The Second Act第二幕

0How to Read This Document如何阅读本文

Evidence labels. Every material claim in this document carries one of four tags. They tell you how much to trust each statement:证据标签。文中每一项实质性论断都带有以下四类标签之一，用来告诉你该论断有多可靠：

• ^F — Fact. Verified against primary sources (original documents or data) or high-quality secondary sources from the last 90 days. The verification log is in Appendix A.——事实：经一手来源（原始文件或数据）或高质量二手来源核实，来源日期在最近 90 天以内。核实记录见附录 A。
• ^C — Consensus/reported. Reported by credible sources, but not independently confirmed by us. Or it comes from what we knew before February 2026, which may have changed since.——共识／报道：来自可信来源，但未经我们独立确认；或来自 2026 年 2 月之前的认知，其后可能已有变化。
• ^I — Inference. Our own reasoning, built from how things work, who wants what, and what has happened before. This is where the alpha (returns beyond what the market hands everyone) lives. It is also where we can be wrong.——推断：我们自己的推理，依据是运作机制、各方动机与历史先例。alpha（超越市场平均的超额收益）来自这里，我们出错的可能也在这里。
• ^S — Speculation. Bets on stories the market has not priced in yet, where the potential win is far larger than the potential loss. We keep these positions small.——投机：押注市场尚未定价的叙事——赢面远大于亏面的非对称押注。这类仓位我们刻意做小。

On monthly precision. This note maps a market path month by month through the end of 2027. We owe you honesty about what that can and cannot mean. Precise price forecasts at that level of detail have almost no real predictive power. Nobody's do. Anyone selling them is selling astrology. What does work is a conditional catalyst lattice — a map of scheduled events and how they connect. The events themselves are facts: Fed meeting dates (FOMC), tariff expirations, elections, earnings, treaty deadlines. The cause-and-effect links between them can be analyzed. The odds of each branch can be honestly estimated. That is what Section 5 delivers: a month-by-month base case (our most likely path), with clear branch points and the early signals that tell you which branch you are on before the consensus catches up. Treat the base case as a spine to deviate from, not a prophecy.关于月度精度。本文把市场路径逐月推演到 2027 年底。这种做法能说明什么、不能说明什么，我们有义务讲清楚。如此精细的点位预测，几乎没有任何真实的预测力。任何人的都没有。兜售这种预测的人，卖的是占星术。真正有用的，是一张条件化催化剂网格——一份日程性事件及其相互关联的地图。事件本身是事实：FOMC（美联储议息会议）日期、关税到期、选举、财报、条约期限。连接这些事件的因果机制可以分析。每条分支的概率可以诚实估计。这正是 §5 交付的内容：一条逐月展开的基准情形（我们认为最可能的路径），附带明确的分支节点，以及让你先于市场共识识别自己身处哪条分支的早期信号。请把基准情形当作可以据以偏离的主轴，而不是一则预言。

One housekeeping line, then no more disclaimers: this is research and scenario analysis, not personal investment advice. Position size, suitability, and execution are yours. That is the last disclaimer in this document.例行说明一句，此后不再重复：本文是研究与情景分析，不是针对个人的投资建议。仓位大小、是否适合、如何执行，均由阁下自行负责。这是本文档中的最后一条免责声明。

1Executive Summary — The Ten Calls执行摘要——十大判断

2State of Play — July 2, 2026局势盘点——2026年7月2日

2.1 · The tape (the market scoreboard)2.1 · 盘面（市场记分牌）

The most telling pattern on this board: the Dow at a record, chip stocks in freefall, and small caps having their best half in 35 years. That is not "the AI bubble popping and taking everything with it." It is a change of leadership inside a bull market that is still running — money is rotating, not fleeing. This is what 1998–99 looked like. It is what every mid-cycle pause to digest capex (capital spending) has looked like since the railroads. Whether it stays a rotation or turns into a 2000-style bust depends on credit (Section 4).这张盘面上最有诊断价值的形态：道指创纪录、半导体自由落体、小盘股录得 35 年来最佳上半年。这不是“AI 泡沫破裂并拖垮一切”。这是一轮仍在进行的牛市内部的领导权换班——资金在轮动，不是在出逃。1998–99 年就是这个样子；自铁路时代以来，每一次周期中段的资本开支（capex，企业的大额投资支出）消化期也都是这个样子。它是保持轮动、还是恶化成 2000 年的翻版，取决于信用这条分支（§4）。

2.2 · What actually happened in the first half of 2026 (the short version)2.2 · 2026 年上半年究竟发生了什么（速览）

The half-year packed three distinct market phases into five months:这半年在五个月里塞进了三个截然不同的市场阶段：

Phase 1 — the double shock (Feb 20 – late March). On February 20 the Supreme Court struck down the IEEPA tariffs (tariffs imposed under an emergency-powers law) by 6-3. Roberts wrote the opinion, joined by Gorsuch, Barrett, and the three liberals^F. Within hours the administration re-imposed a 10% global surcharge under Section 122 of the Trade Act of 1974 — the first use of that law in fifty years. It took effect February 24, with a legal maximum of 150 days^F. Four days later, on February 28, the US-Israel war on Iran began^F. Iran used mines and missiles to effectively close the Strait of Hormuz (a vital oil-shipping channel). The US blockaded Iranian ports in return from mid-April. The IEA estimated the disruption at 14 million barrels a day at its peak^F. Stock indexes avoided an official correction (a drop of 10% or more) in March. But under the surface it was a bloodbath: the average S&P member fell 21% at some point in the first half^F.阶段一——双重冲击（2月20日–3月下旬）。2月20日，最高法院以 6–3 推翻 IEEPA 关税（依据紧急权力法加征的关税）。Roberts 主笔，Gorsuch、Barrett 与三位自由派大法官加入^F。数小时内，政府依据《1974 年贸易法》第 122 条重新加征 10% 全球附加关税——该条款五十年来首次启用——2月24日生效，法定期限上限 150 天^F。四天后，即2月28日，美以对伊朗的战争爆发^F。伊朗用水雷和导弹把霍尔木兹海峡（全球关键的石油运输通道）逼至事实性关闭。美国自 4 月中旬起反向封锁伊朗港口。IEA 估算峰值时供应中断达 1,400 万桶／日^F。3 月股指层面躲过了正式回调（跌幅 10% 以上）。但表面之下是一场屠杀：标普成份股平均个股在上半年某一时点回撤达 −21%^F。

Phase 2 — the war melt-up (April – June 2). A melt-up is a sharp rally driven by fear of missing out. It sounds backwards, but it follows the classic script (1990–91 did the same): the S&P made its first record high of the year on April 15, during the war^F. A hyper-concentrated group led: energy, defense, and the AI-scarcity trade. The war supercharged that trade — a hot war accelerates demand for computing power, and the inflation it caused made "real assets plus pricing power" (memory chips, electric power, semiconductors) the only game in town. Participation was razor thin: only ~17% of S&P members beat the index over the month into early June, a decade low^F. Then first-quarter earnings (late April) delivered the capex shock. The Big-4 tech giants' 2026 spending guidance converged on ~$725B, up 77% year over year — Amazon ~$200B, Microsoft ~$190B, Alphabet raised to ~$190B, Meta raised to $125–145B, citing memory-chip and component inflation^F. Goldman lifted its 2025–30 projection of cumulative hyperscaler capex (the total capital spending of the biggest cloud companies) from $4.5T to $5.3T^F. Memory became ~30% of hyperscaler data-center spend — four times its 2023 share^F.阶段二——战争融涨（4月–6月2日）。融涨（melt-up）指由“怕错过”情绪推动的急涨。听着反直觉，却符合经典剧本（1990–91 年也是如此）：标普在战争进行中的4月15日创下年内首个历史新高^F。领涨的是一个高度集中的板块群：能源、国防与 AI 稀缺性交易。战争进一步为其加力——热战会加速算力需求，而它引发的通胀让“实物资产 + 定价权”（即存储芯片、电力、半导体）成为唯一的选择。参与度极窄：截至 6 月初的一个月内，仅约 17% 的标普成份股跑赢指数，为十年最低^F。随后一季报（4月下旬）投下资本开支冲击：四大科技巨头 2026 年开支指引收敛至约 $725B，同比 +77%——Amazon 约 $200B、Microsoft 约 $190B、Alphabet 上调至约 $190B、Meta 上调至 $125–145B，理由是存储芯片与元器件涨价^F。高盛把 2025–30 年超大规模云厂商（即最大的几家云计算公司）累计资本开支预测从 $4.5T 上调至 $5.3T^F。存储已占其数据中心开支约 30%，是 2023 年份额的 4 倍^F。

Phase 3 — the crack (June 2 – now). From the June 2 record, the AI complex began to fail, piece by piece. First, Broadcom missed on AI networking: ~$4.1B against a ~$4.8B consensus, and it did not raise its 2027 outlook — the cycle's first true demand-side miss. That triggered the Nasdaq's first −4% day and wiped ~$1.3T off chip stocks in a single session^C. Second, the June 17 Fed meeting — new chair Warsh's debut at the FOMC (the Fed's rate-setting committee) — came out hawkish (leaning toward higher rates; "dovish" is the opposite): the easing bias was deleted, and the dot plot (each Fed official's rate forecast) flipped to project a hike^F. That removed the support under valuations. Third, the last ten days brought the thematic top signals. OpenAI filed a confidential S-1 (the paperwork to go public) in June, with advisers already discussing a delay to 2027^F. SpaceX (+xAI) went public in mid-June at a $1.75–2T target valuation, the largest in history^F. Meta announced "Meta Compute" — selling surplus GPU capacity and hosted-model APIs; META rose 8.8% on July 1 while the neoclouds (young AI-focused cloud providers) fell hard on the news: CRWV −14%, NBIS −17%^F. The same day brought reports that OpenAI engineers unlocked big gains in inference efficiency (making AI models much cheaper to run)^C — together, a two-headed scare about future chip demand. Then the July 2 FT report: OpenAI is discussing a 5% equity stake for the US government^F. Semis are now down hard for consecutive sessions, with Micron −10% and −7% back-to-back^F — while the Dow prints records.阶段三——裂缝（6月2日–至今）。自6月2日的纪录高点起，AI 板块开始一块块失灵。第一，Broadcom 的 AI 网络业务不及预期：约 $4.1B 对约 $4.8B 的一致预期，且未上调 2027 年展望——本周期第一次真正的需求端失手。它触发了纳斯达克首个 −4% 交易日，单日抹掉约 $1.3T 芯片市值^C。第二，6月17日 Warsh 首秀的 FOMC（联储议息委员会）转向鹰派（倾向加息；“鸽派”则相反）：删除宽松倾向，点阵图（每位联储官员的利率预测）翻向加息^F。这抽走了估值的支撑。第三，最近十天出现了主题层面的见顶信号：OpenAI 于 6 月保密递交 S-1（上市申请文件），而顾问已在讨论推迟到 2027 年^F；SpaceX（+xAI）6 月中旬以 $1.75–2T 目标估值上市，为史上最大^F；Meta 宣布 “Meta Compute”，出售富余 GPU 算力与托管模型 API——META 7月1日 +8.8%，新兴云（专做 AI 算力的新云计算公司）应声重挫：CRWV −14%、NBIS −17%^F；同日又有报道称 OpenAI 工程师大幅提升了推理效率（让 AI 模型的运行成本大降）^C——两件事合成一场对未来芯片需求的双头恐慌；随后7月2日 FT 报道 OpenAI 正讨论向美国政府让渡 5% 股权^F。半导体已连续数日重挫——Micron 连续两日 −10% 与 −7%^F——而道指同时刷新纪录。

2.3 · The war and its wind-down — the biggest macro force for H22.3 · 战争及其退潮——主导下半年的最大宏观变量

Timeline (all points ^F): war starts Feb 28 → Iran effectively closes the Strait of Hormuz from late Feb → US blockades Iranian ports mid-April → Brent peaks at $126 April 30 → June 11: Trump cancels planned strikes and signals a deal → June 17: a memorandum of understanding (MoU — a preliminary agreement) is signed, with Pakistan's PM Sharif mediating and Qatar assisting. The Strait reopens; the US blockade lifts; a 60-day window opens to negotiate a permanent deal (deadline ~August 16), with Iranian transit toll-free for 60 days → 40–50M barrels of Iranian crude are exported in the first two weeks^F. Tehran claims it is getting premium prices. Market data say otherwise: Iranian Light's discount to Brent has actually widened ($2.50–5, versus ~$1 before the deal), and more than 20M barrels sit unsold off Asia^C. That is a glut signal, not a scarcity signal → Brent ~$71.5 on July 2, below the February 27 pre-war close^F.时间线（各节点均为^F）：2月28日开战 → 2 月底起伊朗事实上关闭霍尔木兹海峡 → 4 月中旬美国封锁伊朗港口 → 4月30日布伦特见顶 $126 → 6月11日 Trump 取消既定打击计划、释放协议信号 → 6月17日签署谅解备忘录（MoU，即初步框架协议），由巴基斯坦总理 Sharif 斡旋、卡塔尔协助。海峡重开；美国解除封锁；开启 60 天永久协议谈判窗口（截止日约8月16日），伊朗过境 60 天免通行费 → 头两周伊朗出口原油 4,000–5,000 万桶^F。德黑兰宣称卖出了溢价。市场数据却相反：Iranian Light 对布伦特的贴水（低于基准的折价）实际在走阔（$2.50–5，协议前约 $1），且超过 2,000 万桶滞销漂在亚洲外海^C。这是过剩信号，不是稀缺信号 → 7月2日布伦特约 $71.5，低于2月27日战前收盘^F。

Three consequences the market has not finished digesting:市场尚未消化完的三个后果：

1. The inflation impulse is mechanically reversing. May CPI came in at 4.2% headline, 2.9% core (core excludes food and energy)^F. The 1.3-percentage-point gap between the two is almost entirely energy. Crude has round-tripped, and the June ISM survey shows wholesale energy back at pre-war levels^F (refined products like gasoline lag, but they follow). So the July–October headline prints inherit a negative energy contribution, plus friendly base effects (this year's numbers get compared against last year's high ones). Every 2025-style "inflation is re-accelerating" story dies between August and November. This is close to simple arithmetic.^I通胀脉冲正在机械性逆转。5 月 CPI：整体 4.2%、核心 2.9%（核心 = 剔除食品与能源）^F。两者之间 1.3 个百分点的差距几乎全部来自能源。原油已完成往返，6 月 ISM 调查显示批发能源价格回到战前水平^F（汽油等成品油滞后，但会跟上）。因此 7 月至 10 月的整体通胀读数将继承负向的能源贡献，外加友好的基数效应（今年的数字对比的是去年的高基数）。每一个 2025 年式“通胀再加速”的故事都会在 8 月至 11 月之间死掉。这近乎简单算术。^I
2. The hawkish Fed pricing is a wasting asset — it loses value with time. One soft jobs report (June: +57K jobs; U-3 unemployment down to 4.2% only because people left the workforce) has already cut September hike odds from 67% to below 50%^F. And the dots that flipped hawkish on June 17 were marked when refined-product prices were at their peak.^I鹰派联储定价是一项随时间贬值的资产。仅一份疲软的就业报告（6 月：新增 +5.7 万；失业率 U-3 降到 4.2%，只是因为有人退出了劳动力市场），就已令 9 月加息概率从 67% 跌破 50%^F。而6月17日翻鹰的点阵图，是在成品油价格的峰值时点写下的。^I
3. The deal is not done. The MoU is an interim framework. The two sides traded blows after signing it: June 26–28 brought a tanker hit, two nights of US strikes on Iranian coastal sites, and IRGC (Iran's Revolutionary Guard) missiles on US bases in Kuwait and Bahrain, before talks resumed in Doha^F. Iran claims sole administrative control of the Strait^F. BIMCO (the big shipping association) still calls transits very risky^F. The ~August 16 permanent-deal deadline is a real yes/no moment. Failure would not necessarily mean full war. But it would re-widen the oil risk premium and delay the disinflation. This is the main near-term fork in the road. Our odds: durable de-escalation 60% — though likely only after the deadline slips (prediction markets price ~74% odds of a permanent deal by December, but barely a coin flip by August^C). Frozen semi-conflict with flare-ups: 30%. Full resumption: 10%.^I协议并未落定。MoU 只是临时框架。双方在签署之后仍互有打击：6月26–28日，先有油轮遇袭，接着美军连续两夜打击伊朗沿海目标，伊朗革命卫队（IRGC）导弹袭击美军驻科威特与巴林基地，随后谈判才在多哈恢复^F。伊朗宣称对海峡拥有唯一行政管辖权^F。航运组织 BIMCO 仍把过境评为高风险^F。约8月16日的永久协议截止日是一个真实的“成或不成”节点。谈崩不必然意味着全面复战，但会重新拉宽石油风险溢价、推迟反通胀进程。这是近期最主要的分叉口。我们的赋值：持久降级 60%——尽管很可能要等截止日延期之后才实现（预测市场对 12 月前达成永久协议定价约 74%，但对 8 月前仅略强于抛硬币^C）。冻结型半冲突、间歇交火：30%。全面复战：10%。^I

2.4 · The Fed under Warsh — read the man, not the dots2.4 · Warsh 治下的联储——读人，不读点阵图

Facts first (all ^F). Warsh was nominated in early 2026. The DOJ's criminal probe of Powell — opened in January over his Eccles-renovation testimony — was dropped in April, unblocking Senate confirmation (May 13, on party lines). He was sworn in May 22 and took the Miran board seat (term to 2040). Powell remains a governor until January 2028 — a competing power center without precedent since 1948. The Supreme Court blocked the Cook firing in June. Bostic resigned in February. The June 16–17 FOMC: a unanimous hold at 3.50–3.75%. The statement was cut to ~130 words. Forward guidance (telling markets in advance where rates are headed) was abolished. The official projections (the SEP): 2026 PCE inflation (the Fed's preferred gauge) 3.6%, core 3.3%, GDP 2.2%. The median dot implies one 2026 hike; 9 of 18 participants see hikes, six see two. Warsh submitted no dot of his own. He launched five task forces (operations, communications, data, productivity/labor, inflation causes), reporting by ~year-end. At Sintra (June 30–July 1) he said inflation expectations have eased and saw "no immediate case for a July hike" — while insisting "prices are too high" and refusing to give forward guidance^F. On the Fed's balance sheet he said only that it is too large — it "borders on fiscal policy" — and that any change will be telegraphed until markets fully understand it^F. (A verification note: contrary to some early reporting, he did not float outright bond sales at Sintra. Active sales — most plausibly mortgage bonds first, toward a smaller portfolio holding only Treasuries — are the street's expectation and long-standing Warsh doctrine, not a live proposal^C.)先摆事实（均为^F）。Warsh 于 2026 年初获提名。DOJ 对 Powell 的刑事调查（1 月因 Eccles 大楼翻修证词立案）4 月撤销，参议院确认随之解锁（5月13日，沿党派线表决）。他5月22日宣誓就任，接任 Miran 的理事席位（任期至 2040 年）。Powell 以理事身份留任至 2028 年 1 月——1948 年以来首见的竞争性权力中心。最高法院 6 月阻止了解雇 Cook。Bostic 2 月辞职。6月16–17日 FOMC：一致同意按兵不动于 3.50–3.75%。声明压缩至约 130 词。废除前瞻指引（提前告诉市场利率去向的做法）。官方预测（SEP）：2026 年 PCE 通胀（联储偏好的通胀指标）3.6%、核心 3.3%、GDP 2.2%。点阵图中位数隐含 2026 年加息一次；18 人中 9 人预期加息、6 人预期加息两次。Warsh 本人没有提交点阵。他启动五个工作组（操作、沟通、数据、生产率／劳动力、通胀成因），约年底前提交报告。在辛特拉（6月30日–7月1日），他说通胀预期已经回落、“没有 7 月加息的迫切理由”——同时坚称“物价太高”，并拒绝给出前瞻指引^F。关于联储资产负债表，他只说规模过大——“近乎财政政策”——且任何变化都会提前充分沟通、直到市场完全理解^F。（核查注记：与部分早期报道相反，他在辛特拉并未提出直接卖债。主动出售——最可能按揭债券（MBS）先行，走向一个更小、只持国债的组合——是市场预期与 Warsh 的长期主张，不是现行提案^C。）

Our read^I: Warsh is running the standard new-chairman playbook — act tough early to build credibility. Three tells (giveaway signs) say the hawkishness is front-loaded theater, not a lasting stance:我们的解读^I：Warsh 在走新任主席的标准剧本——先示强、攒信誉。有三个破绽表明这轮鹰派是前置的表演，而非长期立场：

• He withheld his own dot and let the committee's hawks fill the grid. He owns the option, not the forecast. A chairman who actually plans to hike anchors the median himself.他扣下了自己的点阵，任由委员会的鹰派去填格子。他持有的是选择权，不是预测。真打算加息的主席会亲自锚定中位数。
• His stated framework — look through supply-shock inflation; AI is structurally disinflationary because it raises productivity — is exactly the doctrine you would build in advance if you planned to cut rates into 3% core inflation in 2027 without calling it surrender.他公开的框架——对供给冲击型通胀“看穿不理”；AI 通过提升生产率带来结构性反通胀——恰恰是那种如果你打算在 2027 年顶着 3% 核心通胀降息、又不想被说成投降时，会提前搭好的学说。
• The task forces (especially "causes of inflation" and "productivity") report by year-end. They are the institutional vehicle for a framework revision — one that redefines the Fed's reaction function just as the calendar turns to the cutting year. New frameworks are how Fed chairs change policy without admitting error.年底前交报告的工作组（尤其“通胀成因”与“生产率”）正是框架修订的制度化载体——恰好在日历翻入降息之年时重写联储的反应函数。新框架从来是联储主席不认错却改政策的方式。

The complication is real, though. This FOMC has the most dissents in three decades^C. Powell votes from the back bench. Six members want two hikes. That is not a committee Warsh commands yet. His path of least resistance is exactly what the calendar offers: hold rates while sounding hawkish, let energy disinflation and a softening job market do the work, let the data retire the hike case through the autumn, and let the December task-force reports rebuild the doctrine for 2027 cuts. Trump wanted a rate-cutter, and the June meeting looked like betrayal. The more likely truth is sequencing: credibility is the currency Warsh is accumulating precisely so he can spend it. The Warsh-dovish-inflection thesis from our earlier work is not dead. It was interrupted by a war. It is now back on schedule, two quarters late.不过，复杂性是真实的。这届 FOMC 的异议票是三十年来最多的^C。Powell 在后排投票。六位委员想加息两次。这还不是一个 Warsh 能号令的委员会。他阻力最小的路径恰是日历给的：嘴上鹰派、按兵不动，让能源反通胀和走软的就业市场去干活，让数据在入秋后逐步瓦解加息论，再让 12 月的工作组报告为 2027 年降息重建学说。Trump 想要一个降息者，6 月会议看似背叛。更可能的真相是先后次序：信誉正是 Warsh 刻意积攒的货币——攒下来就是为了花出去。我们早前研究中的 Warsh 鸽派拐点论没有死。它只是被一场战争打断，如今带着两个季度的延迟重回时间表。

One balance-sheet warning worth isolating. The shrink-the-footprint doctrine^C is a live tail risk (a low-probability but severe scenario) for long-term bonds. Suppose Warsh pairs 2027 rate cuts with active QT-by-sales — actually selling bonds off the Fed's balance sheet. "Cut short rates, shrink the footprint" is a coherent Warshian doctrine. In that case the curve bull-steepens violently (short-term rates fall much faster than long-term ones), and the term-premium event — term premium being the extra yield investors demand to hold long-term bonds — arrives earlier than the 2027–28 window we pencil in Section 9. Watch the December task-force language for exactly this combination.值得单独拎出的资产负债表警示。缩表学说^C是长端债券的一个活体尾部风险（概率低但杀伤大的情景）。设想 Warsh 把 2027 年降息与主动卖出式缩表（真的把债券卖出联储资产负债表）配对——“降短端利率、缩表”是一套自洽的 Warsh 式学说。那样收益率曲线将剧烈牛陡（短端利率下行远快于长端），期限溢价事件（期限溢价 = 持有长期债券所要求的额外收益）会比我们在 §9 勾画的 2027–28 窗口更早到来。12 月请紧盯工作组措辞中是否出现这一确切组合。

2.5 · The tariff machine, rebuilt mid-flight2.5 · 关税机器——飞行中重建

The Supreme Court did not end the tariff regime. It forced a rebuild of the plumbing (all points ^F unless noted):最高法院并没有终结关税体制。它只是逼迫整套管道重建（除另注外均为^F）：

• IEEPA struck down Feb 20 (6-3). All the reciprocal and fentanyl tariffs are void from the start (ab initio). $166B was collected and is now contested^F. The Court of International Trade (CIT) ordered universal refunds in March, and CAPE is actually paying them — ~$23B approved and moving — while the government appeals to the Federal Circuit^F. More than 1,000 follow-on suits are queued at the CIT. The refund flow is already real money. It is a 2026–27 fiscal cost for the government and a corporate windfall concentrated in big importers — an underappreciated earnings tailwind for retail and hardware names, and an underappreciated irritant for Treasury supply (the government must borrow more to pay it).IEEPA 于2月20日被推翻（6–3）。所有对等关税与芬太尼关税自始无效（ab initio）。已征收的 $166B 现处争议之中^F。美国国际贸易法院（CIT）3 月下令全面退税，CAPE 确已在付款——约 $23B 已获批并在流转——同时政府向联邦巡回上诉法院上诉^F。另有 1,000+ 起跟进诉讼排在 CIT。这股退税流已是真金白银：它是政府 2026–27 年的财政成本，也是集中于大型进口商的企业意外之财——对零售与硬件类公司是一条被低估的盈利顺风，对国债供给（政府要多借钱来退款）则是一个被低估的扰动。
• The Section 122 bridge: the 10% global surcharge took effect Feb 24. (A raise to 15% was floated; the most recent reporting says it was never formally adopted — verify live before sizing anything on it.) Annex II exempts ~1,100 product codes; USMCA-qualifying goods (under the US-Mexico-Canada trade pact) are exempt; it does not stack on Section 232 tariffs (Section 232 = the US trade-law provision for national-security tariff investigations). Statutory expiry: July 24, 2026. Congress cannot be bypassed and has not extended it. The CIT invalidated it May 7 (2-1, injunction covering plaintiffs only) and the Federal Circuit stayed that ruling — so it keeps collecting until it dies of natural causes in three weeks.第 122 条过桥安排：10% 全球附加税2月24日生效。（曾有上调至 15% 的风声；最新报道显示从未正式落地——据此建仓前请实时核实。）Annex II 豁免约 1,100 个税则码；符合 USMCA（美墨加贸易协定）资格者豁免；不与232条款关税（232条款 = 美国贸易法中的国家安全关税调查条款）叠加。法定到期日 2026年7月24日——国会绕不开，也没有延期。CIT 于5月7日裁定其无效（2–1，禁令仅限原告），联邦巡回上诉法院暂缓执行——因此它会继续征收，直到三周后自然死亡。
• The replacement is already teed up: a USTR (US Trade Representative) Section 301 "forced-labor" framework covering ~60 economies (~99% of import value) — a 10% tier (Canada, Mexico, EU, UK, Taiwan) and a 12.5% tier (China, Vietnam, India, Japan, Korea, Thailand). Comment deadline July 6. It is designed to land just as Section 122 lapses. Exemptions: USMCA-qualifying goods, products already covered by Section 232, pharma, critical minerals. Most trade lawyers read it as legally sturdier than IEEPA — but litigation is certain.替代方案已经架好：USTR（美国贸易代表办公室）的第 301 条“强迫劳动”框架，覆盖约 60 个经济体（约占进口额 99%）——10% 档（加拿大、墨西哥、欧盟、英国、台湾），12.5% 档（中国、越南、印度、日本、韩国、泰国）。评论期截止7月6日，设计好在 122 条失效之际接续落地。豁免：符合 USMCA 资格者、232条款已覆盖品类、医药、关键矿产。多数贸易律师认为它比 IEEPA 法律根基更稳——但诉讼板上钉钉。
• Sector-by-sector escalation continues regardless. Pharma Section 232 tariffs take effect July 31 (100% on patented drugs and APIs — the active ingredients in drugs^F). A semiconductor Section 232 is already in force — 25% on a narrow set of advanced chips since January 15^F — and a robotics/drones Section 232 is at the decision stage (the robotics report passed its statutory deadline in May; a tariff decision is pending^F). Read that list again: the administration is building tariff walls around exactly the Act II and Act III industries — chips and robots. That is industrial policy written in tariff codes. It tells you where they think the strategic value goes next.行业级加码照旧推进。医药232条款关税7月31日生效（对专利药与原料药（API，药品的活性成分）征 100%^F）。半导体232条款关税已然生效——自1月15日起对一小类先进制程芯片征 25%^F——机器人／无人机232条款调查处于决策阶段（机器人报告已于 5 月超过法定期限；关税决定待发^F）。把这份清单再读一遍：政府正在恰好围绕第二幕与第三幕的产业修筑关税墙——芯片与机器人。这是用税则码写成的产业政策。它告诉你他们认为战略价值下一步去哪。

Net result: the trade-weighted average tariff went from 15.3% (pre-ruling) to 8.1% (at the strike-down) to ~11.4% under the Section 122 bridge^F. The Section 301 regime rebuilds a durable floor around 10–12.5%. The tariff level survives; the tariff chaos premium fades. Businesses get a fairly stable 10% world instead of a 10–145% roulette wheel. That quietly helps profit margins, capex planning, and market breadth (how many stocks are actually going up).^I净结论：贸易加权平均关税从 15.3%（裁决前）→ 8.1%（被推翻瞬间）→ 122 条过桥下约 11.4%^F。301 体制将重建一个约 10–12.5% 的耐久下限。关税水平存续；关税混乱溢价消退。企业得到的是一个大体稳定的 10% 世界，而不是 10–145% 的轮盘赌。这在悄悄利好利润率、利好资本开支规划、也利好市场广度（真正在上涨的股票有多少）。^I

2.6 · The AI complex: demand intact, structure changing2.6 · AI 复合体：需求未损，结构生变

The prevailing bear case — "the AI bubble is popping" — mashes three different things together. Separate them:主流看空论——“AI 泡沫正在破裂”——把三件不同的事搅在一起。把它们拆开：

(a) Demand: not popping — still limited by supply. Google Cloud's backlog (revenue already under contract but not yet delivered) hit $460B, doubling in a single quarter^F. Microsoft's commercial RPO (its version of the same backlog measure) is $627B; its AI revenue run-rate (current pace, annualized) is $37B, growing 123% year over year; and CFO Amy Hood attributed $25B of the capex raise to memory and component cost, not volume^F. AWS grew 28%, with Trainium at a $20B run-rate^F. OpenAI is at ~$25B annualized revenue and crossed 1B monthly active users in June^F/C. The sharper datapoint: Anthropic's run-rate hit $47B in May — passing OpenAI outright^C. The revenue side keeps beating the skeptics' models. Micron: its HBM3E and HBM4 chips (high-bandwidth memory, the specialized chips AI systems need) are fully booked through calendar 2027, with visibility into 2028; HBM4 is in high-volume shipment for Vera Rubin; industry supply stays constrained well beyond 2026; and PC/phone makers are being rationed down low-double-digits because AI ate the wafers^F. Nvidia's January quarter: data-center revenue $62.3B, up 75%; networking up 263%^F. This is not 2001 telecom, where traffic grew 4x a year against capacity built for 100x a year. Capacity still trails demand. One honest amendment, though: after Broadcom's June miss and OpenAI's reported inference-efficiency gains^C, the demand case is now a preponderance of evidence rather than a unanimous verdict. We treat it that way in the tell-matrix (§4.1, §12).(a) 需求：没有破裂——仍被供给卡住。Google Cloud 的积压订单（已签约但尚未交付的收入）达 $460B，单季翻倍^F。Microsoft 的商业 RPO（同类的合同积压指标）$627B；AI 收入年化 $37B、同比 +123%；CFO Amy Hood 把资本开支上调中的 $25B 归因于存储与元器件的成本而非用量^F。AWS 增长 28%，Trainium 年化收入 $20B^F。OpenAI 年化收入约 $25B，月活 6 月突破 10 亿^F/C。更锋利的数据点：Anthropic 年化收入 5 月达 $47B，已彻底反超 OpenAI^C。收入端一直在击败怀疑者的模型。Micron：HBM3E 与 HBM4（高带宽存储，AI 系统必需的专用芯片）已全部订满至 2027 自然年、能见度延伸至 2028 年；HBM4 正为 Vera Rubin 大批量出货；行业供给紧张将持续到远超 2026 年；PC／手机的存储配额被削减低双位数——因为 AI 吃掉了晶圆^F。Nvidia 1 月季报：数据中心收入 $62.3B、+75%，网络业务 +263%^F。这不是 2001 年的电信业——那时流量每年涨 4 倍，产能却按每年 100 倍搭建。产能仍落后于需求。但要诚实补一句：经历 Broadcom 6 月的失手与 OpenAI 据报道的推理效率突破^C之后，需求论如今是证据占优，而非一致裁决。我们在信号矩阵（§4.1、§12）中也如此对待。

(b) Financing structure: deteriorating in exactly the way late cycles do. Capex-to-revenue ratios for 2026: Oracle ~86%, Meta ~54%, Microsoft ~47%, Alphabet ~46%^F. Hyperscaler bond issuance is forecast at ~$93B for 2026, with upside^F. Alphabet — a company that generated cash for two decades — priced an $84.75B equity raise on June 2 (including a $10B Berkshire placement), on top of more than $85B of debt issued in a year: over $170B of external funding^F. Veteran tech investors were shocked ("I never thought Google would need to hit the public markets") — and the shock is itself the signal. The buildout has crossed from being funded by cash flow to being funded by markets. Market-funded booms depend on refinancing. Refinancing-dependent booms die of credit events, not demand misses. The fragile nodes remain where we mapped them in January: the neoclouds and the OpenAI-adjacent SPVs (special-purpose vehicles — off-balance-sheet financing entities). And Meta's entry into compute resale^F just attacked the neoclouds' business model from above (CRWV −14%, NBIS −17% in a day^F). The numbers underneath: CoreWeave carries $24.9B of debt, with interest eating roughly half its adjusted EBITDA (a measure of core operating profit)^C. Oracle's five-year CDS (the market price of insuring its debt against default) hit record wides near 200 basis points (2 percentage points); it has cut ~21,000 jobs in a year; and OpenAI is reportedly walking away from expanding its flagship Stargate site with it^C/F. Private credit is cracking at the edges too: a rising default index, the first-ever quarterly outflow from non-traded BDCs (funds that lend directly to private companies), and roughly a quarter of direct-lending books exposed to software borrowers being disrupted by AI^C.(b) 融资结构：正以晚周期特有的方式恶化。2026 年资本开支／收入比：Oracle 约 86%、Meta 约 54%、Microsoft 约 47%、Alphabet 约 46%^F。超大规模云厂商 2026 年债券发行预测约 $93B，且有上行空间^F。Alphabet——一家产生了二十年现金流的公司——6月2日定价 $84.75B 股权融资（含 $10B 的 Berkshire 定向配售），叠加一年内发行的逾 $85B 债务：外部融资合计超过 $170B^F。资深科技投资者为之震惊（“我从没想过 Google 会需要去公开市场融资”）——这份震惊本身就是信号。这轮建设已从现金流自筹跨入市场融资。靠市场融资的繁荣依赖再融资；依赖再融资的繁荣死于信用事件，而非需求失手。脆弱节点仍在我们 1 月标注的位置：新兴云与 OpenAI 关联的 SPV（特殊目的载体，表外融资实体）。而 Meta 进军算力转售^F刚刚自上而下打击了新兴云的商业模式（CRWV 单日 −14%、NBIS −17%^F）。底层数字：CoreWeave 背负 $24.9B 债务，利息吞掉其调整后 EBITDA（核心经营利润指标）约一半^C。Oracle 五年期 CDS（为其债务违约买保险的市场价格）打出接近 200 个基点（约 2 个百分点）的纪录宽点；一年内裁员约 21,000 人；据报道 OpenAI 正放弃与其扩建旗舰 Stargate 站点^C/F。私募信贷的边缘也在开裂：违约指数走高、非交易型 BDC（直接放贷给私营企业的基金）出现史上首次季度净流出、直接放贷账簿约四分之一的敞口在被 AI 冲击的软件借款人身上^C。

(c) Market structure: the mechanics of euphoria are completing. The IPO window is doing exactly what it did in late 1999. SpaceX+xAI listed mid-June at the largest valuation in IPO history (~$2T target, $75B raise, ticker SPCX)^F. OpenAI filed confidentially in June with an $852B private mark ($110–122B February round; Nvidia $30B, SoftBank $30B, Amazon's tranche contingent on an IPO or a formal AGI determination — a finding that AI has reached human-level general intelligence)^F. Anthropic is on the same track^C. Bending Spoons popped ~40% on its July 1 debut^F. FINRA margin debt (money borrowed to buy stocks) hit a record $1.42T in May, up 54% year over year^F. The SPCX debut was the largest retail net-buy day ever recorded — while institutions cut risk (BofA's fund-manager survey showed the equity overweight falling from 50 to 38)^F/C. SPCX itself has round-tripped: −30% from its June 16 high, though still above the issue price^C. And the strangest, most regime-defining datapoint of the week: the July 2 semi selloff was catalyzed partly by OpenAI's government-stake headline^F. The market read state ownership as a cap on the upside, not a subsidy. Correctly, in our view.(c) 市场结构：亢奋的机制正在走完。IPO 窗口正在重演 1999 年末。SpaceX+xAI 6 月中旬以 IPO 史上最大估值上市（目标约 $2T、募资 $75B、代码 SPCX）^F。OpenAI 6 月保密递交、私募估值标记 $852B（2 月轮融资 $110–122B；Nvidia 出资 $30B、SoftBank $30B，Amazon 的份额以 IPO 或 AGI 认定为先决条件——即正式认定 AI 达到人类水平的通用智能）^F。Anthropic 走在同一轨道^C。Bending Spoons 7月1日上市首日暴涨约 40%^F。FINRA 融资余额（借钱买股票的规模）5 月创 $1.42T 纪录、同比 +54%^F。SPCX 上市首日是有记录以来最大的散户净买入日——机构却在同期降风险（BofA 基金经理调查的股票超配从 50 降到 38）^F/C。SPCX 本身已走完往返：较6月16日高点 −30%，但仍高于发行价^C。而本周最诡异、最能定义机制的数据点：7月2日的半导体抛售，部分由 OpenAI 政府持股的头条催化^F。市场把国家持股解读为上行空间的天花板，而非补贴。在我们看来，读对了。

Verdict: what is "popping" is the pricing of scarcity into perpetuity — markets had priced chip scarcity as if it would last forever, and that pricing, the parabolic phase of Act I, is unwinding. The buildout itself is not popping. It has $725B of committed 2026 spend, a $5.3T five-year corridor^F, a sovereign shareholder-in-waiting (the US government), and a bipartisan political floor (nobody runs for office on "lose the AI race to China"). The correct posture is neither calling the top nor blindly buying the dip in the old leaders. It is rotating to where the capital wave lands next (Sections 6–8), renting the old leaders around specific datapoints, and holding cash for the October confluence (the window when several catalysts cluster).裁决：正在“破裂”的，是把稀缺性定价成永续——市场此前把芯片稀缺当成永恒来定价，这层定价（第一幕的抛物线阶段）正在拆除。建设本身没有破裂：它拥有 2026 年 $725B 的已承诺支出、$5.3T 的五年走廊^F、一位候场的主权股东（美国政府），以及两党共同的政治下限（没人会拿“把 AI 竞赛输给中国”当竞选纲领）。正确的姿态既不是喊顶，也不是无脑抄底旧领导者。而是轮动到资本浪潮的下一个落点（§6–8），围绕具体数据点“租用”旧领导者，并为 10 月的共振窗口（多个催化剂扎堆的时点）持有现金。

3The Political Economy Engine — Interests, Actors, and the Path They Force政治经济引擎——利益、行为主体与它们合力逼出的路径

Markets run on stories. Policy runs on interests. This section maps those interests, actor by actor, and shows the buildout path their combined incentives force. The timeline in Section 5 is generated here — everything downstream is mechanism, not vibes.市场靠故事运转，政策靠利益运转。本节逐一梳理各方的利益，再展示这些激励叠加后必然逼出的建设路径。§5 的时间线正是由本节推导而来——下游的一切都是机制，不是感觉。

3.1 · What the administration is actually optimizing3.1 · 政府真正在优化什么

Strip away the rhetoric. The administration is optimizing four things, in this order:剥去言辞包装，本届政府按以下顺序优化四件事：

1. Nominal GDP > interest cost. This is the sovereign-refinancing thesis from our January dossier, unchanged. The debt is so large that the only politically survivable path is to grow the economy faster than the interest bill on that debt. AI capex (capital spending) is the only private-sector engine big enough to matter. Four companies will spend $725B in 2026 — about 2.4% of GDP in direct spending, before any ripple effects^F/I. This is why the administration will never let the buildout die. It is the growth plug for the fiscal hole. Every policy choice below follows from this.名义 GDP 增速 > 利息成本。这是我们一月档案中的主权再融资论点，原封未动。债务太重，唯一在政治上走得通的路，就是让经济增速跑赢债务利息。AI 资本开支（企业投入建设与设备的钱）是唯一体量够大的私人部门引擎——2026 年四家公司直接支出 7,250 亿美元，约占 GDP 的 2.4%，还未计入连带效应^F/I。所以政府绝不会让这轮建设夭折：它是填补财政窟窿的增长塞子。下文每一项政策选择都由此而来。
2. Win the midterms — which now means affordability. The 2025 off-year losses (NJ, VA, Georgia PSC — the state utility regulator) were affordability elections. Electricity is the sharpest edge of it. Residential rates rose 5–7% nationally in 2025 (measure-dependent) and +42% over five years. DC rose +26% and PA +19% in a single year. Five-year increases: +94% (DC), +74% (MD), +73% (ME)^F. In PJM — the largest US regional power grid — capacity prices (what utilities pay to lock in future power supply) have cleared at the cap in three straight auctions, and by PJM's own math data centers added $9.3B (+174%) to those costs^F. The longest-serving senate president in Utah history lost his primary over a 7.5GW data-center campus^F. Pew finds 38% of Americans say data centers are "mostly bad" for home energy costs^F. Trump has gone beyond words. His "Ratepayer Protection Pledge" commits the signing hyperscalers (the giant cloud companies) to build, bring, or buy their own power for new AI facilities, and to fund the grid upgrades they cause^F. A bill advancing in Congress would harden that into law^C. The administration must deliver the buildout without pain in household power bills. That single constraint reshapes the entire power trade (Section 8.1).赢下中期选举——如今这意味着：可负担性。2025 年的非大选年败选（新泽西、弗吉尼亚、佐治亚州公共服务委员会——即州公用事业监管机构）本质上都是"东西太贵"的选举，而电价是其中最锋利的一刃。2025 年全国居民电价上涨 5–7%（因口径而异），五年累计上涨 42%。DC 一年内上涨 26%，宾夕法尼亚（PA）一年内上涨 19%。五年涨幅：+94%（DC）、+74%（马里兰，MD）、+73%（缅因，ME）^F。PJM（美国最大的区域电网）的容量价格（电力公司为锁定未来供电支付的费用）连续三次拍卖触及上限；按 PJM 自己的测算，数据中心为这项成本增加了 93 亿美元（+174%）^F。犹他州史上任职最久的参议院议长，因一座 7.5GW 数据中心园区在初选中落败^F。Pew 调查显示，38% 的美国人认为数据中心对家庭能源成本"弊大于利"^F。Trump 已不止于口头表态："用电户保护承诺"（Ratepayer Protection Pledge）要求签署的超大规模云厂商（最大的几家云计算公司）为新建 AI 设施自建、自带或自购电力，并承担由其引发的电网升级费用^F。国会正在推进的一项法案还可能把它写进法律^C。政府必须在不推高居民电费的前提下完成这轮建设。仅这一条约束，就重塑了整个电力交易（见 §8.1）。
3. Beat China to deployed AI — on American chips and American models. The tools: the AI Action Plan / Genesis Mission architecture^C; "export the whole stack" diplomacy with the Gulf, Japan, and Korea (selling chips, cloud, and models as one package); and now the sharpest new instrument, direct state participation. That means the government taking a cut and a say: revenue shares on export licenses, approval power over frontier-model releases (GPT-5.6^F), export controls that briefly hit even Anthropic's international access^F, and a proposed 5% stake in OpenAI framed as a "Public Wealth Fund"^F. In short: the state is building a toll-booth-plus-shareholder position across the AI frontier. Congress would need to approve the equity piece^F, and the Learning Resources ruling's logic (no revenue-raising without clear authorization from Congress) hangs over the revenue-share tools^C. But the direction is unambiguous, and close to bipartisan: Altman is offering co-option up front, to pre-empt a Sanders-style "seize it" framing^F/I.在落地部署的 AI 上击败中国——而且要用美国芯片、美国模型。手段包括：AI 行动计划/Genesis Mission 的架构^C；向海湾、日本、韩国"整栈输出"的外交（把芯片、云、模型打包卖出去）；以及如今最锋利的新工具——国家直接参股。具体来说：出口许可的收入分成、对前沿模型发布的审批权（GPT-5.6^F）、一度波及 Anthropic 国际接入的出口管制^F，以及以"公共财富基金"名义提出的 OpenAI 5% 股权方案^F。一句话：国家正在整个 AI 前沿构建"收过路费＋当股东"的头寸。股权部分需经国会批准^F；Learning Resources 案的司法逻辑（没有国会明确授权就不得创收）也悬在收入分成工具头上^C。但方向毫不含糊，且带有几分两党共识：Altman 主动献股"招安"，正是要抢在 Sanders 式的"没收"叙事之前^F/I。
4. Keep the war ended. The Iran war wrecked the affordability story (CPI — consumer price inflation — hit 4.2%^F). It bruised approval going into the Beijing trip^F. And it pushed the Fed hawkish (leaning toward higher rates). Every incentive — electoral, fiscal, monetary — now points one way: nail down the permanent deal inside the 60-day window and keep oil cheap through November. Expect maximal flexibility on deal terms and loud "cheap gas" messaging.^I让战争保持"已结束"。伊朗战争重创了可负担性叙事（CPI，即消费者物价通胀，升至 4.2%^F），让支持率在北京之行前受挫^F，还把美联储推向鹰派（倾向加息）。选举、财政、货币三方面的激励如今全部指向一件事：在 60 天窗口内敲定永久协议，并让油价一路压低到 11 月。可以预期：协议条款上最大限度的灵活，以及高分贝的"廉价汽油"宣传。^I

3.2 · Actor-by-actor3.2 · 逐个行为主体拆解

Bessent (Treasury). His interests, in four parts. First: keep the term premium (the extra yield investors demand to hold long-term bonds) suppressed while the government's debt gets refinanced. Second: a stable tariff floor that raises revenue without Supreme Court risk — hence the rebuild on Section 301, an older tariff law aimed at unfair trade practices. Third: stablecoins (dollar-pegged digital tokens) as a structural buyer of T-bills (short-term government debt). The GENIUS stablecoin law takes effect around January 2027, and every dollar of compliant stablecoin float is a bill bid. Fourth: LNG-and-investment diplomacy that recycles Japan/Korea trade-deal money into US energy and chip plants^C. Watch his QRA choices (the Treasury's quarterly borrowing plans, late July and late October). Continued bill-heavy issuance plus buybacks amounts to quiet, light-touch capping of long-term yields — it lets the AI debt wave get priced. His headache: the contested ~$170B tariff-refund liability^F. Expect the refund mechanics to be slow-walked into 2027.^IBessent（财政部）。他的利益诉求有四条。第一，在政府滚动再融资期间，压低期限溢价（持有长期债券所要求的额外收益）。第二，一个既能创收、又没有最高法院风险的稳定关税底座——301条款（针对不公平贸易行为的关税工具）的重建由此而来。第三，把稳定币（与美元挂钩的数字代币）做成短期国债的结构性买盘：GENIUS 法案约在 2027 年 1 月落地实施，每一美元合规稳定币流通量都是一笔国库券（短期国债）买单。第四，与日韩的"LNG＋投资"外交，把贸易协议资金回流到美国能源与芯片厂^C。盯住他的 QRA（财政部季度再融资计划）选择（7 月末、10 月末）：继续偏重短债发行＋回购，等于悄无声息地变相压住长期利率，让 AI 债务浪潮得以顺利定价。他的心病是约 1,700 亿美元、仍有争议的关税退款负债^F——预计退款机制会被一路拖到 2027 年。^I

Warsh (Fed). Covered in 2.4. One addendum on incentives. His legacy play is "the chairman who modernized the Fed" (five task forces^F) — a reformer identity, not an inflation-fighter identity. Reformers change frameworks. And a framework changed in this administration's direction means easier policy dressed in productivity language. His AI-disinflation doctrine^F — the idea that AI pushes prices down — is the bridge between serving the president who appointed him and the price-stability oath he swore on television. Both can be true if AI productivity is real. Conveniently, the buildout narrative requires believing exactly that.Warsh（美联储）。详见 §2.4。这里补充一点激励分析。他给自己押注的历史定位是"让美联储现代化的主席"（五个工作组^F）——这是改革者人设，不是通胀斗士人设。改革者改框架；而顺着本届政府的方向改框架，就等于用生产率话术包装的更宽松政策。他的"AI 去通胀"学说^F（认为 AI 会压低物价）是一座桥：一端连着任命他的总统，一端连着他在电视镜头前立下的物价稳定誓言。只要 AI 生产率是真的，两头就都能成立。而恰好，建设叙事本身就要求你相信这一点。

Lutnick/Greer (Commerce/USTR). They are building tariff architecture meant to last. First, a Section 301 forced-labor tariff floor by late July^F. Then the sectoral stack under Section 232 of US trade law (tariff investigations on national-security grounds): pharma July 31, and semiconductors and robotics in the pipeline^F. The Section 232 robotics investigation is the tell. The administration already views humanoid robots as the next strategic industry to wall off — before Unitree does to it what BYD did to autos. Positioning implication: the US/allied robotics BOM (bill of materials — the parts list) gets a protected market.^ILutnick/Greer（商务部/USTR）。两人正在搭建能长期存续的关税架构。先是 7 月末前落地 301条款的强迫劳动关税底座^F；随后是按行业铺开的232条款（美国贸易法中以国家安全为由的关税调查）组合——制药 7 月 31 日，半导体与机器人已在管道中^F。232条款机器人关税调查就是明牌：政府已把人形机器人视为下一个必须提前筑墙的战略产业——要赶在 Unitree 对它做出 BYD 对汽车业做过的事之前。仓位含义：美国及盟友的机器人物料清单（BOM，即零部件清单）将得到一个受保护的市场。^I

The frontier labs. OpenAI: pre-IPO (about to list on the stock market, marked at $852B^F), revenue $2B/mo^F, with computing contracts in the hundreds of billions — and now offering the state 5%. The incentive is fusion with the state ahead of the listing: convert political risk into political sponsorship. The Musk complex: SpaceX+xAI merged and public at ~$2T^F; Tesla converting Fremont to a 1M/yr Optimus line^F; a joint Terafab chip fab; the Starmind 1M-satellite orbital-compute constellation filed^F; Anthropic renting Colossus at $1.25B/month and Google at $920M/month^F. Musk has vertically integrated launch + compute + robotics + energy, and taken it public at the top of the window. Anthropic/Google: the enterprise-and-distribution axis. The labs share one incentive: keep the capex narrative alive through their listings. So H2 2026 will be structurally rich in bullish AI announcements, regardless of the underlying economics. Discount the newsflow accordingly.^I前沿实验室。OpenAI：即将上市（估值标记 8,520 亿美元^F），月收入 20 亿美元^F，算力合约义务高达数千亿美元，如今主动向国家献出 5%。其算盘是在上市前与国家融合：把政治风险换成政治背书。Musk 系：SpaceX+xAI 合并后以约 2 万亿美元估值上市^F；Tesla 正把 Fremont 工厂改造成年产 100 万台的 Optimus 产线^F；合资兴建 Terafab 芯片厂；Starmind 百万颗卫星的轨道算力星座已提交申请^F；Anthropic 以每月 12.5 亿美元、Google 以每月 9.2 亿美元租用 Colossus^F。Musk 已把发射＋算力＋机器人＋能源垂直整合，并在窗口顶部推向公开市场。Anthropic/Google 则是企业客户与分发渠道这条轴线。各实验室有一个共同激励：在各自上市完成前，维持资本开支叙事不坠。这意味着无论底层经济性如何，2026 年下半年必然密集出现利好 AI 的公告。请据此对新闻流打折。^I

China (Xi). The strategy revealed by the May summit and since: stabilize the US relationship for three years — the "strategic stability" framework^F, Xi in Washington Sept 24^F, and a truce extension before the late-October expiry as the base case^F/C — while decoupling compute demand by decree. The evidence: the H200 chip licenses the US granted this spring, which carry a 25% revenue share to the US Treasury^F, have produced zero deliveries. Beijing told its platforms to pull back. It is steering Alibaba/Tencent/ByteDance to Huawei Ascend (targeting ~600K 910Cs in 2026, though CXMT's supply of HBM — the high-bandwidth memory AI chips need — caps usable packages at ~250–300K^C), to Alibaba T-Head, and to Cambricon. DeepSeek's latest model is optimized for Ascend^F. Read: China has decided the Nvidia performance gap is closable, and is eating the near-term cost. Three consequences. (a) Nvidia's China TAM (total addressable market — the sales it could win there) is politically dead in both directions; stop underwriting its return. (b) The Chinese AI stack becomes a policy-guaranteed domestic monopoly — the convergence trade in Chinese platforms/semis has a state floor under it. (c) Rare-earth leverage remains partially in place^F. The magnet chokepoint (robot BOM!) stays a Chinese card — exactly why Western magnet capacity like MP's carries a strategic premium.^F/I中国（Xi）。从 5 月峰会及其后的动作看，战略很清楚：把对美关系稳定三年——"战略稳定"框架^F、Xi 于 9 月 24 日访问华盛顿^F、基准情形是在 10 月末到期前延长休战^F/C——同时用行政指令让算力需求脱钩。证据：美国今春批出的 H200 芯片许可证（附带 25% 上缴美国财政部的收入分成^F），交付量为零。北京让平台企业收手，把 Alibaba/Tencent/ByteDance 导向 Huawei Ascend（2026 年目标约 60 万片 910C；但 CXMT 的 HBM——AI 芯片所需的高带宽内存——供给，把可用封装量压在约 25–30 万片^C）、Alibaba T-Head 与 Cambricon。DeepSeek 的最新模型已针对 Ascend 优化^F。解读：中国已判定与 Nvidia 的性能差距可以补上，并甘愿承受近期代价。三个后果：(a) Nvidia 的中国 TAM（潜在市场规模）在两个方向上都已政治性死亡——别再为它的回归定价；(b) 中国 AI 技术栈成为有政策保底的国内垄断——中国平台/半导体的收敛交易之下有国家托底；(c) 稀土筹码仍部分在手^F——磁材卡点（机器人 BOM！）依旧是中国的一张牌，这正是 MP 之类西方磁材产能享有战略溢价的原因。^F/I

Japan/Korea. The Nikkei printed records above 72,000 in June, then broke 70,000 in the July 2 memory-chip rout. The KOSPI fell 7.9% in a single session — tripping its "sidecar" automatic trading halt — with SK Hynix −14.6%^F. Korea is a leveraged tracker of the memory-chip cycle: Samsung + SK Hynix are roughly half the index's weight. And the BOJ (Bank of Japan) is raising rates into it — 1.0% now, the highest since 1995, with the yen still at 40-year lows^F. Both countries are recycling trade-deal commitments into US energy and chip plants^C. Both are structurally long (positioned to win from) the memory supercycle. That makes them the geographic expression of Act I's blow-off and correction. The trade: KOSPI/Nikkei relative strength is the real-time global thermometer for the memory cycle.日本/韩国。Nikkei 6 月创下 72,000 点上方的纪录，随后在 7 月 2 日的存储芯片股崩跌中跌破 70,000 点。KOSPI 单日下挫 7.9%——触发 sidecar 熔断（自动暂停交易）——SK Hynix 大跌 14.6%^F。韩国等于一个自带杠杆的存储芯片周期跟踪器：Samsung + SK Hynix 约占指数权重的一半。而 BOJ（日本央行）正对着它加息：目前 1.0%，为 1995 年以来最高，日元却仍在 40 年低位^F。两国都在把贸易协议承诺回流到美国能源与芯片厂^C，且都在结构上做多（押注受益于）存储超级周期。这使它们成为第一幕冲顶与回调的地理表达。交易含义：KOSPI/Nikkei 的相对强弱，就是存储周期温度的实时全球体温计。

3.3 · Does the plan include the whole stack? Yes — and the order of operations is clear3.3 · 计划是否涵盖整个技术栈？是——而且先后顺序清晰可辨

Does the administration want hardware and infrastructure and software and embodied applications (AI put into physical machines — robots, drones, satellites)? In other words, the whole stack? The question now has an empirical answer. Map the policy tools to the layers of the stack:政府是要硬件、又要基础设施、又要软件、又要具身应用（装进实体机器的 AI——机器人、无人机、卫星）吗？换句话说，是要整个技术栈吗？这个问题如今有了经验答案。把政策工具映射到技术栈各层：

3.4 · The midterm calculus and what it forces in Q33.4 · 中期选举的算术及其对三季度的倒逼

Start with the structural pattern. The president's party has lost House seats in 19 of the last 21 midterms. Affordability polling is brutal, and electricity is the localized cutting edge^F. Base case: Democrats take the House (~80–85%), Republicans hold the Senate (~55–60%)^F. Polymarket priced 84% / 57% on July 2, and Sabato just moved Alaska and Ohio to toss-up. The backdrop is brutal: Trump approval ~38%, economic approval 33% — a record low — and cost-of-living the top issue at 37%^F. Two consequences matter for portfolios:先看结构性规律：过去 21 次中期选举，总统所在党有 19 次丢掉众议院席位；可负担性民调惨烈，电价是最具地方杀伤力的锋刃^F。基准情形：民主党拿下众议院（约 80–85%），共和党守住参议院（约 55–60%）^F。Polymarket 7 月 2 日的定价是 84% / 57%，Sabato 刚把阿拉斯加与俄亥俄移入"势均力敌"。背景同样惨烈：Trump 支持率约 38%，经济支持率 33%——创纪录新低——生活成本以 37% 高居议题榜首^F。两条与投资组合相关的后果：

1. Pre-election policy is all affordability, all Q3. Expect: maximal pressure for the Iran permanent deal (cheap gas). A data-center/electricity package — federal guidance on who pays for grid costs, fast-tracked hookups for behind-the-meter power (power generated on-site rather than bought from the grid), and possibly headline "tech pays" rates on large power users^I. Tariff-refund theater. And zero tolerance for a Fed hike — if Warsh raised rates in September, eight weeks before the election, the political cost would approach infinite. One more reason the hike never comes^I.选前政策全是可负担性，全部压在三季度。可以预期：为伊朗永久协议施加最大压力（廉价汽油）；一揽子数据中心/电力方案——联邦层面的成本分摊指引、加速审批的表后自供电（在场址自发电、不从电网购电）并网、可能还有面向大用户的标题式"科技买单"电价^I；关税退款的政治秀；以及对美联储加息的零容忍——Warsh 若在 9 月、选前八周加息，政治代价趋近无穷。这是加息永远不会到来的又一个理由^I。
2. Post-election gridlock is historically bullish. Markets have rallied in the 12 months after every midterm since 1950^C. Gridlock, plus a Fed turning dovish (leaning toward cuts), plus falling energy prices, is the recipe for a melt-up (a sharp, euphoric rally) from Q4 2026 into 2027. If Democrats take the House, the AI complex keeps its bipartisan floor. The party's data-center critique is about ratepayers and siting, not about ceding the race to China. But the "tech pays" regulatory tail gets fatter — which reinforces the behind-the-meter tilt rather than threatening the buildout.^I选后的府会僵局，历史上是利好。自 1950 年以来，每次中期选举后的 12 个月，市场都在上涨^C。僵局＋转鸽（倾向降息）的美联储＋能源价格回落，正是 2026 年四季度延伸到 2027 年的融涨（急速亢奋式上涨）配方。若民主党拿下众议院，AI 板块恰恰保住两党共识的底座：该党对数据中心的批评，焦点在用电户与选址，不是要在对华竞赛上认输。但"科技买单"的监管尾部会变肥——这只会强化表后自供电的倾斜，不会威胁建设本身。^I

4Cycle Position — 1998 × 1990, With 2000 on the 2027 Horizon周期定位——1998 × 1990，2000 隐现于 2027 地平线

4.1 · The three analogs, scored4.1 · 三个历史类比的评分

1990 Gulf War overlay (near-perfect fit for Feb–Dec 2026): The script then: an outside oil shock (Kuwait/Hormuz). Inflation spiked — CPI (the main inflation gauge) hit 6.3% then, versus 4.2% now. The Fed (America's central bank) froze mid-shock in a hawkish stance (leaning toward higher rates). The war resolved faster than feared. Oil prices round-tripped back to where they started. Inflation cooled within two quarters. The Fed then cut rates. Stocks bottomed before the cuts and surged through the following year. The 2026 version is running at roughly 2x speed, because the oil round-trip has already completed^F. What the overlay predicts: the inflation relief arrives August–November; the Fed's hawkish dot plot (each official's rate forecast) ages like milk; and the easing trade starts a quarter or more before the actual cuts. One comfort we refuse: the 1990 recession itself began in July 1990, weeks after the market highs. That branch lives on this map as Scenario E.1990 年海湾战争叠图（与 2026 年 2–12 月近乎完美吻合）：当年的剧本：一场来自外部的石油冲击（科威特/霍尔木兹）。通胀飙升——CPI（消费者价格指数，衡量通胀的主要指标）当年达 6.3%，如今为 4.2%。美联储（美国央行）在冲击中途冻结于鹰派立场（即倾向加息）。战争的化解快于市场担忧。油价往返回落至起点。通胀在两个季度内降温。美联储随后降息。股市在降息之前见底，并在次年一路上涨。2026 年版本正以约 2 倍速运行，因为油价往返已经完成^F。该叠图的预测：通胀缓解在 8–11 月到来；美联储的鹰派点阵图（每位官员的利率预测）会像牛奶一样迅速变质；宽松交易将领先实际降息一个季度或更久。有一点安慰我们拒绝接受：1990 年的衰退本身始于 1990 年 7 月，就在高点之后数周。这一分支在本图谱上对应情景 E。

1998 LTCM skeleton (the base architecture): LTCM was a giant hedge fund whose 1998 collapse threatened the whole financial system. The pattern: a violent, credit-flavored scare erupts inside a long-running tech boom. The Fed cuts rates. The boom's character changes — leadership narrows, then broadens again into new names. An 18-month melt-up (a fast, late-stage surge) follows. It ends in a top built from new supply and renewed tightening. The 2026 mapping: this time the scare is spread out — the memory-stock unwind, BTC/DAT deleveraging (forced unwinding of borrowed money), and an attack on the neocloud business model (the newer AI-cloud providers) — rather than one fund. But the function is the same. It forces the policy put — the belief that officials will step in to stop a crash — into the open. The state-equity discussions, with the government taking ownership stakes, are this cycle's version of the Fed-brokered LTCM rescue. The system is revealing that it will not let its strategic sector have an uncontrolled accident.^I1998 年 LTCM 骨架（基础架构）：LTCM（长期资本管理公司）是一家巨型对冲基金，1998 年倒下时险些拖垮整个金融体系。当年的模式：一场带信用色彩的剧烈恐慌，发生在长期科技繁荣内部。美联储降息。繁荣的性质发生改变——领涨面先收窄，再向新名字重新扩散。随后是一轮 18 个月的融涨（末段快速冲刺行情）。终点是供给放量叠加重新紧缩的顶部。2026 年的映射：这次恐慌是分散的——存储芯片股仓位回吐、BTC/DAT 去杠杆（借来的钱被迫平仓）、neocloud（新型 AI 云服务商）商业模式遭到质疑——而非单一基金出事。但功能相同。它迫使政策看跌期权（市场相信官方会出手托底）公开亮相。国家入股的讨论，就是本轮周期版本的美联储撮合 LTCM 救援。体系借此表明：它不会允许战略行业发生失控事故。^I

2000 telecom (the risk case — for 2027-28, not now): Here is the honest checklist. A thesis that cannot be falsified is not a thesis.2000 年电信（风险情形——针对 2027–28 年，而非当下）：下面是诚实的对照清单。无法被证伪的论点不是论点。

4.2 · The cloud-digestion sub-analog (for the semis specifically)4.2 · 云消化期子类比（专门针对半导体）

2018-19 is the template. Back then, hyperscaler capex — the big cloud companies' capital spending — went flat for ~6 quarters after doubling. SMH, the semiconductor ETF, fell ~23% in H2-2018. Micron halved while reporting record earnings at a 4x price-to-earnings ratio — because those record earnings were themselves the peak of the cycle. Then 2019-21 delivered +180% off the low. The 2026 memory-chip setup rhymes, with one enormous difference: HBM (high-bandwidth memory, the specialized chips AI systems need) turned a spot-priced commodity into a contracted, sold-out product line run by just a few suppliers (an oligopoly). The 2026 contracts are binding, and HBM4 is ramping at 2x the prior generation's pace^F. So the 2018 earnings cliff is dampened for the HBM slice, while ordinary DRAM/NAND memory keeps its full boom-bust cyclicality. The 90–95% quarter-on-quarter jump in contract prices^F is both the bull proof and the top warning. Pricing like that creates its own demand destruction — PC and phone makers are already rationing units^F — and its own supply response: new 2027 capacity, per the Micron/SK/Samsung roadmaps^F/C. What that means in practice: the memory trade's business peak plausibly lands late-2026 to 2027. But the stock peak comes first. The H1 parabola plus this week's action suggest the stock peak is forming now, in this quarter. Rent these positions; don't own them. Full framework in 8.6.2018–19 年是模板。当年，超大规模云厂商（各大云计算巨头）的资本开支在翻倍之后持平了约 6 个季度。半导体 ETF SMH 在 2018 年下半年下跌约 23%。Micron 股价腰斩，同时还在打印创纪录盈利、市盈率仅 4 倍——因为那些创纪录的盈利本身就是周期的顶点。随后 2019–21 年自低点上涨 +180%。2026 年的存储芯片格局与之押韵，但有一个巨大差异：HBM（高带宽内存，AI 系统所需的专用芯片）把一个现货定价的大宗商品，变成了签约锁定、售罄、由少数几家把持（寡头垄断）的产品线。2026 年合约已具约束力，HBM4 正以上一代 2 倍的速度爬坡^F。因此，2018 年那种盈利悬崖机制对 HBM 这一块被削弱；而普通的 DRAM/NAND 内存仍保留完整的周期性。合约价格环比飙升 90–95%^F，既是多头论据，也是顶部警报。这样的定价会制造自身的需求破坏——PC 与手机厂商已在配给式压减出货^F——也会催生自身的供给响应：2027 年新产能，参照 Micron/SK/Samsung 路线图^F/C。落到操作上：存储交易的基本面峰值大概率在 2026 年末至 2027 年。但股价峰值会先到。上半年的抛物线走势加上本周盘面表明，股价峰值正在当下、就在本季度形成。这些仓位只能租、不能长持。完整框架见 §8.6。

5The Master Timeline — July 2026 → December 2027总时间轴——2026年7月 → 2027年12月

How to read this table: Column 2 lists events that are already scheduled or set by law. Those are facts — the skeleton. Column 3 is our base case: the single most likely path (all ^I unless noted). Column 4 names the surprise that would knock that month off the base case, and which scenario in §5.2 it triggers. Market-mood tags: RO = risk-on (markets rising), CH = chop (sideways digestion), RF = risk-off (markets falling).如何阅读本表：第二列罗列已排定或法律规定的事件——这些是事实，构成骨架。第三列是基准情形：最可能的一条路径（除注明外均为^I）。第四列给出当月可能令你偏离基准的意外，及其激活的情景（情景树见 §5.2）。市况标签：RO＝risk-on（风险偏好，市场上涨），CH＝chop（震荡消化，横盘），RF＝risk-off（风险规避，市场下跌）。

5.1 · Month-by-month base case5.1 · 逐月基准情形

Calendar hygiene ^F (check before trading): remaining 2026 FOMC meetings: Jul 28–29, Sep 15–16, Oct 27–28, Dec 8–9. Verify the 2027 schedule once published. Verify MSCI review dates. Verify Treasury QRA dates each quarter.日历纪律^F（交易前核对）：2026年余下FOMC会议：7月28–29日、9月15–16日、10月27–28日、12月8–9日。2027年日程公布后需核实。核实MSCI审议日期。每季度核实财政部QRA日期。

5.2 · Scenario tree (18-month horizon)5.2 · 情景树（18个月视野）

6The Buildout Stack — Where Each Layer Sits on Its S-Curve建设堆栈——每一层在其 S 曲线上的位置

The model is simple. Capital moves through the stack (the layers of the AI buildout) like a wave. Every layer runs the same four phases: scarcity discovery → parabolic repricing → digestion/margin socialization → utility-ization. In plain terms: the market discovers a shortage; the stock reprices almost vertically; then growth gets digested and profits get competed away; finally it becomes a boring utility-like business. The edge: own layers entering phase 2. Only rent (hold short-term) or exit layers entering phase 3.框架很简单。资本像一道波浪，逐层穿过整个堆栈（即 AI 建设的各个层级）。每一层都走同样的四个阶段：稀缺发现 → 抛物线式重估 → 消化期/利润率社会化 → 公用事业化。说白了：市场先发现某处短缺；股价随即近乎垂直地上涨；然后增长被消化、利润被竞争摊薄；最后变成一门平淡的公用事业式生意。可操作的优势就一条：持有正在进入第 2 阶段的层；对进入第 3 阶段的层，只做短线“租用”或退出。

7The Incumbent Question — Mag 7, Disaggregated在位者之问——Mag 7 逐一拆解

7.1 · The FCF-trough machine, quantified7.1 · FCF 触底机器的量化

Here is the math the market is punishing. The Big 4 spent $410B on capex (capital spending) in 2025. Their own guidance (company forecasts) points to ~$725B in 2026. The analyst consensus range for 2027 is ~$0.9–1.05T^F/C. Meanwhile, operating cash flow (the cash the core business generates) is growing only around 15% a year. So free cash flow (the cash left over after all spending, or FCF) bottoms for the group somewhere in the second half of 2026 through the first half of 2027. Wall Street now tracks the Big 4's combined FCF toward roughly $4B in Q3 2026^C — functionally zero. On top of that, depreciation (spreading the cost of all this new equipment over 3–6 years of reported profits) keeps pressuring earnings-per-share optics through 2027–28. One caveat we accept from our own devil's-advocate review: spending guidance is still being raised. So the bottom is marked by the first flattening of capex guidance, not by a calendar date. Enter in stages, and size positions for the risk of being 12–18 months early — the precedent is META in 2022. Two patterns from history decide the trade:市场正在惩罚的，是一笔简单的算术。四大巨头 2025 年的资本开支（购置设备与基建的投入）为 4,100 亿美元。其自身指引（公司预测）指向 2026 年约 7,250 亿美元。2027 年分析师共识区间约 0.9–1.05 万亿美元^F/C。而经营性现金流（主营业务产生的现金）增速只有 15% 上下。因此，集团层面的自由现金流（扣除全部开支后剩下的现金，FCF）将在2026 年下半年至 2027 年上半年之间触底。华尔街目前追踪的四大合计 FCF，在 2026 年三季度约为 40 亿美元^C——实质上等于零。此外，折旧（把新设备成本分摊到 3–6 年的报表利润里）将在整个 2027–28 年持续压低每股收益的账面数字。我们接受一条自我质疑的警示：开支指引仍在上调。所以底部的标志是资本开支指引首次走平，而不是某个日历日期。应分批入场，仓位要为提前 12–18 个月的风险留出余地——先例是 2022 年的 META。两条经验规律决定这笔交易：

1. Stocks in capex cycles bottom 1–2 quarters before reported FCF does. The market prices the change in direction — a slowdown in spending guidance, or proof the spending is earning money — not the level itself. That puts the buy window inside the next two quarters of weakness: the July-earnings selldowns and the October confluence (several catalysts landing at once), not after a visible all-clear.资本开支周期中的股票，见底通常领先报表 FCF 触底 1–2 个季度。市场定价的是方向的变化——开支指引减速，或开支开始赚钱的证据——而不是水平值本身。这意味着买入窗口就在未来两个季度的疲弱期之内：即 7 月财报季的抛售与 10 月的多因素共振（多个催化剂同时到位），而不是等到警报明确解除之后。
2. Proof of monetization (AI actually generating revenue) is the trigger for a re-rating, and it is already stacking up. MSFT reports a $37B annual pace of AI revenue, growing 123%^F. Google Cloud holds a $460B backlog (revenue already under contract but not yet delivered)^F. AWS is reaccelerating^F. When these lines force analysts to model AI revenue as a durable growth business — not an excuse for the spending — the valuation mechanism flips. When, not if.变现证据（AI 真正产生收入）是重估的触发器，而且证据已在不断累积。MSFT 披露的 AI 年化收入达 370 亿美元、增速 123%^F。Google Cloud 在手订单（已签约但尚未交付的收入）达 4,600 亿美元^F。AWS 增速重新加快^F。当这些业务迫使分析师把 AI 收入当作可持续的增长业务、而不再是为开支开脱的托词时，估值机制就会翻转。是“何时”，不是“是否”。

7.2 · Verdicts, name by name7.2 · 逐名判定

• GOOGL — the structural winner; first back to highs. Google designs its own AI chips (TPUs). That gives it the lowest cost per token (each unit of AI output) among the giant cloud providers. Its own CFO makes the argument: in-house chips + models + products = a cost and research advantage^F. Its cloud backlog doubled to $460B in one quarter^F. The stock printed an all-time high on May 13 and traded within a hair of Nvidia's crown^F/C. The four-week slide since (~12% off the high), plus the controversy over its $84.75B share sale^F, is the entry point — not a warning. Raising cheap capital to extend a proven return advantage is exactly what a structural winner should do at this point in the cycle. Expected path: consolidation in the $340–370 zone (currently ~$360^F), then new highs by Q1–Q2 2027. One honest admission: the re-rating (the market paying a higher multiple) already ran in the first half. What remains is compounding growth plus this dip. Hedgeable risks: dilution from the share sale, and the DOJ's Chrome cross-appeal^C. What would prove us wrong: cloud backlog growth below 15% quarter-over-quarter for two straight quarters, or Gemini losing consumer share to OpenAI.GOOGL——结构性赢家；将率先重回高点。Google 自研 AI 芯片（TPU）。这让它在超大规模云厂商（最大的几家云计算公司）中拥有最低的单 token（每单位 AI 输出）成本。其 CFO 层面的论述正是：自研芯片 + 模型 + 产品 = 成本与研发优势^F。云在手订单一个季度内翻倍至 4,600 亿美元^F。股价 5 月 13 日创历史新高，一度距 Nvidia 的王座仅一步之遥^F/C。此后四周回落约 12%，加上 847.5 亿美元股权融资的争议^F——这是入场点，不是警报。在周期的这个位置，募集低成本资金去放大已被验证的回报优势，正是一个结构性赢家该做的事。预期路径：在 340–370 美元区间整固（现价约 360 美元^F），2027 年一至二季度前创新高。坦率地说：这一判断的重估行情（市场给出更高估值倍数）已在上半年走完。剩下的是复利增长，加上这次回调。可对冲的风险：融资带来的稀释、DOJ 对 Chrome 案的交叉上诉^C。证伪条件（能推翻判断的证据）：云在手订单环比增速连续两个季度低于 15%；或 Gemini 在消费端对 OpenAI 出现份额流失。
• MSFT — the trough accumulation (buy the bottom, in stages). Down ~22% year-to-date — the worst of the Mag 7, and ~30% off its high^F. Yet it disclosed $37B of AI revenue at an annual pace, growing 123%, and the market refuses to pay for it^F. The fear is priced hardest here. Management admits it is capacity-constrained. Its contracted backlog (RPO) stands at $627B — with the caveat that a large slice comes from one customer, OpenAI^C. And $25B of its capex increase is cost inflation, not extra ambition^F. Of the whole group, this is the cleanest case of capital spending converting into steady, recurring revenue. Accumulate on the earnings-week selldowns. The second leg should come with the Q4-26/Q1-27 reports. What would prove us wrong: AI revenue growth slowing below 60% year-over-year while capex still accelerates.MSFT——底部分批建仓的标的。年初至今下跌约 22%——Mag 7 中最差，较高点回撤约 30%^F。但它同时披露了 370 亿美元、增速 123% 的 AI 年化收入，市场却拒绝为此付费^F。恐惧在这里被定价得最充分。管理层自认产能不够用。在手合同订单（RPO）达 6,270 亿美元——警示是其中相当大一块来自单一客户 OpenAI^C。而且本轮资本开支上调中，有 250 亿美元属于成本上涨，不是新增野心^F。在这组公司里，它是“资本开支转化为稳定经常性收入”最干净的案例。在财报周的抛售中分批买入。第二段行情应随 2026 年四季度/2027 年一季度财报出现。证伪条件：AI 年化收入同比增速降破 60%，而资本开支仍在加速。
• META — the highest beta (most sensitive stock) on this question, in both directions. The bear case was that Meta has no cloud business to earn back its AI spending. Meta Compute answers that. Announced July 1, it sells surplus GPU (AI chip) capacity plus hosted-model APIs. The stock rose 8.8% that day, then fell 4.9% the next^F. In one move, management admitted the problem and offered the solution. It converts the scariest capex-to-revenue ratio in the group (54%^F) into a second business line — while conceding it over-ordered chips. Own it for the advertising-ROI engine plus this new line. Size the position for volatility. What would prove us wrong: a 2027 capex guide above $180B without a matching Meta Cloud revenue disclosure.META——对这个问题弹性（beta）最高的股票，双向皆然。空头逻辑曾经是：Meta 没有云业务，AI 开支赚不回来。Meta Compute 正面回应了这一点。它于 7 月 1 日官宣，出售富余 GPU（AI 芯片）算力和托管模型 API。当日股价 +8.8%，次日 −4.9%^F。管理层在同一个动作里，既承认了问题，又给出了解法。它把全组最吓人的资本开支/收入比（54%^F）变成第二条业务线——同时承认了芯片超额下单。持有理由：广告投资回报（ROI）引擎，加上这条新业务线。仓位要为剧烈波动留足空间。证伪条件：2027 年资本开支指引超过 1,800 亿美元，且没有对应的 Meta Cloud 收入披露。
• AMZN — solid, unexciting, middle of the pack. $200B of capex. AWS growing +28%. Trainium (its in-house AI chip) at a $20B annual revenue pace^F. Plus an investment in OpenAI with an IPO-contingent structure^F. It wins when the tide rises. The interesting extra: retail margins could get relief from tariff normalization plus IEEPA refunds (tariffs collected under that emergency law being returned)^I.AMZN——扎实、不惊艳、中庸。资本开支 2,000 亿美元。AWS 增速 +28%。自研 AI 芯片 Trainium 年化收入 200 亿美元^F。另有对 OpenAI 的投资（含 IPO 触发条款结构）^F。大潮涨时它就赢。有意思的额外看点：关税正常化，加上 IEEPA 关税退款（依据该紧急状态法征收的关税获退还），有望修复零售利润率^I。
• NVDA — great company, rented stock (trade it, don't marry it). The kingdom is intact. Vera Rubin has been in full production since June 1 and ships to eight named cloud providers in Q3^F. Data-center revenue runs at $62B a quarter^F. Yet the stock lagged the very sector it powers, up just ~7% in the first half^F. And all four ceilings are now visible. Customers are building their own chips at scale. China is politically dead in both directions (the H200 is licensed but frozen^F). The state takes a toll on its export economics. And a company this size runs into the law of large numbers (growth must slow as the base gets huge). Expect a wide, violent trading range. Worth trading around Rubin/GTC datapoints; not worth its 2024-era portfolio weight. What would end even the trade: two straight quarters of data-center revenue growth below 25% year-over-year, or a top-3 customer publicly shifting the majority of its orders to in-house chips.NVDA——伟大的公司，租来的股票（做波段，别长抱）。王国依然完好。Vera Rubin 自 6 月 1 日起全面量产，三季度将向八家具名云厂商出货^F。数据中心单季收入 620 亿美元^F。但这只股票却跑输了它所赋能的整个板块，上半年仅涨约 7%^F。四重天花板如今全部可见：客户大规模自研芯片；中国业务在政治上双向死局（H200 获许可却被冻结^F）；国家对其出口收益抽成；以及大数定律（体量越大，增速必然放缓）。预期一个宽幅、剧烈的震荡区间。值得围绕 Rubin/GTC 数据点做波段；不值得给到 2024 年那种组合权重。证伪条件（即便对波段仓位而言）：数据中心收入同比增速连续两个季度低于 25%；或某前三大客户公开宣布把多数订单转向自研芯片。
• AAPL — the rotation sponge (where money hides when it leaves AI names). It rose 4.8% on the July 2 semiconductor rout — driven as much by foldable-iPhone production reports as by rotation^F. The market has designated it the safe-haven megacap. Fine as ballast (a portfolio stabilizer); not an AI thesis. Two notes, though. Siri now runs on a custom 1.2-trillion-parameter Gemini model^F. And Tim Cook hands the CEO seat to hardware chief John Ternus on September 1^F — a tilt toward AI devices and robotics that could eventually make it one.AAPL——轮动的海绵（资金撤出 AI 股时的藏身处）。7 月 2 日半导体大跌当天，它逆势上涨 4.8%——折叠屏 iPhone 投产报道的贡献不亚于资金轮动本身^F。市场已把它指定为避险巨头。作为压舱石没有问题；但它不是一个 AI 论题。不过有两点值得注意。Siri 现已运行在定制的 1.2 万亿参数 Gemini 模型上^F。而且 Tim Cook 将于 9 月 1 日把 CEO 之位交给硬件负责人 John Ternus^F——这一向 AI 设备与机器人的倾斜，最终可能让它变成一个 AI 论题。
• TSLA — reclassifying itself in real time. Model S/X are dead. The Fremont plant is converting to a 1M-per-year Optimus robot line. The V3 unveil is imminent. Capex tops $25B^F. The company is publicly betting itself on physical AI (AI in robots and machines, not just software). Musk claims $25T outcomes at Davos^F; BofA models 1.2M humanoid robots industry-wide by 2030^C. Against that: management admitted in January that the robots do no “useful work” yet^F. The flirtation with a SpaceX merger^F is a governance overhang. And the stock fell 7.5% on July 2 despite good delivery numbers^F — the market agrees. How to treat it: as an option-like bet (small stake, big possible payoff) on the 2027 physical-AI leg, not a fundamentals position. The cleaner way to play the same idea is the BOM suppliers — bill of materials, the robots' parts list (8.3).TSLA——正在实时给自己重新归类。Model S/X 停产。Fremont 工厂正改造为年产 100 万台的 Optimus 机器人产线。V3 发布在即。资本开支超 250 亿美元^F。这家公司正公开把自己押注在物理 AI（装进机器人和机器里的 AI，而非仅是软件）上。Musk 在达沃斯宣称 25 万亿美元级别的结果^F；BofA 模型预测 2030 年全行业人形机器人达 120 万台^C。反面证据：管理层 1 月承认机器人尚做不了“有用的工作”^F。与 SpaceX 合并的暧昧表态^F是治理层面的悬顶之患。7 月 2 日交付数据不错、股价仍跌 7.5%^F——盘面认同这一判断。处理方式：把它当作面向 2027 年物理 AI 行情的类期权仓位（投入小、潜在回报大），而不是基本面仓位。同一想法更干净的表达，是人形机器人物料清单（BOM，即零部件清单）上的供应商（见 §8.3）。

8Act II Narratives, Ranked by Asymmetry第二幕叙事：按不对称性排序

How we score each idea: crowding (how many investors already own it — how much is already in the price), catalyst density (how many concrete events could move it in the next 2–4 quarters), policy alignment (does government policy help or hurt), and payoff shape (how much you can win versus lose — the "asymmetry" in the title). These map onto our Factor Watch baskets — notes inline. (One structural gap we flag on our own desk: Factor Watch still carries no dedicated Crypto/Digital Assets basket; Section 10 is its founding document.)评分维度：拥挤度（多少投资者已持有——价格已计入多少）、催化剂密度（未来2–4个季度有多少明确事件）、政策一致性（政策是助力还是阻力），以及收益结构（能赢多少、会输多少——即标题里的「不对称性」）。这些叙事与我们的 Factor Watch 篮子一一对应，注释随文标出。（我们在自家席位上标记的一个结构性缺口：Factor Watch 至今仍没有专门的加密／数字资产篮子；§10 即其奠基文件。）

8.1 · The politically-adapted power trade — behind-the-meter everything ★ highest conviction8.1 · 政治适应型电力交易——一切皆表后自供电 ★ 最高确信度

Mechanism: voters are revolting against rising power bills (§3.1). FERC (the US federal energy regulator) and the states are fighting over who pays for giant new data-center loads. Both parties now agree that "tech pays." Together, these forces push each new gigawatt of power off the shared grid. The hyperscalers (the giant cloud companies) are responding rationally — and visibly — by generating their own power: on-site gas (turbines, recip engines, gensets — all flavors of packaged generators), microgrids, firm PPAs (long-term contracts that guarantee power delivery), and eventually SMRs (small modular nuclear reactors). This is "behind-the-meter" power: generated on-site rather than bought from the grid. Meanwhile the merchant IPPs (independent power producers who sell into the shared grid) carry political risk from both sides — heading into an election that is partly about their prices.机制：用电户正在反抗不断上涨的电费（§3.1）。FERC（美国联邦能源监管机构）与各州正为谁来承担数据中心的巨额新增负荷而争斗。两党如今都同意「科技买单」。这些力量合在一起，正把每一吉瓦新增电力逼离共享电网。超大规模云厂商（大型云计算公司）的理性应对已然可见：自己发电——厂区内燃气（燃气轮机、往复式机组、发电机组）、微电网、firm PPA（保证供电的长期购电合同），最终是 SMR（小型模块化核反应堆）。这就是「表后自供电」：电在自家场地发，不从电网买。与此同时，merchant IPP（向共享电网售电的独立发电商）承受着来自两边的政治风险——前方正是一场部分以它们电价为议题的选举。

Expressions: gensets/recip engines & turbines (CAT — already so recognized as an "AI beneficiary" that it can fall 7% in a day^F — CMI, GNRC industrial); switchgear/electrical (ETN, VRT, PWL→POWL, HUBB, nVent); grid and data-center construction plus electricians (PWR, EME — a recent Factor Watch addition — FIX, IESC, MYRG, STRL); transformers (GEV's electrification segment, Hitachi/Siemens Energy ADRs — US-listed versions of the foreign shares). For nuclear, own the fuel cycle, not the SPACs (pre-revenue shell-company listings): CCJ, physical uranium (SPUT), and enrichment (LEU — the scarcity of HALEU, a special reactor fuel now cut off from Russian supply, compounds into 2027–28 contracting) rather than reactor stories with no revenue.标的：发电机组／往复式发动机与燃气轮机（CAT——其「AI 受益股」身份已被市场充分认知，以至于可以单日下跌 7%^F——CMI、GNRC 工业板块）；开关设备／电气（ETN、VRT、PWL→POWL、HUBB、nVent）；电网／数据中心建设与电气承包商（PWR、EME——Factor Watch 新近纳入——FIX、IESC、MYRG、STRL）；变压器（GEV 的电气化业务板块、Hitachi/Siemens Energy 的 ADR，即在美上市的外国股票凭证）。核电通过燃料循环表达，而非 SPAC（尚无营收的壳公司上市）：CCJ、实物铀（SPUT）与铀浓缩（LEU——HALEU 这种特殊反应堆燃料因去俄化而稀缺，将在 2027–28 年签约周期中复合放大），而非没有营收的反应堆故事。

Anti-position: hold less than normal — or avoid — the pure merchant IPPs (VST/CEG/TLN-type exposure) until the post-midterm regulatory shape is known. The same politics that blesses behind-the-meter power can cap what these companies earn in the capacity markets.^I反向仓位：在中期选举后的监管格局明朗之前，低配（低于正常仓位）或回避纯商业化 IPP 敞口（VST/CEG/TLN 一类）。祝福表后自供电的同一套政治，也可能压制它们在容量市场的盈利。^I

Catalysts: the Q3 affordability package^I, hyperscaler self-generation announcements, PJM (the biggest US grid operator) auction politics, post-midterm clarity.催化剂：三季度电价可负担性一揽子方案^I、超大规模云厂商自发电公告、PJM（美国最大电网运营商）拍卖政治、中期选举后的明朗化。

8.2 · Second-derivative capex: thermal, electrical content, optics/fabric ★8.2 · 资本开支的二阶导：散热、电气含量、光模块／网络互联 ★

Mechanism: $725B of 2026 capex^F converts, mechanically, into physical buildings over 12–24 months. "Second-derivative" means we buy what that spending buys: the equipment inside. And the content per megawatt — the dollar value of transformers, switchgear, liquid cooling, and 1.6T optics (ultra-fast fiber-optic links between chips) packed into each megawatt of data center — is rising as racks get denser. Networking grew +263% at Nvidia^F. The fabric layer (the gear that connects the chips) now compounds faster than the compute layer (the chips themselves).机制：2026 年 7,250 亿美元的资本开支^F将在 12–24 个月内机械地转化为落地建筑。「二阶导」的意思是：买这笔开支所购买的东西——楼里的设备。而单兆瓦价值含量——每兆瓦数据中心里变压器、开关设备、液冷与 1.6T 光模块（芯片之间的超高速光纤互联）的美元价值——正随机柜功率密度上升而上升。Nvidia 的网络业务增长 +263%^F。互联层（连接芯片的设备）的复合增速，如今已快于算力层（芯片本身）。

Expressions: VRT, MOD, nVent (thermal/cooling); COHR, LITE (another recent Factor Watch addition — validated by this quarter's trading), FN, CLS, ANET, AVGO/MRVL (fabric/custom chips — both just took 10–13% hits^F, which is the entry mechanism working); and advanced-packaging equipment (machines that stack and wire chips together) as the "memory-becomes-logic" lever — BESI, CAMT/ONTO, AMAT (its advanced-packaging line is framed as potentially doubling DRAM-adjacent equipment spend^C).标的：VRT、MOD、nVent（散热）；COHR、LITE（Factor Watch 另一新近纳入标的——本季度行情已予验证）、FN、CLS、ANET、AVGO/MRVL（互联／定制芯片——两者刚刚遭受 10–13% 的打击^F，这正是入场机制在起作用）；先进封装设备（把芯片堆叠、连线在一起的机器）作为「存储逻辑化」的杠杆——BESI、CAMT/ONTO、AMAT（其先进封装产品线被描述为有望令 DRAM 相关设备开支翻倍^C）。

Entry mechanics: optics is where the correction actually happened — four separate single-day flushes of 8–17% since May (scares over CPO delays, then the July 2 "photonics reset")^F. That makes it the preferred entry among these second-derivative plays. Valuations still carry froth (COHR trades at ~189 times its past year's earnings^C). So buy the flushes (the sharp sell-offs). Never buy the strength.入场机制：光模块是真正发生过回调的领域——5 月以来出现四次独立的 8–17% 单日急跌（CPO 延期恐慌、7 月 2 日的「光子学重置」）^F。这使其成为二阶导标的中的首选入场点。估值仍有泡沫（COHR 约为过去一年盈利的 189 倍^C）。所以只买急跌，绝不追强势。

Catalysts: each hyperscaler capex print; Rubin (Nvidia's next platform) rack architecture disclosures; 1.6T qualification cycles.催化剂：每一次超大规模云厂商资本开支披露；Rubin（Nvidia 下一代平台）机柜架构细节公布；1.6T 认证周期。

8.3 · Physical-AI BOM, behind the Section 232 wall ★ (the front-run)8.3 · 物理 AI 的物料清单(BOM)，藏身232条款关税墙之后 ★（抢跑交易）

Mechanism: BOM means bill of materials — the parts list. The story caught fire in the first half. Optimus V3's unveil is weeks away, with the Fremont plant converting (the 2026 target is 50–100K units; the 1M/yr figure is a year-end run-rate ambition — a hoped-for pace, not installed capacity^F). Figure builds one robot an hour. Unitree shipped 5,500+ robots in 2025 (32% global share) and cleared its STAR-market IPO hearing in a record 73 days. 1X's $20K NEO ships this year. OpenAI launched an in-house robotics division in May^F. Yet the equity supply of pure plays — stocks that give direct exposure — is tiny. And the Section 232 robotics report (a tariff investigation under US trade law) has passed its statutory deadline: a tariff decision is pending^F — a wall waiting for a signature.机制：BOM 即物料清单——零部件清单。这一叙事已在上半年点燃。Optimus V3 发布仅数周之遥，Fremont 工厂正在改造（2026 年目标 5–10 万台；每年 100 万台是年底的运行速率愿景——期望达到的节奏，而非已建成产能^F）。Figure 每小时下线一台机器人。Unitree 2025 年出货 5,500 台以上（全球份额 32%），并以创纪录的 73 天通过科创板 IPO 上会。1X 的 2 万美元 NEO 今年发货。OpenAI 于 5 月成立内部机器人部门^F。但纯正标的的股票供给——能提供直接敞口的股票——极小。而且 232条款机器人调查报告（美国贸易法下的关税调查）已过法定期限：关税决定悬而未决^F——一堵只待签字落笔的墙。

A reality anchor, so no one accuses us of chasing the story: only ~16,000 humanoid robots were installed worldwide in 2025^C. That is precisely the point. Owning the BOM front-runs the deployment wave; it does not depend on this year's shipment counts. The 2027 wave needs: rare-earth magnets, precision actuators and motion parts, sensors (the onsemi/ST class — ST already uses humanoids in its own chip plants^C), test equipment (TER — its robotics segment is growing again^F), and wiring harnesses and assembly.先立一个现实锚点，以免被指责为追逐叙事：2025 年全球人形机器人安装量仅约 16,000 台^C。而这恰恰是关键。持有 BOM 抢跑的是部署浪潮，不依赖今年的出货量。2027 年的部署浪潮需要：稀土磁材、精密执行器与运动部件、传感器（onsemi/ST 一档——ST 已在自家芯片厂使用人形机器人^C）、测试设备（TER——其机器人业务板块重回增长^F），以及线束与组装。

The uncomfortable fact our red team (our in-house devil's advocates) surfaced: ~70% of Optimus V3's components come from China (Tuopu, Sanhua, Leaderdrive, Green Harmonic^C). And China has tightened its licensing of magnet technology — adding the element holmium to close a substitution loophole — even while suspending its broader October export controls until November 10^C/F. A Section 232 tariff wall would tax a parts list the US cannot yet make at home. Which is exactly why the scarce domestic links carry the strategic premium: MP's magnets sit under a $110/kg Defense Department price floor — now in the money, since NdPr (the key magnet material) trades near $90^C — with heavy-rare-earth output starting mid-2026 and the 10X magnet plant in 2028^F. It is also why US robot-OEM stocks (the companies that build the finished robots) are the narrative leg of this trade, not the earnings leg.我们红队（内部唱反调的团队）挖出的一个令人不安的事实：Optimus V3 约 70% 的零部件来自中国（Tuopu、Sanhua、Leaderdrive、Green Harmonic^C）。而且中国已收紧磁材技术许可——新增钬（holmium）元素以堵住替代漏洞——即便其同时将更广泛的 10 月出口管制暂停至 11 月 10 日^C/F。一堵232条款关税墙征税的对象，是一份美国尚无法本土化生产的零部件清单。这正是稀缺的本土环节享有战略溢价的原因：MP 的磁材受 110 美元／公斤的美国国防部价格下限保护——现已进入价内，因为 NdPr（关键磁材原料）市价约 90 美元^C——重稀土产出 2026 年年中启动、10X 磁材工厂 2028 年投产^F。这也是美国机器人整机厂（OEM，造整台机器人的公司）股票只是这笔交易的叙事之腿、而非盈利之腿的原因。

Timing: be positioned before CES (Jan 2027). After it, this is consensus. The Q3-2026 semi-wide weakness is the entry vehicle.时点：须在 CES（2027 年 1 月）之前完成布局。在那之后，这就是共识。2026 年三季度半导体板块的整体疲弱即入场载体。

Discipline: the robot makers' stocks (TSLA, listed Chinese names) are narrative beta — they move on the story, violently. The BOM is where the per-unit economics accrue, no matter which robot wins.纪律：整机厂股票（TSLA、中国上市公司）是叙事贝塔——随故事剧烈起落。无论最终哪家机器人胜出，单位经济的积累都发生在 BOM 环节。

8.4 · Inference monetization / the software flip ★8.4 · 推理变现／软件翻转 ★

Mechanism: the market currently pays for spend. The 2027 regime pays for yield on spend — proof the money is earning a return. Winners: companies with distribution, proprietary data, and an agent surface (the place where AI agents actually do work). That means the Section 7 incumbents; PLTR — whose fundamentals accelerated to +85% growth while the stock de-rated 27% YTD, a valuation reset we welcome^F; and cybersecurity as "agent security" — no longer a barely-priced idea but a real profit story: Palo Alto closed its $25B CyberArk acquisition and shipped agent-lifecycle security, CrowdStrike launched agentic MDR (threat detection and response run by AI agents), and both printed record quarters^F. The engine underneath is the token paradox. Tokens are the units of text AI models consume. Their unit price deflated ~4x in a year, yet enterprise AI spend nearly tripled — because agent workflows burn 50–500x the tokens per task^C. Losers: per-seat SaaS — subscription software billed per employee, the pricing unit AI deletes. Honesty requires saying this is no longer a fresh short. The street has already named it (the "SaaSpocalypse": ~$2T of software market value erased since late 2025; CRM −33%, NOW −36% YTD^F/C), and the countertrend has begun (Atlassian +30% on raised guidance with rising net revenue retention — existing customers spending more; Workday beating^F). The 2027 version of this trade is dispersion — betting on the gap between winners and losers: short the seat-native firms that cannot switch to usage-based pricing, long the converters. ServiceNow already books half its net-new business on consumption pricing; Salesforce just bought a usage-billing platform^F/C. Staffing, BPO (outsourced back-office work), and IT services remain the slower-burning second wave (Accenture's bookings just went negative year over year^F).机制：市场当前为开支付钱。2027 年的范式为开支的收益率付钱——要证明钱花得有回报。赢家：拥有分发渠道、专有数据和智能体入口（AI 智能体真正干活的地方）的公司。也就是 §7 的在位者；PLTR——基本面加速至 +85% 的增长，而股价年内估值下修 27%，这是我们乐见的估值重置^F；以及作为「智能体安全」的网络安全——已不再是几乎无人定价的题材，而是实打实的利润故事：Palo Alto 完成 250 亿美元收购 CyberArk 并推出智能体全生命周期安全产品，CrowdStrike 推出智能体化 MDR（由 AI 智能体执行的威胁检测与响应），两者均交出创纪录季报^F。底层引擎是 token 悖论。token 是 AI 模型处理文本的计量单位。其单价一年内通缩约 4 倍，而企业 AI 开支却近乎增至三倍——因为智能体工作流每项任务消耗 50–500 倍的 token^C。输家：按席位计费的 SaaS——按员工人数订阅收费的软件，而 AI 正在抹除这个计价单元。必须诚实地承认：这已不再是新鲜的空头。华尔街已为其命名（「SaaSpocalypse」，SaaS 末日：2025 年末以来约 2 万亿美元软件市值蒸发；CRM 年内 −33%、NOW −36%^F/C），且反趋势已经开始（Atlassian 上调指引且净收入留存率上升——老客户花得更多了——股价 +30%；Workday 超预期^F）。这笔交易的 2027 年版本是离散度——押注赢家与输家之间的差距：做空无法转向按用量计价的席位原生厂商，做多转型者。ServiceNow 的净新增业务已有一半按消耗定价签约；Salesforce 刚刚收购了一个用量计费平台^F/C。人力派遣、BPO（外包后台业务）与 IT 服务仍是燃烧更慢的第二波（Accenture 的订单刚刚转为同比负增长^F）。

Catalysts: AI-revenue disclosures from Q3-26 onward; enterprise agent deployment counts; the first major SaaS name repricing to usage.催化剂：2026 年三季度起的 AI 收入披露；企业智能体部署数量；第一家大型 SaaS 厂商转向按用量重新定价。

8.5 · The China convergence trade (an old edge, now policy-floored)8.5 · 中国收敛交易（旧优势，如今有政策托底）

Mechanism: Beijing froze H200 chip deliveries to force adoption of the domestic technology stack^F — Huawei Ascend, Alibaba's T-Head at scaled mass production, Cambricon, DeepSeek running on Ascend^F. That is a state-guaranteed revenue pool for China's AI supply chain. Ahead of it sit the Sept 24 Xi-in-DC summit, an expected truce extension around the November 10 suspension expiry^F/C, and the standing CXMT-IPO catalyst^C (status to re-verify). The liquidity-rotation sequence we track on the desk — money moving in order: semis → software → China → crypto — has completed leg one in spectacular fashion. Legs three and four are the Q4-26/2027 trade.机制：北京冻结 H200 芯片交付，以强制推动国产技术栈的采用^F——Huawei Ascend、Alibaba T-Head 已规模化量产、Cambricon、DeepSeek-on-Ascend^F。这为中国 AI 产业链提供了一个国家担保的收入池。前方排着 9 月 24 日的 Xi 访美峰会、11 月 10 日管制暂停到期前后预期中的休战延期^F/C，以及 CXMT IPO 这一常备催化剂^C（状态待复核）。我们席位上跟踪的流动性轮动序列——资金按顺序流动：半导体 → 软件 → 中国 → 加密——已以壮观的方式走完第一棒。第三、第四棒是 2026 年四季度／2027 年的交易。

Expressions: BABA/9988 core, Tencent, KWEB beta. Timed to the Sept/Oct catalysts — not bought blind in July.标的：以 BABA/9988 为核心，Tencent、KWEB 贝塔。按 9／10 月的催化剂择时入场——而非 7 月盲买。

8.6 · Memory endgame protocol (managing a won trade)8.6 · 存储终局操作规程（管理一笔已经赢下的交易）

Framework: split the memory-chip complex into two parts. HBM-contracted — HBM is high-bandwidth memory, the premium chips wired directly to AI processors: an oligopoly, sold out, with HBM4 ramping at 2x^F. And commodity DRAM/NAND — the ordinary memory in PCs and phones: spot-price-driven, with demand destruction already visible in rationed PC/phone units^F, and 2027 supply waves coming. The stocks price the blend. The blow-off priced the commodity spike as if it were permanent.框架：把存储芯片板块拆成两部分。HBM 合约端——HBM 即高带宽内存，直连 AI 处理器的高端芯片：寡头格局、产能售罄、HBM4 以 2 倍规模爬坡^F。以及大宗 DRAM/NAND——电脑和手机里的普通内存：现货价格驱动、需求破坏已在被配给的 PC／手机出货中显现^F、2027 年供给浪潮将至。股票定价的是混合体。冲顶行情却把大宗价格脉冲当作永久来定价。

The protocol: (1) Reduce systematically into strength. A +300%-in-half-a-year move is what distribution (the phase where early winners sell into strength) looks like — and Micron guiding to 86% gross margins (profit after production costs)^F is what cycle peaks smell like. (2) Define an HBM-only re-entry valuation: what MU/SK would be worth if commodity margins fell back to mid-cycle and only the HBM annuity (the steady contracted income) persisted. Buy there during the 2027 digestion — not 15% below the high. (3) Treat SNDK/NAND as pure trading paper — rent it, never own it. (4) Use KOSPI (South Korea's stock index) as the live cycle thermometer. It currently reads fever: −7.9% in a session, SK Hynix −14.6%^F. SK Hynix's Nasdaq ADR (July 10) adds a cleaner US-hours gauge^F. The next structural memory leg is HBM4E/custom-base-die in 2027–28 — a different trade with different winners (foundry and packaging capture more of the stack), which is why 8.2's equipment names are the continuation vehicle.^I操作规程：(1) 系统性地逢强减仓。半年 +300% 一类的走势，正是派发（早期赢家趁涨势出货的阶段）的模样——而 Micron 给出 86% 毛利率（扣除生产成本后的利润率）指引^F，正是周期见顶的味道。(2) 定义一个仅按 HBM 估值的再入场价：若大宗毛利率跌回周期中位、只剩 HBM 年金（稳定的合约收入）存续，MU/SK 值多少钱。在 2027 年消化期到那里去买——而不是在高点下方 15% 处买。(3) 把 SNDK/NAND 当作纯交易性筹码——只租不买。(4) 用 KOSPI（韩国股指）作为实时周期温度计。当前读数为高烧：单日 −7.9%，SK Hynix −14.6%^F。SK Hynix 的纳斯达克 ADR（7 月 10 日）将提供一个更干净的美股时段量表^F。下一段结构性存储行情是 2027–28 年的 HBM4E／定制基础裸片（custom base die）——那是一笔赢家不同的另一种交易（代工＋封装将攫取产业链更大份额），这正是 8.2 的设备标的作为接续载体的原因。^I

8.7–8.10 · Seeded optionality (small, patient)8.7–8.10 · 播种式期权仓位（小仓位、有耐心）

Orbital compute / space: Starmind's 1M-satellite FCC filing (FCC is the US communications regulator), SpaceX's orbital data centers, and $2.17B/month of disclosed compute leasing (Anthropic $1.25B/mo + Google $0.92B/mo — against a Colossus that was only ~11% utilized before the deals, which tells you these are anchor-tenant subsidies as much as real demand^F/C) together make "compute above the grid politics" a fundable 2027–28 story. Consolidation has begun (Rocket Lab's $8B Iridium bid, June 29^F). Expressions beyond SPCX itself: RKLB, ASTS, launch-adjacent suppliers.轨道算力／太空：Starmind 的 100 万颗卫星 FCC（美国通信监管机构）申报、SpaceX 的轨道数据中心，以及每月 21.7 亿美元的已披露算力租赁（Anthropic 每月 12.5 亿美元＋Google 每月 9.2 亿美元——对照一个在交易达成前利用率仅约 11% 的 Colossus，这说明它们既是真实需求、也同样是锚定租户式补贴^F/C），合在一起使「凌驾于电网政治之上的算力」成为一个可获资金支持的 2027–28 年叙事。整合已经开始（Rocket Lab 于 6 月 29 日以 80 亿美元竞购 Iridium^F）。SPCX 之外的标的：RKLB、ASTS、发射链相关供应商。

Defense-AI/drones: AVAV's $500M Army award — for counter-drone systems, note, not drone supply^F — typifies the money flow. Golden Dome (a $185B estimate, with $3.2B of space-interceptor awards already spread across 12 firms^F) compounds through FY-27. Anduril's private shares changing hands near a $100B value, against a $61B funding round in May^C, screams froth in the private market. Own the suppliers the money flows through — not the halo names.国防 AI／无人机：AVAV 获美国陆军 5 亿美元订单——注意是反无人机系统、而非无人机供应^F——是这类资金流的典型。「金穹」（Golden Dome，估算 1,850 亿美元，12 家公司已分获 32 亿美元太空拦截器订单^F）将贯穿 2027 财年持续复合。Anduril 二级转让估值标记接近 1,000 亿美元，而 5 月融资轮仅 610 亿美元^C——一级市场的泡沫呼之欲出。应持有承接订单流的供应商——而非光环本身。

Agentic payments rails: x402/ACP-class standards (protocols that let AI agents pay each other) plus stablecoin settlement in the GENIUS-law era — COIN/Base is the listed pure-play (Section 10).智能体支付轨道：x402/ACP 一类标准（让 AI 智能体之间互相付款的协议）＋GENIUS 法案时代的稳定币结算——COIN/Base 是上市纯正标的（§10）。

Tariff-refund arbitrage: $166B has been collected. About $23B is already approved and being paid via CAPE. The DOJ's appeal is the freeze risk^F. The refunds accrue to big importers' 2026–27 EPS (earnings per share) — Walmart's claim alone is ~$10.2B^C, though it says it will reinvest the money in lower prices. A secondary market in refund receivables (the right to collect a refund, bought and sold like an IOU) is forming^C. The special-situations pocket includes both the claimants and the assignees (the buyers of those claims).^S关税退款套利：已征收 1,660 亿美元。约 230 亿美元已获批并正通过 CAPE 支付。DOJ 的上诉是冻结风险所在^F。退款将计入大型进口商 2026–27 年的 EPS（每股盈利）——仅 Walmart 一家的申索即约 102 亿美元^C，尽管其表示将把这笔钱重新投入降价。退款应收权益（收取退款的权利，可像欠条一样买卖）的二级市场正在形成^C。这一特殊情况机会口袋同时包含申索方与受让方（购买这些申索的人）。^S

9The Ex-AI Complex, Commodities, Rates, and the Dollar非 AI 板块、大宗商品、利率与美元

9.1 · “Will ex-AI be obliterated again?” is the wrong question9.1 · “非 AI 资产会再遭碾压吗？”是个问错了的问题

The fear inside that question describes the first half of 2026. Memory and chip stocks soared while ex-AI equities (everything outside the AI trade) got starved of new money. But look at what actually happened. The Dow had its best first half since 2021. The Russell (an index of smaller US companies) had its best since 1991. Equal-weight (where every stock counts the same) beat cap-weight (where bigger companies count more) by 2.4 points^F. The obliteration already didn’t happen. What happened was a sorting — and the sorting, not a blanket ex-AI verdict, is what continues:这个问题里潜藏的恐惧，描述的是 2026 年上半年。当时存储与芯片股一路狂飙，非 AI 股票（AI 交易之外的一切）拿不到新增资金。但看看实际结果：道指录得 2021 年以来最佳上半年。罗素指数（美国小盘股指数）录得 1991 年以来最佳。等权指数（每只股票权重相同）反而跑赢市值加权指数（大公司权重更大）2.4 个百分点^F。所谓“碾压”并没有发生。真正发生的是一场分拣。延续下去的正是这场分拣——而不是对非 AI 资产的一刀切裁决：

• The ex-AI longs that genuinely work are consolidation stories (companies buying each other), not macro stories. Regional banks rode a merger wave to a seven-year high (KRE +19%^F). Biotech rode $29B of acquisitions in two March weeks (XBI +28.5%^F). Both work in every scenario except E, because the driver is deal math, not the consumer.真正跑得通的非 AI 多头，是行业整合（公司互相并购）故事，而不是宏观故事。区域性银行乘并购浪潮升至七年高点（KRE +19%^F）。生物科技在 3 月的两周内迎来 290 亿美元收购（XBI +28.5%^F）。除情景 E 外，两者在所有情景下都成立，因为驱动力是并购算术，不是消费者。
• The genuine losers are ad-driven stocks and the debt-heavy consumer (communication services −6.2%, consumer discretionary negative^F). Those sectors sit exactly where the affordability election hits. Consumer staples (everyday essentials like food and household goods), the consensus victim, actually sit mid-pack. Their real threat is input costs: Campbell’s cut its earnings-per-share guidance by double digits because Section 232 tariffs (duties under US trade law) raised the cost of its cans^C. The tariff re-plumbing keeps that threat alive.真正的输家是靠广告吃饭的股票和背着重债的消费者（通信服务 −6.2%，可选消费为负^F）。这些板块恰恰暴露在“负担能力”选举争夺的核心议题之下。共识眼中的受害者——必需消费品（食品、日用品等生活必需品）——实际居于中游。它们的真实威胁是投入成本：Campbell’s 因 232条款关税（美国贸易法下的进口关税）推高罐体成本，将每股盈利指引下调双位数^C。关税体系的重新布管让这一威胁持续存在。
• The small-cap trap deserves its own warning. The Russell’s leadership is partly chip stocks in a small-cap costume: 16 of its top 50 performers are semiconductor-linked. The index now trades at a premium to the S&P (26.2x vs 20.9x NTM^F — that is, price versus the next twelve months of expected profits). If Act I’s unwind continues, the Russell does not hedge it. It participates in it. The clean ex-AI expression is the equal-weight industrial/financial/health-care middle, not “small caps” as a class.小盘股陷阱值得单独警告。罗素指数的领涨力量，部分是披着小盘股外衣的芯片股：前 50 大涨幅股中有 16 只与半导体相关。该指数目前相对标普已有溢价（未来十二个月预期盈利的 26.2 倍对 20.9 倍^F）。若第一幕的解杠杆继续，罗素指数无法对冲它，只会参与其中。干净的非 AI 表达，是等权的工业/金融/医疗保健中间层，而不是笼统的“小盘股”。
• Rate-sensitive stocks are hostages to §2’s macro path — which is exactly why you own them now. Homebuilders lag (mortgage rates threatened 7% during the war^F). They are the most leveraged and cheapest way to bet on the path from cooling inflation to rate cuts. If our Fed call is right, the builder ETFs ITB/XHB re-rate before the first cut is delivered. If it’s wrong, they tell you fast.利率敏感型资产是 §2 宏观路径的人质——这正是现在持有它们的意义。住宅建筑商落后（战争期间按揭利率一度逼近 7%^F）。它们是押注“通胀降温 → 降息”路径中弹性最大、也最便宜的工具。如果我们对美联储的判断正确，建筑商 ETF（ITB/XHB）会在首次降息落地前完成重估。如果判断错了，它们会最快告诉你。

The answer by scenario, compressed:按情景条件给出的答案，压缩如下：

The portfolio answer is a barbell across the question itself (holding two opposite exposures at once). One side: Act II AI capex plays (capex = capital spending). The other: rate-sensitive domestic value stocks. It is financed by what we are not holding — broad passive index funds at 20.9x forward earnings, with a thin risk premium (little extra reward over safe bonds) and record concentration.^I落实到组合层面，就是围绕这个问题本身构建杠铃（同时持有两端相反的仓位）。一端是第二幕的 AI 资本开支（capex，即企业的资本性支出）标的。另一端是利率敏感的本土价值股。资金来源是我们不持有的部分——20.9 倍前瞻市盈率、风险溢价（相对安全债券的额外回报）单薄、集中度创纪录的宽基被动指数。^I

9.2 · Commodities: one glut, one metal with a calendar, one fuel cycle9.2 · 大宗商品：一场过剩、一种自带日历的金属、一条燃料循环

Oil is now a supply story, and the supply is winning. The IEA’s June balance shows 2027 supply running more than 5 million barrels a day above demand — a surplus with few precedents outside 2020^F. The EIA sees war-halted output fully restored by Q1-27^F. OPEC+ is adding barrels into that glut, if only symbolically (+188 thousand barrels a day for July)^F. And returning Iranian oil is already selling at widening discounts, with 20M+ barrels unsold off Asia^C. Our band: Brent $58–75 into 2027, fraying lower as the surplus builds — with two asymmetries worth owning rather than fighting. First, speculators are the most bearish on oil since 2009^C. So the deal-failure branch (~Aug 16) is a cheap bet on a price spike, not a reason to be structurally long. Second, refiners capture the round trip: their margins (“cracks” — the gap between crude cost and fuel prices) are wide, and pump prices are sticky^F. The H2 long is the refining margin, not the barrel. Energy stocks broadly: trade them as Scenario-D insurance, not as a thesis.^I石油如今是一个供给故事，而且供给正在获胜。IEA 六月平衡表显示，2027 年供给将超出需求 500 万桶/日以上——除 2020 年外历史上鲜有先例的过剩^F。EIA 预计战争造成的停产将在 2027 年一季度前全面恢复^F。OPEC+ 正象征性地增产（7 月 +18.8 万桶/日），迎头撞上这一过剩^F。回归的伊朗原油已在以不断扩大的折价成交，亚洲外海尚有逾 2,000 万桶未售出^C。我们的区间：布伦特 58–75 美元直至 2027 年，并随过剩累积而向下磨损——其中有两个不对称值得持有而非对抗。第一，投机头寸是 2009 年以来最看空的^C。因此协议破裂分支（约 8 月 16 日）只是一次便宜的“赌油价急涨”机会，不是结构性做多的理由。第二，炼厂赚走了整个来回：裂解价差（原油成本与成品油价之间的炼厂利润）宽阔，终端油价黏性十足^F。下半年该做多的是炼厂利润，不是油桶本身。能源股整体上：当作情景 D 的保险来交易，别当作论点。^I

Gold’s 2026 taught a lesson the 2027 trade depends on. Gold did not peak on war. It peaked in January ($5,595) on the debasement bid (buying driven by fear that money itself is losing value). Then it fell through the entire conflict, because the oil shock forced the Fed hawkish (leaning toward higher rates) and pushed the dollar to a 15-month high^F. Gold’s enemy this year was never geopolitics. It was the hike. Which is exactly why the base case ahead is friendly. The hike threat dies. Real yields (bond yields after inflation) at 2.2% fall back as cuts begin. Central banks are still buying ~850 tonnes a year^F. And the fiscal math — deficits over $2T, short-term bill issuance already past the Treasury advisory committee’s (TBAC) guidance^F/C — is the structural sponsor. Path: gold builds a base at $4,000–4,300 while Fed pricing settles, then breaks through $5,600 into 2027^I. Silver ($61, gold-silver ratio ~67^F) is the higher-octane version once the turn confirms. Platinum’s squeeze already ran and unwound (−44% from January^F) — leave it.黄金的 2026 年教了一课，2027 年的交易正依赖这一课。黄金不是因战争见顶。它在 1 月（5,595 美元）因“货币贬值买盘”（担心纸币本身贬值而买入黄金）见顶。随后在整场冲突期间一路下跌，因为油价冲击迫使美联储转鹰（倾向加息），美元升至 15 个月高点^F。黄金今年的敌人从来不是地缘政治，而是加息。这恰恰是往后基准情形对它友好的原因。加息威胁消亡。2.2% 的实际收益率（扣除通胀后的债券收益率）随降息周期回落。各国央行仍在以约 850 吨/年的速度买入^F。财政算术——赤字超 2 万亿美元、短债发行已超出财政部借款咨询委员会（TBAC）指引^F/C——是它的结构性赞助人。路径：在美联储定价明朗前于 4,000–4,300 美元筑底，随后突破 5,600 美元进入 2027 年^I。白银（61 美元，金银比约 67^F）是拐点确认后波动更大的进攻版。铂金的逼空已跑完并回吐（较 1 月 −44%^F）——不必再碰。

Copper is the premier structural long, with a tariff calendar attached. The long-term shortage math held up under checking. Chilean output is down 9% year on year, with the giant Escondida mine down 16%. Ore grades have halved over two decades. JPM estimates a ~330kt shortfall of refined copper in 2026. The IEA’s map of planned supply covers only ~70% of 2035 demand^F/C. But two near-term cushions are real too: Cobre Panamá is restarting, and the US COMEX exchange holds a record ~650kt of inventory rushed in ahead of tariffs^F/C. Now overlay the policy calendar: the Section 232 copper framework proposes 15% duties from January 2027, 30% from 2028^C. Translation: the US copper price decouples upward from the world price on a legal schedule. Accumulate during the H2 2026 digestion (that inventory pile is your entry mechanism). Express it through the major miners and COMEX-linked names. And treat the COMEX–LME spread (the US-vs-London exchange price gap) as a trade in its own right. This is also the quiet robot trade: a humanoid robot is copper and magnets wearing a face.^I铜是头号结构性多头，且自带一份关税日历。长期缺口的算术经受住了核验：智利产量同比 −9%，Escondida 大矿 −16%。矿石品位二十年间腰斩。JPM 估算 2026 年精炼铜缺口约 33 万吨。IEA 的供给版图仅能覆盖 2035 年需求的约 70%^F/C。但两个近期缓冲同样成立：Cobre Panamá 正在复产，且美国 COMEX 交易所持有创纪录的约 65 万吨抢在关税前运入的库存^F/C。再叠加政策日历：232条款铜关税框架提议自 2027 年 1 月起征收 15% 关税、2028 年起 30%^C。翻译过来：美国铜价将按法定时间表相对世界价格向上脱钩。在 2026 年下半年的消化期建仓（那堆库存正是你的入场机制）。通过大型矿商和 COMEX 挂钩标的做结构性表达。并把 COMEX–LME 价差（美国与伦敦交易所之间的价差）当作一笔独立交易。这也是那笔安静的机器人交易：人形机器人不过是铜和磁材戴上了一张脸。^I

Uranium: the cycle moved to where we said the value was — the fuel cycle (mining, enrichment, and fuel-making, not the reactors). The spot price (for immediate delivery) has gone nowhere ($85.50, rangebound^F). But term prices (long-term contract prices) printed $94/lb, the highest since 2008^F. That is a utility-contracting cycle, not a spot squeeze. It pays CCJ-class producers and physical holders slowly — and pays enrichment scarcity violently. The July 1 confirmation: Centrus signed a $900M DOE task order to commercialize HALEU (a high-grade reactor fuel) after finishing its demo contract early^F. Meanwhile, the quotas allowing Russian fuel step down toward a hard stop in 2028^F. The 2027–28 contracting wave (13 nuclear projects tied to giant cloud companies, >9.8GW committed^C) lands on an enrichment base that cannot expand at spot-market speed. Fuel cycle over reactor SPACs (speculative blank-check reactor startups), unchanged — now with the state literally funding the thesis.^I铀：周期移动到了我们此前指认的价值所在——燃料循环（采矿、浓缩与燃料加工，而非反应堆本身）。现货价（即期交割价格）原地踏步（85.50 美元，区间震荡^F）。但长协价格（长期合同价格）打印出 94 美元/磅，为 2008 年以来最高^F。这是一轮电力公司签长约的周期，不是现货逼空。它会缓慢回报 CCJ 级生产商和实物持有者，却会猛烈回报铀浓缩的稀缺性。7 月 1 日的确认：Centrus 提前完成示范合同后，签下9 亿美元的 DOE HALEU（一种高丰度反应堆燃料）商业化任务订单^F。与此同时，允许进口俄罗斯燃料的豁免配额正逐级递减，2028 年硬性终止^F。2027–28 年的签约浪潮（13 个与云计算巨头相关的核电项目、承诺装机逾 9.8GW^C）将落在一个无法按现货市场速度扩张的浓缩产能基座上。燃料循环优于反应堆 SPAC（借壳上市的投机性反应堆公司），结论不变——如今国家还在真金白银地资助这一论点。^I

And the strategic-minerals sleeper: magnets. NdPr (the rare-earth metals inside high-power magnets) at ~$90/kg has fallen below MP’s $110 floor guaranteed by the Defense Department. The contract-for-difference (which pays MP the gap between market price and floor) is now paying^C. Meanwhile China is tightening its licensing of magnet technology — even inside a truce that suspends export controls, a truce that itself expires November 10^C/F. Every scenario in §5 that includes a robotics Section 232 tariff signing or a truce wobble reprices Western magnet capacity the same direction: up.还有战略矿产中的沉睡者：稀土磁材。NdPr（高性能磁体所需的稀土金属钕镨）价格约 90 美元/公斤，已跌破 MP 与美国国防部约定的 110 美元价格下限。差价合约（补足市价与下限之间差额的合约）已开始付钱^C。与此同时，中国即便身处一份暂停出口管制的休战之中，仍在收紧磁体技术许可——而这份休战本身将于 11 月 10 日到期^C/F。§5 中任何一个包含 232条款机器人关税签署或休战动摇的情景，都会把西方磁材产能朝同一个方向重新定价：向上。

9.3 · Rates: a two-act structure of their own9.3 · 利率：自成一体的两幕结构

Act one (H2 2026): the range holds and the middle of the curve wins. The 10-year Treasury yield sits at 4.46%, mid-band (we carry 4.2–4.75%). The term premium (the extra yield investors demand to hold long-term bonds) is ~0.67%, already a cycle high^C. The MOVE index (bond-market volatility) at 72 says the bond market believes the range even while stocks panic^C. Treasury’s issuance plan protects long-term bonds through the election: bills (short-term government debt) at 21.7% of the total, already above guidance; buyback frequency doubled; dealers expecting no increase in long-term issuance at the August refunding^F. That is Bessent’s lite version of yield-curve control, running as designed. As hike pricing dies (our base case), short-term yields fall first. The 2s10s steepener — a bet that the gap between 2-year and 10-year yields widens (now 0.46 percentage points^C) — is the single cleanest expression of this note’s macro spine. It pays if the Fed cuts (bull steepening). It also pays in the term-premium event (bear steepening). It loses only if a hike is actually delivered — precisely the world where our equity book is handed better entries.^I第一幕（2026 年下半年）：区间守住，曲线中段取胜。10 年期美债收益率 4.46%，位于区间中部（我们给出 4.2–4.75%）。期限溢价（持有长期债券所要求的额外收益）约 0.67%，已是周期高点^C。MOVE 指数（债市波动率指标）72，说明即便股市恐慌，债市依然相信这个区间^C。财政部的发债安排会一路保护长端直到选举：短债（短期国债）占债务比重 21.7%，已高于指引；回购操作频率翻倍；交易商预期 8 月再融资不会增加长债发行^F。这正是 Bessent 的“轻量版收益率曲线控制”，按设计运转。当加息定价消亡（我们的基准情形），短端收益率率先回落。2s10s 陡峭化交易——押注 2 年与 10 年期收益率之差走阔（当前 0.46 个百分点^C）——是本报告宏观主线最干净的单一表达。美联储降息，它赚钱（牛陡）。期限溢价事件爆发，它也赚钱（熊陡）。它只在加息真正落地的世界里亏钱——而那恰恰是我们的股票头寸拿到更好买点的世界。^I

Act two (2027): the term-premium event moves from tail risk to base-case risk. Everything that suppressed long-term yields in 2026 reverses on schedule. Long-term issuance must rise against a FY27 deficit above $2T^C. The tariff-refund drain continues. Cloud giants keep flooding the bond market ($159B YTD already exceeds all of 2025; UBS sees up to $240B^F). And Warsh’s task-force doctrine — a smaller Fed balance sheet, Treasuries only, mortgage bonds run off or sold — lands just as the cut cycle begins^C. “Cut the short rate, shrink the footprint” steepens the curve violently in the benign case. In the malign case it forces the 10Y through 5.25% — our designated window Q3 2027, and the mechanism behind the 2027 correction (§5). The discipline: own duration (long-term bond exposure) tactically in H2 2026, the disinflation window. Refuse it structurally in 2027. Let the steepener carry the through-position.^I第二幕（2027 年）：期限溢价事件从尾部风险升格为基准情形风险。2026 年压制长端的一切因素都将按时间表反转。面对超过 2 万亿美元的 2027 财年赤字，长债发行必须加码^C。关税退款的资金抽离仍在继续。云计算巨头的发债还在滚雪球（年内 1,590 亿美元已超过 2025 年全年；UBS 预计最高 2,400 亿美元^F）。而 Warsh 的工作组信条——美联储资产负债表更小、只持国债、按揭债券缩减乃至出售——恰好在降息周期开启时落地^C。“降短端利率、缩资产足迹”在良性情形下让曲线剧烈牛陡。在恶性情形下，它会把 10 年期收益率推穿 5.25%——我们指定的窗口是 2027 年三季度，这也是 2027 年调整背后的机制（见 §5）。纪律：2026 年下半年（去通胀窗口）战术性持有久期（长期债券仓位）。2027 年结构性拒绝久期。让陡峭化交易承载贯穿始终的头寸。^I

9.4 · The dollar: squeeze first, fade second — and the yen is the sleeper9.4 · 美元：先逼空、后回落——而日元是那个沉睡者

The war episode validated the Milkshake mechanics (the idea that in a crisis, global money gets sucked into the dollar). Risk-off plus an oil shock bid the dollar, exactly as the framework predicts. The aftermath is textbook too: the dollar index (DXY) at a 15-month high (~101.6), squeezing a 14-year-record consensus of bets against the dollar^F. That squeeze is not a trend. It is hike-pricing wearing an FX costume. The sequencing from our Fed call: the squeeze persists while September/October hike odds live. Then the fade resumes into 2027 as the cut cycle reprices. But size it modestly — everyone already owns the weak-dollar thesis. The crowd being early does not make it wrong, only expensive to hold.^I战争插曲验证了“美元奶昔”机制（危机时全球资金被吸进美元）。风险规避叠加油价冲击推高美元，与该框架的预测分毫不差。事后走势也堪称教科书：美元指数（DXY）升至 15 个月高点（约 101.6），逼空一个创 14 年纪录的做空美元共识^F。这场逼空不是趋势。它是披着外汇外衣的加息定价。由我们的美联储判断推出的时序：只要 9 月/10 月的加息概率还活着，逼空就会持续。随后，随着降息周期重新定价，美元回落将在通往 2027 年的路上重启。但仓位要克制——所有人都已持有看空美元的论点。人群来早了不代表它错，只是拿着很贵。^I

The asymmetric currency trade is not the euro (heavy at ~1.14 against a European Central Bank that is hiking^C). It is the yen at ~162 — its weakest in four decades — against a Bank of Japan at 1.0% and still hiking^F. Every month of that gap loads the spring. The yen is the world’s favorite funding currency (borrow cheap yen, buy risky assets elsewhere). A disorderly unwind of that borrowing is a known catastrophe pattern — and it targets exactly the crowded, levered, momentum-heavy positioning this cycle has built (record margin debt, trend-fund longs, everything financed by the carry). We hold long-yen convexity (option-like positions that pay off big if the yen jumps) as a portfolio hedge, not a currency view. Cheap, quiet, and pointed directly at the Q3-2027 correction window and any Scenario-B credit event.^I不对称的外汇交易不是欧元（面对仍在加息的欧洲央行，约 1.14 的欧元显得沉重^C）。而是约 162 的日元——四十年来最弱——对上政策利率 1.0% 且仍在加息的日本央行^F。这一背离每持续一个月，都在上紧发条。日元是全世界最爱用的融资货币（低息借日元，去买别处的风险资产）。这类借贷的无序解杠杆是一种已知的灾难模式——瞄准的正是本轮周期积累的拥挤、加杠杆、重动量头寸（创纪录的融资余额、趋势基金多头、靠套息融资撑起的一切）。我们把做多日元的凸性（类似期权、日元暴涨时收益放大的头寸）当作组合层面的对冲，而非外汇观点。便宜、安静，直指 2027 年三季度的调整窗口和任何情景 B 信用事件。^I

10Bitcoin and the Financial Rails — The Fourth Cycle Ends, the Rails Begin比特币与金融轨道——第四轮周期终结，轨道时代开启

10.1 · Anatomy of the fourth bear, without the cope10.1 · 第四轮熊市解剖——不找借口

Where the cycle stands, nine months after the top: Bitcoin is down 53% from $126K. ETH is down 68%. Bitcoin dominance (its share of the whole crypto market) sits near 55%^F. The leverage is gone — futures open interest (the amount of outstanding leveraged bets) is down 19%, and funding rates (the fee leveraged traders pay to hold positions) are flat^C. Miners are giving up in public. Hashprice (the revenue a miner earns per unit of computing power) is at record lows. Mining difficulty fell 11% in one June adjustment. The average cost of producing a coin is near $87K, against a $60K price. Miners are rebranding themselves as AI-compute hosts^F. And the two big structural buyers of the 2024–25 rally have both flipped. Spot ETFs (funds that buy and hold actual bitcoin) posted their worst month ever in June (−$4.5B out, IBIT −$3.55B)^F. And the treasury-company flywheel — firms that issue stock to buy bitcoin — is now running in reverse. Strategy has paused purchases, sold coins for the first time since 2022, and set up a formal framework to sell up to $1.25B more. Metaplanet fell below 1x mNAV (the company's market price versus the value of the coins it holds) and is pivoting its business model^F.周期现状——距顶部九个月：比特币自 $126K 回撤 −53%。ETH −68%。比特币市值占比（它在整个加密市场中的份额）约 55%^F。杠杆已经出清：期货未平仓合约（市场上未了结的杠杆押注总量）−19%，资金费率（杠杆交易者为持仓支付的费用）持平^C。矿工正在公开投降。Hashprice（矿工每单位算力赚到的收入）创历史新低。6 月一次难度调整下调 11%。平均生产成本接近 $87K，而币价只有约 $60K。矿工纷纷转型做 AI 算力托管^F。同时，2024–25 年上涨的两大结构性买家都已反转。现货 ETF（直接买入并持有真实比特币的基金）6 月录得史上最差单月（净流出 −$4.5B，其中 IBIT −$3.55B）^F。财库公司飞轮——靠发股票买比特币的公司——正在倒转。Strategy 已暂停购币，自 2022 年以来首次卖出比特币，并正式确立了再卖至多 $1.25B 的框架。Metaplanet 的 mNAV（公司市值与其持币价值之比）跌破 1 倍，正在调整商业模式^F。

We should be honest about the part of our own thesis that died. We once expected the bottom to arrive as a forced-selling climax: MSCI (a major index provider) evicting the treasury companies from its indexes, forcing funds to dump the shares. That question was settled in January — MSCI kept them^F. So there will be no index-eviction flush. What replaces it is slower, and in some ways worse for the impatient: a telegraphed, permanent overhang — a large, known seller sitting above the market. Strategy holds 847K coins at an average cost of $75.7K, is $13B underwater, and has an approved plan to sell in a slow drip^F. That supply caps every rally until the cycle genuinely turns. Bear markets that end this way — on an overhang rather than a flush — end on time and exhaustion, not on a single dramatic day. That fits the historical trough window (12–14 months after the top: October–December 2026). It argues for buying on a schedule, not for heroics.先坦白我们论点中已经失效的部分。我们曾预期底部会以被动抛售高潮的形式出现：MSCI（大型指数编制公司）把财库公司剔除出指数，迫使基金抛售。这件事 1 月已有定论——MSCI 保留了它们^F。所以不会有指数驱逐式的清洗了。取而代之的东西更慢，对没耐心的人来说某些方面更糟：一个明牌的、永久性的悬顶压力——市场上方悬着一个人人可见的大卖家。Strategy 持有 847K 枚币，平均成本 $75.7K，浮亏 $13B，还有一份已获批的慢速滴灌式抛售计划^F。在周期真正转向之前，这批筹码会压住每一次反弹。以悬顶压力而非清洗收尾的熊市，靠时间与耗竭结束，而不是某一根 K 线。这与历史底部窗口吻合（顶部后 12–14 个月：2026 年 10–12 月），也支持按时间表买入，而不是豪赌抄底。

10.2 · The trough protocol10.2 · 底部操作规程

• Zone: $42–58K is the main buy zone. The old $45–58K floor assumed ETF inflows would soften the falls; June disproved that^F. The stress floor is $38K if a Scenario-B credit event hits. The low-$30Ks happen only in the worst tail case: rate hikes plus a recession. Note that the top of the zone has already traded ($58.2K in late June) — which is why the protocol starts now, not in October.^I区间：主买入区间 $42–58K。旧的 $45–58K 底部假设 ETF 资金流入能缓冲下跌；6 月证明缓冲不存在^F。若情景 B 的信用事件发生，压力底为 $38K。只有在“加息+衰退”的最差尾部情景中才会跌到 $30K 出头。注意：区间上沿已经成交过（6 月末的 $58.2K）——这正是规程从现在启动、而不是等到 10 月的原因。^I
• Schedule: buy in monthly installments from July to January, with extra weight in October–November (the seasonal low, plus clarity after the midterm elections, plus the most likely trough window). Keep a reserve that only gets spent below $45K. Consensus now also expects an October low^C — one more reason not to bet everything on a date the whole crowd is watching.节奏：7 月到次年 1 月按月分批买入，权重偏向 10–11 月（季节性低点 + 中期选举后局面明朗 + 概率最高的底部窗口）。另留一笔钱，只在 $45K 下方才动用。市场共识现在同样预期 10 月见底^C——这又多了一条理由：别把全部筹码押在一个人人都盯着的日期上。
• Confirmation signals for the turn: ETF flows turn positive for 3+ weeks. Hashprice stabilizes above miners' marginal cost (what it costs to produce one more coin). Bitcoin dominance rolls over as risk appetite returns to altcoins (smaller cryptocurrencies). Funding rates stay neutral through a +20% rally — meaning real spot buying drives it, not leverage. Any three of the four = accelerate buying. None by February = the cycle framework itself goes under review.^I转向的确认信号组：ETF 资金流连续 3 周以上转正。Hashprice 稳定在矿工边际成本（多挖一枚币的成本）之上。随着风险偏好回流山寨币（较小的加密货币），比特币市值占比见顶回落。一轮 +20% 的反弹中资金费率全程保持中性——说明是现货真金白银在买，不是杠杆在推。四项中任意三项成立 = 加快买入。到 2 月一项都没有 = 周期框架本身进入复核。^I

10.3 · Recovery drivers, honestly ranked10.3 · 复苏驱动因素——如实排序

1. The global liquidity turn — liquidity means the amount of money flowing through the financial system, and this is the real driver. It dwarfs everything crypto-specific. Our Fed path (the hike dies → cuts in the first half of 2027), plus a burst of new lending in China, plus a Bank of Japan that raises rates only slowly and under constraint — that is the 2027 cocktail. If the macro call fails, no catalyst below saves the target.^I全球流动性转向——流动性指金融体系中流动的资金量，这才是真正的驱动因素，量级远超一切加密特有因素。我们的美联储路径（加息夭折 → 2027 上半年降息），加上中国的信贷脉冲（新增放贷的一波提速），再加上日本央行受限但渐进的加息正常化，就是 2027 年的配方。若宏观判断落空，下面任何催化剂都救不了目标价。^I
2. GENIUS go-live — the new US stablecoin law taking effect. The statutory outer bound is January 18, 2027, but a "120 days after the rules" clause could pull it forward to November 2026^F. Every compliant dollar of stablecoin float (the total value of stablecoins in circulation) creates two things: built-in demand for T-bills, i.e. short-term government debt (Bessent's channel), and revenue for the payment rails (Coinbase's channel). The float is already $315B, up 25.5% year over year^F.GENIUS 法案落地——美国新的稳定币法生效。法定最晚期限是 2027 年 1 月 18 日，但“规则出台后 120 天”条款可能把时点提前到 2026 年 11 月^F。每一美元合规的稳定币浮存（流通中的稳定币总量）都会带来两样东西：对短期国债的结构性需求（Bessent 的渠道），以及轨道收入（Coinbase 的渠道）。浮存已达 $315B，同比 +25.5%^F。
3. CLARITY — the crypto market-structure bill. Its odds of a House floor pass have decayed to ~48%^F; ethics fights killed the pre-recess window. Treat passage as 2027 upside, not the 2026 base case.CLARITY 法案——加密市场结构法案。其院会通过概率已降到约 48%^F；道德风波毁掉了休会前的窗口。应把它的通过当作 2027 年的上行期权，而非 2026 年的基准情形。
4. Halving anticipation — the halving is the once-every-four-years event that cuts new bitcoin supply in half. Its price multiple shrinks every cycle (92x → 30x → 7.9x → 1.9x^C). It can add fuel to a turn that is already underway. It cannot cause one.减半预期——减半是每四年一次、把新增比特币供应砍半的事件。其价格倍数逐轮衰减（92x → 30x → 7.9x → 1.9x^C）。它只能给一个已经启动的转向添把火，本身不是成因。

Target: $95–150K in H2 2027 — explicitly conditional on driver #1. The four-year clock also runs into our 2028 top-formation window: halving-year euphoria, an election year, and a completed set of topping signs. Sell that strength on a schedule, like everything else on this map.^I目标价：2027 年下半年 $95–150K，明确以驱动因素 #1 为前提。四年时钟同样指向我们的 2028 年顶部构筑窗口：减半年的亢奋 + 大选年 + 一套集齐的见顶征兆。到时按时间表卖出那段强势——和这张地图上的其他一切一样。^I

10.4 · The rails are the higher-conviction trade10.4 · 轨道才是更高置信度的交易

Here is the asymmetry the price drop hides. While the asset was in a bear market, the rails quietly won the standards war. The "rails" are the payment plumbing being built for AI agents — software that buys things on its own. x402, a protocol for those agentic payments, settles them with the USDC stablecoin handling 99.8% of protocol transactions^C. Google's AP2 adopted x402 as its stablecoin extension. Mastercard shipped Agent Pay and is buying BVNK for $1.8B. Amazon's Bedrock agents settle x402 payments on Base — Coinbase's chain — in ~200ms. Visa's stablecoin settlement runs at a $7B annualized rate^F/C. The agentic-commerce future that Act III of this note depends on has already chosen its plumbing. That plumbing is disproportionately Coinbase-adjacent.回撤掩盖了一个不对称。在资产走熊的同时，轨道悄然赢下了标准之战。所谓“轨道”，是为 AI 智能体（能自主下单购物的软件）搭建的支付管道。x402 是这类智能体支付的结算协议，其中 USDC 稳定币处理了 99.8% 的协议交易^C。Google 的 AP2 把 x402 采纳为其稳定币扩展。Mastercard 推出 Agent Pay，并以 $1.8B 收购 BVNK。Amazon 的 Bedrock 智能体在 Base——Coinbase 的链——上以约 200ms 完成 x402 结算。Visa 的稳定币结算已按 $7B 的年化规模运行^F/C。本报告第三幕所依赖的智能体商业未来，已经选好了自己的管道。而这些管道不成比例地围着 Coinbase 转。

The competitive shock proves the same point from the other side. The "Open USD" consortium — 140+ companies including Stripe, Visa, Mastercard, BlackRock and Coinbase itself — launched a stablecoin whose partners keep the reserve yield (the interest earned on the cash backing the coin). Circle fell 39% in a month^F. In stablecoins, whoever owns distribution — the customer relationships — owns the profits. An issuer without distribution is just a money-market fund with extra steps. So avoid CRCL. Own the distribution instead: COIN (perpetual futures — a popular leveraged crypto contract — plus USDC economics, Base, and x402: the listed pure-play on the whole stack), HOOD (its own blockchain layer, tokenized stocks in 120 countries, and a 7% USDG yield that pulls in deposits^C), plus the CME's record institutional volumes (+46% y/y, 24/7 trading live^F). The CME numbers are the boring confirmation: the asset class's market structure kept maturing straight through the bear.竞争冲击从另一侧印证了同一点。“Open USD”联盟——包括 Stripe、Visa、Mastercard、BlackRock 乃至 Coinbase 本身在内的 140 余家公司——推出了一款由合作伙伴保留储备收益（支撑该币的现金所产生的利息）的稳定币。Circle 一个月内跌了 39%^F。在稳定币领域，谁掌握分销（客户关系），谁就拿走利润。没有分销的发行方，只是多绕了几步的货币市场基金。所以回避 CRCL。转而持有分销方：COIN（永续合约——一种流行的加密杠杆合约，加上 USDC 经济利益、Base、x402——覆盖整个技术栈的上市纯正标的）；HOOD（自有区块链层、覆盖 120 个国家的代币化股票、以及吸引存款的 7% USDG 收益率^C）；再加上 CME 的机构成交量纪录（同比 +46%，24/7 交易已上线^F）。CME 的数字是乏味但可靠的佐证：整个熊市期间，这个资产类别的市场结构一直在走向成熟。

11The Playbook — Three Books, Sized for Being Early操作手册——三本账本，按提前入场定仓

11.1 · Architecture11.1 · 架构

Everything above compresses into three books (portfolios), each on its own clock. Two rules bind them — both learned from this quarter's red-team pass (a deliberate attempt to break our own thesis). First: Act I and Act II are one factor at different betas — the same underlying bet, just at different intensities. So we cap their combined weight. Second: entries are events, not prices. The calendar — August weakness, the October confluence (several catalysts landing at once), the period after CES — hands you the chances to add. Chasing strength in a group trading at GE Vernova's 63-times-earnings valuation is how the last buyer in finances everyone else's exit.^I上文的全部内容压缩为三本账本（三个组合），各有各的时钟。有两条纪律把它们绑在一起——都来自本季度的红队推演（刻意攻击自己论点的压力测试）。第一：第一幕与第二幕是同一因子在不同贝塔下的表达——本质是同一个赌注，只是力度不同。所以两者的合计仓位必须设上限。第二：入场看事件，不看价格。日历——8月的疲弱、10月的共振窗口（多个催化剂同时落地）、CES 之后——会把加仓机会递到你手上。在一个以 GE Vernova 63 倍市盈率交易的产业链里追高，等于最后进场的买家替其他所有人的退出买单。^I

11.2 · The books11.2 · 三本账本

Book 1 — The Macro Path (the spine). This book carries the trades from Sections 2, 5 and 9. The 2s10s steepener (a bet that 10-year rates rise relative to 2-year rates). Tactical duration (owning long-term bonds, which gain when yields fall) through the disinflation window. Homebuilders and other rate-sensitive stocks as the leveraged version of the same bet. Gold accumulation at $4,000–4,300 for the 2027 leg. A short lean on the oil glut — expressed by refusing energy exposure and owning refiner margins only. The dollar fade after the squeeze. And yen convexity (an option-like yen hedge that pays more the bigger the move) as the standing hedge. Kill criteria (the signals that force an exit): a rate hike actually delivered — cut duration, keep the steepener; or core CPI re-accelerating above 3.2% on tariff passthrough — kill the early-cut leg entirely.账本一——宏观路径（主轴）。这本账承接 §2、§5、§9 的交易。2s10s 陡峭化交易（押注10年期利率相对2年期利率上行）。在通胀回落窗口内战术性持有久期（即持有长期债券，收益率下行时获利）。地产建筑商等利率敏感股，作为同一赌注的放大版。在 $4,000–4,300 区间为 2027 年那段行情积累黄金。对原油过剩持偏空立场——做法是拒绝能源贝塔，只做多炼厂利润率。美元逼空之后做空美元。再加上日元凸性（一种类似期权的日元对冲，行情越大赔付越多）作为常设对冲。证伪条件（触发离场的信号）：一次加息真正落地——削减久期、保留陡峭化交易；或核心 CPI 因关税传导重新加速至 3.2% 以上——彻底砍掉"提前降息"这条腿。

Book 2 — The Thematic Barbell (the engine). A barbell holds two opposing positions at once. One side: the Act II longs. Behind-the-meter power (power generated on-site rather than bought from the grid) and electrical/thermal equipment, bought on drawdowns (price pullbacks) only. Optics bought on its violent flushes. Data-center construction. The uranium fuel cycle. Copper. Magnets and the robotics bill of materials (BOM — the parts list) ahead of CES. And the China convergence trade, timed to the September–November catalyst run. The other side: the managed Act I tail. Memory chips reduced into strength per the 8.6 protocol. NVDA held as a rented, datapoint-to-datapoint trade. The Mag-7 accumulation tranches (MSFT heaviest, META sized for volatility, GOOGL held), staged into earnings-week selldowns and the October window. Kill criteria per leg are marked in Sections 7–8. The book-level kill: 2027 capex (capital spending) guidance decelerating below +15% year over year. That is the signal the wave itself is cresting — and it de-rates both sides of the barbell at once. Which is exactly why the barbell is capped as one factor.账本二——主题杠铃（引擎）。杠铃即同时持有两端相反的仓位。一侧是第二幕多头：表后自供电（自建电源、不从电网买电）与电气/热管理设备，只在回撤（价格回落）中买入；光模块在其剧烈洗盘中买入；数据中心建设；铀燃料循环；铜；CES 之前布局稀土磁材与机器人物料清单（BOM，即零部件清单）；以及踩着 9–11 月催化剂序列布局的中国收敛交易。另一侧是受管理的第一幕尾仓：按 §8.6 协议在强势中减持存储芯片；NVDA 只作为一笔"租来的"、跟着数据点走的交易；Mag-7（美股七大科技巨头）分批建仓——MSFT 权重最高，META 按波动率定仓，GOOGL 持有不动——分批布置在财报周抛售与 10 月窗口。各条腿的证伪条件见 §7–§8 的标注。账本层面的证伪条件：2027 年资本开支指引减速至同比 +15% 以下。那是浪潮本身正在见顶的信号，且会同时给杠铃的两侧杀估值——这正是把杠铃整体作为单一因子设上限的原因。

Book 3 — Special Situations & Seeds (the options). The BTC tranche schedule plus the rails basket (Section 10). IEEPA refund receivables (tariff money the courts may order returned) and the earnings-per-share surprises at the claimant companies^S. SaaS dispersion pairs: short the vendors that charge per seat, long the ones converting to usage-based pricing. Agent-security cyber. Orbital compute seeds (RKLB-class, entered only after the SPCX halo fades). Defense-AI flow-through suppliers — not the Anduril halo itself. Sized so that any single line going to zero is a bad week, not a bad year.账本三——特殊情形与种子仓（期权性头寸）。BTC 分批建仓时间表，加上轨道（rails）篮子（§10）。IEEPA 退税应收款（法院可能判令退还的关税款）与索赔方公司的每股收益（EPS）惊喜^S。SaaS 分化配对交易：做空按席位收费的厂商，做多转向按用量收费的厂商。面向 AI 智能体安全的网络安全公司。轨道算力种子仓（RKLB 一类，仅在 SPCX 光环退潮后入场）。国防 AI 的传导受益供应商——而不是 Anduril 光环本身。仓位大小的原则：任何单一持仓归零只是糟糕的一周，而不是糟糕的一年。

11.3 · Hedges that pay in B and E11.3 · 在情景 B 与情景 E 中获益的对冲

Credit protection is historically cheap against this map. The HY OAS (the extra yield junk bonds pay over Treasuries) sits at 274 basis points — about 2.74 percentage points — near the tightest of the cycle. It barely blinked during a single session that erased $1.3T of semiconductor value^F. That makes it mispriced insurance against Scenario B's refinancing chain. Oracle's CDS (the cost of insuring its debt against default) is already at record wides — the market whispering where it breaks^C. Pair the credit protection with the yen convexity hedge (§9.4) and with SPX put-spread ladders (tiers of options that pay if the index falls), financed by trimming crowded Act II strength. And note: in Scenario E, duration is the hedge — and Book 1 already owns it for other reasons. The hedge budget is not pessimism. It is what lets us buy the October confluence with both hands if it arrives at Scenario-B prices.^I相对这张情景图，信用保护便宜到了历史级水平。高收益债利差（HY OAS，垃圾债相对国债多付的收益率）仅 274 个基点，约合 2.74 个百分点——接近本轮周期最窄位。在半导体市值单日蒸发 1.3 万亿美元的交易日里，它几乎纹丝不动^F。这就是针对情景 B 再融资链条的错价保险。Oracle 的 CDS（为其债券违约投保的成本）已升至创纪录宽位——那是市场在低声提示断裂点所在^C。把信用保护与日元凸性对冲（§9.4）配对，再加上 SPX 看跌价差梯队（指数下跌时获赔的分层期权组合），资金来自削减拥挤的第二幕强势仓位。还要记住：在情景 E 中，久期本身就是对冲——而账本一出于其他理由已经持有久期。对冲预算不是悲观主义。它是让我们在 10 月共振以情景 B 的价格到来时，能双手满仓买入的底气。^I

12The Falsification Dashboard证伪仪表盘

12.1 · Kill criteria, by call12.1 · 各项判断的证伪条件

12.2 · The weekly tells12.2 · 每周观察信号

Labor comes first: payrolls, revisions, continuing claims. The entire 2026 macro call lives there. Then the rest: the CPI energy contribution (August–November prints); September/October hike pricing; HY OAS and Oracle/CoreWeave spreads (what those two pay to borrow); hyperscaler (giant cloud company) capex language; GEV/VRT multiples vs earnings; BTC ETF flows and hashprice (what miners earn per unit of computing power); the yen; the ~Aug 16 Iran deadline; Section 301 trade litigation; and the KOSPI (South Korea's stock index), which remains the most honest single number in the world for where the memory cycle's temperature sits^F.就业优先：非农、修正值、续请失业金人数——整个2026年宏观判断都系于此。其次是：CPI能源分项贡献（8月至11月数据）；市场对9月/10月加息的定价；高收益债OAS及Oracle/CoreWeave利差（这两家借钱的成本）；超大规模云厂商（各大云计算巨头）的资本开支措辞；GEV/VRT估值倍数与盈利的对比；BTC ETF资金流与hashprice（矿工每单位算力的收入）；日元；约8月16日的伊朗协议期限；「301条款」贸易诉讼；以及KOSPI（韩国股指）——它仍是全球范围内衡量存储周期温度最诚实的单一数字^F。

13The Compression压缩

AAppendix — Verification Log & Method附录——核证记录与方法

Method. Every material claim in this note was re-checked in the week of publication. We ran a research sweep across twelve domains: Fed/rates, oil/Iran, metals, crypto, market breadth (how many stocks are actually rising), the hyperscalers (the giant cloud companies), semiconductors/memory, power/electrical, robotics, political economy, IPO/market structure, and software. Each claim was checked against primary sources or dated secondary sources. We then commissioned five adversarial "red-team" analyses — reviews hired specifically to attack the note's core calls. Where a check failed or corrected the draft, the text above carries the correction. Where the change mattered for positioning, we say so inline rather than silently repairing it.方法。本报告中的每一项实质性论断，都在发布当周重新核证。我们做了一次覆盖十二个领域的研究扫查：美联储/利率、原油/伊朗、金属、加密资产、市场广度（有多少股票真正在上涨）、超大规模云厂商（巨型云计算公司）、半导体/存储、电力/电气、机器人、政治经济、IPO/市场结构、软件。每项论断均对照一手来源或标注日期的二手来源核对。随后我们另请五份对抗性“红队”分析——专门受托来反驳本报告核心判断的审查。凡核证失败或修正了初稿之处，上文均已写入修正。凡修正对仓位安排有实质影响之处，我们在正文中明确说明，而非默默修补。

Material corrections absorbed (selected). The MSCI-deletion story behind the BTC capitulation (the panic-selling flush): refuted — index provider MSCI kept the treasury companies (firms whose main asset is bitcoin) in January. Warsh floating outright bond sales at the Sintra conference: not supported — that was doctrine and street expectation, not a proposal. "Staples and financials are the worst sectors": wrong — communications and discretionary are the laggards. GOOGL "first back to highs": partially played out (all-time high May 13). SOX (the semiconductor stock index) in the first half: +101%, not +80%. Iran's "premium pricing" claim: contradicted by widening discounts on Iranian Light crude. The Busan truce expiry: November 10, not late October. The Section 232 semiconductor tariff (a national-security tariff under US trade law): already in force, not still under investigation. The Section 232 robotics case: report deadline passed, decision pending. The power complex: never needed a correction (GEV hit an all-time high June 30) — the crowding warning in 8.1 (too many investors piled into the same trade) exists because verification demanded it.已吸收的实质性修正（节选）。关于 BTC 投降式抛售（恐慌性抛售）的 MSCI 剔除机制：被证伪——指数公司 MSCI 于1月保留了财库类公司（以比特币为主要资产的公司）。Warsh 在 Sintra 央行年会上提出直接卖出债券之说：证据不支持——那是学理立场与市场预期，并非正式提议。“必需消费与金融是最差板块”：有误——落后的是通信与可选消费。GOOGL“率先收复历史高点”：部分兑现（5月13日创历史新高）。SOX（费城半导体指数）上半年：+101%，而非+80%。伊朗“溢价销售”之说：与 Iranian Light 原油贴水（低于基准价的折价）扩大相矛盾。釜山休战到期日：11月10日，而非10月下旬。半导体232条款关税（美国贸易法下的国家安全关税）：已经生效，而非仍处调查阶段。机器人232条款调查：报告期限已过，决定待定。电力板块：始终无需修正（GEV 于6月30日创历史新高）——§8.1 中的拥挤度警示（太多投资者挤在同一笔交易里），正是核证过程所要求加入的。

Claims we could not verify, so we gave them less weight. The 10-year Treasury yield's 52-week range; memory's exact share of hyperscaler data-center spending; Oracle's ~86% capex (capital spending) to revenue ratio — the direction is supported, but the exact number is not verified; Unitree's 2026 unit target; and the specific on-chain cost-basis support clusters (blockchain data showing the prices where holders bought).无法核证、因而已降权（降低倚重）的论断。10年期美债收益率的52周区间；存储芯片在超大规模云厂商数据中心支出中的确切占比；Oracle 约86%的资本开支/收入比（方向上有支撑，精确数字未经核证）；Unitree 的2026年出货量目标；以及具体的链上成本基础支撑密集区（链上数据显示的持有者买入价位集中区）。

Standing open questions for the next revision. Whether the July 24 tariff handoff produces a measurable bump in CPI (the main inflation gauge). Whether the Section 232 robotics decision lands before or after CES. Whether the GENIUS rules finalize in time for a November go-live. Whether Q3 hyperscaler earnings show the first flattening in capex guidance. And whether the labor-market rollover (weakening jobs data) stabilizes — the one fact that decides between Scenario A and Scenario E.留待下一版解决的未决问题。7月24日的关税交接是否带来可测量的 CPI（主要通胀指标）脉冲。机器人232条款决定在 CES 之前还是之后落地。GENIUS 细则能否赶在11月上线前定稿。三季度超大规模云厂商财报是否首次给出走平的资本开支指引。以及劳动力市场的滚落（就业数据转弱）是否企稳——这是决定情景A与情景E走向的唯一事实。

Predecessor documents: The AI Master Plan (July 1, 2026) — the sovereign-refinancing frame this note builds on; The Warsh Shock (June 17, 2026); The Bessent Bridge (June 24, 2026).前序文件：AI 总体规划（2026年7月1日）——本报告所延展的主权再融资框架；Warsh 冲击（2026年6月17日）；Bessent 之桥（2026年6月24日）。