A-Share Pre-Market Brief | 2026-08-03 Monday

⚠️ Risk warning (up front): This list is nothing more than a pre-market information sweep and observation record. It does not constitute investment advice and contains no ratings, target prices, position sizing or buy/sell recommendations. The labels "priority deep-dive / watch closely / watch only / Pass" used below are purely a ranking of research attention, and imply nothing about future returns.

Data basis disclosed first (please read this section before anything else)

1. Time window: 2026-07-31 15:00 (last Friday's A-share close) → 2026-08-03 07:00 (cut-off of this report). Today is Monday and the first trading day of August; the window spans a full weekend, so news density is materially higher than on an ordinary trading day.
2. Market data sources: Sina full market quotes stock_zh_a_spot (5,533 stocks, timestamp 15:30:02, i.e. the 7/31 close), East Money limit-up pool stock_zt_pool_em / failed-limit-up pool stock_zt_pool_zbgc_em / limit-down pool (7/31 returned an empty table, i.e. no limit-downs that day), East Money dragon-tiger list stock_lhb_detail_em (81 entries on 7/31), THS industry sector summary stock_board_industry_summary_ths, Sina single-stock daily bars stock_zh_a_daily (forward-adjusted, used to compute interval gains and price position), and stock_financial_abstract (financials).
3. ⚠️ Interface gaps (must be disclosed): ① East Money full market quotes stock_zh_a_spot_em returned RemoteDisconnected and stock_individual_info_em errored repeatedly, so free-float market cap in this report comes from the East Money limit-up pool field instead (available only for limit-up stocks), and industry classification exists in two parallel bases — the East Money limit-up pool's "industry" field and the THS sectors — which cannot be compared directly across tables; ② the PE/PB figures in this report are not platform readings — wherever a valuation appears, it was computed independently by the fundamentals-analyst sub-agent using "2026-07-31 close × total share count" and "TTM net profit attributable to parent (2026Q1 + 2025A − 2025Q1)", with the algorithm noted at each location; for stocks not verified by a sub-agent, valuation is uniformly written as "no reliable data available" and is not estimated. 3b. ⚠️ After verification by the fundamentals-analyst sub-agent this edition had three conclusions directly overturned by the data, and the body text has been rewritten accordingly (this is the most important correction of this edition):
   • 中国核电 (China National Nuclear Power, 601985): 2026Q1 net profit attributable to parent −34.19% YoY (RMB 2.064 billion vs RMB 3.137 billion in 2025Q1), ex-non-recurring −35.0%, gross margin down 8.36pct YoY to 39.50%; 2026H1 commercial generation of 119.669 billion kWh, −1.73% YoY. "Approval tailwind" and "current-period earnings decline" are both true at once; the first draft wrote only the former, and the latter has now been added.
   • 云赛智联 (Yunsai Zhilian, 600602): the "cloud computing & big data" segment that carries the computing-power narrative saw 2025 revenue fall rather than rise (RMB 2.998 billion → RMB 2.879 billion, −3.97%), with its share dropping from 53.32% to 45.36%; the company's 2025 revenue growth actually came from systems integration, a business with a gross margin of only 9.95%. Its PE(TTM) is 123.0x / 147.9x ex-non-recurring, PB 4.90x, and ROE only 4.02%. The first draft listed it as "watch closely"; it has been downgraded to "watch only".
   • 太极实业 (Taiji Industry, 600667): engineering general contracting (76.47% of revenue) carries a gross margin of only 1.96%; 2025Q4 alone posted an ex-non-recurring loss of RMB 8.16 million; 2025 operating cash flow was −79.9% YoY; PE(TTM) 78.5x. That 78.5x denominator rests on the unproven assumption that "the 2025 back-payment of income tax was a one-off" — if the tax-incentive qualification has been permanently lost, the valuation picture is entirely different.
4. Dimensions not covered: single-day northbound net flow (the exchanges stopped disclosing it from August 2024; only one data point was obtained, via the dragon-tiger list — northbound net buying of RMB 454 million in 昆仑万维 (Kunlun Tech) alone, which cannot be extrapolated to the whole market), FTSE China A50 futures night session (repeated searches failed to obtain reliable point-in-time data, so "no reliable data available" is written), and the closing level of the Philadelphia Semiconductor Index on 7/31 (only intraday data was obtained; see Section 1, item 15).
5. Two basis conflicts that must be disclosed side by side (no choice is made between them; readers should judge for themselves):
   • July equity ETF net inflow: the intraday basis on 7/31 was "possibly over RMB 540 billion" (East Money, estimate), while the month-end settled basis was RMB 477.836 billion (Sina / Securities Times, settled). This report uses the settled figure of RMB 477.836 billion throughout; the RMB 540 billion figure is an intraday estimate and is not adopted.
   • Strait of Hormuz: Trump said on 8/1 that a framework had been reached and the strait would be "fully opened immediately"; but the Iranian side responded that this was "pure rumour" (reprinted by NetEase). A major geopolitical item where the two parties directly contradict each other; this report has downgraded its impact level from S to A and flagged it in Section 2 as "single-sided source, not confirmed by the counterparty".
6. The single most important structural fact of this edition (it determines the ordering of the entire list): the real leading theme of A-shares on 7/31 was not semiconductors, it was AI applications. On the THS industry basis: IT services +6.29% (net inflow +RMB 5.465 billion), software development +6.15% (net inflow +RMB 9.029 billion), culture & media +5.28% (net inflow +RMB 4.795 billion); whereas semiconductors, despite closing +4.23%, saw net outflows of RMB 20.502 billion, the largest net outflow of any sector in the market, with components also seeing RMB 7.289 billion of net outflow. "Up but sold" and "up and bought" are two different things. The distribution of limit-ups points the same way: 7 limit-ups in advertising & marketing, 5 in specialised engineering, 5 in IT services, 5 in software development, 4 in digital media, and only 2 in semiconductors (盛景微 Shengjingwei and 诚邦股份 Chengbang, neither a core name).
7. The only catalyst in this edition that is "completely unpriced": the 4 nuclear power projects / 8 units approved by the State Council executive meeting on 7/31 (total investment over RMB 170 billion). The counter-evidence is extremely clean — nuclear power stocks showed almost no reaction at all at the 7/31 close: 中国核电(601985) −0.11%, 中国广核 (CGN Power, 003816) −0.72%, 东方电气 (Dongfang Electric, 600875) +0.76%, 江苏神通 (Jiangsu Shentong Valve, 002438) +2.34%, not a single limit-up. That indicates the news only reached the market after the 7/31 close, so today is the first pricing day. ⚠️ A key ownership fact added after sub-agent verification: of the 8 units, Liaoning Zhuanghe Phase I (Units 1 and 2, owner CNNC Datang Zhuanghe Nuclear Power) and Zhejiang Jinqimen Phase II (Units 3 and 4, owner CNNC Xiangshan Nuclear Energy) — 4 units in total — are owned by controlled subsidiaries of 中国核电, and the company issued its own approval announcement on 7/31; the other 4 units belong to CGN (Taipingling Phase III) and SPIC (Laiyang Phase I) and have nothing to do with 601985. In other words: this news is "name-specific" for 中国核电 and merely "industry-level" for everything else — that is the key basis for this edition's ranking.
8. ⚠️ But item 7 comes with two equally clean pieces of counter-evidence, written into Sections 2 and 5:
   • On volume: 8 units were approved this time, below the 10—11 units per year approved in each of the four consecutive years 2022—2025 (10 units in 2022, 10 in 2023, 11 in 2024, 10 in 2025). If the market prices on "normalised approvals", 8 units is actually the lowest in four years, and the expectation gap may point negative.
   • On earnings (sub-agent verified, missed in the first draft): 中国核电 2026Q1 net profit attributable to parent −34.19%, ex-non-recurring −35.0%, H1 generation −1.73%, gross margin down four years running (44.62%→42.91%→41.32%→39.50%). In other words, the market today is being asked to pay, for a tailwind that will not produce electricity until around 2031 at the earliest, for a company whose current-period profit is falling by double digits. Without writing these two side by side, this list would be a one-sided narrative.
9. Every "single-stock benefit path" in this report has been tagged under a four-tier classification: direct order benefit / industry trend benefit / supply-chain indirect benefit / pure concept mapping. Wherever the original company announcement could not be obtained, the location is marked "media-aggregated basis only; the company announcement prevails".

---

0. Today in one sentence

What is being traded today is a market where "capital has already picked a direction, but the direction itself is thin." Last Friday (7/31) A-shares put in a fairly high-quality broad-based up candle: Shanghai Composite +0.72% on turnover of RMB 1,187.7 billion, Shenzhen Component +2.21% on RMB 1,354.3 billion, ChiNext +3.06%, STAR 50 +2.99%, combined Shanghai + Shenzhen turnover of RMB 2.542 trillion (1,187.7 + 1,354.3; separately, this report's sum of turnover across all 5,533 stocks is RMB 2.559 trillion, a basis that includes the STAR Market / BSE — the two must not be mixed), 4,690 advancers / 728 decliners / 879 stocks up more than 5%, 99 limit-ups (East Money pool basis) and 0 limit-downs. This is the first respectable repair candle after July's run of declines.

The strongest catalyst is not one item but a group, and they fall into two completely different categories (⚠️ note: after fundamental verification and QA review, this report has ranked nuclear power ahead of AI applications; the reason is in cold truth No. 6):

• [Ranked 2nd in this report] Already validated but fundamentally unqualified (already priced intraday on 7/31; today tests whether it can continue): AI applications / Agents. On the evening of 7/31 the DeepSeek-V4-Flash official-release API went into public beta (same model architecture as the preview version; purely post-training, delivering a large lift in Agent capability; the company says the V4-Pro official release will come "as soon as possible", with the industry expecting early August); on the same day the NDRC announced it would accelerate the legislative process for the AI Law, and disclosed that "AI-related industry growth exceeded 30% in H1, national intelligent computing power at end-June reached 2.8x the year-earlier level, and the first fully domestic 100,000-card cluster has been put into operation". Capital gave a clear answer: 昆仑万维 net buying on the dragon-tiger list of RMB 944 million (including RMB 454 million of northbound net buying), 蓝色光标 (BlueFocus) RMB 804 million, 中文在线 (Chinese Online) RMB 286 million, 易点天下 (Yidian Tianxia) RMB 272 million, 用友网络 (Yonyou Network) RMB 185 million — the most concentrated buying anywhere on 7/31. ⚠️ But read cold truth No. 6 below immediately afterwards: after fundamental verification, almost none of these most heavily bought main-board names qualify.
• [Ranked 1st in this report] Not reflected before the 7/31 close (today is the first pricing day): nuclear power. On 7/31 the State Council executive meeting approved Liaoning Zhuanghe Phase I, Zhejiang Jinqimen Phase II, Guangdong Taipingling Phase III and Shandong Laiyang Phase I — 8 units in total, with combined investment over RMB 170 billion — yet nuclear power stocks barely moved that day (中国核电 −0.11%, 中国广核 −0.72%, 东方电气 +0.76%; 江苏神通 +2.34% is the only one whose gain clearly exceeded the Shanghai Composite, so strictly speaking it is not "zero reaction"). Among them, Zhuanghe Phase I and Jinqimen Phase II — 4 units combined — are owned by controlled subsidiaries of 中国核电; this is the only catalyst in this edition that names a specific listed company. ⚠️ But "released after the close → no reaction at the close" is a necessary result and carries no incremental information, and this report has cut the expectation-gap score of the nuclear complex accordingly (see the basis correction in Section 4); moreover approvals have been normalised for five consecutive years, so expectations may already have been absorbed over the long run.
• The third item is index-level rather than stock-level: the PBOC's H2 work conference on 8/1 explicitly called for "continuing to implement an appropriately accommodative monetary policy" and to "give full play to the two monetary policy tools that support the capital market", on top of July equity ETF net inflows of RMB 477.836 billion, an all-time monthly high (up 160.39% from the April 2025 peak of RMB 183.507 billion, with broad-based products alone absorbing RMB 315.72 billion in the month). This lifts the valuation centre and risk appetite; it does not map to any particular stock.

Likely direction of capital: the money leaving "AI hardware" went into "AI applications" yesterday, but this report believes the newly opened nuclear power branch is the more worthwhile one to track today — because after verification the main-board AI application names fail on fundamentals across the board (cold truth No. 6), while nuclear power at least has 东方电气(600875) with current-period profit +37.4%. The driver types are industry (large-model iteration) + policy (State Council / PBOC / NDRC) + event (nuclear approvals), three lines stacked, with no earnings driver and no price driver — that is the biggest difference from last Friday's report.

The pre-market state assessment has to include seven cold truths, otherwise this list is just a chase-the-highs list:

1. ⚠️ The limit-up structure on 7/31 was extremely thin: of 99 limit-ups, 89 were first boards, only 10 were consecutive limit-ups, and the highest one is gone as of today. The full ladder: 爱丽家居 (Ailly Home, 603221) 9 boards, 传智教育 (Chuanzhi Education, 003032) 5 boards, 一鸣食品 (Yiming Food, 605179) 4 boards, 高争民爆 (Gaozheng Explosive, 002827) 3 boards, and only 6 names on 2 boards (神雾节能 Shenwu Energy Saving 000820, 普联软件 Pulian Software 300996, 泛微网络 Weaver Network 603039, 税友股份 Servyou Software 603171, 香江控股 Xiangjiang Holdings 600162, 中国宝安 China Baoan 000009). And 爱丽家居, on 9 boards, has announced a trading halt for verification from the open on 3 August (cumulative gain of 135.77% over 7/21—7/31, with the SSE simultaneously placing it under "key monitoring" and applying self-regulatory measures to 137 abnormal trading incidents this week). With the highest board removed, the ceiling drops to 5 boards (传智教育); today's theme ceiling has been forcibly lowered by administrative action — which directly affects the risk appetite of all consecutive-limit-up money.
2. ⚠️ The failed-limit-up rate is glaring: 99 limit-ups versus 107 failed limit-ups — more failures than limit-ups. And the ones that failed were precisely the hard-tech core names: 德明利 (Demingli Electronics, 001309) sealed at the open then turned negative to close −1.06% (turnover RMB 17.355 billion, turnover rate 25.6%), 兆易创新 (GigaDevice, 603986) closed +2.02% (turnover RMB 34.043 billion, the largest in the market), 生益科技 (Shengyi Technology, 600183) broke its limit 17 times to close +5.56%, 深南电路 (Shennan Circuits, 002916) broke its limit 13 times to close +8.26%, 风华高科 (Fenghua Advanced Technology, 000636) broke its limit 3 times to close −2.28% (14.18% intraday range). This is the textbook "index-wide rally, leaders distributing" structure.
3. ⚠️ Semiconductors were the largest net-outflow sector on 7/31 (−RMB 20.502 billion), with components seeing RMB 7.289 billion of net outflow. A green-on-the-screen sector masked capital leaving. Any "semiconductor rebound" narrative today must first explain where that RMB 20.5 billion went.
4. ⚠️ For the strongest theme — AI applications — fundamentals are the weakest link. This report actually pulled the financials of the media/advertising stocks that hit limit-up on 7/31 (stock_financial_abstract): 引力传媒 (Gravity Media, 603598) debt-to-asset ratio 90.71% (2026Q1); 掌阅科技 (iReader Technology, 603533) lost RMB 176 million for full-year 2025 and still lost RMB 34 million in 2026Q1, with operating cash flow negative for 5 consecutive reporting periods; 浙文互联 (Zhewen Interactive, 600986) 2026Q1 net profit attributable to parent −RMB 23 million, and full-year 2025 revenue of RMB 8.349 billion produced only RMB 86 million of profit (net margin about 1.0%); 三维通信 (Sunwave Communications, 002115) lost RMB 13 million in full-year 2025 and RMB 271 million in 2024. Along this chain "share price elasticity" and "profit delivery" are almost unrelated, so it has to be managed with tiered labels rather than with scores.
5. ⚠️ Today has two clear liquidity negatives; one cannot look only at the positives: ① RMB 1,842.0 billion of PBOC open-market reverse repos mature this week, of which RMB 925.5 billion falls due today (Monday) alone, and another RMB 300 billion of 3-month outright reverse repos matures on 8/5; ② today is this week's peak for share unlocks, with RMB 14.042 billion unlocking in a single day (42 companies totalling RMB 23.475 billion this week), including 浙江荣泰 (Zhejiang Rongtai, 603119) unlocking 208 million shares, 43.95% of total share capital.
6. ⚠️⚠️ This is the most important cold truth in the report: the main-board AI application stocks that rose the most yesterday were almost wiped out by fundamental verification. Two fundamentals-analyst sub-agents went item by item through the four main-board AI application stocks that hit limit-up on 7/31 with net buying on the dragon-tiger list; the results follow — this is what pushed the first draft's No. 1 down to No. 11 and its No. 6 down to No. 21:
   • 用友网络(600588): three consecutive years of losses: 2023 net profit attributable to parent −RMB 967 million, 2024 −RMB 2.061 billion, 2025 −RMB 1.389 billion, a cumulative RMB 4.417 billion over three years; another −RMB 723 million in 2026Q1; the 2026 interim guidance still shows a loss of RMB 800—930 million with revenue growth of only +1.9%~+3.3%; high-margin (93.3%) "product licence" revenue is down 29.3% over two years; current ratio 0.66; the controlling shareholder has pledged 49.54% of its holding and made four successive pledge / supplementary pledge filings in June—July. The single most glaring item: in the entire 2025 annual report, "computing power" appears 0 times, "GPU" 0 times, "intelligent computing" 0 times.
   • 利欧股份(002131): on 2026-07-03 the SZSE publicly censured the company, its chairman and its CFO — all three (earnings guidance of RMB 190—250 million → revised to RMB 30—45 million → actual RMB 33.74 million); the 2026 interim guidance shows a first-ever ex-non-recurring loss of −RMB 90 to −130 million; three-year cumulative ex-non-recurring profit of RMB 256 million on revenue of RMB 61.692 billion = net margin 0.41%; 76% of revenue is media agency work at a 3.86% gross margin, while the real gross-profit pillar is water pumps (58.3% of gross profit); at its results briefing the company itself said AI revenue "cannot be clearly broken out", "remains a small share", "has limited near-term impact", and explicitly denied holding equity in Moore Threads / Unitree / DeepSeek.
   • 润建股份(002929): net profit attributable to parent down 91.4% over three years (RMB 439 million → RMB 37.9 million), ROE 0.60%, 2026Q1 revenue −23.5%; accounts receivable of RMB 5.928 billion, 294 days turnover, 93.9% of net assets; debt-to-asset ratio 70.96%, current ratio 1.02, and it is buying back shares using a dedicated share-buyback loan.
   • 智度股份(000676): its statements are the cleanest of the four (debt ratio 13.56%, net cash, 2026H1 ex-non-recurring +10%~+23%), but in the 2025 annual report "Web3" appears 0 times and "crypto" 0 times, while "computing power" appears just 2 times; the real profit engine is traffic monetisation aimed at Yahoo/Google in North America (74.5% of gross profit), unrelated to AI; and on 2026-05-22 it was publicly criticised by the SZSE (retrospective correction of prior accounting errors). Conclusion: along this chain, "share price elasticity" and "profit delivery" are not merely unrelated — they frequently point in opposite directions. If you want exposure to AI applications today, look first at 泛微网络(603039) and 税友股份(603171), which have real profits, rather than at the names that were bought hardest on yesterday's dragon-tiger list.
7. ⚠️ After item-by-item sub-agent verification, the other two hottest narratives of this edition were also discounted by the companies' own financials:
   • 中国核电, the protagonist of the "nuclear approval" story, saw 2026Q1 net profit attributable to parent −34.19% YoY (RMB 2.064 billion vs RMB 3.137 billion), ex-non-recurring −35.0%, 2026H1 commercial generation −1.73%, gross margin down four years running to 39.50%, and interest-bearing debt of roughly RMB 461.7 billion (against cash of only RMB 21.763 billion); moreover of the RMB 17.971 billion of net profit in 2025, only 51.8% was attributable to the parent — minority interests took nearly half. The tailwind is in 2031; the profit decline is now.
   • 云赛智联, the main-board flagship of the "RMB 4 trillion computing-power network" story, is precisely the case where the segment that carries computing power is shrinking: 2025 "cloud computing & big data" revenue RMB 2.998 billion → RMB 2.879 billion (−3.97%), share 53.32% → 45.36%; the whole company's revenue growth came entirely from systems integration at a gross margin of just 9.95%. Three-year revenue +20.6% while net profit attributable to parent rose +0.5% (RMB 193 million → RMB 194 million), ROE locked at 4% year after year, against PE(TTM) 123.0x, 147.9x ex-non-recurring, PB 4.90x. The actual operating data of the past year does not support the sentence "computing power is driving growth."

---

1. News overview

Impact level: S = changes an industry trend or delivers direct order elasticity; A = a clear tailwind for a specific branch benefiting multiple stocks; B = a tailwind but with a long chain requiring verification; C = sentiment stimulus with weak persistence. All times are Beijing time. Anything obtained only on a second-hand aggregated basis, or where sources conflict, is flagged in the "Type" column and down-weighted.

---

2. Strongest tailwind branches, in descending order

⚠️ Ranking-correction statement (please read first): the first draft ranked AI applications 1st and nuclear power 2nd. After fundamentals-analyst verification the two have been swapped — not because the nuclear news got stronger, but because the degree to which the fundamentals of the main-board AI application names were falsified exceeded expectations (用友 has lost RMB 4.417 billion cumulatively over three years while "computing power" appears 0 times in its annual report; 利欧's chairman and CFO were publicly censured by the SZSE; "Web3" appears 0 times in 智度's annual report), while the nuclear branch turned up 东方电气 with 2026Q1 net profit attributable to parent +37.4%. The "Rank" column in the table below is already ordered on the corrected basis, consistent with Section 9.2.

Supplementary judgements on each branch

Branch 1 (AI applications) — why it still ranks first, and why its strength was cut from S to A+. The reason for ranking first has not changed: it is the only branch this edition where news, price and capital flow all validated at once. On 7/31 software development was +6.15% (net inflow RMB 9.029 billion), IT services +6.29% (+RMB 5.465 billion), culture & media +5.28% (+RMB 4.795 billion) and gaming +4.50%; in the limit-up industry distribution there were 7 in advertising & marketing, 5 in software development, 5 in IT services and 4 in digital media; four of the top five names by dragon-tiger net buying belong to this chain (昆仑万维 RMB 944 million including RMB 454 million northbound, 蓝色光标 RMB 804 million, 中文在线 RMB 286 million, 易点天下 RMB 272 million). Is it already fully expected? — Not yet; the V4-Pro official release is still unpublished, and early August is the next definite catalyst point.

⚠️ But the strength must be cut from S to A+, because the fundamental verification results are overwhelmingly negative and concentrated in exactly the hottest names:

• 用友网络(600588) — the largest, best-known main-board flagship of this chain, with RMB 185 million of dragon-tiger net buying on 7/31: three consecutive years of losses totalling RMB 4.417 billion, 2026 interim guidance still showing a loss of RMB 800—930 million with revenue growth of just 1.9%~3.3%, high-margin product licence revenue shrinking 29.3% over two years, current ratio 0.66, and the controlling shareholder having pledged 49.54% of its holding. And in its entire 2025 annual report "computing power" appears 0 times.
• 利欧股份(002131) — on 2026-07-03 the SZSE publicly censured the company, its chairman and its CFO; a first-ever ex-non-recurring loss in the 2026 interim guidance; at its results briefing the company itself said AI revenue "cannot be clearly broken out", "remains a small share", "has limited near-term impact", and denied holding equity in Moore Threads / Unitree / DeepSeek.
• 智度股份(000676) — "Web3" 0 times and "crypto" 0 times in the 2025 annual report; the real profit comes from traffic monetisation with Yahoo/Google in North America.

So the correct use of this branch has changed: short term or medium term? — after verification, only two names in enterprise software, 泛微网络(603039) and 税友股份(603171), have a genuine medium-term earnings-delivery path; the whole marketing & media line is pure sentiment. Is there one-day-wonder risk at highs? — Yes, and 智度股份(000676) has already made a 45-day high (0.0% from the 45-day high), +40.5% over the past 20 sessions, the most extended position of all. Is there anything on the main board to buy? — There are plenty of main-board names, but only 2 survive fundamental verification, and that is where this branch differs most from the first draft's judgement.

Branch 2 (nuclear power) — why it is today's "cleanest" line, and its two traps. Why it is a tailwind: a State Council approval is project establishment in the sense of an administrative licence, making it a direct order benefit (installed capacity for operators, future tenders for equipment makers), and it carries the highest logic robustness of any news item this edition. What makes it harder than an ordinary policy item is that it "names names" — 4 of the 8 units (Zhuanghe Phase I Units 1 and 2, owner CNNC Datang Zhuanghe Nuclear Power; Jinqimen Phase II Units 3 and 4, owner CNNC Xiangshan Nuclear Energy) are owned by controlled subsidiaries of 中国核电, and the company issued its own approval announcement on 7/31; the other 4 units belong to CGN (Taipingling Phase III) and SPIC (Laiyang Phase I). "8 units approved" is news for the whole industry; for 中国核电 it is an announcement. Why it is described as unpriced: nuclear stocks collectively showed zero reaction at the 7/31 close — 中国核电 −0.11%, 中国广核 −0.72%, 东方电气 +0.76%, 江苏神通 +2.34%, not one major nuclear name hit limit-up and nuclear power did not make the top 15 of the limit-up industry distribution, indicating the news only spread after the close. ⚠️ Trap one (volume): the approval cadence has run 6 units in 2019 → 4 in 2020 → 5 in 2021 → 10 in 2022 → 10 in 2023 → 11 in 2024 → 10 in 2025 → 8 in 2026. If the market has taken "10-plus units a year" as its baseline, 8 units is below baseline and this news could perfectly well gap up and fade. Caixin's headline was "high-speed nuclear approval cadence continues" (a positive reading), while the number itself is a four-year low — both readings will show up on the tape today, and the auction decides which wins. ⚠️ Trap two (earnings, added after sub-agent verification): 中国核电 2026Q1 net profit attributable to parent −34.19%, ex-non-recurring −35.0%, gross margin down 8.36pct YoY, 2026H1 commercial generation −1.73%, which the company/sell-side attribute to "7 units in overhaul in Q1 (versus 2 a year earlier) + falling on-grid tariffs + falling composite renewable tariffs". 江苏神通's revenue has been flat around RMB 2.1 billion for three years running (2.133/2.143/2.129), and 2026Q1 was still −9.8%. In other words, not a single company in this branch has current-period earnings that are moving up. Is there an earnings-delivery path: on the operator side you have to wait for units to come online (5—6 years; the 4 units approved this time will not generate power until around 2031 at the earliest); on the equipment side, nuclear island valve tenders usually start 1—2 years after construction begins and revenue recognition lags another 1—2 years, so the corresponding revenue most likely lands in 2028—2030. Not one cent of revenue is delivered within 2026 — that is why it gets an A+ rather than an S. Is there anything on the main board to buy: all of them are main board and most are large-cap central SOEs, so trading recognisability is good but elasticity is poor.

Branch 3 (computing-power network) — it has already had a run, and there is no financial evidence that "computing power is growing". This chain clearly kicked off on 7/31: 宏景科技(301396) hit its 20cm limit, and 美利云(000815) / 云赛智联(600602) / 深桑达Ａ(000032) / 润建股份(002929) / 卧龙电驱(600580) / 麦格米特(002851) all hit limit-up. So today is not day one; it is the day that tests whether there is a second-day handoff. The RMB 4 trillion figure must be down-weighted — the NDRC's exact words were "estimated by relevant institutions", not an official investment plan, and treating it as committed capex is the easiest mistake to make this edition. ⚠️ More importantly, a piece of counter-evidence turned up by the sub-agent takes this branch straight from A to B+: 云赛智联's 2025 "cloud computing & big data" revenue (the segment containing IDC/computing power) fell from RMB 2.998 billion to RMB 2.879 billion (−3.97%), with its share dropping from 53.32% to 45.36%; the company's 11.7% revenue growth that year came entirely from "industry solutions" (systems integration) at a gross margin of only 9.95% — which is exactly why three-year revenue is +20.6% while net profit attributable to parent is only +0.5%, gross margin has fallen four times running, and ROE is locked at 4%. 太极实业 has a different problem: its data centre EPC business is real and it does have a record of winning computing-centre bids, but engineering general contracting, at 76.47% of revenue, carries a gross margin of only 1.96%, and taking more orders does not mean earning more money; on top of that 2025Q4 alone posted an ex-non-recurring loss of RMB 8.16 million and full-year operating cash flow was −79.9% YoY. Conclusion: the narrative in this branch is real, but "the narrative turning into revenue" is something no company's financials can currently prove.

Branch 4 (monetary easing) — this report deliberately sets no recommended stocks. The reason: a PBOC conference affects overall liquidity and the valuation centre, and there are too many links between "appropriately accommodative" and the profit of any particular broker, so any "PBOC easing → buy broker X" mapping is pure concept mapping. And there is a direct offset in the broker sector today — the CSRC fined 6 firms and filed cases against 2 on 7/31. This belongs in position sizing and risk appetite, not in stock selection.

Branch 5 (Hormuz) — conflicting sources, so direction only, no single stocks. Trump's and Iran's statements contradict each other, and the theme has already reversed several times this year (on 7/27 a "US-Iran halt in mutual strikes" sent oil down 8% in a day, with US crude at one point down more than 6% to USD 83.10/barrel). The correct use of this kind of news is "avoid chasing highs in oil & gas", not "buy airlines".

---

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

Note: "main-board buyable" = Shanghai main board (600/601/603/605) or Shenzhen main board (000/001/002); ChiNext (300/301) / STAR Market (688) / BSE (8xx/920) are all marked "No" and only enter the watch zone. ⚠️ Valuation basis: the East Money quote and single-stock information interfaces were both unavailable today (see basis disclosure item 3), so the 4 names carrying PE/PB in the table (中国核电, 江苏神通, 太极实业, 云赛智联) were computed by the fundamentals-analyst sub-agent using "07-31 close × total shares" and "TTM net profit attributable to parent = 2026Q1 + 2025A − 2025Q1"; these are not platform readings; valuations for the remaining stocks were not obtained and are not estimated. "Free-float market cap" is available only for stocks that hit limit-up on 7/31 (from the East Money limit-up pool); for those that did not, "no reliable data available" is written. ⚠️ After verification by two fundamentals-analyst sub-agents this table was substantially reordered relative to the first draft — this is the most important passage in the report, please read it in full:

Stock	First draft	Now	Basis
用友网络	79 pts / No. 1 / priority deep-dive	48 pts / No. 11 / watch only	three consecutive years of losses totalling RMB 4.417 billion; 2026 interim guidance still shows a loss of RMB 800—930 million with revenue growth of only +1.9%~+3.3%; "computing power" appears 0 times in the annual report; current ratio 0.66; controlling shareholder has pledged 49.54% of its holding
利欧股份	68 pts / No. 6 / watch closely	30 pts / No. 21 / Pass	on 2026-07-03 the SZSE publicly censured the company, its chairman and its CFO; first-ever ex-non-recurring loss in the 2026 interim guidance; three-year ex-non-recurring net margin 0.41%; the company itself says AI revenue "cannot be broken out and remains a small share"
润建股份	60 pts / No. 10 / watch closely	43 pts / No. 17 / watch only	net profit attributable to parent −91.4% over three years, ROE 0.60%, 2026Q1 revenue −23.5%; receivables turnover 294 days, 93.9% of net assets; debt ratio 70.96%, current ratio 1.02; buying back shares with a dedicated loan
云赛智联	65 pts / No. 7 / watch closely	48 pts / No. 12 / watch only	the segment carrying computing power saw 2025 revenue −3.97% with its share falling 53.3%→45.4%; PE(TTM) 123x / 148x ex-non-recurring; ROE 4.02%
智度股份	63 pts / No. 8	45 pts / No. 15	"Web3" 0 times and "crypto" 0 times in the annual report; publicly criticised by the SZSE in 2026-05 (retrospective correction of prior accounting errors); at a 45-day high
中国核电	76 pts / No. 2	68 pts / No. 4	2026Q1 net profit attributable to parent −34.19%, H1 generation −1.73%, gross margin down four years running
太极实业	54 pts	50 pts	engineering general contracting gross margin only 1.96%; ex-non-recurring loss in 2025Q4 alone; operating cash flow −79.9%
东方电气	70 pts / watch closely	72 pts / No. 3 / priority deep-dive (↑)	2026Q1 net profit attributable to parent +37.4%, full-year 2025 +31.1%, the only name on the whole list with fast-growing current-period profit
江苏神通	62 pts	66 pts / No. 5 (↑)	nuclear power is 38.10% of core revenue with a 38.48% gross margin, the highest of any segment; debt ratio only 38.84%, net interest-bearing debt RMB 390 million

Table scrolls horizontally

This table in one sentence: the batch of main-board AI application stocks that rose the most and were bought hardest on the dragon-tiger list last Friday (7/31) were almost wiped out by fundamental verification; what actually stands up to scrutiny is the nuclear equipment names and two enterprise software names that barely moved that day.

---

4. Single-stock scoring model (100 points total)

4.1 Mapping rules from score to label (added after the risk-auditor QA review)

The QA review pointed out that the first draft had a "scores do not map to labels" defect (for example the No. 1 name on the overall list, 泛微 at 74 pts, was labelled "watch closely" while 东方电气 at 72 pts was labelled "priority deep-dive"), and that no threshold had ever been defined. The following rules are now added, and the whole table has been re-checked against them:

Note: a high score does not mean a strong label. 泛微 at 71 pts gets "watch closely" rather than "priority deep-dive" because its 2026Q1 revenue grew only +1.2% (the +138% profit growth came from a low base), so its fundamentals dimension does not reach 9 pts; 税友 at 58 pts lands in "watch only" because a PE of 130x triggers "valuation materially above comparable names in the same branch". Only by looking at the two dimensions separately do you avoid treating the score as a strength of recommendation.

Scoring example (用友网络, 79 pts in the first draft → 48 pts after verification):

First draft (wrong, and the arithmetic was wrong too): 13＋14＋10＋12＋5＋9＋9−3 ＝ 69, yet the first draft wrote 79 pts — this is a calculation error that has to be admitted, and it happened to occur on the first draft's No. 1 name. After verification (current): news source 13 (NDRC original text + DeepSeek official) ＋ directness of tailwind 12 (industry trend, with no corresponding revenue line in the annual report) ＋ earnings elasticity 4 (three consecutive years of losses, interim guidance still a loss of RMB 800—930 million) ＋ fundamentals 3 (current ratio 0.66, revenue growth 1.9%~3.3%, product licence revenue −29.3% over two years) ＋ expectation gap 4 ＋ persistence 9 (V4-Pro in early August) ＋ trading characteristics 9 (main board, turnover RMB 1.036 billion) ＋ risk −6 (controlling shareholder has pledged 49.54% of its holding with four successive supplementary pledges −3; only −2.9% from the 45-day high, an elevated position −3) ＝ 48.

Scoring example (中国核电, 76 pts in the first draft → 68 pts after verification):

First draft (wrong, arithmetic also wrong): 15＋18＋4＋13＋9＋5＋8−0 ＝ 72, yet the first draft wrote 76 pts — another calculation error, corrected here as well. After verification (current): news source 15 (State Council + company announcement) ＋ directness of tailwind 18 (approval = direct project establishment, and 4 units are owned by its subsidiaries) ＋ earnings elasticity 4 (no power generated until around 2031 at the earliest) ＋ fundamentals 9 (2026Q1 net profit attributable to parent −34.19%, gross margin down four years running, interest-bearing debt about RMB 461.7 billion, minorities take 48.2% of net profit) ＋ expectation gap 6 (see the basis correction below; cut from 9) ＋ persistence 5 ＋ trading characteristics 9 ＋ risk −2 (8 units is below the norm of the past four years) ＝ 68.

⚠️ On the basis correction for the "expectation gap" dimension (added after the risk-auditor QA review; this report admits the first draft's reasoning here was defective): the first draft's scoring model stipulated that "no reaction on 7/31 = high score (nuclear complex 8—9 pts)", but that reasoning is circular — the State Council news was released after the close on 7/31, so no reaction at the close is a necessary result, carrying no incremental information and implying nothing about a positive expectation gap. Worse, it constitutes a double standard: DeepSeek-V4-Flash was also released on the evening of 7/31, yet this report judged AI applications to have been "already priced intraday on 7/31". This report has therefore cut the nuclear complex's expectation-gap score uniformly from 8—9 to 5—6 (neutral to positive), and changed the wording from "completely unpriced" to "not reflected before the 7/31 close (news released after the close)". A counter-example must also be stated alongside: 江苏神通 closed +2.34% on 7/31, more than three times the Shanghai Composite's +0.72%, so treating it as "zero reaction" is not rigorous; and nuclear approvals have been normalised for five straight years, making them an inherently predictable event, so 中国核电's flat-lining with a 45-day range of just 5.7% can equally be read as expectations having been absorbed over the long run, rather than as unpriced.

---

5. Detailed analysis of key stocks (10 names + 1 appendix)

Note on how this section is composed (different from the first draft, please read first): this section is not simply "the top 10 of the overall list" but "the top 6 of the overall list" + "the 4 names that hit limit-up on 7/31 with the most concentrated capital but were materially downgraded after verification". The reason: those latter 4 (用友网络, 智度股份, 润建股份, 利欧股份) are exactly the stocks that hit limit-up last Friday, had net buying on the dragon-tiger list, and are most likely to be chased today, and what readers most need to see is precisely their verification results, not to have them quietly dropped from the list. For every downgraded name, the first-draft rank, the current rank and the basis are all written out.

⚠️ All financial data in this section comes from stock_financial_abstract (statement basis), with the reporting period noted at each point. Valuations (PE/PB) are missing throughout because of the interface failure, so this section makes no valuation judgements.

1泛微网络603039SH main board　Watch closely

• Related news: same as ① (DeepSeek-V4-Flash official release + acceleration of AI Law legislation).
• Tailwind logic: industry trend benefit. Collaborative office (OA) is one of the most natural landing scenarios for Agents — approval flows, documents and knowledge bases are inherently suited to Agent rework.
• Stage of the industry branch: fermentation stage, and it is one of only 3 names on this chain with 2 boards (the other two being 普联软件 300996 on ChiNext and 税友股份 603171). Within the main board, 泛微 and 税友 are the only consecutive-limit-up ladder in the AI applications branch — these two are the key observation points today for whether the branch can continue.
• Fundamental verification (statement data obtained):
  • 2026Q1: total revenue RMB 333 million, net profit attributable to parent RMB 62.1 million, ex-non-recurring RMB 49.9 million
  • 2025Q1 (same period): total revenue RMB 329 million, net profit attributable to parent RMB 26.1 million → a large YoY improvement in 2026Q1 net profit attributable to parent
  • Full-year 2025: total revenue RMB 2.287 billion, net profit attributable to parent RMB 289 million, ex-non-recurring RMB 268 million
  • Debt-to-asset ratio 37.89% (2026Q1), in a healthy range
  • ⚠️ 2026Q1 operating cash flow per share −RMB 0.478 (2025Q1 was −RMB 0.637; negative Q1 cash flow is a seasonal feature for it, and YoY it is in fact an improvement)
  • Core-business match: yes (collaborative management software is the main business, not a cross-sector pivot)
• Industry position: one of the niche leaders in collaborative office software.
• Technical sentiment: 2 boards on 7/31, limit-up turnover RMB 860 million, sealing orders RMB 108 million, free-float market cap RMB 11.51 billion. ⚠️ +31.5% over the past 5 sessions, one of the largest 5-session gains among main-board names on this list; −24.0% from the 45-day high (indicating a large July drawdown, so this is a rebound rather than a new high).
• Final judgement: watch closely. Its fundamentals are among the best on this chain (real profit, low debt ratio, YoY improvement), but +31.5% in 5 sessions has already used up a lot. Its value lies more in being a thermometer for the branch — whether 泛微 and 税友 can make a 3rd board today directly determines whether AI applications are a "main theme" or a "one-day wonder".

2税友股份603171SH main board　Watch only (⚠️ downgraded from "watch closely" after the risk-auditor QA review; see QA note item 1 at the end)

• Related news: same as ①.
• Tailwind logic: industry trend benefit. Tax & finance SaaS + AI efficiency gains (invoicing, filing, audit and other structured scenarios, highly suited to Agents).
• Stage of the industry branch: fermentation stage, 2 boards on 7/31 (IT services sector, limit-up turnover RMB 962 million, free-float market cap RMB 18.52 billion).
• Fundamental verification (statement data obtained):
  • 2026Q1: total revenue RMB 488 million, net profit attributable to parent RMB 31.3 million, ex-non-recurring RMB 28.8 million
  • 2025Q1: total revenue RMB 449 million, net profit attributable to parent RMB 24.7 million → moderate YoY growth
  • Full-year 2025: total revenue RMB 2.054 billion, net profit attributable to parent RMB 136 million, ex-non-recurring RMB 88.7 million
  • Full-year 2024: total revenue RMB 1.945 billion, net profit attributable to parent RMB 113 million
  • Debt-to-asset ratio 37.78% (2026Q1)
  • ⚠️ 2026Q1 net operating cash flow −RMB 390 million — but this is seasonal, not deterioration: 2025Q1 was −RMB 370 million, full-year 2025 turned positive at +RMB 251 million and full-year 2024 was +RMB 274 million. Collections in tax software are concentrated in H2, so a Q1 net outflow is an industry characteristic and should not be treated as a red flag on its own.
  • Core-business match: yes
• Industry position: a niche leader in tax informatisation (serving both the tax system and the corporate side).
• Technical sentiment: 2 boards, +30.7% over the past 5 sessions, +26.3% over the past 10, −15.4% from the 45-day high.
• ⚠️ Valuation (computed after the QA review; this is the decisive basis for the downgrade): PE(TTM) 130.0x (RMB 45.62 × 406.35 million shares = market cap RMB 18.54 billion ÷ TTM net profit attributable to parent of RMB 142.6 million). This report previously judged 云赛智联(600602) "watch only" over a PE of 123x — 税友 at 130x is higher, so keeping 税友 as a recommendation would be a double standard.
• ⚠️ Profit quality (checked after the QA review): 2025 ex-non-recurring profit of RMB 88.7 million was only 65% of the RMB 136 million attributable to parent, so non-recurring items are not a small share; this report did not obtain the breakdown and marks it as to be verified.
• Final judgement: watch only (downgraded). Its balance sheet is clean (debt ratio 37.78%) and the fit between its core business and AI scenarios is real, but a 130x PE corresponds to a company that earned RMB 136 million attributable to parent in 2025 and is growing gently. The downgrade is not because new bad news emerged, but because this report had applied a looser standard to it than to 云赛智联 — the risk-auditor flagged this, and it is corrected here.

3东方电气600875SH main board　Priority deep-dive (⚠️ upgraded from "watch closely")

• Related news: State Council approval of 8 units on 7/31 (same as ④).
• Tailwind logic: supply-chain indirect benefit. According to public reporting, the company supplies core main equipment for Hualong One and Guohe One and ranks highly in reactor pressure vessels and steam generators (Cailianshe supply chain review⁴⁵). ⚠️ But this must be stated clearly: this approval is only project establishment; equipment tendering has not started, and the company has no award announcement of any kind relating to these 8 units. Equating "approval" directly with "东方电气 has won orders" is wrong.
• Stage of the industry branch: day one, only +0.76% on 7/31, closing at RMB 25.10 with turnover of RMB 1.283 billion, no limit-up.
• ✅ Fundamental verification (pulled by this report via stock_financial_abstract, statement basis — it is the only name on the whole list with fast-growing current-period profit):
  • 2026Q1: total revenue RMB 17.470 billion, net profit attributable to parent RMB 1.585 billion (2025Q1: RMB 1.154 billion, +37.4% YoY), ex-non-recurring RMB 1.179 billion (2025Q1: RMB 1.058 billion, +11.5%)
  • Full-year 2025: net profit attributable to parent RMB 3.831 billion (2024: RMB 2.922 billion, +31.1% YoY), ex-non-recurring RMB 3.192 billion (2024: RMB 1.980 billion, +61.2%), revenue RMB 78.615 billion (2024: RMB 69.695 billion, +12.8%)
  • Debt-to-asset ratio 69.47% (2026Q1), broadly flat against 70.39% at end-2025
  • ⚠️ 2026Q1 operating cash flow −RMB 2.698 billion — but this is seasonal, not deterioration: 2025Q1 was −RMB 3.293 billion (an improvement this year), full-year 2024 was +RMB 10.059 billion and full-year 2025 +RMB 2.014 billion. This company's Q1 cash flow is negative every year, a payment-cycle characteristic of the engineering equipment industry.
  • ⚠️ There is a 26pct gap between +37.4% YoY for net profit attributable to parent and +11.5% YoY ex-non-recurring, indicating a sizeable share of non-recurring items in 2026Q1 profit; this report did not obtain the breakdown and marks it as to be verified.
• Industry position: one of the three main suppliers of nuclear main equipment (alongside 上海电气 and China First Heavy Industries, on a public-reporting basis); it is also a leader in thermal / hydro / wind main equipment — ⚠️ nuclear power is only one of its businesses, and this report did not obtain its segment revenue split, so nuclear power's contribution to its revenue "cannot be confirmed at this time".
• Technical sentiment: closed at RMB 25.10 +0.76% on 7/31 with turnover of RMB 1.283 billion, no limit-up (unpriced).
• Final judgement: priority deep-dive (upgraded). There is only one reason for the upgrade, but it is a hard one: on this list it is the only main-board name where "the catalyst is unpriced (only +0.76% on 7/31)" and "current-period profit is growing fast (2026Q1 net profit attributable to parent +37.4%, full-year 2025 +31.1%)" hold simultaneously. By comparison, 中国核电 is "named by the catalyst but with profit −34%", and the four AI application names are "hot with money but fundamentally unqualified across the board". Two limits must be remembered at the same time: ① its directness of benefit is lower than the operators' — project approval ≠ equipment award, so please do not rank it ahead of 中国核电 just because "equipment stocks have more elasticity"; ② the share of nuclear power in its total revenue could not be obtained this edition, and it is first and foremost a diversified power equipment maker, with nuclear only one piece.

4中国核电601985SH main board　Priority deep-dive

• Related news: on 7/31 the State Council executive meeting approved 4 projects — Liaoning Zhuanghe Phase I, Zhejiang Jinqimen Phase II, Guangdong Taipingling Phase III and Shandong Laiyang Phase I — comprising 8 units with total investment over RMB 170 billion (China Energy News¹, Shanghai Securities News³). All three central SOEs — CNNC, CGN and SPIC — have projects approved.
• ⭐ Ownership relationship (this is the only reason it ranks first in the nuclear branch): 4 of the 8 units are owned by controlled subsidiaries of 中国核电 — Liaoning Zhuanghe Phase I (Units 1, 2; owner CNNC Datang Zhuanghe Nuclear Power Co., Ltd.) and Zhejiang Jinqimen Phase II (Units 3, 4; owner CNNC Xiangshan Nuclear Energy Co., Ltd.); the company issued its own approval announcement on 7/31 (Sina Finance⁴⁶). The other 4 units belong to CGN (Taipingling Phase III) and SPIC (Laiyang Phase I) and are unrelated to this company. ⚠️ The announcement did not disclose shareholding percentages, investment amounts or quantified earnings impact, and this report makes no estimates.
• Tailwind logic: direct order benefit (for an operator = installed capacity targets locked in). Approval means greater certainty of capacity additions over the next 5—6 years, directly affecting long-dated generation and the profit centre. This is the item with the highest logic robustness of all the news this edition — it is not an expectation, not an estimate, not a media paraphrase; it is an administrative licence plus a company announcement.
• Stage of the industry branch: day one (today is the first pricing day). Evidence: at the 7/31 close 中国核电 was −0.11%, 中国广核 −0.72%, 东方电气 +0.76% and 江苏神通 +2.34%; not one nuclear stock hit limit-up and nuclear power did not enter the top 15 of the limit-up industry distribution. The news spread only after the close and the market has yet to price it. There is no limit-up ladder and no flagship — which is both an opportunity and a risk: no ladder means that if it moves today, the ladder is being built from zero; and if it does not move, it will not even qualify as a one-day wonder.
• ⚠️ Fundamental verification (fundamentals-analyst sub-agent; this item points the opposite way from the tailwind above and must be read alongside it):
  • 2026Q1: total revenue RMB 18.925 billion (−6.65% YoY), net profit attributable to parent RMB 2.064 billion (−34.19% YoY), ex-non-recurring RMB 2.037 billion (−35.0% YoY), gross margin 39.50% (−8.36pct YoY)
  • 2025A: revenue RMB 82.075 billion (+6.2%), net profit attributable to parent RMB 9.304 billion (+6.0%); 2024A net profit attributable to parent RMB 8.777 billion; 2023A RMB 10.624 billion
  • Gross margin down four years running: 44.62% (2023) → 42.91% (2024) → 41.32% (2025) → 39.50% (26Q1)
  • 2026H1 commercial generation of 119.669 billion kWh, −1.73% YoY; the nuclear portion was 97.972 billion kWh, −1.89% YoY (company announcement dated 2026-07-02)
  • Reasons for the decline (company / sell-side basis): 7 units in overhaul in Q1 (versus only 2 a year earlier) + falling on-grid tariffs + falling composite renewable tariffs
  • Core-business match: extremely high — nuclear power of RMB 65.453 billion is 79.75% of revenue, and including nuclear-related services roughly 81.5% (2025 annual report basis); the company has no computing-power / IDC business whatsoever
  • 27 units / 26.2120 million kW in operation, 18 units / 20.6750 million kW under construction and approved-pending-construction (as at 2026-06-30, excluding the 4 units approved this time)
  • Valuation (computed by the sub-agent from RMB 9.06 × 20.568 billion shares and TTM net profit attributable to parent of RMB 8.232 billion): market cap RMB 186.35 billion, PE(TTM) 22.6x, PB(MRQ) 1.54x
• ⚠️ Financial red flags (three, all sub-agent verified): ① minority interests of RMB 116.459 billion ≈ parent equity of RMB 120.976 billion; of the RMB 17.971 billion of net profit in 2025 only 51.8% was attributable to the parent, so if new projects keep bringing in minority capital the parent-level figure will be continuously diluted; ② interest-bearing debt totalling about RMB 461.7 billion (short-term borrowings RMB 26.585 billion + non-current liabilities due within one year RMB 32.397 billion + long-term borrowings RMB 390.650 billion + bonds payable RMB 12.077 billion), against cash of only RMB 21.763 billion, with long-term borrowings up 28.4% within a year; ③ construction in progress of RMB 239.367 billion (end-2025, +11.3%), with operating cash flow of RMB 37.4 billion unable to cover capex, so free cash flow is negative (a structural feature and the root cause of continuously rising leverage).
• Industry position: leader — 26.21 million kW of installed capacity in operation makes it the largest nuclear operator in the A-share market, forming a duopoly with 中国广核.
• Technical sentiment: no limit-up on 7/31, closing at RMB 9.06 with turnover of RMB 1.002 billion. Position: +1.8% over the past 5 sessions, +4.1% over the past 20, only −0.8% from the 45-day high, but with an extremely narrow 45-day range (high 9.13 / low 8.64, a range of just 5.7%) — this is a stock that has flat-lined to the extreme: high in position but not crowded, the least high-looking high-position stock on the list.
• ⚠️ Counter-evidence one (volume): the 8 units approved this time are the fewest since 2022. Approvals by year: 6 units in 2019 → 4 in 2020 → 5 in 2021 → 10 in 2022 → 10 in 2023 → 11 in 2024 → 10 in 2025 → 8 in 2026. If the market's anchor is "normalised approvals of 10-plus units a year", then 8 is below the anchor and the expectation gap may point negative. Caixin's report was headlined "high-speed nuclear approval cadence continues", a positive reading; but the number itself is a four-year low — both readings will show up on the tape today, and the auction decides which wins.
• ⚠️ Counter-evidence two (timing mismatch — the most damaging one): the 4 units approved this time take roughly 5—6 years from approval to commercial operation, so power will not flow until around 2031—2032 at the earliest, contributing zero to the 2026—2030 P&L, while first adding construction in progress, capex and financing needs. In other words: the market today is being asked to pay for a 2031 tailwind by buying a company whose 2026Q1 profit is running at −34%.
• Final judgement: priority deep-dive, but it must be understood as "event-driven" rather than "fundamental improvement". It is today's only name that is "hard on news + named in an announcement + completely unpriced", and those three points are worth A-level attention; but its current-period earnings are clearly heading down, the denominator behind the 22.6x PE already contains that −34% Q1, and if tariff and overhaul pressure persists through the year the actual forward PE will be higher than 22.6x. Hard logic ≠ big elasticity: a large-cap central SOE + a 5—6 year delivery cycle + 8 units below the norm + falling current-period profit means it is more likely to be a 3—5% up candle than a limit-up.

5江苏神通002438SZ main board　Watch closely

• Related news: State Council approval of 8 units on 7/31.
• Tailwind logic: supply-chain direct order benefit, but with a very long lag. Whether the owner is CNNC, CGN or SPIC, nuclear-grade valves for all 8 units must be tendered to qualified domestic suppliers, and the company's products already cover Hualong One and Guohe One (precisely the two reactor types used here), high-temperature gas-cooled reactors and fast reactors — the full range. ⚠️ But nuclear island valve tenders usually start 1—2 years after construction begins and revenue recognition lags another 1—2 years → the revenue corresponding to this approval will most likely land in 2028—2030, with essentially no effect on the 2026—2027 statements. This is an "order visibility" benefit, not an "earnings" benefit. The company has no award announcement of any kind relating to these 8 units.
• Stage of the industry branch: day one, +2.34% on 7/31, no limit-up.
• ✅ Fundamental verification (fundamentals-analyst sub-agent — this is the healthiest set of statements in the nuclear branch):
  • Nuclear power really is in the core business and is the most profitable segment: 2025 nuclear industry revenue of RMB 811 million, 38.10% of the total (2024: RMB 743 million / 34.69%, +9.1% YoY), with a gross margin of 38.48%, the highest of any business segment; the rest is energy-saving services 21.23%, metallurgy 17.42%, energy 15.82%
  • 2026Q1: revenue RMB 521 million (−9.8% YoY), net profit attributable to parent RMB 85 million (−4.5%), ex-non-recurring RMB 80 million (−11.4%); but gross margin was 36.58%, +3.31pct YoY (a mix improvement)
  • 2025A: revenue RMB 2.129 billion, net profit attributable to parent RMB 285 million, gross margin 35.21%
  • Debt-to-asset ratio down from 49.08% (2022) to 38.84% (26Q1), with net interest-bearing debt of only about RMB 390 million — healthy leverage
  • Orders (media / field-research basis, no original announcement obtained as support): new orders of RMB 2.289 billion in 2025 (Shentong Nuclear Energy accounting for 40.72%), order backlog over RMB 3 billion; the high-end valve intelligent manufacturing project plans to release capacity progressively in H2 2026
  • Valuation (computed by the sub-agent from RMB 12.70 × 508 million shares and TTM net profit attributable to parent of RMB 281 million): market cap RMB 6.45 billion, PE(TTM) 23.0x, 24.6x ex-non-recurring, PB(MRQ) 1.69x
• ⚠️ Financial red flags: ① three straight years of zero revenue growth: RMB 2.133 billion (2023) → 2.143 billion (2024) → 2.129 billion (2025), with 2026Q1 still −9.8% — nuclear order growth has not yet converted into revenue, and this is its central contradiction; ② operating cash flow is extremely seasonal (87% of the RMB 420 million for full-year 2025 was concentrated in Q4; 26Q1 was −RMB 12 million); ③ accounts receivable of RMB 1.292 billion versus quarterly revenue of RMB 521 million, with 110—123 days of turnover; ④ customers are highly concentrated in CNNC, CGN and SPIC, limiting bargaining power.
• Industry position: over 90% market share in nuclear-grade butterfly valves and ball valves (media / research basis) — the niche leadership claim holds. ⚠️ But this figure conflicts with 中核科技 (CNNC Sufa Technology, 000777)'s claim of "53% share of nuclear-grade butterfly valves"; the two may use different statistical bases (nuclear class 2/3 vs all nuclear classes), and this report flags the conflict but does not score on it. Other competitors: 纽威股份 (Neway Valve, 603699) and Dalian High Pressure Valve.
• Technical sentiment: ⚠️ turnover of only RMB 85 million on 7/31, the worst liquidity on this list; position: −23.0% over the past 45 sessions, −25.6% from the 45-day high (low position).
• Final judgement: watch closely. A low position + a genuine niche leader + nuclear power at 38.1% of core revenue with the highest margin + a debt ratio of only 38.84% + a 23x PE make it the only name in the nuclear branch whose statements can carry the story. Two hard flaws: ① turnover of only RMB 85 million means that being easy to push up also means extremely thin follow-through; ② revenue has been flat for three years and Q1 is still negative, so the lag between order visibility and earnings delivery (2028—2030) is this stock's entire contradiction.

6中国广核003816SZ main board　Watch closely

• Related news: State Council approval of 8 units on 7/31, of which the 2 units of Guangdong Taipingling Nuclear Power Phase III (Units 5, 6) belong to the CGN system.
• Tailwind logic: direct order benefit (for an operator = installed capacity targets locked in), the same nature as 中国核电, only with half as many units (2 vs 4).
• Stage of the industry branch: day one, closing at RMB 4.13 −0.72% on 7/31 with turnover of RMB 425 million, no limit-up.
• Fundamental verification: not obtained this edition — the fundamentals-analyst sub-agent's verification scope this time covered 中国核电, 江苏神通, 太极实业 and 云赛智联, and did not cover 中国广核, so this report makes no statement about its financials, valuation or orders. All that is known are qualitative facts: it is one of the two domestic nuclear operating oligopolists (dual-listed in A and H shares, Hong Kong ticker 1816.HK).
• Technical sentiment: at RMB 4.13 it is the lowest-priced name in the nuclear branch and the one with the poorest per-unit volatility elasticity; turnover RMB 425 million.
• Final judgement: watch closely. Its problem is not the logic but the trading characteristics — for a large-cap central SOE just over RMB 4, even if the news delivers, both the upside and the recognisability are clearly weaker than 中国核电's. And it only got 2 units this time, half of 中国核电's.

7用友网络600588SH main board　Watch only (⚠️ listed as "priority deep-dive" No. 1 in the first draft; cut two levels after verification)

This is the biggest change from the first draft and the one most worth two minutes of the reader's time. The first draft put 用友网络 at No. 1 on the overall list (79 pts) on the grounds that "news + capital + a follow-up catalyst date are all in place". After fundamentals-analyst verification, the score fell to 48 pts, the rank to No. 11 and the label from "priority deep-dive" to "watch only". All the supporting evidence follows.

• Related news: on 7/31 the DeepSeek-V4-Flash official-release API entered public beta with a large jump in Agent capability, and V4-Pro is expected in early August (ITHome⁷); on the same day the NDRC announced it would accelerate the legislative process for the AI Law (Securities Times¹⁰).
• Tailwind logic: industry trend benefit, not a direct order. The path runs "large-model Agent capability improves → enterprise management software embeds Agents into ERP/finance/HR modules → higher ticket size and subscription penetration". The chain has to pass three gates — productisation → sales → revenue recognition — so visible revenue delivery within 2026 is close to impossible.
• Stage of the industry branch: between kickoff and fermentation. 用友 is the highest-profile, largest-cap main-board name on this chain; it hit limit-up on 7/31 with RMB 185 million of dragon-tiger net buying — the capital flow really is real, and the first draft was not wrong about that.
• ⚠️ Fundamental verification (sub-agent verified; every item below is a reason for the downgrade):
  • Three consecutive years of losses, cumulatively RMB 4.417 billion (attributable to parent) / RMB 4.739 billion (ex-non-recurring): 2023 net profit attributable to parent −RMB 967 million, 2024 −RMB 2.061 billion, 2025 −RMB 1.389 billion, 2026Q1 −RMB 723 million (ex-non-recurring −RMB 733 million)
  • 2026 interim guidance (announced 2026-07-15): revenue RMB 3.65—3.70 billion, only +1.9%~+3.3% YoY; still a loss of RMB 800—930 million attributable to parent, ex-non-recurring loss RMB 818—948 million. Items the company itself cites include amortisation of capitalised R&D intangibles of about RMB 650 million and severance compensation of about RMB 170 million (+RMB 40 million YoY)
  • The revenue mix is degrading (the most critical item): high-margin (93.3%) "product licence" revenue has shrunk over two years from RMB 3.309 billion to RMB 2.341 billion (−29.3%), costing about RMB 900 million of gross profit on that line alone; low-margin (33.5%) "technical services and training" has stepped in to hold up scale. Product strength is degrading and the company is trading manpower services for revenue.
  • Net assets shrank from RMB 11.803 billion (end-2023) to RMB 8.506 billion (2026Q1), down 27.9% in two years and three months; ROE has been negative double digits for three straight years (−18.73% in 2025)
  • Debt-servicing pressure: short-term borrowings of RMB 3.969 billion + non-current liabilities due within one year of RMB 909 million = RMB 4.878 billion, against cash of only RMB 2.497 billion; current ratio 0.66 (it was 1.16 in 2023)
  • Valuation: PE is meaningless given the losses; PB of 5.49x is the highest among peers (Glodon 2.56x, Baosight Software 4.65x, Sangfor 4.75x), while within its net assets goodwill of RMB 1.496 billion + intangibles of RMB 3.932 billion = RMB 5.428 billion, or 63.8% of the RMB 8.506 billion of net assets. ⚠️ A basis conflict readers must judge for themselves: dividing the market cap of RMB 37.86 billion by parent net assets of RMB 8.506 billion back-solves to a PB of 4.45x, inconsistent with the 5.49x the sub-agent took from East Money's stock_value_em (the difference may stem from the total-share-count / minority-interest basis). This report discloses both figures side by side and makes no choice between them; either way the conclusion that "over 60% of net assets is goodwill and intangibles" does not change — of every RMB 1 of net assets, RMB 0.64 is goodwill and (largely R&D-capitalisation-derived) intangibles; PS of 4.09x is also the highest among peers
• ⚠️ The most counter-intuitive piece of evidence: the sub-agent ran keyword counts over the full 2025 annual report (208,000 characters) — "computing power" appears 0 times, "GPU" 0 times, "intelligent computing" 0 times; "intelligent agent" 10 times, "Agent" 5 times, "large model" 13 times, a fair share of which appear in the terminology glossary at the front of the report; there is no standalone AI revenue line item in the segment disclosures. In other words: 用友 really is building enterprise-service large models such as YonBIP/YonGPT (it genuinely benefits at the industry-trend level), but as of the latest annual report and interim guidance there is no verifiable data showing AI has brought incremental revenue — for now AI is a cost item for 用友 (RMB 650 million of amortisation per half year), not a revenue item.
• ⚠️ Governance red flags: controlling shareholder Beijing Yonyou Technology holds 26.96%, with a cumulative pledge of 49.54% of its holding; four successive pledge / supplementary pledge announcements on 16 June, 25 June, 27 June and 3 July 2026 — supplementary pledges are usually triggered by share price falls and are a pressure signal. In addition an H-share listing is in progress (filed 2025-06-27, hearing materials updated a second time on 2026-07-15), and if it succeeds there is dilution risk for A-share holders.
• Industry position: domestic ERP leader (in the first tier alongside Kingdee) — the leadership is real; the profitability is not.
• Technical sentiment: limit-up on 7/31, dragon-tiger net buying RMB 185 million, turnover RMB 1.036 billion, market cap RMB 37.86 billion. ⚠️ Elevated position: +23.5% over the past 5 sessions, +25.6% over the past 10, only −2.9% from the 45-day high. It is a "first board that has already risen for 5 sessions", not a first board from a low base.
• Final judgement: watch only. The downgrade logic is simple: a company with revenue growth of only 1.9%~3.3%, three straight years of losses totalling RMB 4.4 billion, a current ratio of 0.66 and a controlling shareholder that has pledged nearly half its holding with successive supplementary pledges — and whose annual report does not contain the words "computing power" even once — can be the best sentiment vehicle in the AI applications sector, but it should not be the top recommendation in an evidence-driven list. The first draft mistook the capital flow (dragon-tiger net buying of RMB 185 million) for fundamentals, and that is corrected here.

8智度股份000676SZ main board　Watch only

• Related news: as above, AI applications / digital marketing.
• ⚠️ Tailwind logic graded: pure concept mapping — the basis is the sub-agent's keyword count over the full 2025 annual report (189,800 characters): "Web3" appears 0 times, "crypto" 0 times, "stablecoin/NFT/GPU/intelligent computing" 0 times each, "computing power" only 2 times, "Agent" 0 times; "blockchain" 56 times and "metaverse" 29 times, but all in qualitative narrative, with no blockchain / metaverse revenue line item in the segment disclosures. The Web3 label the market has attached to it has no counterpart in the latest annual report.
  • The only genuine crypto asset exposure (original wording of the annual report's accounting policy): the Bitcoin arising from cloud computing-power services purchased by subsidiary Zhidu Hong Kong from BitMaintech, with an opening carrying original value of RMB 55.0882 million, 1.23% of the RMB 4.469 billion of net assets, booked as an "intangible asset with an indefinite useful life" (not amortised, only tested for impairment, and generating no profit when the coin price rises).
  • The most persuasive counter-evidence: the latest investor relations activity record of 2026-07-24 (9 institutions attending, hosted by the deputy general manager and board secretary) contains 4 questions and 2,101 characters in total, and not once does it contain the words "AI", "computing power", "Web3", "blockchain" or "metaverse" — the questions were about cash flow, the balance sheet, customer mix and shareholder returns. The company itself does not mention a single one of these labels in its latest institutional briefing.
  • The company's own account of its share price (same record): "as public opinion settled and the AI media sector improved, combined with the company's positive earnings pre-announcement, the share price rebounded to a degree." — the company attributes the rebound to the sector, not to delivery in its own AI business.
• ✅ Fundamental verification (sub-agent verified — note: "the company is good" and "the theme is hollow" are two separate things here and must be stated separately):
  • Its balance sheet is the cleanest of the four main-board AI application stocks: debt-to-asset ratio 13.56%, current ratio 5.44, net cash position (interest-bearing debt of only RMB 361 million against cash and equivalents of RMB 1.484 billion at end-2025)
  • 2026 interim guidance: net profit attributable to parent RMB 135—145 million (+63.69%+75.81%), ex-non-recurring RMB 78—87.5 million (+10.06%+23.47%) — ⚠️ but the parent-level growth includes a one-off investment gain of about RMB 67 million from transferring a 22.4128% stake in Shanghai Yiyan, 48% of the RMB 140 million guidance midpoint. On a genuine operating basis (ex-non-recurring) growth is +10%~+23%, steady but not spectacular. It is the only one of the four whose core business is solidly profitable and growing in 2026H1.
  • ⚠️ But over a longer horizon: revenue is +40.6% over three years (RMB 2.929 billion→4.119 billion) while net profit attributable to parent is −55.9% over three years (RMB 322 million→142 million), with gross margin 19.97%→16.44% — the textbook case of "trading low-margin business for scale"; ROE 3.37%
  • The real profit engine has nothing to do with AI: "internet media services" generated RMB 1.440 billion of revenue at a 34.26% gross margin, contributing 74.5% of the company's gross profit — per the company's IR record this business is "mainly aimed at North America, with major customers including Yahoo, Google and other globally known enterprises", essentially overseas search traffic monetisation; while digital marketing, at 60.24% of revenue, carries a gross margin of only 3.52%
  • Valuation: PE(TTM) 65.9x, ex-non-recurring PE 78.4x, PB 2.20x
• ⚠️ Financial red flags: ① SZSE public criticism on 2026-05-22 plus a regulatory letter issued to the then deputy general manager (Company Department Regulatory Letter [2026] No. 74, based on Guangdong CSRC administrative regulatory measure decision [2026] No. 71), arising from the company's announcement of 2025-04-29 on the retrospective correction of prior accounting errors and incomplete disclosure of a related-party transaction in the 2020 disposal of a subsidiary; ② operating cash flow has long diverged from profit (2023 −RMB 259 million / 2025 −RMB 94 million, narrowing to −RMB 14.7 million in 2026Q1, a 91.56% YoY improvement, which needs the interim report to confirm it persists); ③ accounts receivable of RMB 1.172 billion, up 66.7% in two years, with 2025 credit impairment of −RMB 118 million (consuming 83% of that year's RMB 142 million of net profit attributable to parent); ④ there are unrecovered losses and no cash dividends in recent years.
• Technical sentiment (this is the direct reason for the downgrade): limit-up on 7/31, turnover RMB 1.020 billion, market cap RMB 9.44 billion. ⚠️ 0.0% from the 45-day high — it is the 45-day high itself; +40.5% over the past 20 sessions and +30.0% over the past 45. This is the most extended and most crowded name on the whole list.
• Final judgement: watch only. The call on this stock needs two sentences to be accurate: ① at the company level, it has the cleanest statements of this batch of main-board AI application stocks and is the only one whose core business is growing, and a PB of 2.20x is not expensive — what is being passed on is the price and the theme, not the company; ② at the theme level, its profit comes from North American Yahoo/Google traffic monetisation with no verifiable connection to AI or Web3, while 65.9x TTM PE / 78.4x ex-non-recurring PE has already fully priced the "AI media" imagination. Add "the most extended position on the list + a 45-day high + public criticism as recently as 2026-05", and the risk/reward of chasing it today is among the worst on the list.

9润建股份002929SZ main board　Watch only (⚠️ downgraded from "watch closely")

What makes this stock special: it is the only company among all the names in this report that can produce a standalone financial line item for "computing-power revenue" — and it is also the one with the highest balance sheet risk. These two facts must be read together.

• Related news: the NDRC's estimate of RMB 4 trillion of new direct investment in the computing-power network during the 15th Five-Year Plan; the State Council executive meeting advancing the "Six Networks".
• ✅ Tailwind logic graded: direct order benefit (the only one on the list) — but small in scale and thin in margin. The positive evidence the sub-agent found is the most substantive of the four:
  • The "computing-power network business" is a separately disclosed revenue segment: RMB 308 million in 2023 → RMB 529 million in 2024 → RMB 810 million in 2025, 7.83% of total revenue, up 163% over three years (CAGR 62%)
  • There are physical assets behind it: the annual report states that RMB 1.09 billion was raised via a public convertible bond issue to build the Wuxiang Yungu intelligent computing centre, "the highest-grade and largest-scale intelligent computing centre in Guangxi to date"; it has launched the "Xingsuan Yunchi" product with billing models such as tokens and card-hours, moving to a "token factory" model
  • "Computing power" appears 82 times in the 2025 annual report, "artificial intelligence" 50 times, "large model" 30 times and "intelligent computing" 16 times — the highest density of AI/computing-power narrative among all names in this report, and the only one with matching segment revenue
  • Order backlog of about RMB 22.2 billion as at end-2025
• ⚠️ But the negative evidence is equally hard, and overwhelming:
  • The computing-power gross margin collapsed from 16.1% (2023) to 6.2% (2024) and 6.3% (2025). RMB 810 million of revenue produced only RMB 51.3 million of gross profit. This is not an Nvidia-style computing business; it is a thin-margin sub-leasing / managed-operations business.
  • Computing power is only 7.83% of revenue and 4.05% of gross profit — even if computing power doubled again, it could not change the company's P&L. Yet a market cap of RMB 13.74 billion is paying a premium for the "token factory" narrative.
  • The 92% that is the core business is collapsing across the board: communications network gross margin 19.9%→16.9%, energy network 14.6%→10.5%, digital network 11.6%→7.1%. All four segments' gross margins fell together, which says this is an industry-wide price war / worsening payment terms, not a structural upgrade.
  • An earnings cliff: revenue is +17.3% over three years (RMB 8.826 billion→10.349 billion), while net profit attributable to parent is −91.4% over three years (RMB 439 million→37.9 million), ROE has collapsed from 8.94% to 0.60% and gross margin from 17.92% to 12.23%; 2026Q1 revenue −23.5% YoY (RMB 1.820 billion vs 2.379 billion) and net profit attributable to parent −81.8%
  • ⚠️ The annual report uses a non-standard basis: "net profit excluding the impact of share-based payment was RMB 280 million, up 13.66% YoY" — while statutory net profit attributable to parent was only RMB 37.9 million, a gap of RMB 240 million. Share-based payment is a real cost to shareholders; please use the statutory basis.
  • 2026 interim guidance: not disclosed (the Shenzhen main board only mandates guidance for losses, turnarounds or YoY changes beyond ±50%). On that rule the sub-agent back-solved that 2026H1 net profit attributable to parent may fall in the RMB 20—59 million range, but explicitly labelled this an [inference, not a fact], unconfirmed by the company and subject to the late-August interim report. This report adopts the same labelling and does not treat it as a conclusion.
• 🚩 Financial red flags (the highest balance sheet risk on the list):
  1. Accounts receivable of RMB 5.928 billion, with 294 days of turnover (199.7 days in 2023), close to 3.3x the RMB 1.820 billion of single-quarter revenue in 2026Q1 and 93.9% of the RMB 6.316 billion of net assets
  2. Leverage keeps climbing: debt-to-asset ratio 59.76% (2023) → 64.71% → 69.37% → 70.96% (2026Q1); total assets ballooned from RMB 14.853 billion to RMB 21.747 billion (+46.4%), while net assets rose only from RMB 5.977 billion to RMB 6.316 billion (+5.7%) — the growth was entirely debt-driven
  3. Short-term solvency: short-term borrowings of RMB 4.725 billion + non-current liabilities due within one year of RMB 223 million = RMB 4.948 billion, against cash of RMB 3.084 billion, giving a current ratio of 1.02 with almost no buffer
  4. Extreme operating cash flow: 2026Q1 −RMB 755 million, 2025H1 −RMB 2.527 billion, 2025Q3 −RMB 2.016 billion, with the full year turning positive to +RMB 288 million only on concentrated Q4 collections
  5. Two years of impairments totalling RMB 485 million, 12x the RMB 37.9 million of 2025 net profit attributable to parent
  6. The two sides of the buyback: the 2026-07-15 plan is to repurchase RMB 150—300 million and by 7/31 it had reached 1% of total share capital — fast execution; but the funding includes self-raised money plus a dedicated share-buyback loan, while the debt ratio is already 70.96% and the current ratio 1.02 — buying back shares with borrowed money adds pressure to an already stretched balance sheet; and the buyback is for equity incentives / employee ownership rather than cancellation, so it does not enhance EPS
• Valuation: TTM net profit attributable to parent −RMB 19 million (turned negative), trailing PE 362.3x, ex-non-recurring PE about 2457x; PB 2.17x, PS 1.40x.
• Technical sentiment: limit-up on 7/31, turnover RMB 745 million, market cap RMB 13.74 billion. ⚠️ The most oversold on the list: −30.4% over the past 20 sessions, −34.4% over the past 45, −34.7% from the 45-day high (45-day high 73.48 → current price 47.97).
• Final judgement: watch only (downgraded from "watch closely"). It deserves its own section because it proves one thing: in this computing-power cycle, "having real computing-power revenue" and "being worth buying" are two completely different things. 润建 is the only company in the market that can put computing-power revenue into its segment reporting, but that RMB 810 million of revenue earns only RMB 51.3 million of gross profit, while the core business is losing profit at a rate of over −40% a year, accounts receivable are 93.9% of net assets, and it is still borrowing money to buy back shares. Its rally is more "oversold bounce" than "logic-driven" — on a stock down 34% in 45 sessions, any policy news gets amplified into a reason to bounce.

10利欧股份002131SZ main board　Pass (⚠️ the highest governance risk on the list; ranked No. 6 / "watch closely" in the first draft)

The first draft ranked it No. 6 (68 pts) with a "watch closely" label, on the grounds of "the lowest position + the largest turnover + dragon-tiger net buying". After verification it falls to 30 pts, No. 21 and a Pass label. The merits of its trading structure are real, but they mask what follows.

• Related news: as above, AI applications / digital marketing.
• ⚠️ The tailwind logic has to be regraded as "pure concept mapping" — on the basis of the company's own answers. In the 2026-05-15 results briefing record (the sub-agent extracted the original announcement text), investors directly pressed on whether the AI business could be disclosed separately, what share of revenue it actually accounts for, and whether it was "just a concept story"; the company answered:
  • "The company's AI capability is mainly embedded indirectly into its digital marketing business... the related revenue is still difficult to break out clearly"
  • "This type of business currently remains a small share of the company's overall revenue... with limited near-term impact on operating performance"
  • In the same record the company also explicitly denied three market rumours: "the company does not hold equity in Moore Threads", "the company does not hold equity in Unitree or deepseek", and "it is not yet involved in optical modules, optical fibre, memory or similar fields" When a company itself says "revenue cannot be broken out, the share remains small, the impact is limited", this should not be counted as an industry trend benefit.
• ⚠️ Fundamental verification (sub-agent verified):
  • Three-year cumulative ex-non-recurring profit of only RMB 256 million against three-year cumulative revenue of RMB 61.692 billion — an ex-non-recurring net margin of 0.41%
  • Reported profit is essentially driven by stock trading, unrelated to the core business: fair value change gains/losses of +RMB 1.255 billion in 2023 / −RMB 873 million in 2024 / −RMB 169 million in 2025; in 2023 net profit attributable to parent was RMB 1.966 billion while ex-non-recurring was only RMB 55 million (a gap of RMB 1.910 billion)
  • 2026 interim guidance: net profit attributable to parent RMB 110—150 million (−68.63%~−77% YoY), ex-non-recurring loss of RMB 90—130 million (a first). The company disclosed: Li Auto shares' fair value change plus disposal totalling −RMB 357 million (impact on parent net profit −RMB 268 million), financial assets such as Xinfengguang contributing about +RMB 500 million, and the machinery manufacturing segment loss-making in the period. Strip out the stock trading and the core business is loss-making.
  • The truth about the business mix: 76% of revenue is media agency work at a gross margin of just 3.86% (traffic buying and resale, a working-capital pass-through business); the real gross-profit pillar is machinery manufacturing (water pumps) — RMB 4.141 billion of revenue contributing RMB 1.050 billion of gross profit, 58.3% of the company total. 利欧 is essentially still a water pump company plus an ultra-low-margin advertising traffic agency. And the 2026H1 guidance explicitly says: the sole gross-profit pillar (machinery manufacturing) was loss-making in the period.
  • Valuation: TTM net profit attributable to parent −RMB 43 million (turned negative), trailing PE 935.2x; PS of 1.55x is the most expensive among digital marketing peers (蓝色光标 0.70x, 省广集团 SPGC 0.53x) — using PS to value a pass-through business with a 3.86% gross margin does not stand up in the first place.
• 🚩 Governance red flags (the decisive reason for the Pass):
  1. SZSE public censure (2026-07-03, Shenzhen Listing [2026] No. 938): 2026-01-31 guidance of RMB 190—250 million of 2025 net profit attributable to parent → revised on 2026-04-24 to RMB 30—45 million → actual in the annual report RMB 33.7444 million. The SZSE found "material discrepancies" and publicly censured the company, chairman and general manager Wang Xiangrong and CFO Yang Hao simultaneously, entering it in the securities and futures market integrity file. This is a direct regulatory ruling on the credibility of the financial reports, not a market rumour.
  2. The sequence of the stake sale and the guidance cut: from 2025-10-16 it made successive disclosures on the progress of reducing repurchased shares, announced on 2026-04-14 that the reduction had reached 1%, completed it on 4-15 — and 9 days later (4/24) issued the large downward guidance revision. The company replied that there was "no connection" between the two; the connection is unproven, but the sequence is a matter of public record.
  3. A surge in shareholder accounts: 547,500 on 2025-12-31 → 978,400 on 2026-03-31 (+78.70%). The chips were extremely widely dispersed to retail investors at the 2026Q1 high (a peak of RMB 10.40 in the period), after which the share price fell to RMB 4.66.
  4. Accounts receivable of RMB 5.824 billion, 45.4% of net assets; 2026Q1 operating cash flow −RMB 600 million; a RMB 145 million impairment on the long-term equity investment in *ST Chuangxing in 2025 (the main reason for the guidance cut).
  5. A ten-year-old unsettled performance compensation: in the 2016 acquisition of Zhiqu Advertising the original shareholders committed net profits of RMB 58 million / 75.40 million / 98.02 million for 2016—2018, versus actuals of only RMB 32.24 million / 36.69 million / 7.70 million — all three years missed; arbitration was filed in 2019, the compensation obligors have gone into bankruptcy liquidation, and the compensation remains unfulfilled to this day (2026-08).
• Technical sentiment: limit-up on 7/31, dragon-tiger net buying RMB 116 million, turnover RMB 3.762 billion. Its position really is low: −22.1% over the past 45 sessions, −23.6% from the 45-day high.
• Final judgement: Pass. The trading structure (low position + big turnover + net buying) is its only merit, and in an evidence-driven list that merit cannot outweigh the three items "chairman and CFO publicly censured by the exchange + first-ever ex-non-recurring loss in interim guidance + the company itself says AI revenue cannot be broken out". Ranking it No. 6 in the first draft was wrong, and that is corrected here.

Appendix: 云赛智联(600602) 　 SH main board 　Already downgraded from "watch closely" to "watch only" (full analysis retained)

This entry was No. 7 in Section 5 of the first draft. After fundamentals-analyst sub-agent verification its composite score fell from 65 to 48 and it dropped out of the top ten (now No. 12 on the overall list), but because it is the most typical sample of "computing-power network narrative diverging from financials", the analysis is retained in full so readers can judge the quality of the whole computing-power chain.

• Related news: the NDRC's estimate on 7/31 of RMB 4 trillion of new direct investment in the computing-power network during the 15th Five-Year Plan (Cailianshe¹³); the State Council executive meeting advancing the "Six Networks".
• Tailwind logic: industry trend benefit (real, but of questionable strength). As the IT platform under Shanghai INESA Group, it is the master integrator for the Shanghai Big Data Centre resource platform, operates the Songjiang intelligent computing centre (phase one) and holds a minority stake in Shanghai Intelligent Computing Technology Co., Ltd. The chain is real, not pure concept riding.
• Stage of the industry branch: already through one round of fermentation, not day one. The chain moved as a group on 7/31 (宏景科技 20cm, 美利云, 深桑达Ａ, 润建股份 and 云赛智联 all at limit-up).
• ⚠️ Fundamental verification (fundamentals-analyst sub-agent — this is the most narrative-contradicting finding of this edition and the entire reason for the downgrade):
  • The "cloud computing & big data" segment that carries computing power (covering IDC infrastructure, internet access, IaaS and computing-power business) saw 2025 revenue fall from RMB 2.998 billion to RMB 2.879 billion, −3.97% YoY, with its share dropping from 53.32% to 45.36%
  • Over the same period "industry solutions" (systems integration) revenue rose from RMB 2.381 billion to RMB 3.115 billion, +30.83%, but at a gross margin of only 9.95% — almost all of the company's 11.7% revenue growth in 2025 came from this low-margin business
  • This directly explains: gross margin 19.29% (2024) → 16.76% (2025) → 15.75% (26Q1); three-year revenue RMB 5.264 billion → 6.348 billion (+20.6%) while net profit attributable to parent went RMB 193 million → 194 million (+0.5%); ROE locked around 4% year after year (4.02% in 2025)
  • 2026Q1: revenue RMB 1.449 billion (+2.76%), net profit attributable to parent RMB 45 million (+5.77%), ex-non-recurring RMB 43 million (+10.09%) — positive YoY, but tiny in absolute terms
  • Valuation (computed by the sub-agent from RMB 17.62 × 1.368 billion shares and TTM net profit attributable to parent of RMB 196 million): market cap RMB 24.10 billion, PE(TTM) 123.0x, ex-non-recurring PE 147.9x, PB(MRQ) 4.90x
  • The absolute amount and share of intelligent computing / IDC revenue: not separately disclosed by the company, cannot be confirmed at this time; the stake percentage in Shanghai Intelligent Computing Technology and whether it is consolidated: to be verified
• ⚠️ Financial red flags: ① three years of revenue growth without profit growth; ② gross margin down four times running; ③ operating cash flow is extremely seasonal and persistently negative in the first three quarters (2025 Q1 −4.37 / Q2 +1.44 / Q3 −0.30 / Q4 +9.11, i.e. essentially all of the RMB 589 million full-year figure came in Q4; 26Q1 −RMB 290 million); ④ inventory of RMB 1.618 billion > single-quarter revenue of RMB 1.449 billion, with inventory growing (+32%) far faster than revenue (+11.7%); ⑤ cash fell from RMB 3.934 billion at end-2023 to RMB 1.433 billion in 26Q1 (−63.6%) while the debt-to-asset ratio went 39.06% → 42.31%, and where the money went remains to be verified; ⑥ on the positive side: it has almost no interest-bearing debt (about RMB 420 million) and remains in a net cash position.
• Industry position: not a niche leader. It is a regional (Shanghai state-owned) systems integrator plus IDC operator with resource advantages in local Shanghai government / SOE informatisation, but no scale advantage nationally. Competitors: on the IDC side GDS Holdings, 光环新网 (Sinnet, 300383), 数据港 (Shanghai Athub, 603881), 润泽科技 (Runze Technology, 300442); on the integration side 太极股份 (Taiji Computer, 002368), 东华软件 (DHC Software, 002065).
• Technical sentiment: limit-up on 7/31, turnover RMB 746 million, free-float market cap RMB 18.93 billion; +1.9% over the past 5 sessions, −4.2% over the past 10 (it had been falling right up to the limit-up), −12.9% from the 45-day high.
• Final judgement: watch only. ⚠️ The reason for the downgrade is not "the company is bad" but "the narrative diverges from the financials": the market treats it as the main-board flagship of the computing-power network, while its computing-power segment's 2025 revenue actually shrank and its share is falling, with growth coming from integration work at a sub-10% gross margin. A 123x PE (148x ex-non-recurring) against a company with three years of zero profit growth and a 4% ROE can only be supported by a narrative of asset revaluation or asset injection; it cannot be explained by current earnings power. The next validation point is the 2026 interim report (expected late August) — whether the "cloud computing & big data" segment can return to growth is this stock's life-or-death line.

---

6. Pass list

A note on the "non-main-board" category in the Pass list: 昆仑万维, 蓝色光标, 中文在线 and 易点天下 were among the top names by dragon-tiger net buying on 7/31 and are the genuine leaders of the AI applications branch. This report follows the skill's ironclad rule (main board first) and keeps them out of the recommended list, but that does not mean they are unimportant — quite the opposite: to judge the strength of the AI applications branch today, look at them first, then decide what to do about the main-board names. See the sector signals in Section 8.

---

7. Intra-branch rankings

7.1 AI applications / Agents (today's strongest branch)

• Hardest single stock: 泛微网络(603039) — the only name within the main board on the AI applications chain that combines "real profit (2026Q1 net profit attributable to parent RMB 62.1 million, a large YoY improvement) + low debt (37.89%) + an existing consecutive-limit-up ladder (2 boards)".
• Best single stock for main-board trading: 泛微网络(603039) — after verification and QA review it is the only main-board name left that can stay in the recommended zone in this branch (PE 35.4x, ex-non-recurring 93% of net profit attributable to parent, debt ratio 37.89%, already on 2 boards). ⚠️ 税友股份(603171), with a PE(TTM) of 130.0x above the 123x on which this report rejected 云赛智联, has also been downgraded to "watch only".
• ⚠️ The biggest correction in this branch relative to the first draft: the first draft listed 用友网络 as the main-board flagship with "priority deep-dive" and 利欧股份 as "watch closely"; after verification the two are cut to watch only and Pass respectively. Their capital flows are real (dragon-tiger net buying of RMB 185 million and RMB 116 million), their fundamentals are not — the first draft mistook the former for the latter.
• Who is following along: 浙文互联, 天娱数科, 三维通信, 三人行 — no verifiable causal relationship between their revenue and AI; pure sector co-movement.
• Who gets a Pass: 利欧股份 (governance red flags), 引力传媒 (debt ratio 90.71%), 掌阅科技 (sustained losses).

7.2 Nuclear power (today's only unpriced branch)

• Hardest single stock: 中国核电(601985) — the only name that is "a direct order benefit + 79.8% core-business match + named in an announcement + completely unpriced". But please understand it as an event-driven name, not a fundamentals-improvement name: its 2026Q1 net profit attributable to parent is −34.19%.
• Best single stock for main-board trading: 东方电气(600875) (better elasticity than the operators, ample turnover of RMB 1.283 billion, and 2026Q1 net profit attributable to parent +37.4%). ⚠️ But three qualifications must be remembered at the same time: ① its directness of benefit is lower than the operators' — project approval ≠ equipment award, and there is no award announcement relating to this approval; ② the share of nuclear power in its total revenue could not be obtained this edition, and it is first and foremost a diversified power equipment maker; ③ the +37.4% is on the parent basis while ex-non-recurring is only +11.5% YoY, and that 26pct gap says profit contains a sizeable share of non-recurring items whose composition this report did not obtain — if any one of these three is falsified, it should not rank No. 1.
• Who is following along: 应流股份, 中核科技 (no financial or turnover data was obtained this edition to support them; participation with missing data is not advised).
• ⚠️ A basis conflict that must be disclosed: 江苏神通's ">90% market share in nuclear-grade butterfly/ball valves" and 中核科技's "53% share of nuclear-grade butterfly valves" come from different sources with possibly different statistical scopes (nuclear class 2/3 vs all nuclear classes); they contradict each other, both are on a media/research basis with no annual-report support obtained, and this report does not score on them.
• ⚠️ Risks common to the whole branch: ① 8 units is below the norm of 10—11 per year in 2022—2025; ② not one company has an award announcement relating to this approval; ③ current-period earnings are falling at both the operator and the valve end (中国核电 2026Q1 net profit attributable to parent −34.19%, 江苏神通 −4.5%) — the sole exception is 东方电气 (net profit attributable to parent +37.4%), but its growth has no causal link to this approval (it happened before 7/31) and its ex-non-recurring growth is only +11.5%; ④ nuclear island valve tenders start 1—2 years after construction begins and revenue recognition lags another 1—2 years, so equipment-side revenue most likely lands in 2028—2030.

7.3 Computing-power network / the "Six Networks"

• Hardest single stock: not one name in this branch can be called "hard". The RMB 4 trillion is a third-party estimate; the only one with a real award record is 太极实业 (Huahong FAB9B, RMB 3.778 billion), but that is a semiconductor cleanroom rather than a computing centre, and its engineering general contracting gross margin is only 1.96% — order-winning ability is proven; money-making ability is not.
• ⚠️ There is no "best single stock for main-board trading" in this branch — after verification, all three main-board names, 云赛智联, 太极实业 and 润建股份, are "watch only", and no recommendation is set for this branch. 润建股份 may be the only company in the market that can put computing-power revenue into its segment reporting, but with a computing-power gross margin collapsed to 6.3%, net profit attributable to parent −91.4% over three years, receivables at 93.9% of net assets, and share buybacks funded by a dedicated loan, that is not a reason to participate.
• ⚠️ The core conclusion of this branch (the opposite of the first draft): 云赛智联 is downgraded from "watch closely" to "watch only" — it is the name most widely treated as the "main-board flagship of computing power" on this chain, yet its computing-power segment's 2025 revenue actually shrank, and a 123x PE can only be supported by an asset revaluation / injection narrative, not explained by current earnings power.
• Who gets a Pass: 美利云 (quarterly revenue below RMB 100 million), 深桑达Ａ (a deep downtrend).

---

8. Opening validation signals for today

8.1 Call auction signals (09:15—09:25)

• Nuclear power is today's only branch starting from zero, and the auction is its only evidence: focus on the auction gap-up magnitude for 中国核电(601985), 中国广核(003816) and 东方电气(600875).
  • A gap up of 2%—4% with expanding auction volume → the news is being priced normally and the branch holds;
  • A gap up of >6% → ⚠️ beware the "8 units below the norm" expectation gap being digested intraday; the probability of gapping up and fading rises materially;
  • A gap up of <1% or a flat open → the market has concluded that 8 units missed expectations, and this branch is falsified for the day; there is no need to wait further.
• For the AI applications branch, watch the two main-board names on 2 boards: the auction for 泛微网络(603039) and 税友股份(603171). If both gap up more than 3% and seal quickly → the branch enters its main advance; if they gap up and fade or the auction volume shrinks → yesterday was a broad sentiment rally rather than a main theme.
• ⚠️ Today's highest board is absent: 爱丽家居(603221) is halted, so the top board falls to 传智教育(003032) on 5 boards. Watch its auction — if 传智教育 is weak in the auction, market-wide consecutive-limit-up money has no compass today and first boards will be harder to sustain.
• One-word boards / sealing orders: the largest sealing order values on 7/31 were 蓝色光标(300058) RMB 389 million, 利通电子 (Litong Electronic, 603629) RMB 306 million, 利欧股份(002131) RMB 262 million, 宏景科技(301396) RMB 250 million and 昆仑万维(300418) RMB 179 million — watch today for whether they still show one-word boards or large sealing orders; this is the fastest indicator of whether the money is still in the game.

8.2 Sector signals (09:30—10:30)

• What confirms AI applications: at least one of the three sectors — software development / IT services / culture & media — produces 3 or more fast limit-ups, and 昆仑万维(300418) and 蓝色光标(300058), the two leaders, trade heavy volume without breaking their limits (they are not main board, but they are the flagships of this chain, and if the flagships do not follow, the main-board names have no source).
• What confirms nuclear power: the operators (中国核电/中国广核) and the equipment names (东方电气/上海电气/江苏神通) strengthen together. ⚠️ If only a small-cap name like 江苏神通 hits limit-up while 中国核电 does not move, it means hot money is trading the concept rather than capital endorsing the logic, and persistence will be poor.
• ⚠️ Divergence-return signal: on 7/31 there were 107 failed limit-ups versus 99 limit-ups. If failed limit-ups again exceed limit-ups before 10:00 today, follow-through is still insufficient and yesterday's up candle was only a repair, not a reversal.
• The reverse signal for semiconductors: semiconductors saw RMB 20.502 billion of net outflow on 7/31. If semiconductors surge again today (dragged up by Amazon's +15% in the US) but 德明利(001309) and 兆易创新(603986) break their limits again, it confirms that capital is using the rebound to exit and AI hardware is not participable.

8.3 Single-stock signals

• Follow-through in the top 5: if 泛微网络(603039) (the only main-board AI application stock this report recommends) and 税友股份(603171) (already downgraded, used only as a branch indicator) can hold near the top of the gainers list in the first 30 minutes on rising volume without fading, main-board money is handing off successfully. ⚠️ At the same time watch 用友网络(600588) and 利欧股份(002131) in the opposite direction — they had dragon-tiger net buying of RMB 185 million and RMB 116 million yesterday but their fundamentals have been falsified, so if they spike and fade today while 泛微/税友 strengthen, capital is rotating from theme stocks into earnings stocks, which is good for branch quality; if it is the other way round (they keep limiting up while 泛微/税友 lag), today is still purely sentiment-driven.
• Fast sealing / heavy volume without breaking: 泛微网络(603039) and 税友股份(603171) (both on 2 boards) — a second board turning into a third is the only hard marker of the AI applications branch upgrading from "broad rally" to "main theme". ⚠️ Note: 税友 has already been downgraded to "watch only" by this report on valuation (PE 130x), but it remains a valid indicator of branch strength — eligibility as a candidate and value as a signal are two different things.
• Large-order net inflow: watch whether the software development sector can sustain the +RMB 9.029 billion net inflow of 7/31. If net inflow turns negative, treat it as distribution even if the sector is green on the screen.
• Being blocked by a stronger name: ⚠️ 智度股份(000676) is already at a 45-day high (0.0% from the 45-day high, +40.5% over the past 20 sessions), so if the AI applications branch strengthens today while it spikes and fades instead, that is a textbook signal of capital rotation within the branch. ⚠️ The first draft originally wrote here that "capital will flow to the lower-positioned 利欧股份"; that judgement is void — after verification 利欧股份 was cut to Pass over governance red flags (chairman and CFO publicly censured by the SZSE), so a low position is not a reason to participate; if capital does rotate, this report believes it should look towards 泛微网络(603039) and 税友股份(603171), which have real profits.

8.4 Risk signals (any two of these appearing means the day's trend does not hold)

1. Gap up then dive: the index gaps up and turns negative before 10:00 (US stocks were mildly up overnight so a gap up is likely; the bigger the gap, the more dangerous).
2. A lone limit-up: only 1 nuclear name hits limit-up with no ladder following.
3. Back-row retreat: pure-concept back-row names such as 浙文互联, 天娱数科, 三维通信 and 引力传媒 break their limits first — when the back row breaks first, the branch has reached its divergence phase.
4. Flagships do not follow: 昆仑万维 and 蓝色光标 open lower or spike and fade.
5. Weak index: RMB 925.5 billion of PBOC open-market reverse repos mature today (RMB 1,842.0 billion this week); if tighter liquidity shows up in rates, growth stocks come under pressure.
6. Yesterday's strong theme fades: if the advertising & marketing sector (7 limit-ups on 7/31, the most of any sector) sees its limit-up count halve to 3 or fewer today, the main theme is falsified.
7. ⚠️ The halt effect: 爱丽家居 is halted and the SSE applied self-regulatory measures to 137 abnormal trading incidents this week → the regulatory stance is clearly on the tight side, and expectations for high-position consecutive-limit-up stocks should be lowered across the board today.
8. Unlock pressure: RMB 14.042 billion of shares unlock today in a single day (this week's peak), with 浙江荣泰's unlock equal to 43.95% of total share capital.

---

9. Final recommendation conclusions

9.1 The 5 main-board stocks most worth watching today

⚠️ This list has swapped out two names relative to the first draft. The first draft's No. 2, 用友网络(600588), and No. 4, 利欧股份(002131), were cut to "watch only" and "Pass" respectively after fundamentals-analyst verification. Subsequently, in the risk-auditor QA review, 税友股份(603171) — which had briefly been added to this list — was also removed, because back-solving its PE(TTM) gives 130.0x, higher than the 123x for which this report judged 云赛智联 "watch only", which is a double standard (see the QA note at the end). The current list is: 东方电气, 中国核电, 泛微网络, 江苏神通, 中国广核. The basis for every swap is written out in Section 5 and the QA note at the end; please read them together.

⚠️ One sentence about these 5: their natures are completely different and they should not be understood as a portfolio. Nos. 1, 2, 4 and 5 are "unpriced events" (nuclear power, taking 4 of the 5 slots); only No. 3, 泛微网络, is "a validated trend that needs a handoff" (AI applications). The former bets on first-day pricing today, the latter bets on no fade on day two — the ways these two fail are completely different. ⚠️ Also note: the fundamentals of No. 5, 中国广核, are entirely unverified by this report; it made the list purely on "clear ownership + unpriced", so discount it for information completeness. And note one counter-intuitive result: 3 of these 5 barely moved last Friday (东方电气 +0.76%, 中国核电 −0.11%, 江苏神通 +2.34%), while the names that hit limit-up last Friday and were bought hardest on the dragon-tiger list are all absent from this list.

9.2 Today's 3 strongest branches

⚠️ Note on the ranking change: the first draft ranked AI applications 1st and nuclear power 2nd. The two were swapped after verification, not because the nuclear news got stronger, but because the degree to which the fundamentals of the main-board AI application names were falsified exceeded expectations (see Section 5, items ⑦⑧⑨⑩), while the nuclear branch turned up 东方电气 with current-period profit +37.4%. This is an evidence-driven ranking correction, not a wobble of opinion.

9.3 Directions not worth chasing today, and why

1. Memory / AI hardware (德明利, 兆易创新, 生益科技, 深南电路, 风华高科) — four pieces of counter-evidence hold at once: ① on 7/31 the semiconductor sector saw RMB 20.502 billion of net outflow and components RMB 7.289 billion, the two largest outflows in the market; ② across the day 107 failed limit-ups > 99 limit-ups, and the ones that failed were precisely the core names on this chain (德明利 sealed at the open then turned negative −1.06%, 兆易 broke its limit, 生益 broke its 17 times, 深南 13 times); ③ US memory stocks dived intraday on 7/31 (SanDisk +10%→−6%, Micron +6%→−4%); ④ TrendForce: Q3 DRAM contract prices +13%~18% QoQ and NAND +10%~15%, versus +58%~63% / +70%~75% in Q2 — the slope of price increases is converging sharply. 兆易创新's RMB 1—2 billion buyback is a genuine positive, but what it props up is a valuation floor, not the trend.

2. High-position consecutive-limit-up stocks (传智教育 on 5 boards and anything on 3 boards or more) — the regulatory signal is unusually clear today: 爱丽家居, on 9 boards, is halted for verification, and the SSE applied self-regulatory measures to 137 abnormal trading incidents this week and placed it under "key monitoring". The day the top board is administratively removed is not a day for trading consecutive limit-ups.

3. The pure-concept back row of marketing & media (浙文互联, 天娱数科, 三维通信, 引力传媒, 三人行) — the financials have already given the answer: 引力传媒's debt ratio is 90.71%, 浙文互联 lost money in 2026Q1 with a net margin around 1.0%, and 三维通信 has lost money in two consecutive years. There is no verifiable causal relationship between these companies' revenue and AI; they rise on sector beta and they fall the same way.

4. Oil extraction / tanker shipping (中国石油, 中国海油, 招商轮船) — a reopening of the Strait of Hormuz points negative; but because Iran called Trump's statement "pure rumour", the direction itself is uncertain, making this a position where neither long nor short should participate.

5. Airlines (chased because of the fuel surcharge cut) — the fuel surcharge cut (to RMB 70/person for routes over 800 km from 8/5) is a revenue reduction for airlines, not a positive; cost pass-through from falling oil prices takes 1—2 quarters. Buying airlines on this as a positive gets the direction backwards.

6. The two computing-power names with the widest narrative-versus-financials divergence — 云赛智联(600602) and 太极实业(600667) — this item is where this report changed most from the first draft after fundamentals-analyst verification: 云赛智联 is treated as the main-board flagship of computing power, yet the "cloud computing & big data" segment that carries computing power saw 2025 revenue −3.97% with its share falling from 53.3% to 45.4%, and the company's growth actually came from systems integration at a 9.95% gross margin; three-year revenue +20.6% while net profit attributable to parent rose +0.5%, with a 4.02% ROE, against PE(TTM) 123.0x / 147.9x ex-non-recurring / PB 4.90x. 太极实业 is a case of "order-winning ability proven, money-making ability not": engineering general contracting, at 76.47% of revenue, carries a gross margin of only 1.96%, 2025Q4 alone posted an ex-non-recurring loss of RMB 8.16 million, full-year operating cash flow was −79.9% YoY, accounts payable of RMB 18.341 billion is 52% of total assets, and the PE(TTM) of 78.5x rests on the unproven assumption that "the 2025 back-payment of income tax was a one-off". The problem with these two is not that the companies are bad; it is that the price has already told a story the financials have yet to deliver.

7. 厦门象屿(600057) — the Qingdao Xiangyu warehouse fire on 8/2, with the loss amount undisclosed; there may be adverse moves today, so this is one to avoid rather than participate in.

8. ⚠️ The three names on the AI applications chain that were bought hardest on yesterday's dragon-tiger list — 用友网络(600588), 利欧股份(002131), 智度股份(000676) — this is a direction to avoid that was added after verification, and the one that diverges most from the first draft. On 7/31 they saw net buying of RMB 185 million and RMB 116 million respectively, plus a limit-up on RMB 1.020 billion of turnover, making them the most likely to be chased today; and the verification results are: 用友 has lost a cumulative RMB 4.417 billion across 2023—2025, lost another RMB 723 million in 2026Q1, still guides to a loss of RMB 800—930 million at the interim stage with revenue growth of only 1.9%~3.3%, has "computing power" appearing 0 times in its annual report, and has a controlling shareholder that has pledged nearly half its holding with four successive supplementary pledges; 利欧's chairman and CFO were just publicly censured by the SZSE, its interim guidance shows a first-ever ex-non-recurring loss, and the company itself says AI revenue cannot be broken out; 智度 has "Web3" appearing 0 times in its annual report, its real profit comes from North American Yahoo/Google traffic monetisation, and it is already at a 45-day high. The capital flow is real, the fundamentals are unqualified — the two holding true on the same day is precisely the most dangerous combination.

9.4 Final one-sentence judgement

The one sentence most worth remembering today: the batch of main-board AI application stocks that rose the most and were bought hardest on the dragon-tiger list last Friday were almost wiped out by fundamental verification (用友 lost a cumulative RMB 4.417 billion across 2023—2025 with "computing power" appearing 0 times in its annual report, 利欧's chairman and CFO were publicly censured, "Web3" appears 0 times in 智度's annual report), while what actually stands up to scrutiny is precisely the nuclear power names that barely moved that day — above all 东方电气, up only +0.76% on 7/31 yet carrying 2026Q1 net profit attributable to parent of +37.4%, and 中国核电, named in a State Council announcement for 4 units yet closing −0.11% that day. So the correct reading today is "a structural rebalancing from validated-but-fundamentally-unqualified AI applications towards unpriced nuclear power with real profits", while firmly avoiding AI hardware (RMB 20.5 billion of semiconductor net outflow on 7/31, 107 failed limit-ups > 99 limit-ups) and high-position consecutive limit-ups (爱丽家居 on 9 boards is halted for verification today, and the SSE applied self-regulatory measures to 137 abnormal trading incidents this week); against a backdrop of RMB 925.5 billion of reverse repos maturing today, RMB 14 billion of unlocks and a mildly higher US session overnight (Nasdaq +1.00%, Amazon +15% but Apple −7%), a gap-up open for the index is the high-probability case, and whether it can hold after that gap, whether 泛微网络(603039) can make a 3rd board, and whether the four nuclear names (东方电气 600875 / 中国核电 601985 / 江苏神通 002438 / 中国广核 003816) can strengthen on rising volume together are the only three things worth watching today.

---

10. QA notes (risk-auditor red flags and how this report handled them)

This section is the biggest difference between this report and an ordinary research note: it publicly records the defects found in itself, which were fixed and which were not. Before publication this report successively invoked two fundamentals-analyst sub-agents (fundamental verification) and one risk-auditor sub-agent (red flag QA). The QA raised 14 red flags in total; each is addressed below. Items not adopted are listed too, with nothing hidden.

10.1 Adopted and already fixed (9 items)

10.2 Partially adopted (2 items)

10.3 Not adopted (3 items, with reasons)

10.4 Items the QA explicitly endorsed (9 items, for readers to gauge this report's confidence interval)

The items the risk-auditor explicitly judged "passed" are: ① the distinction between "project approval ≠ equipment award"; ② the down-weighting of "an estimate ≠ a plan" (the RMB 4 trillion); ③ the self-correction on "capital inflow ≠ fundamentals" (the 用友 section); ④ the handling of "intraday data ≠ closing data" (writing "no reliable data available" for the SOX close, and using the settled RMB 477.836 billion for ETFs rather than the intraday estimate of RMB 540 billion); ⑤ disclosing conflicting sources side by side (Hormuz, nuclear-grade valve market share); ⑥ the contrarian judgement that "a fuel surcharge cut is a revenue reduction for airlines"; ⑦ using the companies' own replies as primary evidence (the 利欧 and 智度 IR records); ⑧ risk warnings at both the start and the end; ⑨ strict descending order of overall rank versus total score, with ties handled correctly.

⚠️ One last thing that must be made clear: the risk-auditor itself declared a knowledge cut-off of 2026-05, so it cannot verify from memory the truth of any event within this report's window (the State Council nuclear approvals, DeepSeek-V4-Flash, the various interim guidance items, the SZSE sanctions, etc.); its QA conclusions cover only the report's internal consistency, the logic of its bases and the match between evidence and conclusions. For the truth of external facts, company announcements and exchange disclosures prevail.

---

⚠️ Risk warning: This list is nothing more than a pre-market information sweep and observation record; it does not constitute investment advice. Volatility risk in A-shares is extremely high, automatically generated content may contain stale information or errors in industry-chain mapping, and it must not be used directly as a basis for trading.