1The ten-second answer十秒答案
| Candidate候选人 | During the correction (now → Oct)调整期内（当下 → 10月） | Out of the low (Oct → 2027)低点之后（10月 → 2027） |
|---|---|---|
| META | The asymmetric-haven candidate — the only one. By mechanics still a violent two-way trade (July 29 is a binary); a graded disclosure converts it into the refuge mid-correction. Strength without disclosure is distribution.不对称避风港候选人——独一位。就机制而言仍是剧烈的双向交易（7月29日为二元事件）；一份合格的披露能让它在调整途中变成避难所。没有披露的强势即是派发。 | #1 asymmetry.不对称性第一。 The Google-analog trade: build on the disclosure calendar, add on the capex-fear flush, tranche discipline. Conditional band +25–50% into 2027 on execution [S].谷歌类比交易：沿披露日历建仓、在资本开支恐惧的冲洗中加仓、分批纪律。执行到位的条件区间：至 2027 年 +25–50% [S]。 |
| AAPL | The consensus shelter.共识庇护所。 The best down-day record on the board — earned. But it costs a ~53% premium to its own decade average, at the all-time high.全场最好的下跌日战绩——名副其实。但代价是较其十年均值溢价约 53%，且位于历史高点。 | The exit is the trap: when the last leader emerged, the consensus shelter drew down 13.8% funding it (Dec–Jan, one episode). Least upside of the five.出口即陷阱：上一次领袖现身时，共识庇护所回撤 13.8% 为其供血（12–1月，单一案例）。五者中上行空间最小。 |
| NVDA | Falls with its complex — less than memory, but it falls. A $118B buyback is a floor, not a shield.与所属板块同跌——跌幅小于存储，但仍会跌。1,185 亿美元回购是地板，不是盾牌。 | The quality rebound vehicle: cleared-out positioning, ~20–24x forward, Rubin cycle. Trade it; the house view stands.优质反弹载具：仓位已出清、前瞻约 20–24 倍、Rubin 周期。做波段；本台原判定不变。 |
| GOOGL | Quality ballast. Proof printed, dilution behind it, $340–370 is the accumulation zone.优质压舱石。证据已打印、稀释已落地，340–370 美元是积累区。 | Steady compounder to new highs by 2027 — but the explosive rerating already happened.2027 年前稳步复利创新高——但爆发式重估已经发生过了。 |
| TSLA | Not a haven — last autumn’s “safety” run was a high-beta chase that fell 2x the index inside the stress window.不是避风港——去年秋天的「避险」行情实为高贝塔追逐，压力窗口内跌幅两倍于指数。 | A scheduled encore: the year-end robot-expectations leg, rented from the October low as beta, option-sized, sold into CES. See the card in §5.一场排定的返场：年末机器人预期行情——自十月低点作为贝塔租入、按期权定仓、CES 前后了结。详见 §5 卡片。 |
Dark horses, for completeness: MSFT is the contrarian trough accumulation (worst-performing megacap, cleanest AI-revenue proof); AMZN rides at half weight; AVGO is the custom-silicon winner wearing semis beta. All three are covered in §5.补充几匹黑马：MSFT 是逆向的低谷建仓标的（表现最差的巨头、AI 收入证据最干净）；AMZN 以半权重随行；AVGO 是穿着半导体贝塔外衣的自研芯片赢家。三者详见 §5。
2Two different jobs两份不同的工作
Job one is shelter: absorbing capital on the way down. Its mechanics are boring — low beta, low correlation, a buyback bid under the stock, no financing needs, no binary events on the calendar. Job two is rerating leadership: being the name underweight managers are forced to chase once fear breaks. Its mechanics are the opposite — a fresh classification change, cleared-out positioning, and a proof event that turns a discount into a premium. The asymmetric haven is the fusion: a stock holding both roles at once, which happens only when its proof event lands during everyone else’s drawdown. That is what Google became in Q4 2025 — and, importantly, it was the most-hated megacap in the room when it happened, not the consensus pick. To keep the label falsifiable rather than a story told after the fact, this desk defines the candidate ex ante by three tests: bottom-tier positioning on the 13F crowding desk before the move; a dated disclosure calendar inside the correction window; and the confirmation signature — absorbing capital on index-down days after a disclosure event, not before. One name passes the first two tests today.第一份工作是庇护：在下跌途中吸纳资金。它的机制很无聊——低贝塔、低相关性、股价下方有回购买盘、没有融资需求、日历上没有二元事件。第二份工作是重估领导权：当恐惧散去，低配的基金经理被迫追买的那只股票。它的机制恰好相反——全新的类别重划、已被出清的仓位、以及一个把折价变成溢价的证据事件。不对称避风港就是两者的融合：一只股票同时担任两个角色——这只在它的证据事件恰好落在别人的回撤期间时才会发生。谷歌在 2025 年四季度正是如此——而且重要的是，事发时它是全场最被憎恨的巨头，绝非共识之选。为了让这个标签可证伪、而不是事后编的故事，本台以三条事前检验来定义候选人：起跑时在 13F 拥挤度台处于最低档仓位；调整窗口之内有一份带日期的披露日历；以及确认签名——在披露事件之后的指数下跌日持续吸纳资金，而非之前。今天恰有一个名字通过前两条检验。
Our own engines make the split visible right now. The Mag-7 basket is the only basket on the board positive month-to-date (+3.9%) while Memory & HBM bleeds −12.3%F — rotation into perceived quality is already running, exactly the 2018/2000 sequence The AI Master Plan flagged on July 1. Within that rotation, the tape is discriminating: on QQQ-down days since June 1, Apple beat the index 77% of the time with +1.7% mean excess; Meta beat it 54% with +1.1%; Nvidia beat it 31% with −0.4%; Tesla 38% with −0.7%F. The consensus-haven vote is not hypothetical — it is being counted daily, and Apple is winning it. But the consensus vote and the asymmetric role are different elections: Google was losing the down-day vote (55% beat rate, below Microsoft and Apple) in the very quarter it became the refugeF. Winning down-days is the shelter’s resume; being chased upward through them is the asymmetric haven’s.本站引擎已经把这条分界线照了出来。月初至今，Mag-7 是全板块唯一收正的篮子（+3.9%），而存储与 HBM 篮子失血 −12.3%F——向「感知中的优质」轮动已在进行，正是《AI 总体规划》7月1日标记的 2018/2000 序列。而在这轮轮动内部，盘面在做区分：6月1日以来的 QQQ 下跌日中，Apple 以 +1.7% 的平均超额跑赢指数 77% 的时间；Meta 为 54%、+1.1%；Nvidia 为 31%、−0.4%；Tesla 为 38%、−0.7%F。共识避险的投票不是假设——它每天都在计票，而 Apple 正在赢。但共识投票与不对称角色是两场不同的选举：谷歌在成为避难所的那个季度里，下跌日投票其实是输的（跑赢率 55%，低于微软和 Apple）F。赢下跌日是庇护所的履历；被资金追着穿越下跌日，才是不对称避风港的履历。
3The Google template, measured谷歌模板的度量
Three details from the measurement matter more than the headline. First, the haven behavior was inflow-driven, not defensiveness. On QQQ-down days that quarter, Google beat the index only 55% of the time — less than Microsoft (72%) or Apple (62%)F. What made Google the haven was that it kept rising through the drawdown: fresh money was arriving daily, so the dips didn’t stick. A haven is made by flows, not by beta. Second, its beta halved — from 2.58 to 1.26 in one quarterF — which is what a classification change looks like in the covariance data. Third, the proof predated the price. The rerating’s catalysts were all disclosure events: the September 2 antitrust ruling that took forced breakup off the table (+9% in a session, ~$234B of market capF), the November 14 Berkshire 13F revealing a $4.9B stakeF, Gemini 3 on November 18, and a cloud P&L that was already profitable with a $240B printed backlogF. The multiple went from ~17x trailing at the March-2025 trough to ~29x by year-endC because each leg of the story was documented before the market paid for it.度量结果中有三个细节比标题数字更重要。第一，避险行为是资金流入驱动的，不是防御性。那个季度 QQQ 的下跌日里，谷歌跑赢指数的比例只有 55%——低于微软（72%）和 Apple（62%）F。让谷歌成为避风港的，是它在回撤中持续逆势上行：新钱每天都在进场，下跌根本站不住。避风港由资金流铸成，而非贝塔。第二，它的贝塔腰斩——一个季度内从 2.58 降到 1.26F——这就是「类别重划」在协方差数据里的模样。第三，证据先于价格。那轮重估的催化剂全是披露事件：9月2日反垄断裁决排除强制拆分（单日 +9%、约 2,340 亿美元市值F）、11月14日伯克希尔 13F 披露 49 亿美元持仓F、11月18日 Gemini 3 发布，以及一个已经盈利、且在手订单 2,400 亿美元已经打印的云业务F。估值从 2025 年 3 月低谷的约 17 倍升至年末约 29 倍C，因为故事的每一条腿都是先有文件，市场才付钱。
The template has an epilogue, and it is the part most worth carrying into this autumn. First, the leader’s own dips were real but did not stick: GOOGL took an −8.2% air pocket of its own inside the quarter, troughing December 17F — the asymmetric haven is not a drawdown-free asset; it is one whose drawdowns get bought. Sizing has to survive an −8% pocket without shaking out. Second, the consensus shelter paid for the leader’s coronation. Apple entered December 2025 at its all-time high, freshly crowned as the defensive megacap — and over December 1 to January 31 it fell 8.3%, with a −13.8% peak-to-trough inside the window, while GOOGL gained +7.4% and META +11.9%F. One episode does not make a law — Apple’s current down-day record is genuinely the best in the group — but the mechanism is general and shows up at every leadership turn: the shelter trade is, by construction, the crowded trade, and it is the natural funding source the moment a rerating leader gives managers something better to chase. A shelter bought at its all-time high, at a ~53% premium to its own decade average, protects against the index and not against the rotation.模板还有一段尾声，而这恰是最值得带进今秋的部分。其一，领袖自己的回撤真实存在，只是站不住：GOOGL 在那个季度里自己也挨了一记 −8.2% 的气穴，12月17日见底F——不对称避风港不是没有回撤的资产，而是回撤会被买起来的资产。仓位必须扛得住一记 −8% 的气穴而不被震出局。其二，共识庇护所为领袖的加冕买了单。Apple 以历史新高、新科防御巨头的身份进入 2025 年 12 月——而 12月1日至1月31日 它下跌 8.3%、窗口内峰谷回撤 −13.8%，同期 GOOGL +7.4%、META +11.9%F。单一案例不成定律——Apple 当下的下跌日战绩确实全组最佳——但这个机制是普遍的，在每次领导权更替处都会现身：庇护交易在构造上就是拥挤交易，一旦重估领袖给了基金经理更好的追逐对象，它就是天然的供血来源。在历史高点、以较自身十年均值溢价约 53% 的价格买入的庇护所，防的是指数，防不了轮动。
4Meta’s auditionMeta 的试镜
First, what genuinely rhymes with the template. Meta enters this audition underowned: our ownership desk has it at “low” crowding, held by 17 of 51 tracked funds after 13 cut it in Q1, with Tepper’s Appaloosa among the trimmersF — the same cleared-out starting point Google had. It enters discounted: −23.5% below its August-2025 high of $788 after a −33% derate to the March troughF, on roughly 17–19x the consensus 2027 earnings lineS (consensus figures are single-source this week; treat as indicative). And it has the freshest classification change in the market: Meta Compute — reported July 1, Bloomberg — would make it the first hyperscaler to flip from compute buyer to compute sellerC; Muse Spark 1.1 opened the first paid Meta model API on July 9 at $1.25/$4.25 per million tokens, roughly a quarter of rival list pricesC; subscriptions began rolling out May 27C. The stock answered: +8.8% on July 1 while its own suppliers fell, its best week in years into July 10F, and a 20-day beta to QQQ that collapsed from 1.80 in March to ~0.5F. The narrative inversion — from “ad company overspending on AI” to “AI platform funded by ads” — is underway. On this much, the external briefs and our data agree.先说与模板真正押韵的部分。Meta 带着低配仓位走进试镜：本站持仓台给它的拥挤度评级为「低」，51 家跟踪基金中仅 17 家持有、一季度有 13 家减持，Tepper 的 Appaloosa 也在减持之列F——与谷歌当年同样「已被出清」的起点。它带着折价：较 2025 年 8 月 788 美元的高点仍低 23.5%，此前一路 −33% 跌到 3 月低谷F，对应 2027 年共识盈利约 17–19 倍S（本周共识数字为单一来源，仅作参考）。它还握着全市场最新鲜的类别重划：Meta Compute——彭博 7月1日报道——将使它成为第一家从算力买家翻转为算力卖家的超大规模厂商C；Muse Spark 1.1 于 7月9日开放首个付费 Meta 模型 API，每百万 token 定价 $1.25/$4.25，约为对手报价的四分之一C；订阅已于 5月27日开始铺开C。股价给出了回应：7月1日 +8.8% 而其供应商齐跌、7月10日收出多年最佳单周F、对 QQQ 的20日贝塔从 3 月的 1.80 坍缩到约 0.5F。从「烧钱做 AI 的广告公司」到「用广告养 AI 的平台」的叙事翻转已经启动。到这里为止，外部简报与我们的数据是一致的。
The monetization case is a stack of six levers, and they are not equal — what separates them is the state of the evidence. Ranked from printing to policy-contingent:变现论据是一座六杠杆的堆栈，而六条杠杆并不等价——区分它们的是证据状态。按「正在打印」到「取决于政策」排序：
| Lever杠杆 | Evidence state证据状态 | What the record shows记录显示 |
|---|---|---|
| 1 · AI → ads (RecSys)1 · AI → 广告（推荐系统） | Printing now正在打印 | Q1 ads +33% with impressions +19% and price-per-ad +12% — the fingerprint of better targeting, not inventory stuffingF. The market books it as “ads grew” and pays no multiple for itC — the single most underpriced lever.一季度广告 +33%，且展示量 +19% 与单价 +12% 同时发生——这是定向变准的指纹，不是塞库存F。市场把它记作「广告涨了」，不给任何估值倍数C——最被低估的一条杠杆。 |
| 2 · Subscriptions2 · 订阅 | Live since May 275月27日起已上线 | IG/FB Plus $3.99, WhatsApp Plus $2.99 rolling outC; the FoA “Other” line was already $885M in Q1, +74% y/y, before launchC. First disclosed metrics possible as soon as July 29.IG／FB Plus $3.99、WhatsApp Plus $2.99 铺开中C；FoA「其他」收入一季度已达 8.85 亿美元、同比 +74%，且发生在订阅上线之前C。最早 7月29日 就可能出现首批披露指标。 |
| 3 · Cost base3 · 成本基数 | Scheduled / in hand已排定／在手 | 8,000 layoffs in May flow into 2H opexC; expense guidance held at $162–169B while committing 2026 operating income above 2025C — modest in size, load-bearing for the credibility anchor.五月 8,000 人裁员将体现在下半年费用中C；费用指引维持 1,620–1,690 亿美元，同时承诺 2026 年营业利润高于 2025C——体量不大，却是可信度之锚的承重件。 |
| 4 · Iris custom silicon4 · Iris 自研芯片 | Scheduled: September已排定：九月 | Broadcom-built, TSMC-fabbed, manufacturing start September per a Reuters-reviewed memo; partnership through 2029C. The TPU-style cost-curve bend — with the caveat that first-generation silicon proves nothing until yields and utilization print.博通设计、台积电代工，据路透社审阅的备忘录九月投产；合作延至 2029C。TPU 式的成本曲线拐点——但需注意：第一代芯片在良率与利用率打印之前，什么也证明不了。 |
| 5 · Meta Compute + API5 · Meta Compute + API | Reported, unscheduled已报道，未排期 | Compute resale in development (Bloomberg, Jul 1), no launch date, pricing or revenueC; Muse Spark API live at $1.25/$4.25C — a wedge, not yet a P&L line. The biggest lever and the least documented.算力转售筹备中（彭博，7月1日），无上线日期、定价与收入C；Muse Spark API 已上线，定价 $1.25/$4.25C——是楔子，还不是损益表科目。杠杆最大，文件最少。 |
| 6 · US-model sovereignty6 · 美国模型主权 | Policy-contingent取决于政策 | Chinese models hold 30–46% of routed API tokens (DeepSeek 17.6%, Qwen 13.9%) on 60–90% cheaper pricingC; the State Department is reported preparing restrictions on corporate use of Chinese AI, with a House investigation open (Jul 8)C. Muse Spark is the only US model priced in the gap — 3–4x DeepSeek, ~25% of US frontierC. Zero booked revenue today; a formal rule would convert it from color to catalyst.中国模型占路由 API token 的 30–46%（DeepSeek 17.6%、Qwen 13.9%），定价便宜 60–90%C；据报道国务院正筹备限制企业使用中国 AI，众议院已开启调查（7月8日）C。Muse Spark 是唯一定价落在该缺口中的美国模型——约为 DeepSeek 的 3–4 倍、美国前沿模型的约 25%C。今日入账收入为零；一纸正式规则会把它从背景色变成催化剂。 |
Now the disqualifiers — not for the trade, but for the haven job as it stands today. They come in four layers.接下来是不合格项——不是否定这笔交易，而是否定此刻的避险资格。共有四层。
| Template requirement模板要求 | Google, entering Q4 2025谷歌，2025年四季度入场时 | Meta, entering the fall of 2026Meta，2026年秋入场时 |
|---|---|---|
| De-risking event去风险事件 | Done — Sep 2 ruling removed breakup risk已完成——9月2日裁决排除拆分风险 | Pending — FTC appeal of the dismissed monopoly case is live未完成——FTC 对被驳回垄断案的上诉仍在进行F |
| Printed proof已打印的证据 | Profitable cloud, $240B backlog disclosed云业务已盈利，在手订单 2,400 亿美元已披露 | Meta Compute: no launch date, no pricing, no revenue, no backlog; company declined commentMeta Compute：无上线日期、无定价、无收入、无在手订单；公司拒绝置评C |
| Capital mechanics资本机制 | Buybacks running through the window窗口期内回购持续进行 | Buybacks $0 in Q1 (vs $12.75B a year ago); ~$64B equity raise under consideration一季度回购为 0（去年同期 127.5 亿美元）；约 640 亿美元股权融资在酝酿C |
| Validation buyer认证买家 | Berkshire 13F, Nov 14伯克希尔 13F，11月14日 | None disclosed; next 13F window is August 14尚无披露；下一个 13F 窗口为 8月14日F |
| Stress-test record压力测试记录 | +5.4% through the QQQ drawdownQQQ 回撤期间 +5.4% | −21.6% through the same drawdown — worst of the group同一回撤期间 −21.6%——全组最差F |
The reflexivity trap is the deepest layer. Meta’s new story is not a hedge against the AI-capex trade — it is the AI-capex trade, restated as a vendor. Management raised the 2026 capex guide to $125–145B and named memory prices as the driverF; the compute it would lease is the same compute whose scarcity pricing the correction would deflate; the neoclouds it would undercut are the canaries of the credit tape. If semis and memory correct because AI-capex credibility cracks — the modal Phase-1 path in The AI Master Plan, memory −30–45% from highs — Meta’s option on compute resale reprices down with the complex, not against it. The Second Act called Meta “the highest-beta platform on this question, in both directions,” and July proved the both-directions clause: +8.8% on July 1, −4.9% on July 2F. That is the signature of a rerating candidate, and the opposite of a shelter’s.反身性陷阱是最深的一层。Meta 的新故事不是对 AI 资本开支交易的对冲——它就是那笔交易本身，只是换成了卖方的说法。管理层把 2026 年资本开支指引上调至 1,250–1,450 亿美元，点名存储器价格是主因F；它要出租的算力，正是这轮调整将会压缩其稀缺定价的那批算力；它要降价打击的新兴云，正是信用行情的金丝雀。如果半导体与存储因 AI 资本开支可信度破裂而调整——即《AI 总体规划》阶段一的众数路径，存储自高点 −30–45%——Meta 的算力转售期权将与板块同向下修，而非反向。《第二幕》称 Meta 为「在这个问题上双向弹性最高的平台」，而七月验证了「双向」二字：7月1日 +8.8%，7月2日 −4.9%F。这是重估候选人的签名，恰是掩体的反面。
Which means compute-deal headlines must be graded, not cheered. A deal announced in August or September is the fastest route to the fusion — a contracted number, printed before the de-risking, is precisely the Google-backlog analog. But the same headline can anti-confirm. Deals struck now are priced off scarcity rates — the xAI-era anchors near $35–50B per GW-yearC — that a memory/semis correction actively deflates; and a company with genuine internal compute scarcity does not sell capacity at all. Zuckerberg’s own tell cuts both ways: companies approach Meta “almost weekly,” with short-term deals at large premiums reportedly on the tableC — short-term-and-premium is what a seller signs when it doubts durable demand. The grading rubric, ex ante: term (multi-year confirms; month-to-month anti-confirms), counterparty (an investment-grade or frontier-lab anchor confirms; distressed neoclouds anti-confirm), pricing (locked at or near scarcity anchors confirms; discounts into a falling curve anti-confirm), and accounting (a disclosed backlog/RPO line confirms; a press release without numbers is a press release).这意味着算力交易的头条必须评级，而不是欢呼。八九月间官宣的一纸交易，是通往「融合」的最快路径——在去风险之前打印出来的合同数字，正是谷歌订单簿的类比物。但同一条头条也可能反向确认。此刻签的合同，定价锚在稀缺费率上——xAI 时代的锚约为每 GW·年 350–500 亿美元C——而存储／半导体的调整正在压缩这条曲线；且一家真正算力内需紧张的公司，根本不会外售产能。扎克伯格自己的信号也是双刃：企业「几乎每周」找上 Meta，短期高溢价交易据报已摆上桌面C——「短期＋高溢价」正是卖方怀疑需求持久性时才签的东西。评级标准，事前定死：期限（多年期为确认；月付即反向确认）、交易对手（投资级或前沿实验室锚定租户为确认；困境新兴云为反向确认）、定价（锁定在稀缺锚位附近为确认；沿下行曲线打折为反向确认）、入账方式（披露在手订单／RPO 科目为确认；不带数字的新闻稿只是新闻稿）。
July 29 is a binary, not a checkpoint. Q2 earnings land after the close that dayC. The bull needs: revenue near the top of the $58–61B guide, the 2026 operating-income commitment defended, capex held, and any concrete Meta Compute or subscription disclosure. The bear needs only one sentence: a third capex raise with no revenue attached — or confirmation of the equity raise into a weak tape. Note also what Zuckerberg told an internal town hall on July 2: AI-agent progress “hasn’t really accelerated in the way that we expected”C — the model-quality leg of this story is not yet load-bearing. Our read: the probability-weighted outcome of the print is negative for the haven framing and constructive for the entry framing. A post-earnings selldown is tranche one of the house buy-list, already on the books in The Relapse; the dilution event, if it comes, marks the low’s neighborhood, not the thesis’s death.7月29日是二元事件，不是例行检查点。二季报当日盘后发布C。多头需要：收入贴近 580–610 亿美元指引上限、守住 2026 年营业利润承诺、资本开支按兵不动、外加任何一条具体的 Meta Compute 或订阅披露。空头只需要一句话：第三次上调资本开支而不附带收入——或在弱势盘面中确认股权融资。还要注意扎克伯格 7月2日在内部全员会上的话：AI 智能体的进展「并没有像我们预期的那样加速」C——这个故事里「模型质量」这条腿尚未承重。我们的判读：这份财报按概率加权的结果，对「避险」叙事是负面的，对「入场」叙事是建设性的。财报后的下杀正是本台买入清单的第一批——《复燃》已经记账；而稀释事件若发生，标记的是低点的街区，不是论点的死亡。
5The full cast, scored全体候选人打分
| Name标的 | Correction resilience调整期抗跌 | Rerating fuel重估燃料 | Printed AI proof已打印的 AI 证据 | Capital return资本回报机制 | Positioning edge仓位优势 | Policy fit政策契合 | Verdict判定 |
|---|---|---|---|---|---|---|---|
| META | The asymmetric-haven candidate — fusion gated on disclosures不对称避风港候选人——融合取决于披露 | ||||||
| AAPL | Consensus shelter — earned, priced, exit is the trap共识庇护所——名副其实、已定价、出口即陷阱 | ||||||
| NVDA | Coiled rebound vehicle, not shelter蓄势反弹载具，非庇护所 | ||||||
| GOOGL | Quality ballast; second run in 2027优质压舱石；2027 年第二程 | ||||||
| MSFT | Contrarian trough accumulation逆向低谷建仓 | ||||||
| TSLA | Different contest — 2027 physical-AI option另一场比赛——2027 物理 AI 期权 |
The tape crowned it weeks ago, and the crown is earned: +4.8% on the July 2 semiconductor routF, a 20-day beta near zero, correlation ≤0.19 to every other megacap, and just −0.6% from its all-time high while the group averages −19%F. The mechanics are the full haven kit: ~$143B of expected FY26 free cash flow, a fresh $100B buyback, and a demonstrated ability to pass memory-inflation through prices — Japan iPhone prices +10%, Macs and iPads raised, Cook calling further increases “unavoidable”C — plus the Gemini-Siri deal that quietly deleted the AI-laggard discountF. Institutions added ~1.3B shares in Q2 filingsF. Two caveats define the position’s shape. The entry: ~34x forward, a ~53% premium to the decade average, crowding rated “elevated” — the most concentrated of the groupF — insurance bought at the top of its premium range. And the exit: the template’s epilogue (§3) — the consensus shelter at its all-time high drew down 13.8% funding the last leader’s coronationF, one episode but a general mechanism. The position earns its keep into the correction; the discipline is rotating it toward the asymmetric candidate on the first graded disclosure, not after the crowd does.盘面数周前就已为它加冕，而这顶王冠名副其实：7月2日半导体大跌当天 +4.8%F、20日贝塔近零、与其他巨头相关性 ≤0.19、距历史高点仅 −0.6%（同组平均 −19%）F。机制上是全套避险装备：FY26 预期自由现金流约 1,430 亿美元、新批 1,000 亿美元回购、且已示范把存储涨价转嫁进售价——日本 iPhone 提价 10%、Mac 与 iPad 已涨价、库克称进一步涨价「不可避免」C——外加 Gemini-Siri 协议悄悄删除了「AI 掉队」折价F。机构在二季度申报中增持约 13 亿股F。两条注记决定这笔仓位的形状。入口：前瞻约 34 倍、较十年均值溢价约 53%、拥挤度评级「偏高」——全组最集中F——这是在保费区间顶部买入的保险。出口：模板的尾声（§3）——上一次领袖加冕时，处于历史高点的共识庇护所回撤 13.8% 为其供血F，虽是单一案例，机制却是普遍的。这笔仓位在进入调整的路上挣它的饭钱；纪律在于：在第一份合格披露出现时就向不对称候选人轮动，而不是等人群动了再动。
The popular framing — a safety net, because it never rode the memory bubble — is half right. NVDA sat out the melt-up: roughly +10–13% YTD against SOXX +90%, dead last in its own indexC, because China went to ~zero, custom silicon took ~28% of hyperscaler AI-chip capex, and 3,872 institutions trimmed against 1,286 addingC. That cleared-out state, ~20–24x forward with a PEG near 0.45, a $118B buyback stack, and the Rubin ramp make it the highest-quality rebound vehicle once the complex flushesC. But a safety net is not a shelter: on down days it still loses to the index (31% beat rate, −0.4% mean excessF), a $450M insider-sell cluster printed July 9F, and its vendor-financing web — the CoreWeave backstop, the OpenAI loop — is direct exposure to the very credit tape the house watches for the bottom signalC. In 2018, the analog year, the king of the prior cycle halved after the memory names topped. Trade it around Rubin datapoints; buy it with conviction at the credit-tape low, not before.流行的说法——「安全网，因为它没搭存储泡沫的车」——对了一半。NVDA 确实缺席了这轮融涨：年内约 +10–13%，对比 SOXX +90%，在自家指数里垫底C——因为中国业务归零、自研芯片吃掉约 28% 的超大规模 AI 芯片开支、且 13F 中 3,872 家机构减持对 1,286 家增持C。这种出清状态，加上前瞻约 20–24 倍、PEG 约 0.45、1,185 亿美元回购弹药与 Rubin 放量周期，使它成为板块冲洗后最优质的反弹载具C。但安全网不等于庇护所：下跌日它仍跑输指数（跑赢率 31%、平均超额 −0.4%F），7月9日刚打印 4.5 亿美元的内部人卖出集群F，而它的供应商融资网——CoreWeave 兜底、OpenAI 回环——正是本台盯底信号的那条信用行情的直接敞口C。在类比年 2018，上一轮周期之王在存储见顶之后腰斩。围绕 Rubin 数据点做波段；带信念的买点在信用行情的低点，不在之前。
Round two is a different trade from round one. The proof is now overwhelming — Cloud +63% to $20B at a 32.9% operating margin, a $467.6B backlog, TPUs going externalF — but the market has paid for it once (+108% over 52 weeksF) and the June $85B equity raiseC plus the DOJ’s cross-appeal seeking divestitureF put the dilution and legal cards back on the table. Berkshire, notably, took the other side of the raise with a $10B private placementC. Our 13F desk shows funds cut GOOGL hardest in Q1 (−15 funds) even as Buffett addedF — the crowd left, the anchor stayed. The house calls stand: accumulation zone $340–370, new highs by Q1–Q2 2027, and its down-drawdown record (+5.4% through the last one) still makes it the best-behaved ballast after Apple.第二程与第一程是不同的交易。证据如今已压倒性——云业务 +63% 至 200 亿美元、营业利润率 32.9%、在手订单 4,676 亿美元、TPU 开始外销F——但市场已经为此付过一次钱（52 周 +108%F），而 6 月 850 亿美元股权融资C加上 DOJ 寻求拆分的交叉上诉F，把稀释与法律两张牌重新放回桌面。值得注意的是，伯克希尔以 100 亿美元私募配售站在了融资的另一侧C。本站 13F 台显示一季度基金对 GOOGL 减持最狠（−15 家）而巴菲特在加仓F——人群离场，锚还在。本台原判不变：340–370 美元为积累区、2027 年一二季度前创新高；且它上一轮回撤中 +5.4% 的记录，仍使它是 Apple 之后行为最端正的压舱石。
Candidate lists usually skip the name this framework scores second-best for the October deployment. Microsoft is the worst-performing megacap of 2026 (−20% YTD, −29% from its highF) while carrying the cleanest printed proof in the group: a $37B annualized AI revenue line growing 123%C. It also beat the index on 62–72% of down days across our two measurement windowsF — quiet shelter behavior nobody credits. The overhangs are known: a $146B capex year, OpenAI concentration inside the RPO, and no fresh classification story. It lacks Meta’s narrative spark, which is exactly why the entry is cheap. Tranche it with the house schedule.候选名单通常漏掉本框架中十月部署期得分第二的名字。微软是 2026 年表现最差的巨头（年内 −20%、距高点 −29%F），却握着全组最干净的已打印证据：370 亿美元年化 AI 收入、增速 123%C。在我们两个测量窗口里，它还在 62–72% 的下跌日跑赢指数F——无人记账的安静庇护行为。悬顶之患众所周知：1,460 亿美元的资本开支年、RPO 里的 OpenAI 集中度、以及缺一个新鲜的类别故事。它没有 Meta 的叙事火花——而这恰恰是入场便宜的原因。按本台时间表分批。
Every institutional-haven test fails: institutions were net sellers in Q1 while retail boughtC; the stock lost to the index on down days (−0.7% mean excessF); it trades ~181x forward on a 4.2% operating marginC; and it fell 7.5% on July 2 despite a +25% delivery beatF. The folk memory that Tesla “acted like a haven” last autumn does not survive the price store: the Sep–Dec 2025 run was real (+36.5%, all-time high $490 on Dec 16F) — but inside it TSLA fell 14.4% through the quarter’s QQQ stress window, twice the index and right in line with PLTR (−21.7%) and HOOD (−26.7%)F, took a −16.5% air pocket into Nov 21F, and then gave back 24% from the December top when the robot promise met January realityF. That was a rate-cut-era high-beta chase wearing a robot costume — beta, never shelter.机构避险的每一项测试都不及格：一季度机构净卖出、散户净买入C；下跌日跑输指数（平均超额 −0.7%F）；前瞻约 181 倍、营业利润率 4.2%C；7月2日交付量超预期 +25%，股价仍跌 7.5%F。「Tesla 去年秋天像个避风港」的民间记忆，过不了价格库这一关：2025 年 9–12 月的行情是真的（+36.5%，12月16日历史新高 490 美元F）——但行情内部，TSLA 在该季度 QQQ 压力窗口中下跌 14.4%，两倍于指数，与 PLTR（−21.7%）、HOOD（−26.7%）同列F；11月21日前挨了一记 −16.5% 的气穴F；随后当机器人承诺撞上一月的现实，又从十二月顶部回吐 24%F。那是一场降息时代的高贝塔追逐，披着机器人戏服——始终是贝塔，从来不是庇护。
The encore is on the calendar, though. Low-volume Optimus production is slated for Fremont in late July–August (Model S/X lines already convertedC), the twice-slipped V3 reveal is pendingC, first B2B sales are guided to late 2026 at ~$100K+C, and the house map already schedules the physical-AI ignition for the post-midterm leg — Optimus in the wild, the CES robot deluge, “after CES it’s consensus.” A year-end robot run is therefore the rhyme of last year’s, with the same seasonal slot and a better macro tailwind (the post-low, post-pivot chase). Two things are different this time. The hype now has falsifiable objects — production units, deliveries, a V3 that either does useful work or does not (“entirely forward-valued” is the standing sell-side cautionC) — so the run either converts into a reclassification or repeats the January upset. And the competition now ships: Unitree is delivering consumer humanoids at $7–8K against a 20,000-unit 2026 target, UBTech launched at ~$16.5K with 13,000+ ordersC. On Washington: no evidence, in either direction, of the administration touching the release timing — the searched null is explicit; policy is pro-robotics. The dated policy risk sits in Beijing, not Washington: China’s stricter rare-earth-magnet export controls are suspended only through November 10, 2026C — days after the midterms, mid-run — and Chinese robot makers source magnets domestically while everyone else queues for licenses. Treatment unchanged in kind: rent the ignition from the October low as the high-beta recovery leg, sized as an option; the cleaner expression remains the robot bill-of-materials suppliers — rare-earth magnets included, which pay on both branches of the export-control fork; and take the 2025 lesson into CES, where promise gets marked against product.不过，返场已经写在日历上。Optimus 低量产排定于 7 月底–8 月在弗里蒙特开始（Model S/X 产线已完成改造C），两度推迟的 V3 发布仍待落地C，首批 B2B 销售指引在 2026 年末、单价约 10 万美元以上C；而本台地图本就把「物理 AI 点火」排在中期选举后的那一段——Optimus 实地亮相、CES 机器人洪流、「CES 之后即成共识」。因此年末的机器人行情是去年那场的押韵：同一个季节槽位，外加更好的宏观顺风（低点后、转向后的追逐）。但这一次有两点不同。炒作有了可证伪的对象——量产机、交付量、一台要么能干活要么不能的 V3（「完全靠远期估值」是卖方现成的警语C）——所以行情要么兑现成类别重划，要么重演一月的翻车。而且对手已经在出货：宇树的消费级人形机器人售价 7–8 千美元、2026 年目标两万台，优必选以约 1.65 万美元起售、累计订单超 1.3 万台C。至于华盛顿：没有任何证据显示政府在任一方向上干预发布时点——这一「无发现」是明确检索过的；政策立场是支持机器人的。带日期的政策风险在北京、不在华盛顿：中国更严格的稀土磁材出口管制仅暂停至 2026年11月10日C——中期选举后数日、行情中段——且中国机器人厂商在境内采购磁材，其他所有人都要排队等许可证。处理方式性质不变：从十月低点起租用这场点火行情，作为高贝塔复苏段、按期权定仓；更干净的表达仍是机器人物料清单供应商——包括稀土磁材，它在出口管制分叉的两条枝上都有报酬；并且带着 2025 年的教训走进 CES——在那里，承诺要跟产品对账。
Briefly, the rest of the field. AMZN: AWS reaccelerated to +28%, the fastest in 15 quarters, but $200B of capex crushed trailing FCF toward zeroC — rides at half weight, per the house book. AVGO: the custom-silicon winner (AI revenue +143%, the OpenAI “Jalapeño” co-designC) — but it fell 10% during the last QQQ drawdown and trades with its sector; it is Act-I beta wearing an Act-II story. ORCL is disqualified from haven duty outright: −57% from its high with negative FCF and a $130B+ debt stackC — it is the credit canary the house watches, not a shelter. PLTR, −36% from its high at ~200x trailing, is a sentiment gauge for the speculative complex, nothing more here.其余名单从简。AMZN：AWS 重新加速到 +28%、15 个季度最快，但 2,000 亿美元资本开支把滚动自由现金流压到近零C——按本台账本以半权重随行。AVGO：自研芯片赢家（AI 收入 +143%、与 OpenAI 合作设计「Jalapeño」C）——但上一轮 QQQ 回撤中它跌了 10%，且与板块同步交易；它是披着第二幕故事的第一幕贝塔。ORCL 直接丧失避险资格：距高点 −57%、自由现金流为负、债务超 1,300 亿美元C——它是本台盯着的信用金丝雀，不是庇护所。PLTR，距高点 −36%、约 200 倍市盈率，在这里只是投机板块的情绪仪表，仅此而已。
6The Washington question华盛顿之问
What the evidence supports. The alignment of interests is documented and unusually literal. The state now holds ~9.9% of Intel via converted CHIPS grants and ~15% of MP MaterialsF; on July 2 a proposal surfaced for a 5% federal stake in OpenAI (~$43B) inside a “Public Wealth Fund” concept under which the major AI companies — Meta and Google included — would cede stakes to a sovereign-style vehicleF. The Treasury Secretary publicly cheerleads the “$750B AI buildout” and projects AI doubling productivityC; the AI Action Plan’s permitting, power and export machinery keeps grinding (EO 14318 datacenter fast-tracking, the FERC tariff action of June 18)F; and Section 232 chip tariffs carve out US datacenters — a structural subsidy to whoever consumes compute domesticallyF. An administration this long the trade — institutionally, rhetorically, and in at least one disclosed case personallyF — has every incentive to want confidence held into November 3. That much is real.证据支持的部分。利益的对齐有据可查，而且罕见地「字面化」。联邦政府现经由 CHIPS 拨款转股持有英特尔约 9.9%、持有 MP Materials 约 15%F；7月2日浮出一项提案：联邦持有 OpenAI 5% 股份（约 430 亿美元），置于一个「公共财富基金」概念之下——主要 AI 公司（包括 Meta 与谷歌）都将向这一主权式载体让渡股份F。财政部长公开为「7,500 亿美元 AI 建设」摇旗，并预言 AI 将使生产率翻倍C；《AI 行动计划》的许可、电力与出口机器持续运转（EO 14318 数据中心快速通道、6月18日 FERC 电价行动）F；232 条款芯片关税给美国数据中心开了豁免——对在境内消耗算力者的结构性补贴F。一个在制度上、言辞上、且至少在一例披露中以个人身份F如此深度做多这笔交易的政府，当然有动机希望信心撑到 11月3日。这部分是真的。
Where the evidence ends. There is no documented act of stock selection, no rotation management, and no official statement linking the AI trade to the midterms — our policy sweep looked and logged the nulls explicitly. The contradictions are sharper than the folklore: the DOJ cross-appealed in February seeking Chrome divestiture from the market’s favorite haven; the FTC’s Meta appeal is live; the Apple case is headed to trial; Nvidia’s Groq licensing deal is under inquiryF. A government cannot coherently “pick” a haven it is simultaneously trying to break up. Meanwhile the Fed it just installed — Warsh, confirmed 54–45 in May — is holding at 3.5–3.75% with CPI at 4.2% and three FOMC members hinting the next move is a hikeF. And inside Treasury itself, a career-analyst draft comparing the AI complex to the dotcom bubble was publicly dismissed as “unvetted” the same weekF — the state is not even united on whether to believe its own trade.证据止步之处。没有任何有据可查的选股行为、没有轮动管理、也没有把 AI 交易与中期选举挂钩的官方表态——我们的政策扫描找过，并把这些「无发现」明确记录在案。矛盾比传言更锋利：司法部 2 月交叉上诉，要求从市场最爱的避风港手中剥离 Chrome；FTC 对 Meta 的上诉在进行中；苹果案正走向庭审；Nvidia 的 Groq 授权交易正被调查F。一个政府不可能一边试图拆分某座避风港、一边「钦点」它。与此同时，它刚安插的美联储——沃什，5 月以 54 比 45 险过——在 CPI 4.2% 的背景下把利率钉在 3.5–3.75%，且三位 FOMC 委员暗示下一步是加息F。而财政部内部，一份把 AI 板块类比互联网泡沫的职业分析师草稿，同一周被公开斥为「未经审核」F——这个国家机器甚至对要不要相信自己的交易都未达成一致。
The house law applies, and it cuts against the comfortable version. The AI Master Plan already adjudicated this: every administration wants markets up into midterms, yet midterm-year lows cluster in September–December anyway, because the Fed plus the cycle outrank political will — 2018 being the controlling analog: a new hawkish Fed chair, a memory-supercycle peak, a president publicly demanding higher prices, and a −20% Q4 regardless. Worse for the naive version: a legible support schedule gets front-run. If everyone knows Washington needs the tape green until November 3, the rational move is to distribute before November — which is precisely what the June–July tape looks like. The administration’s calendar marks the exit-liquidity windows, not the floor. The floor arrives where it always does: at the Fed’s capitulation (the October 27–28 FOMC is the earliest plausible venue, six days before the election) or at the credit event that forces it — and if that event comes, the state’s documented playbook is not to save shareholders but to buy strategic assets at marked-down prices through the sovereign-fund machinery it is assembling in plain sight.本台之法在此适用，而且切向不舒服的那一边。《AI 总体规划》早已裁定：每届政府都想让市场撑到中期选举，但中期选举年的低点照样扎堆在 9–12 月，因为美联储加周期的权力高于政治意志——2018 年是决定性类比：新任鹰派联储主席、存储超级周期顶部、一位公开喊话要股价的总统，四季度照样 −20%。对天真版本更不利的是：一份清晰可读的支撑时间表会被抢跑。如果所有人都知道华盛顿需要盘面绿到 11月3日，理性选择就是在 11 月之前派发——这正是 6–7 月盘面的模样。政府的日历标记的是出货流动性窗口，不是地板。地板出现在它一贯出现的地方：美联储投降之时（10月27–28日 FOMC 是最早的可信场合，距选举六天），或逼出投降的信用事件——而若该事件到来，国家机器有据可查的剧本不是拯救股东，而是通过它正在光天化日之下组装的主权基金机制，以折价收购战略资产。
One policy channel deserves separate billing, because it is the only one that could hand a specific company revenue directly: model sovereignty. Chinese models — DeepSeek at 17.6% of routed API tokens, Qwen at 13.9%, the Chinese-origin share peaking at 46% — took that share on pricing 60–90% below US frontier ratesC. Washington is moving against exactly that: federal device bans are in force at agency level with statutes advancing, several states have banned Chinese models from government networks, and on July 8 the State Department was reported preparing restrictions on corporate use of Chinese AI, with two House committees opening a joint investigation the same weekC. If a formal rule lands, roughly a third of the token market needs a new low-cost home — and Muse Spark, priced at 3–4x DeepSeek but ~25% of US frontier, is the only American model sitting in the vacated gapC. This is the cleanest case where administration interest and a single stock’s revenue line actually intersect — and it still is not stock-picking: it is a restriction on adversary products whose commercial benefit falls to whoever priced for it. Today it books zero revenue; the tell that converts it from thesis-color into a catalyst is a signed rule or executive order, not a press report.有一条政策通道值得单独记账，因为它是唯一可能把收入直接递到某一家公司手里的：模型主权。中国模型——DeepSeek 占路由 API token 的 17.6%、Qwen 占 13.9%、中国系份额一度触及 46%——靠低于美国前沿模型 60–90% 的定价拿下了这块份额C。华盛顿正对准这一点行动：联邦机构层面的设备禁令已经生效、立法在推进，多个州已禁止政府网络使用中国模型，而 7月8日 据报国务院正筹备限制企业使用中国 AI，同周两个众议院委员会开启联合调查C。若正式规则落地，约三分之一的 token 市场需要一个新的低价归宿——而定价为 DeepSeek 3–4 倍、却仅为美国前沿模型约 25% 的 Muse Spark，是唯一站在这个真空缺口里的美国模型C。这是「政府利益」与「单一股票收入科目」真正相交的最干净案例——但它依然不是选股：它是对敌对方产品的限制，商业利益落给恰好为此定了价的人。今日它入账收入为零；把它从论据背景色变成催化剂的信号，是一纸签署的规则或行政令，不是一篇报道。
7The two-phase answer两阶段答案
Three operational notes bind the calendar to the book. One: the correction is not yet confirmed — our risk radar still reads RISK_ON at 79/100 as of July 10, trend and volatility gates greenF, and the index sits within a few percent of its high. The two-phase plan does not require calling the top; it requires refusing to spend October’s cash in July. Two: the memory complex has already broken first, on schedule — all four names peaked June 18–25 and sit 12–28% below those peaksF with the Samsung/SK Hynix session of July 7 as the KOSPI tellC — and the house discipline stands: bounces there are for selling, and the trade re-enters on the credit tape at the Phase-2 low, not on DRAM prints. Three: the freshest sentiment reads are already half-way to the fear the window needs — AAII bears at ~42%, FMS cash still thin at 4.1%, “long semis” the most crowded trade at 80% consensusC. The crowd knows which trade is crowded; it has not yet been made to pay for it. That payment is the October entry.三条操作注记把日历和账本钉在一起。其一：调整尚未确认——本站风险雷达截至 7月10日 仍为 RISK_ON、79/100，趋势与波动率闸门皆绿F，指数距高点仅数个百分点。两阶段计划不要求你猜顶；它只要求你拒绝在七月花掉十月的现金。其二：存储板块已按剧本率先破位——四只标的全部在 6月18–25日 之间见顶、现距峰值 12–28%F，7月7日三星／SK 海力士的 KOSPI 熔断日即为信号C——本台纪律不变：那里的反弹用来卖；再入场看第二阶段低点的信用行情，不看 DRAM 报价。其三：最新情绪读数已经走到窗口所需恐惧的半路——AAII 空头约 42%、FMS 现金仓位仍低至 4.1%、「做多半导体」以 80% 共识度列最拥挤交易C。人群知道哪笔交易拥挤；只是还没被迫为此付账。那笔账单，就是十月的入场券。
8Falsification dashboard证伪仪表盘
| Claim判断 | What kills it什么能杀死它 | What confirms it什么能确认它 |
|---|---|---|
| Meta = the asymmetric-haven candidateMeta = 不对称避风港候选人 | 2027 capex guided vertical with zero compute-revenue disclosure (house kill, unchanged); subscription rollout reversed on user backlash; Watermelon-class next model slips into 2027 with Spark adoption stalling; or deal news failing the grading rubric — short-term, discounted, distressed-counterparty leases read as an internal-demand admission, not proof.2027 资本开支指引拉垂直且无算力收入披露（本台原判，不变）；订阅因用户反弹而回撤；下一代旗舰模型拖入 2027 且 Spark 采用停滞；或交易头条未通过评级标准——短期、打折、困境对手方的租约应读作内需不足的自供状，而非证据。 | A rubric-passing contracted number — multi-year term, quality counterparty, scarcity-anchored pricing, disclosed backlog accounting — or a subscriber print; capex guide held at Q2; equity raise resolved (done at a level, or shelved). Fusion confirmation: absorbing capital on index-down days after the disclosure.一份通过评级标准的合同数字——多年期、优质对手方、锚定稀缺费率、披露在手订单科目——或一个订户数字；二季报维持资本开支指引；股权融资了结（定价完成或搁置）。融合确认：披露之后在指数下跌日持续吸纳资金。 |
| AAPL = the consensus shelter (and its exit trap)AAPL = 共识庇护所（及其出口陷阱） | For the shelter role: iPhone 18 US price hikes visibly dent September preorders; a China demand shock. For the exit-trap warning: AAPL continuing to lead after a Meta disclosure lands would refute the rotation-funding mechanism.就庇护角色而言：iPhone 18 美国提价明显打击九月预订；中国需求冲击。就出口陷阱警示而言：若 Meta 披露落地之后 AAPL 仍持续领涨，即推翻「轮动供血」机制。 | Shelter: continued ≥70% down-day beat rate; the September event passing price hikes without unit guidance cuts. Exit trap: two consecutive weeks of down-day underperformance vs QQQ once a leader emerges — the Dec–Jan pattern repeating.庇护：下跌日跑赢率维持 ≥70%；九月发布会完成提价且不下调销量指引。出口陷阱：领袖现身后对 QQQ 的下跌日表现连续两周转弱——12–1月的图案重演。 |
| NVDA = rebound vehicle, not shelterNVDA = 反弹载具，非庇护所 | DC growth <25% y/y two straight quarters, or a top-3 customer publicly moving majority orders in-house (house kill, unchanged); a neocloud credit event triggering its backstops turns the floor into a trapdoor.数据中心同比增速连续两季 <25%，或前三大客户公开转向自研（本台原判，不变）；新兴云信用事件触发其兜底义务，地板变活门。 | Rubin Q3 shipments to the eight named clouds on schedule; positioning washout completing (insider-sell clusters ending, 13F re-adds).Rubin 三季度按期向八家具名云厂商出货；仓位出清完成（内部人卖出集群结束、13F 重新增持）。 |
| The October window itself十月窗口本身 | The ~20% grind-higher path: soft Jul/Aug CPI + tariff relief + no credit accident + a clean $1T IPO, all four together. Or Scenario C: earnings re-ignite Act I immediately — in which case strength is distribution fuel, not refutation.约 20% 概率的「缓步走高」路径：7/8 月 CPI 温和 + 关税放松 + 无信用事故 + 1 万亿美元 IPO 顺利，四者齐备。或情景 C：财报立即重燃第一幕——彼时强势是派发燃料，不是反驳。 | A hike delivered in Sep/Oct, or the credit tape cracking (Oracle CDS, neocloud refi, ABS spreads) — either drags the low deeper but confirms the window’s location.9／10 月落地加息，或信用行情破裂（Oracle CDS、新兴云再融资、ABS 利差）——两者都会加深低点，但都确认窗口的位置。 |
| Washington = tailwind, not selector华盛顿 = 顺风，非选股人 | A documented federal act that singles out and supports a specific listed equity outside a crisis (not a sector program) — that would upgrade the coordination claim from folklore to fact.出现有据可查的联邦行为：在非危机状态下点名支持某只特定上市股票（而非行业计划）——那将把「协调」说从传言升格为事实。 | The pattern holding: pro-buildout programs + live antitrust suits against the same names + a hike-leaning Fed; state equity entries occurring only at distress prices.当前图案延续：支持建设的计划 + 对同批公司的反垄断诉讼 + 倾向加息的美联储；国家入股仅发生在危机折价处。 |
There are two havens in every correction: the one the crowd already owns, and the one proof builds mid-storm. Apple is paid for not losing; the asymmetric haven is paid when the paper prints. Google showed the order of payment — and showed that the consensus shelter picks up the bill. Meta’s invoices start dating from July 29; what lands on paper between then and October decides whether it takes the role through the storm or collects it at the low. 每一轮调整里都有两种避风港：一种是人群早已持有的，另一种是证据在风暴中途现造的。Apple 挣的是「不输」的钱；不对称避风港挣的是「纸面兑现」的钱。谷歌示范了付款顺序——也示范了最后由共识庇护所来买单。Meta 的发票从 7月29日 起开始落款；从那时到十月之间写上纸面的东西，决定它是顶着风暴接下角色，还是在低点领取角色。
Sources & method资料来源与方法
Method方法: This report was built from a twelve-lane evidence sweep (local price stores, the ownership/risk/basket engines, and web-source dossiers on Meta, Alphabet, Nvidia, Apple, Tesla, the memory cycle, administration policy, model-sovereignty restrictions, and midterm statistics), tested against four independent adversarial reviews — one attacking the Meta-as-haven thesis, one attacking the administration-coordination claim, one auditing cross-source factual consistency, and one attacking the asymmetric-haven framing itself (its strongest objections — the single-episode nature of the Dec–Jan exhibit, the two-sided reading of compute-deal news, and the unretired bear facts — are incorporated in the text rather than argued away). Claims are labeled [F] store- or filing-checked fact, [C] corroborated across independent sources, [I] inference, [S] desk estimate. Numbers that failed the consistency audit (single-source July 9–10 closes, one vendor’s Micron revenue figure, conflicting consensus-EPS prints) are either hedged in text or excluded.本报告基于十二条证据通道的扫描（本地价格库、持仓／风险／篮子引擎，以及关于 Meta、Alphabet、Nvidia、Apple、Tesla、存储周期、政府政策、模型主权限制与中期选举统计的网络来源档案），并经四路独立对抗性审查检验——其一攻击「Meta 即避风港」论点，其二攻击「政府协调」说，其三审计跨来源事实一致性，其四攻击「不对称避风港」框架本身（其最有力的异议——12–1月案例的单一性、算力交易头条的双向读法、以及未解除的空头事实——均已写入正文，而非辩解掉）。声明标注为 [F] 价格库或申报文件核实、[C] 多来源交叉印证、[I] 推断、[S] 本台估计。未通过一致性审计的数字（单一来源的 7月9–10日收盘价、某供应商的美光营收数字、互相矛盾的共识每股盈利）已在正文中加以对冲或剔除。
Local data本地数据: data/stocks and massive_stock_day parquet price stores (last bars 2026-07-08 / 2026-07-02); site risk_state.json (as-of 2026-07-10); smartmoney 13F desk (holdings as-of 2026-03-31 — one quarter stale, next window Aug 14); baskets.json (as-of 2026-07-09).data/stocks 与 massive_stock_day parquet 价格库（末根 K 线 2026-07-08／2026-07-02）；站内 risk_state.json（截至 2026-07-10）；smartmoney 13F 台（持仓截至 2026-03-31——滞后一个季度，下一窗口 8月14日）；baskets.json（截至 2026-07-09）。
Primary & press一手来源与报道: Meta Q1-2026 results and call (capex guide, memory-cost citation); Bloomberg (Meta Compute, Jul 1); Meta AI blog (Muse Spark 1.1 API, Jul 9); FT/Bloomberg (equity-raise consideration, Jun 5); Reuters (Iris/Broadcom September production; Zuckerberg town-hall remarks, Jul 2); Alphabet Q1-2026 8-K (Cloud, backlog) and the Sep 2 2025 remedies ruling; NVDA 8-K/10-Q FQ1-FY27; Apple Q2-FY26 8-K and Gemini-Siri partnership coverage; Tesla Q2 delivery 8-K; White House AI Action Plan and EO 14318; CNBC (OpenAI stake proposal, Jul 2; Trump disclosure analysis; State Department corporate-use restriction reports and OpenRouter token-share data, Jul 7–8); NOTUS (Treasury draft, Jul 6); Congress.gov (H.R.1121, No Adversarial AI Act); trade.gov / Federal Register (American AI Exports Program, Apr–Jun 2026); Unitree/UBTech shipment and pricing reporting + MOFCOM rare-earth-magnet announcements 18/61/62/70 (robot supply chain); BofA Global FMS (Jun 2026); Cboe/AAII sentiment series.Meta 2026 年一季报及电话会（资本开支指引、存储成本表述）；彭博（Meta Compute，7月1日）；Meta AI 博客（Muse Spark 1.1 API，7月9日）；FT／彭博（股权融资酝酿，6月5日）；路透（Iris／博通九月投产；扎克伯格全员会讲话，7月2日）；Alphabet 2026 年一季度 8-K（云业务、在手订单）及 2025年9月2日 救济裁决；NVDA FQ1-FY27 8-K／10-Q；Apple FY26 二季度 8-K 及 Gemini-Siri 合作报道；Tesla 二季度交付 8-K；白宫《AI 行动计划》与 EO 14318；CNBC（OpenAI 入股提案，7月2日；特朗普持仓披露分析；国务院企业使用限制报道与 OpenRouter token 份额数据，7月7–8日）；NOTUS（财政部草稿，7月6日）；Congress.gov（H.R.1121、《无敌对 AI 法案》）；trade.gov／联邦公报（美国 AI 出口计划，2026年4–6月）；宇树／优必选出货与定价报道 + 商务部稀土磁材第 18/61/62/70 号公告（机器人供应链）；美银全球基金经理调查（2026年6月）；Cboe／AAII 情绪序列。
Sibling reports姊妹报告: The Second Act《第二幕》 · The Relapse《复燃》 · The AI Master Plan《AI 总体规划》 — this report extends their rulings to the safe-haven question and changes none of them.本报告将上述裁定延伸至避险问题，且不改动其中任何一条。
Three externally generated AI briefs on the same question were used as reference opinions only; their factual claims were re-derived independently. Where they disagreed with our stores (Meta’s July closes, the shape and dating of Google’s rerating), the stores and filings won. This is a scenario research memo, not financial advice; probabilities marked [S] are desk estimates.关于同一问题的三份外部 AI 简报仅作参考观点使用；其事实性声明均经独立重验。凡与本站数据库冲突之处（Meta 七月收盘价、谷歌重估的形态与日期），以数据库与申报文件为准。本文为情景研究备忘录，非投资建议；标注 [S] 的概率为本台估计。