A-Share Pre-Market Brief | 2026-08-13 Thursday

Window: 2026-08-12 15:00 (A-share close) → 2026-08-13 07:00. All A-share data is on an 8/12 close basis; all US data is on an 8/12 US-Eastern close basis (after 04:00 Beijing time on 8/13).

---

0. Today in one sentence

The strongest catalyst is Lumentum's overnight earnings, not the indices. The three major US indices were close to flat overnight (Dow −0.04%, S&P +0.26%, Nasdaq +0.54%), but the Philadelphia Semiconductor Index rose +2.49%, and optical communications plus the AI server chain produced extreme single-stock moves: Lumentum +13.63%, CoreWeave +19.28%, Super Micro +19.02%, Dell +9.87%, Coherent +8.24%, Seagate +7.03% — this was a narrow but extreme AI hardware move, not a broad rally.

The strongest branch is optical communications/CPO, but it is already in its second leg rather than at the start. A-share communication network equipment and components saw main-force fund inflows on two consecutive days from 8/11 to 8/12 (+RMB 3.435 billion → +RMB 9.178 billion, a 2.67x amplification, same sub-sector basis), clearing the "two consecutive days in the same direction" gate set by yesterday's recap; the broader communication equipment industry basis was +RMB 13.094 billion on 8/12, ranking first among industries, but this piece did not obtain its 8/11 figure on the same basis, so all continuity judgments are made strictly on the sub-sector basis; overnight, Lumentum delivered hard demand-side evidence with FY27Q1 revenue guidance of US$1.225–1.275 billion and an operating margin of 39.5%–40.5%, which is the third layer of confirmation.

Likely direction of capital: redistribution within the AI hardware chain, not a new direction. This piece judges that today's biggest expectation gap does not lie in optical modules themselves (they already rose on 8/12 and already spread to 20cm catch-up names), but in two positions that have been left behind: ① Super Micro explicitly attributed its revenue shortfall to power, cooling and networking bottlenecks in customer projects, yet A-share liquid cooling concept saw only +RMB 1.172 billion of main-force inflow on 8/12 and copper high-speed connectivity +RMB 2.361 billion, severely lagging the +RMB 13.696 billion into optical modules; ② the memory chain has completely diverged between China and the US — overnight Micron +4.92%, SanDisk +5.76%, Seagate +7.03%, while all four A-share memory module makers (江波龙 Longsys, 佰维存储 Biwin Storage, 德明利 Demingli, 深科技 Shenzhen Kaifa) saw main-force net outflows on 8/12.

Driver type: earnings-driven above all, with the demand side confirmed simultaneously at home and abroad. Overnight Cisco disclosed US$4 billion of AI orders from hyperscale cloud providers and Q1 guidance above consensus across the board, and named optical module product design wins it expects to secure in the next six months; within the same window, Tencent's Q2 earnings call explicitly stated it "has already stepped up compute procurement and will substantially increase compute procurement in Q3," with Q2 capex payments of RMB 59.3 billion — this is one confirmation each from offshore and onshore demand inside the same window, which is harder evidence than simply "US stocks went up." The policy side is not blank either: SASAC Chairman Cheng Fubo visited China Telecom at 22:38 on 8/12 and called for coordinated advancement of a nationally integrated computing power network. The order side, however, really is empty: a full-field scan of 1,053 Shanghai/Shenzhen/Beijing exchange announcements inside the window found only 2 large new-order announcements, so there is no order-elasticity catalyst today.

Pre-market state: a higher open is likely, but this is a "divergence day," not an "acceleration day." Three suppressing factors must be written out side by side: ① optical communications already spread to 20cm catch-up names on 8/12, which usually marks the middle-to-late stage of a main line; ② part of Lumentum's gain comes from expectations of share transfer created by FCC restrictions on Chinese optical modules, and that part is a negative rather than a positive for A-shares (see 2.1); ③ Broadcom (AVGO) closed flat overnight at −0.01%, the only large-cap in the sample listed in section 2.2 of this piece that did not participate.

---

1. News overview

The "blanks" inside this window are information too: ① a full-field scan of 1,053 Shanghai/Shenzhen/Beijing announcements in the window on "winning bid / signing / order / framework agreement / procurement contract / supply / major contract" found only 2 large new-order announcements (CSPC Innovation receiving the US$30 million upfront payment from AstraZeneca, and 灵康药业 Lingkang Pharmaceutical winning an inter-provincial alliance centralized procurement), and not a single order announcement in the AI hardware chain; ② there are no new share subscriptions on 8/13, and lock-up expiries cover only 爱玛科技 Aima Technology (603529) and 共创草坪 Gongchuang Turf (605099), together worth less than RMB 80 million, so there is no unlock pressure today; ③ 中际旭创 Innolight's (300308) H1 results will be reviewed and released by the board on 21 August (per its 8/12 HKEX "Date of Board Meeting" announcement, which takes precedence over the "8/24" previously reported by media), with an interim dividend also under consideration; 新易盛 Eoptolink's (300502) interim report is on 8/25. The results pre-announcements of 天孚通信 T&S Communications (300394) and Eoptolink were published in mid-to-late July, so they are old news and cannot be treated as today's driver.

One announcement that is easily misread: the RMB 108.3 million hydrogen fuel cell contract announced by 和顺电气 Heshun Electric (300141) on 8/12 merely changes the paying entity from CRRC Electric Vehicle to Sichuan China Railway; the total contract value is unchanged, and it is not a new order.

---

2. Strongest catalyst branches in descending order

2.1 One thing that must be spelled out: Lumentum's surge cuts both ways for A-shares

This is the most important judgment in this piece, and also the place most easily led astray by the simplified reading of "US optical comms surged → A-share optical modules follow."

Lumentum's +13.63% breaks into two parts:

So the correct reading is: what can be extrapolated is demand (the total volume and margin of 1.6T/CPO); what cannot be extrapolated is share. A list that treats Lumentum's entire +13.63% as a positive for A-share optical modules will systematically overstate the hardness of today's optical communications catalyst.

At the same time the current true status of the FCC matter must be recorded: as of this piece, the FCC has not issued any formal restrictive document. Innolight's official response of 2026-08-05 reads: "Upon verification, the US Federal Communications Commission has not issued any restrictive document in the relevant field… the company does not comment on the related rumors" (Sina Finance²²). This is policy risk at the rumor stage, not an accomplished fact — it can neither be treated as an enacted negative nor treated as nonexistent.

2.2 The real shape of the overnight move: narrow, and with one clear gap

Indices and single stocks must be looked at separately, or this move will be read as far too strong:

The Dow closed down and the Nasdaq was up only +0.54%, while these AI hardware names rose 5%–19% — money was pulled from elsewhere to buy AI hardware; this was not incremental money lifting everything. Broadcom closing flat is the single line in this table that most deserves attention: as a core name in AI ASICs and networking chips, it did not participate, which says the overnight main line was optical interconnect + memory + whole machines, not the broader concept of "AI chips." Using Broadcom's flat close to question the whole chain would be excessive, but using it to cool down the claim of "a broad AI breakout" is reasonable.

2.3 Capital has only endorsed two-thirds of the "bottleneck beneficiary" logic

Super Micro named three bottlenecks: power, cooling, networking. Mapping those three words back to A-shares and checking the actual 8/12 fund flows one by one gives an inconsistent result — this divergence must be written out, or "bottleneck beneficiary" will be taken as a logic that can be bought wholesale:

Conclusion: in A-shares, "bottleneck beneficiary" stands on only two legs — cooling and connectivity; the power leg is empty. This divergence is itself informative — it shows the market has not taken Super Micro's remark as an undifferentiated sector positive, but is already picking. This piece therefore elevates 英维克 Envicool (002837) to the first representative stock of the branch, and excludes the power direction from recommendations entirely.

⚠️ But this logic has two hard flaws that must be written next to the conclusion, not only in the risk column:

Flaw one: Super Micro's earnings are not "overnight new news"; A-shares already had a full trading day to price them and did not buy. They were released post-close US Eastern 8/11 = approx. 04:20 Beijing time on 8/12, i.e. before the A-share open on 8/12. So "the liquid cooling/connectivity sectors have not reacted yet" cannot be phrased as "the news has not transmitted through"; it can only be phrased as: A-shares already had a full trading day to price it, and capital chose not to buy. These two phrasings point to completely different probabilities — the former is a lag, the latter is a judgment. The draft of this piece used the former, and that was wrong.

Flaw two: the most on-point US name in this chain barely moved that night. Vertiv (VRT) is the leading US name for data-center power and cooling, and it closed 8/12 at just +2.32%, far below Dell's +9.87%, Arista's +6.39% and Super Micro's +19.02% over the same period. If the market really read "power/cooling bottlenecks" as a positive for bottleneck-link suppliers, Vertiv should have been the most direct beneficiary, and it did not rise. This is direct counter-evidence to the "bottleneck → supplier positive" inference, and the draft of this piece placed Vertiv alongside Arista and Dell in the supporting evidence, which was selective use of the same table.

This section's conclusion is therefore downgraded: "bottleneck beneficiary" is this piece's own inference, not a pricing logic the market has already validated; it is weakened both by a full trading day of A-share inaction and by the non-move of the on-point US name. The branch strength is downgraded from A to B+, and 英维克 Envicool (002837) has been withdrawn from the 5-stock recommendation list in section 9.1 and moved to the watch list. The only reason for keeping it as an object of observation is the stock-level capital profile (a main-force share of 10.57%, far above the sector's 0.73%, with a turnover rate of only 3.50%), which is single-stock evidence and cannot be elevated into a sector conclusion.

2.4 [Self-correction] The "China-US divergence in the memory chain" is not an expectation gap; I got the comparables wrong

The draft of this piece treated "US memory names surging overnight while all four A-share memory module makers tracked here saw main-force net outflows on 8/12" as one of today's biggest expectation gaps. On fundamental verification, that judgment is wrong and is corrected here — the error is in the comparable relationship, not in the data.

Micron, SanDisk, Western Digital and Seagate, which rose overnight, are IDMs and media makers. The only A-share counterpart to them is 长鑫科技 CXMT (688825), followed by 兆易创新 GigaDevice (603986), which has in-house and outsourced brand production. Whereas 江波龙 Longsys (301308), 佰维存储 Biwin Storage (688525) and 德明利 Demingli (001309) are memory module makers, whose global peers are the likes of Phison, ADATA and Netac, not Micron. Using the SOX's +2.49% to linearly extrapolate to A-share module stocks is picking the wrong comparable. A-share module makers seeing outflows on 8/12 and US IDMs rising overnight can both be true at once without contradiction — they were never on the same curve to begin with.

More importantly, the money-making mechanisms of the three types of company are fundamentally different:

The "price-change-rate type" must be spelled out separately, because its risk is asymmetric:

The sequential increases in DRAM/NAND contract prices are decelerating off a cliff — far more severely than the "marked slowdown" written in the draft of this piece. Going back to the primary source (TrendForce, published 2026-07-03, not the 8/8 Sina secondhand write-up):

The first derivative of price is still positive; the second derivative has already turned sharply negative. For distributors and module makers that earn the timing spread of "buy low, sell high," this means the cost base of inventory built at Q2 highs is catching up with selling prices. Prices need not fall; they only need to stop rising for a 58.90% module gross margin to revert toward a normal 15%–20% (Longsys's 2025Q1 gross margin was only 10.35% and Biwin's 1.99% — that is the normal level of this business).

And their inventory leverage looks like this: Longsys inventory of RMB 25.78 billion = 1.39x net assets attributable to parent, with 340.5 turnover days (≈4.7 quarters); Biwin inventory of RMB 12.07 billion = 1.42x net assets, 282 turnover days; Shannon Semiconductor inventory grew from RMB 2.60 billion to RMB 6.97 billion within 9 months, inventory/net assets 1.43x, gearing ratio 75.40% (the highest in the group). A 10%/20% inventory write-down would hit Longsys's net assets attributable to parent by about −13.9%/−27.8% (estimated by this piece, not disclosed by the company).

At the same time, the valuation profile of these three is a textbook sample of "low PE percentile + high PB percentile":

Mechanism: the E in the PE denominator is stretched to its limit by peak profits at a cycle top, while the PB denominator is net assets, which does not jump with a single quarter of windfall profit. So a systematic divergence between PE percentile and PB percentile is itself a direct reading of "E is at a top." A 6.5x annualized PE is not cheap — cyclicals always look cheapest at the top.

One external corroboration: overseas IDMs trade on a forward PE of just 3.6–6.3x (Micron 6.04, SanDisk 6.30, SK Hynix 3.64, Samsung 3.97, on a stockanalysis 2026-08-12 basis), far below their respective TTM PEs. The global market is not paying a valuation for peak earnings; it is pricing on "this E is not sustainable."

Revised conclusions (summarized in 2.6): ① memory branch strength is downgraded from A to B+; ② 香农芯创 Shannon Semiconductor (300475) is downgraded from "watch closely" to "watch only" — it is the "price-change-rate type" name with the highest PB percentile, and there has already been an empirical test on 2026-07-13: after disclosing on 7/10 a pre-announced 21–24x increase in net profit attributable to parent, it fell 20CM limit-down the next day, so the "good news fully priced" script has already been played out once; ③ 澜起科技 Montage Technology (688008) is the only name in the group whose PE and PB percentiles do not diverge (77.4% vs 78.8%), which means its PE can be compared historically — and the conclusion of that comparison is "expensive," not "cheap" — and its H1 ex-non-recurring profit is up only +14.5%~+32.9% (the company itself wrote "slight increase" in its pre-announcement). Its defensive case holds, but "defensive" does not mean "undervalued."

On 长鑫科技 CXMT (688825) — this piece does not put it on any recommendation list, and the reasons can now be made concrete: it listed on 2026-07-27 and has traded only 17 sessions through 8/12, with an issue price of RMB 8.66 and a current price of RMB 53.52 (+518%); free-float shares of 4.503 billion / total share capital of 66.881 billion = a free float of only 6.73%, with all 60.193 billion pre-IPO shares (90.0%) under lock-up. So for the reading "main-force net inflow of RMB 3.603 billion, first in the whole market," the denominator must be the free-float market cap of RMB 241.0 billion rather than the total market cap of RMB 3,579.5 billion: the former gives 1.50%, the latter 0.101%, a 14.8x difference between the two bases. Compare GigaDevice's free float of 95.26% and Biwin Storage's 100.00% — CXMT's 6.73% is an absolute outlier in this group, and its "first in capital inflow" mainly reflects scarcity of tradable shares rather than the strength with which capital is pricing fundamentals.

2.5 [The most important item in this piece] The capital choices of 8/12 are negatively correlated with whether AI revenue has actually landed in the financials

Grading the PCB/materials/optical chip group by "whether a verifiable AI-related revenue figure can be found in public disclosure," and then comparing against their actual 8/12 performance, produces a highly counterintuitive result:

How to read it: the two strongest movers on 8/12 (Kinwong at limit-up and Lianrui at +15.01%) happen to sit in the third tier, with the lowest degree of realization; while Yuanjie Semiconductor, with the highest degree of realization, had the only main-force participation an order of magnitude lower than the rest (0.7% vs 11%–14.5% for the others).

This does not mean the market is wrong — the value of a pre-market list lies precisely in spelling out this dislocation: if capital in a direction is flowing toward the positions with the best story and the emptiest financials, then its durability rests on sentiment rather than earnings, and the moment the interim report lands is the moment of falsification. And that moment is already on the calendar: Lianrui New Material's interim report is due 15 August (this Friday) and it issued no results pre-announcement; Kinwong Electronic's interim report is due 22 August, likewise with no pre-announcement, and its Q1 profit declined.

One more thing that must be said alongside this: the entire AI hardware chain's 8/12 rally is a rebound after a deep drawdown, not a new high. Checking each name's 2026 high one by one:

Nine of the ten are still 24%–45% below their June highs. This fact suppresses both extreme readings at once: ① "the main line has run to its final stage" — by price position it has not even recovered the prior high, so it is hardly a final stage; ② "this is a new main advance" — by time structure, the deep drawdown after the June top has yet to be repaired, so this is the second leg of a rebound. The only one inside a new-high structure is Kinwong Electronic (just 1.8% below its high, +49.8% over the past 20 sessions), and it happens to be the weakest on fundamentals (Q1 net profit attributable to parent −28.37%, with both PE and PB percentiles at 100.0%, i.e. the most expensive since 2018).

2.6 Final revisions to the strength of three branches (relative to this piece's draft)

---

Data-sourcing record: a cross-market timing mismatch that was nearly written in wrong

The search first hit Sina Finance's piece published 2026-08-13 00:01, "US semiconductor and memory chip stocks rise across the board," which said "the Dow rose 0.28%, the S&P 500 rose 0.46%, the Nasdaq rose 0.88%, the Philadelphia Semiconductor Index rose more than 3%, and Lumentum rose more than 7%."

That article was published at 00:01 Beijing time on 8/13 = 12:01 US Eastern on 8/12, a midday intraday snapshot, not the close. Checking the close against the CNBC quote interface (timestamps 2026-08-12 16:00–17:15 ET), the true close was: Dow −0.04% (not +0.28%), S&P +0.26%, Nasdaq +0.54%, SOX +2.49% (not "more than 3%"), Lumentum +13.63% (not "more than 7%").

Deviations existed in both directions at once: the indices were overstated (the Dow's +0.28% actually closed at −0.04%), while Lumentum was severely understated (+7% → +13.63%). Taking the midday snapshot directly would have produced both the wrong conclusion of "a broad overseas rally" and the wrong conclusion of "a moderate rise in optical communications," precisely inverting the main direction of this piece's judgment. Everything has been switched to the CNBC close basis.

---

3. Overall single-stock catalyst strength ranking (descending by stock)

Board price limits: Shanghai/Shenzhen main boards ±10% / ChiNext ±20% / STAR Market ±20% / Beijing Stock Exchange ±30%. All price and capital data are on an 8/12 close basis.

---

4. Single-stock scoring model

Note on the maximum score (which must be stated first, or the scores cannot be interpreted): the seven positive dimensions cap out at 15+20+15+15+10+10+10 = 95 points in total, with risk deductions of 0 to −15, so the actual range in this piece is −15 to 95, not 0 to 100. This piece does not rescale; please read it on a 95-point scale.

One re-scoring that must be made public: the draft gave 中际旭创 Innolight (300308) 12 points for "earnings elasticity," with the reason given as "the H1 interim report is not out, so it is inferred from industry prosperity" — this directly violates the table's own stated basis (that dimension specifies that "expectation only" can score just 0–5), and it contradicts section 5.1's own statement that "management commentary is not data and cannot count as earnings validation." After strict re-scoring on the stated basis, Innolight's total falls from 87 to 69, below Eoptolink's 77, so the two swap places in the ranking. That outcome is what the model should look like: under the same news flow, results that have been disclosed and are still accelerating on a single-quarter basis should rank ahead of results for which no figure has been disclosed at all.

Worked example — 新易盛 Eoptolink (300502), total 77: news source 14 (the Lumentum earnings release) + directness 16 (the same sub-sector supply side) + earnings elasticity 13 (H1 pre-announced net profit attributable to parent of RMB 7–8 billion already announced, with Q2 alone +78.1%~+120.3% YoY, no lower than Q1 — the only one of the big three not decelerating) + fundamentals 14 (No. 2 globally, optical interconnect at 99.72% of revenue) + expectation gap 6 + durability 9 (communication network equipment and components saw inflows on 2 consecutive days, amplified 2.67x) + trading characteristics 10 − risk 5 (all-time PB percentile 95.4%, −3; overseas revenue 96.16% with the largest tariff exposure, −2) = 77.

Worked example — 中际旭创 Innolight (300308), total 69: news source 14 + directness 16 + earnings elasticity 4 (H1 not disclosed, no pre-announcement, so on this table's basis "expectation only" can score just 0–5) + fundamentals 15 (about 50% global share in 800G silicon photonics, No. 1 globally per LightCounting) + expectation gap 6 (already rose on 8/12 with capital already in, partly priced) + durability 9 + trading characteristics 10 (1st in turnover market-wide with a turnover rate of only 2.80%, stable share ownership) − risk 5 (FCC tail risk −3, late diffusion −2) = 69.

Worked example — 光迅科技 Accelink (002281), total 50: news source 14 + directness 16 + earnings elasticity 8 + fundamentals 11 + expectation gap 2 + durability 4 + trading characteristics 2 (turnover of RMB 8.61 billion with main force of only +RMB 19 million) − risk 7 (it lagged while the sector rose broadly, showing capital is actively avoiding it) = 50. On the same news and with the same directness (both 14+16), trading characteristics and capital validation pull it 19 points below Innolight — this is precisely the quantified expression of "the concept is right but capital does not endorse it."

---

5. Detailed analysis of the Top 10 stocks

51 中际旭创 Innolight300308ChiNext ±20%

• Related news: post-close US Eastern on 2026-08-12, Lumentum reported FY26Q4 revenue of US$1.01 billion (+109% YoY, +23% QoQ) and adj EPS of US$3.23, with FY27Q1 guidance of revenue US$1.225–1.275 billion and a non-GAAP operating margin of 39.5%–40.5%, all far above consensus; the stock rose +13.63% (Benzinga¹). The core drivers are AI data-center optical interconnect and demand for 1.6T and CPO lasers.
• Catalyst logic: indirect but from the same source. Lumentum is a laser/optical chip supplier and Innolight is an optical module maker; both sit on the same AI optical interconnect demand curve. What is affected is revenue expectations, not sentiment — Lumentum's 40% operating margin guidance shows this chain is currently not in a price-war structure. But the share side must be deducted: part of Lumentum's gain embeds expectations of share transfer from FCC restrictions on Chinese optical modules, and that part cannot be extrapolated (see 2.1).
• Stage of the industry branch: fermentation / diffusion phase, leaning middle-to-late. On 8/12 five optical communications names hit limit-up (永鼎股份 Yongding, 通宇通讯 Tongyu Communication, 通鼎互联 Tongding Interconnection, 炬光科技 Focuslight, 长进光子 Changjin Photonics) while the leader rose only 3.84% — money is in the leader and elasticity is in the catch-up names, the classic shape of a diffusion phase. There is a core group (this stock + Eoptolink + T&S Communications) and there is a ladder, but no name has yet reached 3 or more consecutive limit-ups.
• Fundamental verification: the core business is high-speed optical modules, a perfect match with AI optical interconnect rather than a concept mapping — optical communication modules were 97.58% of 2025H1 revenue, with overseas revenue at 90.58%. Ranked No. 1 in LightCounting's 2025 global optical module vendor TOP10. 2026Q1 revenue of RMB 19.496 billion (+192.1%), net profit attributable to parent of RMB 5.735 billion (+262.3%), gross margin 46.06% (highest in the group), operating cash flow of RMB 3.368 billion (OCF/net profit 0.59), gearing ratio 32.64%.
• [Correction to the interim report date] H1 results come on 21 August, not the 24 August reported by media. The basis is the company's "Date of Board Meeting" announcement released on HKEX on 2026-08-12 (inside this piece's window), which reads: "The board meeting will be held on Friday, 21 August 2026 to… consider and approve the interim results of the Company and its subsidiaries for the six months ended 30 June 2026 and their publication, and to consider the recommendation of an interim dividend." Company announcements take precedence over media accounts. The company has also issued no results pre-announcement, and the recent earnings-call remark of being "very confident in the half-year operating situation and industry demand" is management commentary rather than data and cannot count as earnings validation.
• [One valuation reading that must be corrected] A PB of 28.3x at the 96.2% percentile is a stale figure and cannot be used to argue it is expensive. That PB uses book net assets as of 2026-03-31 and does not include at all the H-share proceeds received at the end of July: the company's 54.5 million H shares listed on the HKEX main board on 2026-07-30 (stock code 3308) at an issue price of HK$980, with net proceeds of about HK$52.891 billion, the largest Hong Kong IPO of 2026 to date. A rough pro-forma net asset figure is about RMB 91.8–93.8 billion, giving a pro-forma PB of about 11.5–11.7x, less than half the value at listing. It also corresponds to EPS dilution of about 4.7%.
• But the H-share matter also has a contrary side: it broke issue on its first trading day on 30 July, closing down 2.04%, while the A shares fell 9.15% the same day. Hong Kong pricing anchors and pressures the A shares, which is the flip side of the same coin as "the H-share raise strengthens the balance sheet."
• Industry position: global leader.
• Technicals and sentiment: ChiNext ±20%. On 8/12 it closed at RMB 921.00, +3.84%, with turnover of RMB 28.41 billion (1st in the whole market), main-force net inflow of +RMB 2.596 billion, and a turnover rate of only 2.80%. Being first in turnover market-wide with a turnover rate of only 2.80% means this is normal turnover for a large cap, with a stable share-ownership structure rather than a hot-money battleground. No limit-up, no one-word board, no one-day-wonder characteristics.
• Final call: priority deep-dive. It is the only name in this chain with all three supports at once — global leadership + two consecutive days of large capital inflow + overnight supply-chain earnings validation. But the key today is not whether it rises, it is how much it rises — if we again see "the leader +3% while second-tier names hit limit-up," that is a divergence signal marking the end of the diffusion phase.

52 新易盛 Eoptolink300502ChiNext ±20%

• Related news: same as 5.1. In addition, in mid-to-late July 2026 the company pre-announced 2026H1 net profit of RMB 7–8 billion, +77.56%–102.93% YoY, with Q2 net profit +51.80%–87.77% sequentially (THS²³). That pre-announcement is old July news and not today's driver, but it is the only one of the big three that is "also accelerating on a sequential basis" — a sequential basis speaks to the current state of the cycle better than a year-on-year basis.
• Catalyst logic: direct, same as 5.1.
• Stage of the industry branch: same as 5.1.
• Fundamental verification: 800G LPO modules cut power consumption by 50%, and public reports say it has won orders from top customers such as Meta and Amazon. This piece could not independently confirm "winning orders" at the level of company announcements; it appears only in industry-media accounts, and on an evidence grading basis it should be recorded as "media account, not confirmed by announcement." The interim report is expected on 25 August.
• Industry position: No. 2 in LightCounting's 2025 global optical module TOP10, overtaking Coherent for the first time. Optical interconnect products were 99.72% of 2025 revenue, with overseas revenue at 96.16% (the highest of the big three) — which is both a purity advantage and the largest tariff/export-control exposure.
• [Valuation: one misuse that must be called out] PE (TTM) 55.6x, at an all-time percentile of just 49.8% (near the median), while PB is 29.1x at an all-time percentile of 95.4%. The mathematical meaning of that combination is that ROE is at a historical extreme — full-year 2025 ROE reached 72.75%. The reason PE looks "not expensive" is that the denominator E is propped up to a historic high by peak ROE; PB is the real constraint here. So the argument that "Eoptolink is only on 55x PE, half as expensive as T&S (120x)" is invalid. This does not mean the cycle has topped (Q2 is still accelerating), but the valuation position must be read via PB.
• Cash flow is the soft spot: 2026Q1 operating cash flow / net profit attributable to parent was only 0.25 (0.81 for full-year 2025). Two consecutive years of very weak Q1 cash conversion, so the cash content of profit is weaker than at Innolight and T&S Communications.
• Technicals and sentiment: on 8/12 it closed at RMB 428.20, +3.11%, with turnover of RMB 18.75 billion, main force +RMB 1.966 billion, and a turnover rate of 3.49%. The shape matches Innolight's — a mild gain with large capital inflow. It is −29.9% from the year's high (610.49, 06-25).
• Final call: priority deep-dive. The difference from Innolight is that it has an announced results pre-announcement that is accelerating sequentially as a floor; the risk is that the RMB 7–8 billion range has been thoroughly discussed, so an interim report at the low end would be a negative.

53 生益科技 Shengyi Technology600183Shanghai main board ±10%

• Related news: no company-level news. The branch news is the 8/12 printed circuit board sector main-force net inflow of RMB 4.318 billion and components +RMB 2.784 billion, plus overall strength in the overnight AI server chain (Dell +9.87%, Super Micro +19.02%).
• Catalyst logic: one layer indirect. AI server shipment growth → demand for high-multilayer/high-frequency high-speed CCL → this stock. What is affected is revenue, not sentiment.
• Stage of the industry branch: fermenting; it is the second echelon of the main line.
• Fundamental verification: domestic CCL leader. This piece did not obtain reliable data on the specific share of revenue from high-frequency high-speed CCL for AI servers, and makes no inference — this is the key gap in assessing the stock's "directness of benefit," see section 10.2.
• Industry position: domestic leader.
• [AI revenue has already landed at the profit level, and this is its hardest point] The H1'26 pre-announcement (2026-07-14) put net profit attributable to parent at RMB 3.099–3.298 billion, +117%~+131% YoY, with ex-non-recurring profit of RMB 2.719–2.918 billion. The company explicitly attributes it: in CCL, "demand for high-end products continues to grow rapidly… driving timely release of expanded capacity"; in circuit boards, "subsidiary Shengyi Electronics benefits from the high-prosperity demand of downstream fields such as AI compute and high-speed communications… the high-multilayer HDI board project for intelligent compute centers is landing efficiently on schedule and rapidly releasing benefits." The scale of profit (RMB 3 billion+) and the attribution language are both hard; but the specific amount and share of AI-related revenue were not separately disclosed in the annual report — not obtained.
• [A structural fact: 16% of the market cap is a mapping of a subsidiary] Shengyi Technology holds 59.85% of 生益电子 Shengyi Electronics (688183). Shengyi Electronics had a market cap of RMB 93.726 billion on 8/12, so its attributable equity value is about RMB 56.1 billion, roughly 16% of Shengyi Technology's RMB 347.9 billion market cap. In other words, the purest AI server board exposure in this chain has already been separately priced at the subsidiary level. A further 3.08% of shares sits in a trust account securing exchangeable bonds, which constitutes a potential future channel for share reduction.
• [Three risks that must be listed side by side, none of which the draft wrote]
  • Valuation: PE (TTM) 88.56x, PB 21.77x, both at all-time percentiles of 98.7%.
  • The company has already issued its own risk notice: because the three-day gain over 8/4–8/6 deviated by more than 20%, the company published a "Share Trading Risk Notice" on 8/8, acknowledging that "the 8/7 close of RMB 139.98 was about 37% above RMB 102.03 on 8/3" and that "the company's latest price-to-book ratio of 17.24 is above the industry (CSI printed circuit board) figure of 9.85." By 8/12 PB had risen to 21.77, widening the deviation further.
  • The chairman's share-reduction plan is inside its execution window: chairman Chen Renxi disclosed a reduction plan on 2026-05-30 (no more than 335,800 shares, 25% of the holding), which as of 8/7 had not been executed, with the window running to roughly mid-to-late September per the announcement.
• [Timing] The interim report is due 15 August (this Friday). Because the pre-announcement already gives a +117%~+131% range, the room to beat is capped by the top of that range, making this a validation that matures within two days.
• Technicals and sentiment: on 8/12 it closed at RMB 143.23, +5.86%, with turnover of RMB 10.74 billion, main force +RMB 1.213 billion, a turnover rate of only 3.17%, and main-force net / turnover of 11.3%. Low turnover rate plus large net inflow, the same shape as the optical module leaders. It is −25.4% from the year's high (191.88).
• Final call: watch closely (downgraded one notch from the draft's "priority deep-dive"). The reason for the downgrade is not deteriorating fundamentals — its H1 profit is among the most fully realized in the group — but that a 98.7% valuation percentile + the company's own 8/8 risk notice on a price-to-book above the industry + an unfinished chairman share-reduction window + an interim report landing 8/15 with the pre-announcement range already capping the upside, taken together, mean "priority deep-dive" would assert more certainty than the evidence supports.

54 天孚通信 T&S Communications300394ChiNext ±20%

• Related news: same as 5.1. It has pre-announced 2026H1 net profit of RMB 1.124–1.304 billion, +25%–45% YoY, with Q2 +28.45%–60.98% sequentially.
• Catalyst logic: direct. The company describes itself as "an optical components maker, with optical module makers as our downstream customers" (Sina Finance 2026-08-05²²) — this positioning makes it closer to Lumentum's business nature than the optical module makers are, and one of the most direct comparables in this round.
• Stage of the industry branch: same as 5.1.
• New announcements inside the window (two, pointing in opposite directions, which must be listed side by side):
  • [Positive · S] The "2026 Restricted Share Incentive Plan (Draft)" released 2026-08-12 19:48: granting 2.2887 million shares at RMB 120 per share (just half the 8/12 close of RMB 240.05), covering 745 people. Company-level performance targets use 2025 ex-non-recurring net profit as the base: net profit growth of 150% in 2027 (trigger value 120%), 400% in 2028 (320%), and 700% in 2029 (560%). This is management expressing confidence through three-year targets, and it is a falsifiable hard commitment rather than a qualitative statement on an earnings call — in scale terms, +700% by 2029 means ex-non-recurring net profit must reach 8x the 2025 level.
  • [Note · B] An "Announcement on the Proposed Change of Accounting Firm" was released at the same time. This piece did not obtain the reason for the change; it is a governance item that requires reading the original text, and no inference is made.
• Fundamental verification (on a single-quarter basis, which is the key): H1 pre-announced net profit attributable to parent of RMB 1.124–1.304 billion, +25%–45% YoY; but backing out from Q1 net profit of RMB 492 million, Q2 alone is only +12.5%~+44.5% YoY, whereas Q1 was +45.8% — only the top of the range barely matches Q1, and the midpoint decelerates markedly. The company gives three reasons in the pre-announcement: "a shortage of certain materials had some impact on production ramp for related products", higher finance costs from FX losses, and the tax shield from equity incentive vesting (positive). The material bottleneck has been named by IR: minutes from the Thailand plant dated 2026-07-09 state that "affected by tight supply of 200G EML chips in the first half of 2026, the active product line is currently running only small-volume production," with month-by-month improvement expected in the second half.
• Its CPO linkage is the most concrete of the big three: IR states that "the company supplies FAU and ELS products for CPO products" and that "the FAU products supplied for CPO have completed design finalization and are in the order-delivery stage"; on NPO it is working with multiple customers on development. If today's main line is CPO rather than 1.6T optical modules, T&S is the name with the most concrete linkage.
• Industry position: an optical components platform (2025H1 active optical 63.8% + passive optical 35.1%, 98.9% combined). Note that it is not an optical module maker, so it is not on LightCounting's global optical module TOP10 — that is a different track, not a ranking decline; do not misread it.
• Valuation: PE (TTM) 120.6x, PB 43.7x, both at all-time percentiles of 96.6%. Gearing ratio 14.25% and goodwill of RMB 30 million — the cleanest balance sheet of the big three.
• Technicals and sentiment: on 8/12 it closed at RMB 240.05, +7.89%, with turnover of RMB 12.72 billion, main force +RMB 1.221 billion, and a turnover rate of 4.91%. Its gain was the largest of the big three (Innolight +3.84%, Eoptolink +3.11%). It is −32.3% from the year's high (354.67, 06-03).
• Final call: watch closely, not a priority deep-dive. It carries a genuine set of contradictions, all three of which must be laid out: the business positioning that best fits the overnight catalyst (CPO FAU already shipping) + the most aggressive equity incentive targets (+700% by 2029) + but Q2 alone is the only clear deceleration of the three, with PE/PB percentiles both at 96.6%. It rose the most on 8/12, showing the market has already paid a premium for "best fit"; whether that is delivered depends on when the 200G EML shortage clears, and on that point there is currently only the company's own account, unverified.

55 兆易创新 GigaDevice603986Shanghai main board ±10%

• Related news: overnight Micron +4.92%, SanDisk +5.76%, Seagate +7.03%, Western Digital +3.69%; Korean semiconductor exports for 8/1–8/10 were US$10 billion, +155.4% YoY, an all-time high, with semiconductors rising from 26.7% to 46.8% of exports (Korea Herald⁹).
• Catalyst logic: direct. As an in-house memory producer, rising DRAM/NAND contract prices flow straight into gross margin, unlike a distributor that only earns a spread.
• Stage of the industry branch: this is the item requiring the most caution in this piece. The prosperity data (Korean exports +155.4%) is extremely hard, but the slope of price increases is slowing: TrendForce expects 2026Q3 conventional DRAM contract prices +13%–18% QoQ and NAND +10%–15%, a marked slowdown from Q2 (Sina Tech 2026-08-08¹³). The absolute level of the boom is high while the marginal change is decelerating — both are true at once.
• Fundamental verification: 2026Q1 revenue of RMB 4.188 billion (+119.4%), net profit attributable to parent of RMB 1.461 billion (+522.8%), gross margin 57.08% (37.44% a year earlier), operating cash flow of +RMB 1.783 billion (OCF/net profit 122%, good conversion of profit into cash), a gearing ratio of just 8.13% and inventory/net assets of 0.13x — the cleanest balance sheet in the group. Memory chips were 71.34% of 2025 revenue and microcontrollers 20.75%. Its own-brand DRAM is foundried by CXMT, with products being niche DDR3L/DDR4; it has no HBM business.
• [The most unexpected finding in this piece] GigaDevice holds about 1.80% of the equity of CXMT (688825). On CXMT's total market cap of RMB 3,579.5 billion as of 8/12, that stake is worth about RMB 64.4 billion, or 22.2% of GigaDevice's own total market cap of RMB 289.9 billion. Yet GigaDevice's total assets as of 2026-03-31 were only RMB 27.77 billion, so the stake is clearly still carried far below market value, with a huge unrecognized gain. Excluding that portion, the core business is worth about RMB 225.5–232.0 billion. There is an imminent trigger for this: GigaDevice's interim report is due 19 August — if the stake is measured at FVTPL, CXMT's +518% since its 7/27 listing will produce a one-off non-cash gain on the order of tens of billions of yuan in Q3, thoroughly distorting headline net profit; if it is measured at FVOCI, it goes only to other comprehensive income and not to profit or loss. The two accounting treatments point in completely opposite directions for interpreting "GigaDevice's Q3 results."
• A blemish in profit quality (which must be written): H1 pre-announced net profit attributable to parent of about RMB 6.90 billion and ex-non-recurring profit of about RMB 4.85 billion → non-recurring items of RMB 2.05 billion, 29.7% of net profit attributable to parent, which the company attributes to "an increase in the period-end fair value of securities investments held." Backing out Q2 alone, non-recurring items were 36.8% of net profit attributable to parent (only RMB 51 million in Q1). Looking at GigaDevice's results requires looking at the ex-non-recurring figure.
• Valuation: PE (TTM) 100.86x, PB 11.46x. A 3-year PE percentile of 24.5% vs a PB percentile of 90.9% — a moderately strong divergence, showing this round of pricing has departed from its own recent range. Q2 annualized PE is 13.3x.
• Industry position: domestic memory design leader, top three globally in NOR Flash. It has completed an A+H structure (Hong Kong listing on 2026-01-13).
• Technicals and sentiment: on 8/12 it closed at RMB 413.17, +2.70%, with turnover of RMB 18.75 billion, main force +RMB 881 million, and a turnover rate of 6.82%.
• Final call: watch closely. Its position in the memory chain is clear — "price-level type," with Q3 average prices still above Q2, so sequential earnings growth most likely continues, unlike module makers that depend on the rate of change of prices. What really needs watching is not price increases but the accounting classification of the CXMT stake in the 8/19 interim report, whose impact on headline profit far exceeds that of the core business itself.

56 澜起科技 Montage Technology688008STAR Market ±20%

• Related news: same as 5.5.
• Catalyst logic: direct. Memory interface chips ramp with DDR5 penetration and rising AI server memory capacity, which is an independent volume logic separate from "memory price increases" — it earns per interface chip unit, not on die prices. This makes it less sensitive to "slowing price increases" than GigaDevice and the memory module makers.
• Stage of the industry branch: fermenting.
• Fundamental verification: one of the global duopoly in memory interface chips. Detailed financials are in section 10.2.
• Industry position: one of a global duopoly, with high barriers.
• Technicals and sentiment: on 8/12 it closed at RMB 215.43, +4.22%, with turnover of RMB 9.44 billion, main force +RMB 515 million, and a turnover rate of 3.86%.
• Final call: watch closely. Of the memory names listed in this piece it is the one whose logic depends least on price increases, making it the most defensive in a "price increases have topped" risk scenario. The risk is that its valuation is already high.

57 景旺电子 Kinwong Electronic603228Shanghai main board ±10%

[This section's conclusion has been downgraded from the draft's "watch closely · 75 points" to "watch only · 58 points."] The downgrade is not because it hit limit-up on 8/12, but because the results of fundamental verification contradict the draft's assumptions. The original text is retained for comparison.

• Related news: no company-level news. On 8/11 it published an abnormal-movement announcement because the cumulative three-day deviation over 8/6, 8/7 and 8/10 exceeded 20%, confirming on self-examination that "production and operations are normal," that "there is no material information that should have been disclosed but was not," and that "directors, supervisors, senior management and the controlling shareholder did not trade during the abnormal-movement period." The only substantive matter disclosed in that announcement is that the company is advancing an offshore H-share issue and listing on the HKEX main board, has obtained CSRC filing, and has re-submitted its application to the HKEX, with implementation still uncertain.
• Catalyst logic: one layer indirect. But the "AI server PCB" logic finds no revenue support in its public disclosure: the 2025 annual report contains no breakdown of AI/server board revenue whatsoever, and the relevant passages are almost entirely citations of Prismark industry forecasts (a 2024–2029 data infrastructure PCB CAGR of 16.4%, etc.) plus product capability descriptions ("AI accelerator cards, general servers, AI servers, 400G and 800G"), with no amount or share of the company's own AI-related revenue.
• Fundamental verification (this is the core of the downgrade):
  • 2026Q1 net profit attributable to parent of RMB 233 million, −28.37% YoY, the only one of the six in the group with declining profit; yet revenue was +16.41%, i.e. revenue growth without profit growth, with a Q1 gross margin of only 18.76%.
  • The core business is printed circuit boards at RMB 14.373 billion (93.89% of the total), within which the domestic-sales gross margin is only 9.69% versus 28.01% for exports.
  • There is still no H1 results pre-announcement; the interim report is due 22 August.
• Valuation: PE (TTM) 86.34x, PB 7.39x, with both all-time and 3-year percentiles at 100.0% — i.e. the most expensive position since 2018.
• Price position: only −1.8% from the 2026 high, +49.8% over the past 20 sessions. Among the ten AI hardware names in the group, it is the only one inside a new-high structure; the other nine remain in deep drawdowns of 24%–45% from their June highs (see 2.5).
• Technicals and sentiment: Shanghai main board ±10%; on 8/12 it hit limit-up +10.00% at RMB 98.21, with sealing orders of RMB 445 million (2nd largest in the whole market), turnover of RMB 6.14 billion, main force +RMB 716 million, a turnover rate of 6.77%, and 0 breaks of the limit. The seal quality is indeed high — the draft did not get that wrong, but it describes trading intensity, not fundamentals.
• Final call: watch only. Laying the three items side by side makes the conclusion clear: the valuation is at its most expensive since 2018 (PE/PB percentiles both 100%), profit is the only decline in the group (−28.37%), AI revenue cannot be found in the financials, and there is no results pre-announcement as a floor — while its interim report lands on 22 August. This is the most typical sample of what section 2.5 calls "fund flows negatively correlated with the realization of AI revenue." The 8/12 limit-up is real and the fundamental void is also real; the two do not conflict, and what a pre-market list should say is the latter.

58 香农芯创 Shannon Semiconductor300475ChiNext ±20%

[This section's conclusion has been downgraded from the draft's "watch closely · 73 points" to "watch only · 56 points."] The draft treated it as "the expectation gap in the memory chain"; after verification the judgment reverses, for the reasons below.

• Related news: same as 5.5. But the key point is: Micron, SanDisk, Western Digital and Seagate, which rose overnight, are IDMs and media makers, while Shannon Semiconductor is a distributor — they are not peers (see 2.4).
• Catalyst logic: indirect, and the most fragile of the three mechanisms. It earns the timing spread of "buy low, sell high," and its profit function tracks the rate of change of price Δ, not price itself.
• Fundamental verification:
  • The core business in 2025 was electronic component distribution at RMB 33.223 billion (94.25% of the total), with a gross margin of only 2.86%; the own brand "海普存储 Haipu Storage" was RMB 1.658 billion (4.70%). Overseas revenue was 85.30%. It holds SK Hynix and MediaTek agency rights and AMD distributor qualification.
  • Elasticity mechanism: gross margin went 1.98% (25Q1) → 3.97% (25FY) → 9.12% (26Q1). Going from 2% to 9% looks mild, but for a business with a 1% net margin it is a dozens-fold amplification of net profit (that is how the +7835% YoY figure arises). That amplification works just as well in reverse.
  • Risk symmetry: prices need not fall; they only need to stop rising for the excess gross margin to go straight to zero. And Q3 contract price increases have already fallen from Q2's +58%+63% to +13%+18% (see 2.4).
  • At the same time it levered up heavily to stockpile in 2026Q1: inventory of RMB 2.60 billion (end-2025) → RMB 6.97 billion (26Q1), inventory/net assets 0.73x → 1.43x, gearing 65.58% → 75.40%, the highest in the group.
• Valuation profile: a PE (TTM) percentile of 36.3% against a PB percentile of 92.7% — near an all-time high. Q2 annualized PE is only 7.9x. This is precisely the "low PE percentile + high PB percentile" combination, a direct reading that E is at a cycle top, not that it is cheap.
• [There has already been one complete empirical test, and this is the most persuasive item]: on 2026-07-13, after disclosing on 7/10 a pre-announced 21–24x increase in net profit attributable to parent, the stock fell 20CM limit-down. The "good news fully priced" script has already been played out in full on this name, and it happened just a month ago.
• Technicals and sentiment: on 8/12 it closed at RMB 163.80, +3.67%, with turnover of RMB 5.81 billion, main force +RMB 807 million, a capital share of 13.89% (the highest in the memory chain), and a turnover rate of 7.99%.
• Final call: watch only. The draft read "8/12 capital inflow + the overnight surge in US memory" as an expectation gap, and that was wrong: mismatched comparables (its counterpart is not Micron), the most fragile mechanism (it earns Δ, not price), a valuation profile with cycle-top characteristics (PB percentile 92.7%), and the highest inventory leverage (1.43x net assets, gearing 75.40%) — plus it fell limit-down just a month ago precisely because results far exceeded expectations. The inflow that day is a fact, but it is the weakest piece of evidence on this name and is not enough to outweigh the four items above.

59 太辰光 Taichenguang300570a downgraded inference, filed together with the reasons for the downgrade · ChiNext ±20%

• Related news: Super Micro (SMCI) FY26Q4 revenue was US$11.2 billion, below the market's expected US$11.73 billion, and the company explicitly attributed the gap to delays in customer projects across three links — power, cooling and networking; despite the revenue miss, gross margin doubled from 9.5% to 17.5%, EPS beat, and the stock rose +19.02% (Yahoo Finance²⁴).

• Catalyst logic: how to read this news is one of this piece's core judgments. "Revenue below expectations" is usually a negative, but when the cause of the gap is "customers cannot build fast enough" rather than "customers no longer want it," it is a positive for suppliers at the bottleneck — Super Micro itself named power, cooling and networking as the bottlenecks and the stock then rose +19.02%, showing the market read it that way too. Copper high-speed connectivity, where Taichenguang sits, corresponds exactly to the "networking" link; that concept saw main-force net inflow of RMB 2.361 billion on 8/12 with a main-force share of 5.98%, and this stock led the concept.

• Stage of the industry branch: a lagging start. For comparison: on 8/12 the optical module concept saw main-force +RMB 13.696 billion and CPO +RMB 12.457 billion, while copper high-speed connectivity was only +RMB 2.361 billion and the liquid cooling concept only +RMB 1.172 billion (a main-force share of 0.73%, almost no capital). On the same AI hardware chain, optical interconnect is already fully priced while power/cooling/connectivity are not.

• Fundamental verification: a niche optical connectivity maker. This piece did not obtain reliable data on how far its "AI-related revenue" is realized in the financials, which is this name's biggest uncertainty — it may be a genuine beneficiary or merely a concept mapping. Per the ironclad rules, no grade above B+ is given before verification.

• Industry position: niche, not a leader.

• Technicals and sentiment: on 8/12 it closed at RMB 149.76, +6.63%, with turnover of RMB 3.12 billion, main force +RMB 134 million, and a turnover rate of 10.96%. The turnover rate is on the high side and the inflow amount is not large.

• Final call: watch closely (a watch slot; withdrawn from the recommendation list). The draft of this piece listed the "bottleneck beneficiary" logic as the biggest expectation gap and put this branch's names into the 5-stock recommendation list. The final version retracts that, for three reasons, each harder than the draft's logic:

  1. The timing does not hold: Super Micro's earnings were released at approx. 04:20 Beijing time on 8/12, so A-shares could price them all day on 8/12. So "the sector has not reacted" is not a lag; it is that capital looked and did not buy.
  2. The on-point US name did not rise: Vertiv (VRT), the leading US name for data-center power and cooling, closed that night at only +2.32%, far below Dell's +9.87% and Arista's +6.39%. The premise that "the market reads the bottleneck as a positive for suppliers" did not even hold in the US market itself.
  3. The mapping chain is too long: the "networking" Super Micro refers to means customer-side data-center network delivery, which is at least two layers removed from domestic MPO/fiber connector supply.

  The only reason for keeping it as an object of observation is the capital profile (a 5.98% main-force share in the copper high-speed connectivity concept, with this stock leading it), which is trading-level evidence and cannot be elevated into an industry conclusion. 英维克 Envicool (002837, Shenzhen main board ±10%) in the same branch has a better profile (main-force share 10.57%, turnover rate only 3.50%), but is likewise listed only as a watch slot for lack of AI/liquid cooling revenue share data — the criterion is the self-imposed constraint at the end of section 10.2: any name whose AI revenue share has not been obtained does not enter the recommendation list.

510 光迅科技 Accelink002281a counter-example that must be written into the Top 10 · Shenzhen main board ±10%

• Related news: exactly the same as 5.1. It is a front-rank optical module maker with no difference in news coverage from Innolight.
• Catalyst logic: direct, on the same level as the leader.
• Stage of the industry branch: same branch.
• Technicals and sentiment: on 8/12 it closed at RMB 185.03, +0.52%, with turnover of RMB 8.61 billion, main-force net of only +RMB 19 million, and a turnover rate of 5.87%.
• Why it deserves a Top 10 slot: on 8/12 five optical communications names hit limit-up and the leader took in RMB 2.596 billion of net inflow, yet this stock turned over RMB 8.61 billion with almost no net inflow and a gain of +0.52%, close to flat. Same news, same branch, same front-rank status, and capital gave a completely different answer. This shows the market is not speculating on the optical communications concept indiscriminately but is differentiating — a plus for the whole main line (a differentiated move is healthier than an undifferentiated rally) and a minus for this stock.
• Final call: watch only. And it is one of today's most valuable validation signals: if optical communications continues to strengthen today while Accelink still does not rise, capital is still differentiating and the main line is of high quality; if even Accelink hits limit-up today, that is an undifferentiated catch-up rally, which usually appears in the final leg of a main line.

---

6. Pass list

---

7. Intra-branch ranking

7.1 Optical comms / CPO / optical chips

• Hardest single stock: 中际旭创 Innolight (300308) — best on all three of position, capital and share ownership.
• Highest trading distinctiveness: 中际旭创 Innolight (300308), turnover of RMB 28.41 billion topping the whole market with a turnover rate of only 2.80%.
• Who is chasing: 永鼎股份 Yongding (600105), 通宇通讯 Tongyu Communication (002792), 通鼎互联 Tongding Interconnection (002491) — limit-ups but weak fundamental linkage to AI optical interconnect.
• Who to Pass: 光迅科技 Accelink (002281, capital does not endorse it), 源杰科技 Yuanjie Semiconductor (688498, bad position), 通宇通讯 Tongyu Communication (002792, excessive turnover).

7.2 Memory

• Hardest single stock: 澜起科技 Montage Technology (688008). The reason is that its growth comes from the number of interface chips rather than memory prices, so under the "slowing price increases" risk already confirmed by industry data, it is the only one not directly hit.
• The core contradiction of this branch: the capital directions of ranks 1–3 and 5–8 are opposite. The in-house/interface/distribution side all saw net inflows on 8/12, while the module side all saw net outflows. "Memory" is not a branch that can be bought or sold as a whole; it must be split.

7.3 PCB / CCL / materials

• Hardest + most distinctive: 生益科技 Shengyi Technology (600183) — the one in this branch with the fullest combination of "gain not yet overextended + large net inflow + low turnover rate."
• Who to Pass: 联瑞新材 Lianrui New Material (688300) and 菲利华 Feilihua (300395); the logic holds but they already rose 15.01% and 8.02% respectively on 8/12, and in a pre-market list, a stock that has already run its first leg is not a good starting point for observation.

---

8. Validation signals at today's open

8.1 Auction signals

• The size of the opening gap in the optical communications leaders is today's first thing to watch. 中际旭创 Innolight (300308) and 新易盛 Eoptolink (300502) are on ChiNext at ±20%; a 3%–5% gap up would be reasonable pricing of the overnight positive; a gap up of more than 8% would mean the overnight positive is cashed out all at once in the auction, and an open-high-close-low session is likely.
• 景旺电子 Kinwong Electronic (603228, Shanghai main board ±10%) had sealing orders of RMB 445 million and 0 breaks of the limit on 8/12: watch whether it opens one-word or sharply higher, and whether the sealing orders hold.
• 永鼎股份 Yongding (600105) had 7 breaks of the limit on 8/12: if it opens lower in the auction, or gaps up clearly less than the sector, that is the first signal that the optical communications ladder is breaking.

8.2 Sector signals

• For optical communications to continue, the leaders themselves must accelerate, rather than second-tier names continuing to hit limit-up. 8/12 was already the "leader +3%, second tier 20cm" diffusion shape; a repeat of that shape today is judged a divergence, not a continuation.
• 3 or more fast limit-ups appearing in optical communications or PCB counts as main-line confirmation; if limit-ups concentrate in the catch-up names that already hit limit-up yesterday, it counts as a stock-of-money game.
• The memory chain is today's most critical sector signal: US memory surged overnight while all A-share memory modules saw outflows on 8/12. Watch whether 江波龙 Longsys (301308), 佰维存储 Biwin Storage (688525) and 德明利 Demingli (001309) can turn positive today accompanied by main-force net inflows — this is the direct test of whether "overseas drivers" can override "domestic capital's judgment."
• Today is a decision day for semiconductors: 8/12's +RMB 5.450 billion is the 1st sample, and per the rule established yesterday, another net inflow today would constitute 2 consecutive days in the same direction and permit formal ranking; an outflow would make 8/12 a single-day pulse.

8.3 Single-stock signals

• 中际旭创 Innolight (300308): whether it can post a 3rd consecutive day of net inflow and stop lagging the second tier in its own gain.
• 生益科技 Shengyi Technology (600183): whether it can post a 2nd consecutive day of net inflow and turn PCB/CCL from a "follower" into an "independent branch."
• 光迅科技 Accelink (002281) is a reverse probe: if it still does not rise today, capital is still differentiating and the main line is healthy; if it hits limit-up, we have entered an undifferentiated catch-up rally and the main line is near its end.
• 百花医药 Baihua Pharmaceutical (600721): whether it can reach 8 boards. The company has issued a risk warning, the main force has already flowed out, and there were 11 breaks of the limit; if it can still seal the limit under those three pressures, hot money sentiment is extremely strong; a broken board would mean the pharma ladder and the market-wide consecutive-limit-up height step down at the same time.

8.4 Risk signals

• Open high, close low is the shape most in need of guarding against today. Overnight was a narrow move of "indices barely moving + a few AI hardware names surging" (Dow −0.04%), and this structure easily plays out in A-shares as a gap up followed by a volumeless fade.
• Broadcom (AVGO) closed flat overnight at −0.01% — if A-share AI chip/ASIC names are clearly weaker than optical modules today, that is normal differentiation consistent with the overnight structure and need not be read as falsification of the main line.
• The performance of 中鼎股份 Zhongding Sealing Parts (000887) and 巨轮智能 Greatoo (002031) today is a test of "whether clarification announcements work": both rose on 8/12 and were then contradicted by the companies themselves. If they remain strong today, the market is in a phase of euphoria that ignores fundamentals, and the position-sizing implication of the entire list should be reduced.
• Yesterday's recap already recorded: all 5 "previous-day consecutive limit-up / limit-up" type indices listed in yesterday's recap showed large gains + main-force net outflows on 8/12 (thematic stocks −RMB 1.868 billion, previous-day limit-ups including one-word boards −RMB 1.709 billion), i.e. the rise in consecutive limit-up names is driven by a relay of existing hot money. If that structure continues today, avoid orphan names with no main-line affiliation.
• Applied Materials (AMAT) reports FY26Q3 tonight (post-close US Eastern 8/13), which is event risk for the semiconductor equipment direction. Intraday today is "pre-event," so it is unwise to bet heavily on direction there.

---

9. Final recommendations

9.1 The 5 stocks most worth watching today

9.2 Today's 3 strongest branches

9.3 Directions not to chase today

1. The four memory module makers — 江波龙 Longsys (301308), 佰维存储 Biwin Storage (688525), 德明利 Demingli (001309), 深科技 Shenzhen Kaifa (000021). US memory surged overnight, but all four saw main-force net outflows on 8/12. Do not participate until the conflict between the "overseas positive" and "domestic capital" is resolved; they are objects of observation today, not objects of purchase.
2. Limit-up names contradicted by their own companies — 中鼎股份 Zhongding Sealing Parts (000887, clarified that the Tencent Cloud cooperation has no substantive content), 巨轮智能 Greatoo (002031, clarified that RV reducers are only 1% of revenue). When the reason for a limit-up is overturned in writing by the issuer, that is the hardest basis for a Pass.
3. High-level consecutive limit-up names as a whole — 百花医药 Baihua Pharmaceutical (600721, 7 boards + risk warning + main-force net outflow + 11 breaks of the limit), 开开实业 Kaikai Industry (600272, 5 boards + risk warning + turnover rate 39.43%). Yesterday's recap already confirmed that all consecutive limit-up indices show the "large gain + main-force net outflow" structure of an existing-money relay.
4. Humanoid robots — the highest news grade (global H1 shipments +272%, China at 97%) with the least capital endorsement (concept main force of only +RMB 647 million, 159th among 300 concepts), and the leader disclosed that night that revenue share is 1%. "Big news + no money + self-denial by the leader" all hold at once.
5. Passive components / MLCC — the sector was −RMB 1.534 billion on 8/12 and 洁美科技 Jiemei Technology (002859) broke its run of limit-ups; the one-day wonder is confirmed, and no second attempt should be made.
6. Materials names that already ran their first leg on 8/12 — 联瑞新材 Lianrui New Material (688300, +15.01%), 菲利华 Feilihua (300395, +8.02%). The logic holds but the position is already high.
7. Pharma / CRO — pharma and biotech saw main-force net outflows of RMB 3.440 billion on 8/12, first among industries; the ebb is under way.

9.4 The final one-sentence call

A higher open is likely today, but this is a "divergence day" rather than an "acceleration day": the overnight driver was extremely strong yet extremely narrow (Dow −0.04%, Broadcom flat, while Lumentum +13.63% and CoreWeave and Super Micro both +19%), and A-share optical communications has already reached the late-diffusion stage of "leader lagging, second tier at 20cm" — so what is really worth betting on today is not "whether the main line is still there," but "whether money will move from optical modules that have already run their course toward the power/cooling/high-speed connectivity that Super Micro named by name yet no one wants, and toward the memory that the overnight US session confirmed yet domestic capital abandoned."

---

10. Data sources, bases, and unverified items

10.1 Sources and bases of the main data in this piece

10.2 Known data gaps (read the corresponding conclusions as low-confidence)

The following items could not be obtained by this piece; they are explicitly flagged at the relevant places in the text and have not been filled in by inference:

1. The AI/liquid cooling and AI connectivity revenue shares of 英维克 Envicool (002837) and 太辰光 Taichenguang (300570) on a latest-reporting-period basis. This is the thinnest evidence link in the "bottleneck beneficiary" branch, and it is the direct reason both are excluded from the 9.1 recommendation list and listed only as watch slots.
2. The share of revenue from AI server boards/packaging substrates at 生益科技 Shengyi Technology (600183), 景旺电子 Kinwong Electronic (603228) and 兴森科技 Fastprint (002436) — none of the three annual reports breaks this out.
3. Global or domestic market share figures for the six AI materials/PCB companies and the five optical communications companies — not one discloses a verifiable share in its annual report.
4. The 800G/1.6T revenue split by speed for the optical module companies — none of the five provides it in periodic reports; all current statements about 1.6T come from qualitative management remarks in investor communications, not audited data.
5. Customer concentration for Innolight and Eoptolink consists of media accounts of annual report figures, not checked against the originals; customer concentration for T&S Communications and Advanced Fiber Resources was not obtained.
6. The latest review milestone for Advanced Fiber Resources' acquisition of 安捷讯 Anjiexun, together with the specific split of "shares/convertible bonds/cash" and the lock-up period.
7. The original text of the share lock-up commitments in 长鑫科技 CXMT's (688825) prospectus — the tiered unlock schedule for its 60.193 billion pre-IPO shares (90.0% of total share capital) was not obtained; the 6.73% free float and the listing date cited in the main text have been independently verified.
8. A complete list of trading suspensions and resumptions on the A-share calendar for 8/13 — this piece covers only what has been confirmed from announcements (兆日科技 Zhaori Technology 300333 resuming trading, 逸飞激光 Yifi Laser 688646 resuming and removing its warning label, 鸿特科技 Hongte Technology 300176 suspended from 8/13 for rights-issue payment), and completeness is not guaranteed.
9. Any new announcements released between 08:00 and 08:30 on 8/13 are not covered by this piece. The latest item inside the window was 康盛股份 Kangsheng (002418) at 23:04 on 8/12 (the chairman and general manager being released on bail pending trial).

One methodological self-constraint: the gaps above do not constitute permission to keep awarding high scores. This piece has imposed a hard constraint on names whose fundamentals are unverified — any name whose AI-related revenue share has not been obtained does not enter the 5-stock recommendation list in section 9.1 (Envicool and Taichenguang were withdrawn on that basis and moved to watch slots). "A disclosed gap" is not the same as "a gap that has been dealt with."

---

⚠️ Risk warning: this list is pre-market information organization 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 supply-chain mapping errors; it must not be used directly as a basis for trading.

Data-sourcing and run record (not sent to clients)

1. Cross-market timing mismatch (detailed in the internal block of section 2.2) Sina Finance's US market article of 2026-08-13 00:01 was a US midday snapshot rather than the close; the indices were overstated and Lumentum severely understated. Everything was switched to and checked against the CNBC quote interface (timestamps 2026-08-12 16:00–17:15 ET). Lesson: any US market article "published between 00:00 and 03:59 Beijing time on 8/13" is an intraday piece and cannot be used as the close; close pieces appear no earlier than 04:00.

2. Two cases of "old news as new news" that were caught

• TrendForce's "AI server shipments raised to nearly 31%": Sina Finance's early-session article on 8/12 reported it as that day's "major hard-tech positive," but checking against the Cailian Press original confirmed the research was published 2026-08-07 12:13, making it 5-day-old news. It is flagged in section 1 as "old news re-fermenting" and was not used as today's driver.
• The results pre-announcements of the optical module big three: searches repeatedly returned content about "results pre-announcements beating expectations," and verification confirmed that the Eoptolink and T&S Communications pre-announcements were released in mid-to-late July 2026, with interim report dates of 8/25 and 8/24 respectively. Flagged as old news.

3. Status of the akshare interfaces

• stock_lhb_detail_em succeeded this time in retrieving 57 dragon-tiger list entries for 8/12 (yesterday's recap could not get them because it is generated at 15:30; this run at 07:00 is past the 18:00 publication time). The scheduling problem raised in item 6 of section 6.4 of yesterday's recap simply does not exist in the pre-market slot — the pre-market slot can get the previous day's complete dragon-tiger list.
• stock_zt_pool_em is normal.
• Interfaces going through push2.eastmoney.com/api/qt/clist/get still require /tmp/em.py (browser UA + multi-host rotation) to get around blocking; akshare's default UA still fails. That fetcher is still in /tmp and will be lost on reboot; it should be consolidated into the repository's tooling layer.
• Per-stock single-point queries (a fs=i:0.CODE loop) timed out within 2 minutes and were switched to "paginated whole-market pull + local filtering," which is an order of magnitude more efficient. The latter should be adopted uniformly going forward.

4. Data not obtained

• The official close change for the Philadelphia Semiconductor Index was not obtained from Chinese media; the CNBC .SOX quote was used instead (12,399.376, +2.49%, timestamp 17:15 ET).
• The A-share calendar for 8/13 (a complete list of new share subscriptions/lock-up expiries/suspensions and resumptions) was not obtained: the JRJ 8/13 calendar page 404s and searches returned only 8/12 and 8/3. The suspension and resumption information in section 1 of this piece covers only the two confirmed from announcement summaries (Zhaori Technology, Yifi Laser 688646) and completeness is not guaranteed.
• 英维克 Envicool's (002837) 8/12 capital data was not obtained (it is not among the top 600 by turnover), and this is flagged as such in the section 2 table.

5. Sub-agent status: all returned, and they substantively overturned two of the draft's judgments

Three fundamentals-analyst agents (optical communications chain / memory and semiconductor chain / PCB materials and optical chip chain), one general-purpose announcement scan, and one risk-auditor quality check all returned and have been merged into the main text. This time the sub-agents did not add detail; they changed conclusions:

• The announcement scan saved a substantive error: the draft listed 城地香江 Chengdi Xiangjiang (603887) as a representative stock of the liquid cooling/AI server ancillary branch, whereas the company's announcement at 20:07 on 8/12 explicitly stated it "has no compute leasing business as yet," with an H1 loss forecast of RMB 7.2–10.8 million and outstanding external guarantees of RMB 4.261 billion, 121.87% of net assets. Without scanning announcements, this stock would have entered the main text as a "representative stock." It has been deleted.
• The same agent recovered 8 S/A-grade news items inside the window, all missed by the draft: Cisco's US$4 billion of AI orders, Tencent's Q2 earnings-call remarks on compute procurement, T&S Communications' equity incentive draft, SASAC's compute network visit, Yuanjie Semiconductor's RMB 4.268 billion capacity expansion, Sugon DataEnergy's liquid cooling subsidiary, Counterpoint's YMTC NAND third-globally item, and the passing of Comrade Zhu Rongji. The draft's sentence that "no new ministry-level industrial policy was found inside this window" was wrong and has been corrected.
• The memory chain agent overturned the draft's core framework: the draft treated "US memory surging vs A-share module capital outflows" as an expectation gap, when it is actually a mismatch of comparables (Micron/SanDisk/Western Digital/Seagate are IDMs, while A-share module makers correspond to the likes of Phison/ADATA). It has been rewritten as section 2.4.
• The PCB chain agent produced the most valuable item in this piece: the 8/12 capital choices are negatively correlated with "the degree to which AI revenue is realized in the financials." Written up as section 2.5.
• The quality-check agent caught 4 foundation-level errors, all now fixed: ① the communication equipment fund flow of "3.435 → 13.094" was a splice across two bases (on the sub-sector basis it should be 3.435 → 9.178, ×2.67); ② the seven dimensions of the scoring model cap at 95, not 100, and Innolight's 87 points did not add up; ③ giving Innolight 12 points for "earnings elasticity" violated the model's own basis (undisclosed results can score only 0–5), and after re-scoring, Innolight's 69 points fall below Eoptolink's 77, swapping 1st and 2nd on the recommendation list; ④ semiconductors' +RMB 5.450 billion was contaminated by 长鑫科技 CXMT alone (+RMB 3.603 billion, a newly listed stock with only a 6.73% free float), leaving about +RMB 1.85 billion once excluded.
• The quality check also flagged two cases of "inference presented as fact": ① the "share transfer" portion of Lumentum's gain is this piece's own inference; ② the timing on "bottleneck beneficiary" was wrong — Super Micro's earnings were released at 04:20 Beijing time on 8/12, so A-shares could price them all day, and "the sector has not reacted yet" can only be phrased as "capital chose not to buy"; moreover Vertiv rose only +2.32% that night, which is direct counter-evidence, yet the draft placed it among the supporting evidence. The branch has been cut from A to B+, and Envicool has been withdrawn from the recommendation list and moved to a watch slot.

Process suggestion: the quality check was started while the main text was still being rewritten, so its line-number snapshot did not match the final version and it read several already-fixed problems. Next time it should be started only after the fundamentals and announcement agents have all returned and the main text is final.

6. Two known limitations (unfixed, filed for the record)

• The quality check noted that third-party sources show a +13.94% close for Lumentum, 0.31 percentage points from the +13.63% used in this piece from CNBC (timestamp 16:00 ET). The CNBC basis is retained and the source is stated in 10.1; not every US single stock was re-checked one by one.
• Any new announcements between 08:00 and 08:30 on 8/13 that are not covered have been disclosed to clients in item 9 of section 10.2.

7. A script error that nearly caused an incident (must be remembered) This run used s[s.index(A):s.index(B)] to take the segment to be replaced, but B (the closing tag of the internal marker) appears twice in the text, and index hit the earlier one, making the slice an empty string; s.replace('', new) then inserted new between every two characters, and the file ballooned from 54KB to 90MB. It was fully restored via s.replace(new,''). Lesson: slice positioning must use rindex or first assert end > start, and before replacing one must assert that old is non-empty.