does not hold.

---

⚠️ Risk warning: this list is a pre-market information review and observation only and does not constitute investment advice. A-share volatility risk is extremely high, and automatically generated content may contain timeliness gaps or industry-chain mapping errors; it must not be used directly as a basis for trading.

Data-Sourcing and Run Log (not sent to clients)

1. This report executed the 7 input items left by §6.5 of yesterday's recap

2. Record of self-retractions before finalization (7 in total, 3 of them from sub-agents)

4 found by the main slot's own checks:

1. The entire rationale for Fudan Microelectronics was wrong. The draft treated "attributable profit +338.58%" as the core expectation gap. Tracing to the first disclosure date showed the 7/07 pre-announcement had already given a +313%~+416% range and had disclosed the RMB 470 million SJ Semiconductor fair value gain, and the actual RMB 849 million still landed at the lower bound of the range. Rewritten as "the only new information is the ex-non-recurring figure and single-quarter structure".
2. SJ Semiconductor. The draft at one point wanted to write "Fudan Microelectronics' paper gain on its SJ Semiconductor stake" as a standalone industry-chain catalyst. Verification showed SJ Semiconductor listed on 2026-04-21, so the paper gain is a Q2 fair value revaluation, not a new event. Deleted.
3. Musk's "memory bottleneck". Several media outlets used it as a headline catalyst. Tracing to the source confirmed it is a three-word reply to someone else's view ("Few realize this"), with no new data; downgraded from S to B with a separate downgrade note.
4. SanDisk's 80% gross margin target. The draft leaned toward writing it up as "an overly aggressive red flag". Verification confirmed the target is below SanDisk's currently achieved level, i.e. conservative guidance; the bearish statement was deleted.

3 corrected by sub-agents (all substantive; the main slot would not have found them):

5. ⚠️ The P/E basis was wrong throughout (the highest-impact item). The draft labelled East Money's f9 field as "P/E (TTM)", when it is in fact the dynamic P/E (latest reporting period annualized). A sub-agent verified this name by name by back-calculating "market cap ÷ single-quarter net profit ×4". This error would have made the three memory module makers read as "the cheapest in the market", when the true TTM is 1.9–3.3x the dynamic value; more seriously, for Fudan Microelectronics the dynamic 27.6 versus TTM ex-non-recurring 125.8 is a 4.6x gap, and the draft used "lowest P/E" as one of the core reasons for putting it in the top recommendation slot. A "valuation basis statement" was added in §3, the whole table was converted to two columns (dynamic / TTM ex-non-recurring), and a "distorted valuation basis" deduction item was added in §4.
6. Fastprint's "AI substrate" logic was falsified. The draft used "AI packaging → ABF substrates" as its benefit path and put it 3rd in the Top 5. From the 2025 annual report's revenue breakdown and the original text of the 2026-05-08 investor relations record, a sub-agent found: IC packaging substrate gross margin −16.06% (loss-making), FCBGA only in small-batch production, asked twice whether it supplies Nvidia/AMD and answering "striving for a breakthrough" (i.e. not yet), and zero of the RMB 3.9 billion placement going to FCBGA/ABF. Removed from the Top 5, with the benefit path rewritten as "memory price increases → CSP substrates".
7. Fastprint's institutional net buying was miscalculated. The draft said RMB 207 million, missing one institutional seat of +RMB 53.9 million; the correct figure is RMB 261 million (76% of the buy side). Corrected.

4 additional contributions from the memory-module specialist agent (none found by the main slot):

8. Independently reproduced the P/E basis error — two agents using different methods (one back-calculating packaging names, one back-calculating module makers) reached the same conclusion, so this correction is doubly confirmed.
9. Demingli's Q2 single-quarter attributable profit −17.7% QoQ: revenue +25.5% QoQ while attributable profit fell 17.7%, with net margin compressed 15.3pct in one quarter; the agent also ruled out one by one the three technical explanations of share-based payment (only about a 4pct drag), impairment reversal (Q1 was a provision of +RMB 19.97 million rather than a reversal) and non-recurring items (Q1 ex-non-recurring ≈ attributable), judging it to be genuine operating compression. "The lowest P/E in the market" and "the only QoQ decline" land on the same company — the complete mechanism of a low-P/E trap. Demingli was accordingly downgraded from "watch only" to Pass.
10. Corrected the main slot's crude classification that "all three module makers have peaked": Longsys was +73.9% QoQ and Biwin +50.1%, so two are still accelerating; the peaking signal is not in the income statement but in the balance sheet and cash flow statement. The body text has been rewritten accordingly.
11. Dispelled the HBM conceptual illusion: all 41 occurrences of "HBM" in Biwin's 2025 annual report sit in the industry overview chapter; 3 of Longsys's 5 occurrences are comparisons with SK Hynix's HBM. None of the three produces HBM.

Three further upgrades/downgrades brought by sub-agents: JCET was raised from 7th to 3rd in the Top 5 (the only one of the three packagers with non-recurring items at about 4%); TongFu Microelectronics was lowered from 4th to 5th (non-recurring 56%, P/E 86–94x on an ex-non-recurring basis rather than 40x); Demingli was downgraded from "watch only" to Pass.

An honest self-assessment: this report involved 11 self-retractions + 11 QA corrections, of which only 4 were found independently by the main slot, 7 came from fundamentals-analyst and 11 from risk-auditor. The ratio is unbalanced — especially on the P/E basis, which simultaneously contaminated the entire §3 table, three sections of §5 and the ranking in §9, and yet the main slot felt no doubt whatsoever when writing the header "P/E (TTM)". This is consistent with yesterday's recap's judgment that "recaps are naturally easier to self-falsify than pre-market reports": the pre-market slot's self-checking ability has not improved along with recap quality, and dependence on sub-agents is a current structural fact, not an accident of this particular report.

3. Data-sourcing failures and interface status

• push2.eastmoney.com still returns intermittent 502s; industry/concept sectors were switched to the clist interface on push2delay.eastmoney.com, which worked.
• Limit-up pool validation passed: the official push2ex pool has tc=106 == len(pool)=106.
• The Dragon-Tiger List was available (published after 18:00), 60 entries. stock_lhb_stock_detail_em returned TypeError: NoneType for CETC Chip Technology, Chipown and Changyingtong — verification showed these three did not meet the listing threshold and never entered the Dragon-Tiger List; it is not an interface failure.
• ⚠️ The interim pre-announcement interface RPT_PUBLIC_OP_NEWPREDICT has unreliable data and has been abandoned. Spot checks found Puya Semiconductor's pre-announced net profit shown as RMB 3.950 billion, while its 2026Q1 single-quarter net profit was only RMB 251 million, so H1 could not possibly reach that magnitude; Montage Technology's pre-announced value also back-calculates to a 2025H1 that contradicts disclosed single-quarter data. No figure from this interface is cited anywhere in the body text; the guidance-basis data for JCET/TongFu/Huatian in §5 came from sub-agents reading the original company filings and are separately sourced.
• CNBC @CL.1 (WTI) gave +0.56%, while the USO crude ETF was +2.91% and Trading Economics +3.09%. Two against one, so @CL.1's change_pct baseline was judged faulty, and the body text adopts +3.09% with the source noted. This is another reproduction of "an ETF proxy detecting a silently stale spot quote".
• CNBC .KS11 (KOSPI) last_time was stuck at 2026-08-14, initially judged to be dead data. Verification confirmed it was the Korean market holiday on 8/17 (Liberation Day substitute holiday), not interface lag — and that dug out the lead verification point in §8.1. This "suspected bad data" turned out to be the highest-value discovery in the report.
• stock_zh_a_disclosure_report_cninfo threw a TypeError, so the CNINFO channel was unused and filings were switched to East Money's np-anotice-stock interface (799 entries).
• The ex-non-recurring YoY field in East Money's interim database RPT_LICO_FN_CPD returned NaN, so ex-non-recurring data was taken from stock_financial_abstract_ths (single quarter) and stock_financial_abstract, cross-checked and consistent.

4. Back-calculation checks on key figures

• Longsys operating cash flow: East Money's interim database gives "operating cash flow per share −7.449". Back-calculating share count from EPS 25.20 and attributable profit RMB 10.577 billion = 419.7 million shares, which multiplies out to about −RMB 3.13 billion. The body text flags this as a back-calculated value.
• CXMT market cap: f116 gives RMB 4.13 trillion, consistent with "66.881 billion total shares × RMB 61.80"; the dynamic P/E of 41.7 matches Q1 net profit of RMB 24.762 billion annualized. f162/f167 must be divided by 100; the first reading was mistakenly recorded as P/E 4173 / PB 2966 and has been corrected.
• T&S Communications ex-non-recurring: the single-quarter table gives Q2 ex-non-recurring +13.60% YoY, but H1 cumulative RMB 164.32 million vs RMB 168.62 million in the prior-year period = −2.55%. The two bases point in opposite directions. The body text adopts the H1 cumulative basis and labels it explicitly, on the grounds that "the interim ex-non-recurring decline" is the fact that contrasts with the +20% share price move — but this selection is indeed selective, and it has been submitted to the QA agent for dedicated review.
• TTM ex-non-recurring P/E recomputation: summing the four most recently disclosed single quarters of ex-non-recurring profit. Fudan Microelectronics RMB 372 million → 125.8x; Longsys RMB 11.305 billion → 15.9x; Demingli RMB 4.077 billion → 24.2x; Biwin RMB 3.817 billion → 31.7x; T&S Communications RMB 277 million → 158.8x; Zhongwei Semiconductor RMB 252 million → 74.5x.

5. QA (risk-auditor) results and disposition

QA issued 14 red flags in total. 11 were corrected in the final draft, as listed below:

The 3 undisposed items (stated honestly):

• The second half of red flag 5: §1 says the Musk item was "traced to the primary source", yet the only link given is the second-hand 24/7 Wall St. round-up, and the X post permalink, posting timestamp and Diamandis's original text were never obtained. Per this repository's lesson that "Chinese-language sources corroborating each other is an echo, not verification", there is likewise only a second-hand source here; the phrase "traced to the primary source" is not warranted and should be changed to "relayed via second-hand sources, the original post not obtained" — not changed due to time constraints, recorded as a defect.
• Red flag 7: Fastprint's "single-day Dragon-Tiger List institutional net buying" is still one of the grounds for its inclusion, while §4 of this report has just laid down the rule that single-day flow extremes must not be used for ranking. It has been removed from the Top 5 and its benefit path rewritten as "memory price increases → CSP substrates", but the metric was not annotated with "a single-day reading, with a high probability of mean reversion the next day".
• The PB control group in red flag 14: PB is disclosed only for the three module makers plus 4 reference names; JCET, Shenzhen Kaifa, Shannon Chip Creation, Naura, Montage, AMEC, Piotech, Nata Opto-electronic, Huazheng New Material and Jinjian Cereals all lack PB, so the cross-stock PB comparison lacks a complete control group.

6. Known incomplete / pending verification

• Both fundamentals-analyst sub-agents returned and produced substantive corrections (see §2), but three items remain unobtained: first-hand data on Fastprint's IC substrate capacity and customer qualifications, CXMT's primary financial statements and absolute net asset figure (so the denominator of PB 29.66 cannot be verified), and CXMT's lock-up expiry schedule.
• Longsys's contract liabilities of RMB 2.217 billion for 2026Q1 come from a single source, Tonghuashun; the interim report discloses only RMB 1.330 billion at period end and RMB 355 million at the start of the year. The forward-looking signal "contract liabilities −40% QoQ" is therefore a single-source conclusion; it is not used as an argument in the body text and is noted here for reference only.
• Q2 2026 single-quarter data for CXMT, GigaDevice, Montage Technology, Shannon Chip Creation and Shenzhen Kaifa are all unobtained (interim reports not disclosed).
• The JPMorgan food crisis report in the seed branch could not be traced to a primary source or release date, so it was treated as grade C and annotated to note that all the facts it cites predate the scan window.
• Demingli's Q2 single-quarter attributable profit of −17.7% QoQ is derived from "guidance midpoint − Q1 actual" (the guidance range is RMB 5.7–6.5 billion, and the entire range implies a QoQ decline, with even the upper bound at −5.7%, so the conclusion is robust across the range), but the specific cause of the margin compression requires confirmation from the formal interim report (8/28).

6. Methodological notes

Two structural findings from this report are worth recording:

1. The Korean market holiday in §8.1 came from an investigation into "data that looked broken" (.KS11 timestamp lagging 3 days), not from any news search. Same category as last week's "CBOE _DJI is dead data": push freshness checks down to symbol granularity, and the bad data itself will carry information.
2. The distribution of first-seal times in the limit-up pool: all semiconductor names sealed after 11:00 (TongFu 11:00, Tztek 13:00, CETC Chip 13:21, Fastprint 13:33, Jinlu 13:38, Olead 14:42), while the morning limit-ups were agriculture and junk names. This can only be read from first-seal timestamps; no sector performance data can show it. But §8.6 already flags it: this structure supports both the "not fully priced" and the "late-session chasing" readings simultaneously, and must not be used in one direction only.

The third, and the one most worth remembering from this report: the P/E basis error was something the main slot could never have found on its own — both the dynamic P/E and the TTM P/E are "normal numbers", with no anomalous feature to trigger suspicion; it was exposed only when a sub-agent back-calculated from market cap. Layered on top of the conclusion that "memory modules look cheapest", it was enough to invert the entire valuation ranking of the report. From now on, any valuation field taken directly from a quote interface must have its basis verified by back-calculation first.