The AI Master PlanAI 总体规划

1What the plan actually is计划到底是什么

Lever 1 — Maximize physical throughput杠杆 1 — 最大化物理吞吐量

Executive orders accelerate federal permitting for large-scale AI data centers — streamlining environmental reviews, expanding FAST-41 coverage, and promoting federal and contaminated lands for project siting — plus a Commerce financial-support initiative (loans, grants, tax incentives) for data centers needing 100+ MW of new generation where sponsors commit at least $500M in capex. The February 24, 2026 State of the Union added a hyperscaler power pledge: signatories build, bring, or buy new generation sufficient to meet their data-center demand and cover the full cost of delivery infrastructure, negotiating separate rate structures with utilities — that last part exists because electricity affordability was becoming a midterm liability.行政令通过简化环境审查、扩大 FAST-41 覆盖范围、推动联邦土地和受污染土地用于项目建设，加快大型 AI 数据中心的联邦许可审批；同时，商务部推出金融支持计划（贷款、补助、税收激励），针对需要新增 100MW 以上发电能力、且发起方承诺至少 5 亿美元资本开支的数据中心项目。2026 年 2 月 24 日的国情咨文进一步加入了超大云厂商的电力承诺：签署方必须建设、带入或购买足够满足自身数据中心需求的新发电能力，并承担交付基础设施的全部成本，与公用事业公司谈判单独的费率结构——最后这一点之所以存在，是因为电价可负担性已经开始成为中期选举负担。

Lever 2 — Force the capex onshore with a tariff stick杠杆 2 — 用关税大棒迫使资本开支回流本土

The January 14, 2026 Section 232 proclamation is the sharpest instrument and widely misread. It imposes 25% ad valorem duties on a narrow category of advanced chips defined by TPP and DRAM-bandwidth thresholds (H200/MI325X class) — but with broad exemptions for U.S. data centers, startups, R&D, and public sector, meaning it deliberately does not tax the domestic buildout. Phase 2, after trade negotiations conclude, holds "broader tariffs at a rate of duty that is significant" plus a tariff-offset program rewarding companies investing in U.S. semiconductor production — a loaded gun pointed at Taiwan, Korea, and Japan to force fab investment stateside.2026 年 1 月 14 日的 Section 232 公告是最锋利、也最容易被误读的工具。它对一小类先进芯片征收 25% 从价税，这类芯片以 TPP 与 DRAM 带宽门槛定义（H200/MI325X 级别）；但同时对美国数据中心、初创企业、研发与公共部门提供广泛豁免。这意味着它有意不对美国国内建设征税。第二阶段在贸易谈判结束后，将保留"税率显著的更广泛关税"，并设立关税抵扣计划，奖励投资美国半导体生产的公司——这是一把指向台湾、韩国和日本的上膛枪，目的在于迫使晶圆厂投资迁往美国。

Lever 3 — Turn the state into a toll collector on the AI trade杠杆 3 — 把国家变成 AI 贸易的收费站

H200 sales to China were approved with the U.S. government taking 25% of sales; China cleared ByteDance, Alibaba and Tencent for 400,000+ units, while Blackwell remains restricted. The export-license structure even requires chips to route through the U.S. for third-party testing, triggering the 25% Section 232 tariff and aligning it with the 25% China sales fee. Add tariff revenue, and the government has built itself a royalty stream on the boom.美国批准向中国销售 H200，同时美国政府抽取 25% 销售额；中国方面已批准字节跳动、阿里巴巴和腾讯购买 40 万多颗 H200，而 Blackwell 仍受限制。出口许可结构甚至要求芯片经由美国进行第三方测试，从而触发 25% Section 232 关税，并与 25% 的中国销售分成形成一致。再加上关税收入，政府已经给自己搭建了一个从繁荣中收取版税的收入流。

Lever 4 — State-capitalism scaffolding杠杆 4 — 国家资本主义脚手架

Trump signed an executive order calling for a federal sovereign wealth fund, and the administration has already taken stakes in Intel, IBM and other quantum and critical-mineral companies. Intel is the template: a 9.9% stake worth roughly $8.9B, largely by converting CHIPS Act grants into equity. As of June 2026 this is extending to the labs themselves: Trump said the federal government may take direct equity stakes in leading AI companies, naming OpenAI, Anthropic, and xAI — under a framework where OpenAI would donate equity rather than sell it, seeding a "Public Wealth Fund," with proceeds distributed to Americans Alaska-style.特朗普签署行政令，要求建立联邦主权财富基金；政府也已经持有英特尔、IBM 以及其他量子和关键矿产公司的股权。英特尔是模板：美国政府持有约 9.9% 股份，价值约 89 亿美元，很大程度上是把《芯片法案》补助转换成股权。到 2026 年 6 月，这种做法正在延伸到 AI 实验室本身：特朗普表示联邦政府可能直接持有领先 AI 公司的股权，点名 OpenAI、Anthropic 和 xAI——在这一框架下，OpenAI 将捐赠股权而不是出售股权，用于种子化一个"公共财富基金"，并以阿拉斯加式分红把收益分配给美国人。

On the demand side: the Genesis Mission EO (Nov 2025); in May 2026 the Department of War announced agreements with eight leading AI companies to deploy their capabilities on classified networks; and a June 2026 national-security memorandum directed rapid onboarding of advanced models and buildout of high-security computing facilities.需求侧则有：2025 年 11 月的 Genesis Mission 行政令；2026 年 5 月，"战争部"（原文 Department of War）宣布与八家领先 AI 公司达成协议，在机密网络上部署其能力；2026 年 6 月，一份国家安全备忘录又要求快速接入先进模型，并建设高安全级别计算设施。

Lever 5 — Narrative and political management杠杆 5 — 叙事与政治管理

Sacks (who stepped down as czar in March 2026): "In Q1, AI was already 75% of GDP growth… stopping progress in AI would be equivalent to halting the U.S. economy," while acknowledging "polls may show that AI is not popular, but economic growth is." The Public Wealth Fund is the answer to that unpopularity — pre-distribution to buy consent, against a backdrop of roughly 49,000 jobs already lost to AI in 2026 through April, on top of ~55,000 in 2025.Sacks（2026 年 3 月卸任 AI czar）表示："一季度，AI 已经贡献了 GDP 增长的 75%……阻止 AI 进步等同于让美国经济停摆。"同时他承认，"民调可能显示 AI 不受欢迎，但经济增长受欢迎。"公共财富基金就是对这种不受欢迎的回应——用预先分配收益来购买同意，背景则是 2026 年截至 4 月，约 4.9 万个岗位已因 AI 流失，加上 2025 年的约 5.5 万个。

The five levers at a glance五个杠杆一览

The unifying logic统一逻辑

This is a physical-stock strategy, not an equity-price strategy. The objective function is gigawatts, fabs, and installed compute on U.S. soil, plus toll and equity rights over the global stack. Bessent stated the calculus outright: "the biggest risk when it comes to AI is China getting ahead of the US." Under that objective, overbuilding is strictly preferred to underbuilding, and equity prices matter only as the financing mechanism.这是一套"物理存量"战略，不是一套"股价"战略。它的目标函数是千兆瓦、晶圆厂，以及安装在美国土地上的算力，同时在全球技术栈上获取收费权与股权权利。Bessent 把计算方式说得非常直白："谈到 AI，最大的风险是中国领先美国。"在这个目标函数下，过度建设严格优于建设不足；股价只是在融资机制上重要。

2Was the bubble part of the plan?泡沫是计划的一部分吗

Not authored不是原创者

The melt-up predates the second term. There is no evidence anyone in the administration forecast hardware economics at current levels — Micron's numbers verify as nearly beyond parody. No policy document anticipated 86% memory margins (unverified as intent).这轮暴涨早在第二任期之前就已经开始。目前没有证据显示政府中有人预判了当前水平下的硬件经济性——美光的数据几乎荒诞到像讽刺。没有任何政策文件预期过 86% 的内存毛利率（作为意图尚未验证）。

Amplified — documentable through revealed preference被放大——可通过显性偏好记录下来

Every lever adds demand or removes friction:每一个杠杆都在增加需求或移除摩擦：

• Permitting reform removes supply friction;许可改革移除供给摩擦；
• The tariff exempts the domestic buildout while taxing everything else;关税豁免国内建设，同时对其他路径征税；
• Reopening China (H200) adds demand; Gulf full-stack export packages add demand; federal procurement mandates add demand;重新开放中国 H200 销售增加需求；海湾国家全栈出口包增加需求；联邦采购指令增加需求；
• Bessent's claim that financial deregulation has "unlocked some $3 trillion in lending capacity" widens the credit pipe funding the capex.Bessent 声称金融监管放松"释放了约 3 万亿美元贷款能力"，这又扩大了为资本开支融资的信贷管道。

Not a single policy dampens AI demand or leans against valuations. The government's own balance sheet now benefits from froth: 25% of China chip revenue, tariff receipts, a 9.9% Intel position, and a prospective donated stake in OpenAI ahead of an IPO targeting up to $1 trillion as early as September 2026.没有任何一项政策在压低 AI 需求或反制估值。政府自己的资产负债表也已经从泡沫中受益：25% 的中国芯片收入、关税收入、9.9% 的英特尔持股，以及 OpenAI 在最早可能于 2026 年 9 月、目标估值最高 1 万亿美元 IPO 前，潜在捐赠给政府的股权。

The precise answer更精确的答案

They were expecting — engineering — a scarcity-driven capex supercycle onshore, and accepted froth as the cost of speed. The specific hardware-pricing bubble (memory at 86% GM, RAM prices surging +90% in Q1 2026) is an emergent consequence they now have to manage, because it crossed from asset prices into consumer prices — Apple raised product prices June 26 due to chip costs, triggering another KOSPI trading suspension.他们预期并工程化了一轮由稀缺驱动、在美国本土发生的资本开支超级周期，并接受泡沫作为速度的成本。具体的硬件价格泡沫（内存 86% 毛利率、2026 年一季度 RAM 价格暴涨 +90%）是一个"涌现后果"，现在他们必须管理它，因为它已经从资产价格传导到了消费者价格——苹果在 6 月 26 日因芯片成本上升而提高产品价格，引发 KOSPI 又一次交易暂停。

3The pop — mechanism & the toolkit破裂 —— 机制与工具箱

Hyperscaler capex / FCF reality (2026)超大云厂商资本开支 / 自由现金流现实（2026）

Google, Amazon, Microsoft, and Meta plan $725 billion of 2026 capex, +77% from $410B in 2025, with Goldman projecting $5.3 trillion cumulative FY25–FY30. The free cash flow that once funded this is already gone.Google、Amazon、Microsoft 与 Meta 计划在 2026 年投入 7250 亿美元资本开支，较 2025 年的 4100 亿美元增长 +77%；高盛预计 2025–2030 财年累计资本开支为 5.3 万亿美元。曾经为此提供资金的自由现金流已经消失。

With internal cash exhausted, funding has shifted to leverage and equity issuance: Alphabet netted ~$85 billion from a secondary offering; Nvidia raised $25 billion in its first debt sale of the AI boom; SpaceX followed its $85.7B IPO with a $25B bond offering days later; and roughly half of Morgan Stanley's estimated $3 trillion in 2025–28 datacenter spend is covered by private credit. Most telling: $49 billion of Q1 2026 hyperscaler "other income" came explicitly from equity stakes in private companies — the largest share of earnings from other income in at least a decade.内部现金枯竭后，融资已经转向杠杆与股权发行：Alphabet 通过二次发行净筹约 850 亿美元；Nvidia 在 AI 繁荣中首次发债，融资 250 亿美元；SpaceX 在 857 亿美元 IPO 后数日又发行 250 亿美元债券；同时，Morgan Stanley 估计 2025–2028 年数据中心支出约 3 万亿美元，其中大约一半由私人信贷覆盖。最能说明问题的是：2026 年一季度超大云厂商“其他收入”中的 490 亿美元明确来自对私人公司的股权投资——这是至少十年来利润中来自其他收入占比最高的一次。

June 2026 was the dress rehearsal2026 年 6 月是彩排

June 5: a jobs report (172,000 vs. 80,000 expected) sent the Nasdaq to its worst week in over a year (−4.7%) — good news as bad news, the classic late-regime signature. June 23: KOSPI halted trading to prevent a crash; Samsung and SK Hynix −12% in a morning; the slump went global June 24–25, with Oracle closing its worst week since the dot-com bubble at −19%. Micron and SanDisk dropped ~13% in a single session — a leverage-driven unwind, not a demand collapse — after which Micron's earnings and guidance appeared to put doubts about the AI rally at ease.6 月 5 日：就业报告（17.2 万 vs 预期 8 万）让 Nasdaq 创下一年多以来最糟糕一周（−4.7%）——好消息变坏消息，典型的周期末期信号。6 月 23 日：KOSPI 暂停交易以防崩盘；三星与 SK 海力士一上午跌 12%；跌势在 6 月 24–25 日蔓延全球，Oracle 创下自互联网泡沫以来最差周表现，收跌 19%。美光与 SanDisk 单日跌 约 13%——这是杠杆驱动的去杠杆，而非需求崩塌——随后美光财报与指引似乎平息了市场对 AI 反弹的怀疑。

The Fed put is structurally disabled — for now“Fed put”目前在结构上失灵

Warsh was confirmed 54–45 — the closest in the modern era — and sworn in May 22, 2026. Trump appointed him to cut. Instead: since the Iran war began in late February, inflation flared to CPI 4.2% in May — the highest since April 2023. At his first meeting, nine officials signaled support for higher rates, six for two quarter-point increases — a sharp change from March. The dot plot confirmed even one 2026 cut is no longer the base case. As of June 30 at Sintra: "we've all looked around, and we've seen that prices are too high." See The Warsh Shock for the full account.Warsh 以 54–45 票获确认——现代历史上最接近的票数——并于 2026 年 5 月 22 日宣誓就职。特朗普任命他是为了降息。但自 2 月下旬伊朗战争开始后，通胀升温，5 月 CPI 达到 4.2%，为 2023 年 4 月以来最高。第一次会议上，9 名官员支持更高利率，6 名官员支持加息两次、每次 25 个基点——相比 3 月出现明显转变。点阵图确认，即便一次 2026 年降息也不再是基准情形。到 6 月 30 日辛特拉会议时，Warsh 表示：“我们都环顾四周，看见价格太高了。” 详见 The Warsh Shock。

Administration toolkit — deployment order政府工具箱——部署顺序

What they cannot fix他们修不了什么

Hyperscaler ROI. If inference economics don't cover depreciation plus capital cost, no export approval, procurement contract, or wealth fund restores a $725B private capex run-rate. The one honest sentence in the entire administration posture is Sacks's — "a reversal would risk recession" — which is an admission, not a policy. See The Bessent Bridge for how the credit-channel plumbing is actually being routed.超大云厂商的投资回报率。如果推理经济性无法覆盖折旧加资本成本，那么无论出口许可、采购合同还是财富基金，都无法恢复 7250 亿美元的私人资本开支年化速度。整个政府姿态中最诚实的一句话正是 Sacks 的话——“反转会带来衰退风险”——这是一种承认，不是一项政策。信贷渠道管道的实际布设方式，详见 The Bessent Bridge。

4The real thesis — refinancing the United States真正的论点 —— 为美国再融资

Start from the constraint nobody in the administration frames as the constraint. Gross national debt hit $38.09 trillion, up $2.18 trillion in a year, with deficits running 6–7% of GDP in non-crisis conditions. Foreign official demand for duration is structurally impaired. The genuine problem facing this administration is not "win the AI race" — it is how does the United States refinance itself at tolerable rates while running wartime deficits in peacetime.从政府没有公开表述的约束条件开始看。美国国债总额已达 38.09 万亿美元，一年增加 2.18 万亿美元；在非危机环境下，赤字仍运行在 GDP 的 6–7%。海外官方部门对长期久期的需求在结构上受损。摆在本届政府面前的真正问题不是"赢得 AI 竞赛"，而是：在和平时期运行战时赤字的情况下，美国如何以可承受利率为自己再融资。

The four legs of the refinancing machine再融资机器的四条腿

The growth leg. AI capex is not a sector story; it is the business cycle itself. Sacks stated it without euphemism: AI was 75% of GDP growth in Q1 2026, and business investment has displaced the consumer as the top GDP contributor. Nominal GDP running hot (5.9% spending growth) is what keeps tax receipts up and debt/GDP arithmetic alive. The bubble is the denominator strategy.增长腿。 AI 资本开支不是一个行业故事；它就是商业周期本身。Sacks 毫不委婉地说过：AI 在 2026 年一季度贡献了 GDP 增长的 75%，企业投资已取代消费者成为 GDP 最大贡献项。名义 GDP 保持高温（支出增长 5.9%），正是维持税收收入和债务／GDP 算术可持续的东西。泡沫就是分母战略。

The funding leg. Bessent became the first Treasury Secretary to publicly position stablecoins as a future pillar of U.S. debt financing, projecting the sector grows tenfold to $3 trillion by 2030 and stating GENIUS-regulated stablecoins will "lead to a surge in demand for US Treasuries." GENIUS reserve rules hard-wire this: reserves must be cash, deposits, repo, or Treasuries with ≤93 days remaining maturity — every dollar of stablecoin growth is a mandatory bill purchase. The BIS has quantified the mechanism: a $3.5B stablecoin inflow lowers 3-month bill yields by up to 4bp within 10 days, with effects that increase under Treasury market stress. The Fed ended QT in December 2025 and is again growing its balance sheet through reserve-management purchases of T-bills. Layer on Bessent's claim that financial deregulation has "unlocked some $3 trillion in lending capacity."融资腿。 Bessent 成为第一位公开把稳定币定位为未来美国债务融资支柱的财政部长。他预计该行业到 2030 年增长十倍至 3 万亿美元，并表示 GENIUS 监管下的稳定币将"带来美国国债需求激增"。GENIUS 储备规则把这个机制硬编码：储备必须是现金、存款、回购或剩余期限不超过 93 天的美国国债——每一美元稳定币增长都是强制性国库券购买。BIS 已经量化这一机制：35 亿美元稳定币流入会在 10 天内使 3 个月国库券收益率最多下降 4 个基点，而且在国债市场压力下影响还会增强。美联储在 2025 年 12 月结束 QT，并通过储备管理性质的国库券购买重新扩张资产负债表。再叠加 Bessent 声称金融监管放松"释放了约 3 万亿美元贷款能力"。

The dollar leg. This is the Dollar Milkshake implemented as deliberate statecraft rather than emergent flow. Gulf sovereigns recycle petrodollars into U.S. AI equity — MGX closed a $49 billion AI fund backing OpenAI and Anthropic. Allies buy "full-stack AI export packages." China pays a 25% Treasury toll on every H200. Global savings are being channeled into dollar-denominated AI assets by policy design, sustaining dollar demand despite the deficit trajectory. Bessent's Greenspan framing is not analysis; it is term-premium management — he needs the market to believe the 1990s analogy so the long end stays anchored while the rest runs.美元腿。 这就是"美元奶昔"从自然流动变成有意的国家战略。海湾主权资本把石油美元回流到美国 AI 股权——MGX 完成了一个 490 亿美元 AI 基金，支持 OpenAI 与 Anthropic。盟友购买"全栈 AI 出口包"。中国为每一颗 H200 向美国财政部支付 25% 通行费。全球储蓄正在被政策设计导入美元计价 AI 资产，从而在赤字轨迹恶化的情况下仍维持美元需求。Bessent 的格林斯潘框架不是分析，而是期限溢价管理——他需要市场相信 1990 年代类比，才能让长端保持锚定，同时其他部分继续运行。

The participation leg. The state has made itself long its own bubble: 25% of Nvidia's China revenue, Section 232 tariff receipts, the 9.9% Intel stake, and negotiated donated equity from OpenAI into a Public Wealth Fund ahead of a ~$1T IPO.参与腿。 国家已经做多自己的泡沫：Nvidia 中国收入的 25%、Section 232 关税收入、9.9% 的英特尔持股，以及 OpenAI 在约 1 万亿美元 IPO 前向公共财富基金谈判捐赠的股权。

5Warsh decoded — the put moved沃什解码 —— put 换了位置

Warsh campaigned as a dove — arguing the Fed was "backward-looking" and too slow to cut, with PIMCO expecting 2–3 cuts in 2026 toward 2.75–3%. The Iran war, 4.2% CPI, and memory-driven goods inflation forced the hawkish posture at his first meeting. But his structural agenda never changed.Warsh 竞选／游说时是一个鸽派——他认为美联储"向后看"、降息太慢；PIMCO 预期 2026 年降息 2–3 次，政策利率降至 2.75–3%。伊朗战争、4.2% CPI 以及内存驱动的商品通胀，迫使他在第一次会议上采取鹰派姿态。但他的结构性议程从未改变。

The structural agenda that never moved从未改变的结构性议程

• A new Treasury–Fed Accord under which "the Fed chair and the Treasury secretary can describe to markets plainly… 'This is our objective for the size of the Fed's balance sheet.'"推动新的财政部–美联储协议，让"美联储主席和财政部长可以清楚地向市场描述……'这就是我们对美联储资产负债表规模的目标'"。
• A shift toward shorter duration, potentially involving Fannie and Freddie.转向更短久期，可能涉及房利美和房地美。
• The abolition of forward guidance.废除前瞻指引。

Former Fed officials see what this is. Lacker warns of "a less constructive agreement that lets the Treasury use the Fed's balance sheet to bypass Congress"; one former senior official said that, followed to its logical conclusion, "the Fed could lose control of its balance sheet." Lacker's proposed division — Fed does monetary policy, Treasury does credit policy — is the actual operating doctrine being installed.前美联储官员看得很清楚。Lacker 警告，可能出现"一种不那么建设性的协议，使财政部能够利用美联储资产负债表绕过国会"；另一位前高级官员表示，如果逻辑走到尽头，"美联储可能失去对其资产负债表的控制"。Lacker 提出的分工——美联储做货币政策，财政部做信贷政策——正是正在安装的实际操作原则。

Read as realpolitik用现实政治来读

Warsh's front-end hawkishness is the credibility purchase that protects the long end while Treasury runs the absorption complex. Trump's serenity — "We have a very good guy over there now… I'm guided by what he wants" — reflects Bessent's tutoring ("the bond market has taken out more governments than howitzers") and a shared expectation the hawkish posture is seasonal: Bessent is already scripting the pivot ("gas prices will come back down, inflation will come back to target"), and Warsh has pre-built his exit ramp by framing AI as ultimately deflationary. (inference)Warsh 的短端鹰派姿态，是为保护长端而购买的信誉，以便财政部运行短端吸收复合体。特朗普的平静——"我们现在那边有一个很好的人……我听他的"——反映出 Bessent 的辅导（"债券市场推翻的政府比榴弹炮更多"）以及一个共同预期：鹰派姿态是季节性的。Bessent 已经在写转向脚本（"汽油价格会回来，通胀会回到目标"），而 Warsh 也通过把 AI 描述为长期通缩力量，预先修好了退路。（推论）

Operational consequence: who responds first操作后果：谁先响应

In a crack, first responders are not rate cuts. Crisis management is migrating from the technocratic Fed to the political Treasury, converting "who gets rescued" from a rules question into a loyalty question. The order of first responders:当市场出现裂缝时，第一响应者不是降息。危机管理正在从技术官僚式的美联储转移到政治性的财政部；"谁被救"从规则问题变成忠诚问题。第一响应的顺序：

This doctrine is the connective tissue of the regime. For the shock that installed the hawkish posture, see The Warsh Shock; for the fiscal architecture Treasury is building underneath it, see The Bessent Bridge.这一教义是整个体制的连接组织。关于安装鹰派姿态的那次冲击，见 The Warsh Shock；关于财政部在其下方构建的财政架构，见 The Bessent Bridge。

6The China track — two lists, one negotiation中国路线 —— 两张名单，一场谈判

Track one (loud): the Pentagon's 1260H list expanded June 8 by 65 entities — capturing Alibaba, Baidu, BYD, CATL and NIO — with procurement bans effective June 30, 2026 and product bans June 30, 2027. Hawk theater with real teeth. BABA is now a designated Chinese Military Company, with the entity-level procurement ban having taken effect as of this report.第一轨（高调）：五角大楼 1260H 名单在 6 月 8 日扩容 65 家实体——覆盖阿里巴巴、百度、比亚迪、宁德时代、蔚来——采购禁令于 2026 年 6 月 30 日生效，产品禁令于 2027 年 6 月 30 日生效。这是鹰派剧场，但有真实杀伤力。BABA 现在已被指定为中国军事公司，实体级采购禁令已在本报告发布时生效。

Track two (quiet): the Commerce Entity List — the instrument that actually severs supply chains — has not been updated since October 2025, the longest gap in more than a decade, while CXMT, DeepSeek and 100+ other Chinese firms have cleared interagency review but sit unpublished, attributed to "extensive negotiations with China." A former Commerce official states trade policy is overriding the national-security tool. The February trial balloon — CXMT/YMTC posted as removed from 1260H, withdrawn within an hour after hawk backlash — and their restoration in June, weeks after a Trump–Xi Beijing summit, shows the live tug-of-war between the dealmaker faction (Bessent/Lutnick) and congressional hawks, who are simultaneously pushing a two-year statutory Blackwell ban.第二轨（低调）：商务部实体清单——真正会切断供应链的工具——自 2025 年 10 月以来一直没有更新，这是十多年来最长的间隔；与此同时，CXMT、DeepSeek 以及另外 100 多家中国企业已通过跨部门审查，却仍未公布，理由是正在与中国进行"广泛谈判"。一位前商务部官员表示，贸易政策正在压过国家安全工具。2 月的试探气球——CXMT / YMTC 一度被公布为从 1260H 名单移除，但在鹰派反弹后不到一小时撤回——以及它们在 6 月恢复上榜（特朗普–习近平北京峰会后数周），展示了交易派（Bessent / Lutnick）与国会鹰派之间的现场拉扯，后者同时推动两年法定 Blackwell 禁令。

The bargain这笔交易

America sells previous-generation compute — the H200, two generations behind Blackwell/Rubin — for cash, dependency, and a 25% Treasury toll; withholds the frontier; freezes the Entity List as the standing hostage; and keeps CXMT's Entity List addition in its pocket. China accepts the toll — 400K+ H200 units cleared for ByteDance, Alibaba and Tencent — while sprinting on domestic substitution.美国出售上一代算力——H200，落后 Blackwell / Rubin 两代——换取现金、依赖性和 25% 财政部通行费；保留前沿产品；冻结实体清单作为常设人质；把 CXMT 加入实体清单的选项留在口袋里。中国接受通行费——字节跳动、阿里、腾讯获准购买 40 万多颗 H200——同时全力推进国产替代。

Given that today's Commerce report explicitly empowers Trump to modify the semiconductor tariff, and given four Pennsylvania toss-up House seats are contested substantially on electricity/affordability, there are meaningful odds — ~35–40% (estimate) — of consumer-memory relief action before the midterms: CXMT accommodation, tariff carve-outs, or public jawboning.考虑到今天的商务部报告明确赋予特朗普修改半导体关税的能力，又考虑到宾夕法尼亚州四个摇摆众议院席位很大程度上围绕电价 / 可负担性竞争，合理概率——约 35–40%（估计）——是：中期选举前会出现消费内存缓释行动——CXMT 妥协、关税豁免，或公开口头施压。

7Where the losses land — the transmission map损失落在哪里 —— 传导图

Oracle's total debt exceeds $130 billion against $248 billion in new lease commitments; its 5-year CDS has risen ~310% to a sixteen-year high; Moody's has it on negative watch; and by January 2026 JPMorgan was struggling to find investors for the Stargate debt syndication. CoreWeave's $7.5B GPU-collateralized facility at ~11% variable began repayments in January 2026 precisely as collateral value declined sharply. The datacenter ABS market stands at ~$25 billion against projected take-out needs approaching $300 billion — a funding cliff, not a gap.Oracle 总债务超过 1300 亿美元，同时有 2480 亿美元新租赁承诺；其 5 年期 CDS 已上涨 约 310%，至 16 年高点；Moody's 将其列入负面观察；到 2026 年 1 月，JPMorgan 已难以为 Stargate 债务银团找到投资者。CoreWeave 75 亿美元、以 GPU 为抵押、利率约 11% 浮动的贷款工具，恰好在抵押品价值大幅下跌时，于 2026 年 1 月开始还款。数据中心 ABS 市场规模约 250 亿美元，而预计的接盘需求接近 3000 亿美元——这不是资金缺口，而是融资悬崖。

Loss routing: the pain is architected to land in the shadows损失路径：痛苦被设计成落在阴影之中

Per the Fed's own studies: banks' direct AI-adjacent exposure is only 0.8% of assets — the banking system has been deliberately insulated — but up to a quarter of bank lending to nonbanks flows to private credit firms (versus 1% in 2013), major life insurers hold nearly $1 trillion in private credit, and state pensions (NY, PA) are invested in the Blue Owl vehicles behind Meta's SPVs and Oracle's datacenter financings. The pain is architected to land in NBFI space — private credit, insurance, pensions — exactly where Warsh's doctrine says the Fed should not go and where Lacker's division assigns responsibility to Treasury.按美联储自己的研究：银行对 AI 邻近资产的直接敞口仅占资产 0.8%——银行系统被有意隔离；但银行对非银机构贷款中，多达四分之一流向私人信贷公司（2013 年仅 1%），大型寿险公司持有近 1 万亿美元私人信贷，纽约州、宾夕法尼亚州等州养老金也投资于 Blue Owl 工具，而这些工具支持了 Meta 的 SPV 与 Oracle 的数据中心融资。痛苦被设计成落在非银金融机构——私人信贷、保险、养老金——这正是 Warsh 教义中美联储不应介入、而 Lacker 分工中交给财政部负责的地方。

The scheduled fuse is accounting, not demand预定引信是会计，不是需求

Goldman's May 2026 analysis identified GPU useful life as the single variable with the largest impact on capital requirements through 2031: shortening from five years to three pushes cumulative 2026–31 depreciation from ~$3 trillion to ~$4 trillion. Burry's arithmetic — $176 billion of understated depreciation 2026–28, Oracle and Meta operating income more than 20% above economic reality — collides with Nvidia's annual architecture cadence (Hopper '22, Blackwell '24, Rubin '26, Rubin Ultra '27). Amazon already broke ranks, shortening lives citing accelerating AI pace; Meta extended to 5.5 years — proof useful life is a management lever, not an engineering fact. The FY2026 10-K season (Jan–Feb 2027) is when auditors and the SEC comment-letter process force this into the open.高盛 2026 年 5 月的分析指出，GPU 有用寿命是影响 2031 年前资本需求的最大单一变量：从五年缩短到三年，会把 2026–2031 年累计折旧从约 3 万亿美元推至约 4 万亿美元。Burry 的算术——2026–2028 年低估折旧 1760 亿美元，Oracle 与 Meta 经营收入比经济现实高出 20% 以上——与 Nvidia 的年度架构节奏相撞（Hopper 2022、Blackwell 2024、Rubin 2026、Rubin Ultra 2027）。Amazon 已率先打破队形，以 AI 迭代加速为由缩短使用寿命；Meta 则延长至 5.5 年——这证明有用寿命是管理层杠杆，而不是工程事实。2026 财年 10-K 季（2027 年 1–2 月）将是审计师与 SEC 评论函流程迫使其公开化的时点。

8The predictive pathway预测路径

Why a blow-off into the midterms is the wrong base case为何"冲进中期选举"是错误的基准情形

Argument one — the market path is set by the Fed plus the cycle, not by the administration's incentive. Every administration wants the market up into the midterms; midterm years nonetheless produce their lows in the September–December window (October modal: 2022, 2014, 2002, 1998, 1990, 1966; 2018's came December 24, just after the election). The controlling analog is 2018 itself — a new hawkish Fed chair tightening into a memory supercycle peak, a China trade conflict, and a president publicly demanding higher stock prices — and the S&P still fell ~20% into late December. MU peaked in late May 2018, roughly two quarters before its own revenue peaked, then fell ~56%. Political will did not override the Fed plus the cycle then; the current setup is harsher, because this Fed has hikes on the table.论点之一——市场路径由美联储加周期决定，而非由政府的动机决定。每一届政府都希望股市在中期选举前走高；然而中期选举年的低点仍集中在 9 月至 12 月的窗口（10 月为众数：2022、2014、2002、1998、1990、1966；2018 年的低点出现在 12 月 24 日，恰在选举之后）。决定性的类比就是 2018 年本身——新任鹰派美联储主席在存储超级周期顶部加息、中美贸易冲突、以及一位公开要求股价走高的总统——而标普 500 仍在 12 月底之前下跌 约 20%。美光（MU）在 2018 年 5 月底见顶，约比自身营收峰值早两个季度，随后下跌 约 56%。当时政治意志并未压过美联储加周期；当前格局更为严峻，因为这届美联储手上还摆着加息选项。

Argument two — a legible support schedule gets front-run. A telegraphed schedule ("support withdrawn after November 3") invites distribution before the withdrawal, dragging the top forward. The late-June/early-July tape is consistent with distribution beginning months ahead of the administration's timetable — the one contingency the plan cannot control.论点之二——一份清晰可读的支撑时间表会被抢跑。一份被预告的时间表（"11 月 3 日之后撤走支撑"）会招来在撤走之前的派发，从而把顶部往前拖。6 月底至 7 月初的行情，与派发早于政府时间表数月的判断一致——这正是该计划无法掌控的唯一变量。

The distribution evidence (as of July 1–2, 2026)派发证据（截至 2026 年 7 月 1–2 日）

The signature is everywhere at once: record highs sold on no new information, a speculative complex already in its own bear market, leadership narrowing into perceived quality, the regime-break event arriving early, and the memory forward-supply derating switching on. Per the historical pattern, memory-stock declines of 50–60% lead the fundamental downturn by 1–2 quarters — the market has begun applying trough multiples to peak earnings, which is how every memory-equity top has formed.派发的特征同时出现在各处：在没有任何新信息的情况下卖出创纪录高点、投机板块已陷入自身的熊市、领导权向被视为优质的名字收窄、机制断裂事件提前到来，以及存储前瞻供给的估值下修被开启。按历史规律，存储股 50–60% 的下跌会领先基本面下行 1–2 个季度——市场已开始对峰值盈利套用谷底估值倍数，而每一次存储股顶部都是这样形成的。

Phase 1 — now through October 2026 (distribution & decline)阶段 1 —— 当下至 2026 年 10 月（派发与下跌）

The hardware/memory complex cycle top is in, or completes on one more policy-driven retest — ~60–65% (estimate). The broad index is in distribution: either the June high stands, or one marginal rotation-driven high prints into the September OpenAI IPO window — ~55–60% combined for a decline into a Q4 low. (The 2018 pattern: SOX/MU peaked May–June, SPX made a marginal new high in late September before breaking; 2000: SPX retested its March high in early September while the speculative complex was already down 30–40%.)硬件/存储板块的周期顶部已经形成，或在再一次由政策驱动的回测后完成——约 60–65%（估计）。大盘指数正处于派发之中：要么 6 月高点即为最终高点，要么在 9 月 OpenAI IPO 窗口前打出一个由轮动推动的边际新高——合计约 55–60% 的概率跌向第四季度低点。（2018 年的模式：费城半导体指数/美光在 5–6 月见顶，标普在 9 月底打出边际新高后破位；2000 年：标普在 9 月初回测其 3 月高点，而投机板块彼时已下跌 30–40%。）

Policy headlines — the pending Commerce tariff-modification decision, the July 24 Section 122 tariff expiry/replacement, wealth-fund announcements, China deals — now function as squeeze risk within a downtrend and exit liquidity for the administration's schedule, the 2018 trade-truce-headline pattern. The grind-higher-into-the-election path is only ~20%, requiring all four of: soft Jul/Aug CPI, Commerce tariff relief, no credit accident, and a successful $1T IPO. Immediate nonlinear cascade (KOSPI leverage + neocloud credit unwind with no rotation phase): ~15–20%. Drawdown zones if the modal path completes: SPX −12 to −20% from high (midterm-year median ~17%); memory/semis −30 to −45% from their highs.政策头条——商务部待定的关税修改决定、7 月 24 日第 122 条款关税到期/替换、主权财富基金公告、对华交易——如今的作用是下跌趋势中的逼空风险，以及政府时间表的出货流动性，即 2018 年贸易休战头条的模式。"缓步走高至选举"的路径仅有 约 20% 的概率，需要同时满足全部四项：7/8 月 CPI 温和、商务部关税放松、无信用事故、以及一场成功的 1 万亿美元 IPO。即时的非线性级联（KOSPI 杠杆 + 新云信贷去杠杆、无轮动缓冲）：约 15–20%。若众数路径兑现，回撤区间：标普自高点 −12 至 −20%（中期选举年中位数约 17%）；存储/半导体自高点 −30 至 −45%。

Phase 2 — October 2026 through January 2027 (the low, conditional on the pivot)阶段 2 —— 2026 年 10 月至 2027 年 1 月（低点，取决于转向）

Pure seasonality says the low arrives around the midterms; the 2018 analog says the low requires the Fed's capitulation (Powell's came January 4, 2019). This cycle's low therefore needs the Warsh pivot, which needs either two soft CPI prints (post-Iran gas declines make Aug/Sep plausible; ADP already printed 98K on July 1, showing labor cooling) or enough market damage to guarantee them. Low window: October 2026 – January 2027, with the October 27–28 FOMC (six days pre-election) the earliest plausible softening venue and December–January more likely if inflation stays sticky.纯季节性表明低点出现在中期选举前后；2018 年类比则表明，低点需要美联储投降（鲍威尔的投降出现在 2019 年 1 月 4 日）。因此本轮的低点需要沃什（Warsh）转向，而这需要两次温和的 CPI 数据（伊朗事件后油气回落使 8/9 月具备可能性；ADP 已在 7 月 1 日录得 9.8 万，显示劳动力降温），或足够的市场破坏来确保其发生。低点窗口：2026 年 10 月 – 2027 年 1 月，其中 10 月 27–28 日 FOMC（选举前六天）是最早可能的转鸽场合，若通胀顽固，则 12 月至 1 月的可能性更大。

The post-election political reset runs concurrently: if the GOP loses the House — a live risk given 75 projects worth $130B blocked in Q1 2026, four PA toss-ups fought on electricity prices, and ~60% of Trump voters worried about AI — expect investigations, dead preemption, and a hostile appropriations posture. Historically, the midterm low is the single best entry of the four-year cycle; this framework treats it as the primary re-entry window, gated on the credit tape (Oracle CDS, neocloud refinancing, ABS spreads) rather than the equity tape.选举后的政治重置同步展开：若共和党丢掉众议院——考虑到 2026 年一季度有 价值 1300 亿美元的 75 个项目被搁置、宾州四个胶着席位围绕电价开打、以及约 60% 的特朗普选民对 AI 感到担忧，这是一个真实存在的风险——预计将出现调查、抢先权（preemption）失效，以及敌对的拨款姿态。从历史看，中期选举低点是四年周期中最佳的单一入场点；本框架将其视为主要的再入场窗口，以信用行情为闸门（甲骨文 CDS、新云再融资、ABS 利差），而非股市行情。

Phase 3 — 2027 (post-pivot rally into the structural wall)阶段 3 —— 2027 年（转向后的反弹撞上结构性高墙）

The blow-off belongs after the midterm low: post-pivot, H1 2027, pre-election-year seasonality — the strongest year of the cycle — running directly into the fuses this document mapped: the FY2026 10-K depreciation season (Jan–Feb 2027), the ~$300B ABS take-out wall, mid-2027 arrival of new memory supply (SK Hynix M15X, Micron Idaho), and the CXMT/YMTC ramps. Rally into a wall — the post-LTCM 1999 → 2000 compression, sped up. The structural resolution fork:暴涨属于中期选举低点之后：转向后的 2027 年上半年，选举前一年的季节性——周期中最强的一年——将径直撞上本文件所勾勒的引信：2026 财年 10-K 折旧季（2027 年 1–2 月）、约 3000 亿美元的 ABS 接盘墙、2027 年年中到来的新存储供给（SK 海力士 M15X、美光爱达荷厂），以及长鑫/长江存储的放量。反弹撞墙——即 LTCM 之后 1999 → 2000 的压缩，只是加速版。结构性结局分叉：

9Portfolio translation组合转译

MU (Micron) — the equity top is now, and the canary is a neocloudMU（美光）——股价顶部就在当下，而报警的金丝雀是一家 neocloud

The bull case is intact but the clock has run out on the equity. Memory stocks lead the fundamental peak by 1–2 quarters, and the fundamental peak — the guided $50B revenue quarter at ~86% gross margin — is set for Aug–Nov 2026. Run the arithmetic backward and the equity-cycle top window lands now, on the current tape: Q2–Q3 2026.多头逻辑仍然成立，但股价的时钟已经走完。存储股领先基本面峰值 1–2 个季度，而基本面峰值——指引中的 500 亿美元收入、约 86% 毛利率的那个季度——被定在 2026 年 8–11 月。把这套算术倒推，股价周期的顶部窗口就落在当下、落在眼前的行情上：2026 年第二至第三季度。

Worse, the forward-multiple derating is already activated. With CXMT at ~17% of global DRAM capacity by end-2026, Google qualifying CXMT, YMTC's 2H26 ramp, and glut warnings pointing to late 2027, the market is applying trough multiples to peak earnings. The practical consequence is brutal and counterintuitive: earnings beats from here are distribution events, not trend resumption — every spectacular print is a chance for someone to sell to you.更糟的是，远期估值的下调已经启动。随着 CXMT 到 2026 年底占全球 DRAM 产能 约 17%、Google 正在对 CXMT 进行验证、YMTC 于 2026 下半年爬产、以及供给过剩警告指向 2027 年底，市场正在对峰值盈利套用谷底估值。其现实后果既残酷又反直觉：从这里开始的盈利超预期都是派发事件，而非趋势重启——每一份亮眼的财报，都是别人向你派货的机会。

But the drawdown shape differs from prior cycles, and this is the crux. The ~$100B RPO carries floors above all prior cycle peaks, and new supply is physically unable to arrive before mid-2027. Those two facts rule out a 2018-style earnings collapse and argue instead for a multiple-compression grind punctuated by credit shocks. The trade has changed species: the stock now crashes with credit events, not spot DRAM prices.但这一轮回撤的形态不同于以往周期，而这正是关键。约 1000 亿美元的 RPO 带有高于以往所有周期峰值的底价，而新增供给在 2027 年年中之前根本无法到位。这两个事实排除了 2018 年式的盈利崩塌，转而指向一种由信用冲击穿插的估值压缩式磨底。这笔交易的物种已经改变：股价如今随信用事件而崩，而非随 DRAM 现货价格。

The three tripwires are unchanged; only their framing has sharpened now that this is a credit trade:三个触发器保持不变；既然这已成为一笔信用交易，只是它们的定性更加锐利：

Positioning. Remaining upside is policy-squeeze rallies, not fundamental re-rating — de-risk into strength. The primary re-entry is the Phase 2 midterm-low window, and the trigger is the credit tape — Oracle CDS, neocloud refis, ABS spreads — not the DRAM contract prints, which will still look spectacular at the low. Buying the beat is buying the distribution; buying the credit thaw is buying the bottom.仓位安排。剩余的上行空间是政策逼空式反弹，而非基本面重新估值——逢强减仓。主要的再入场点是第 2 阶段的中期低点窗口，触发条件是信用行情——Oracle CDS、neocloud 再融资、ABS 利差——而不是 DRAM 合约报价，后者在底部依然会相当亮眼。买超预期就是买派发；买信用解冻才是买底部。

COIN / BTC / stablecoins — a Treasury-policy volume tailwindCOIN／BTC／稳定币——一股由财政部政策驱动的交易量顺风

The refinancing-regime thesis is structurally significant: GENIUS makes Circle and Tether quasi-fiscal institutions whose growth is now Treasury policy (Bessent's $3T target). That is a secular volume tailwind independent of crypto beta.再融资制度假说具有结构性意义：GENIUS 使 Circle 与 Tether 成为准财政机构，其增长如今就是财政部政策（Bessent 的 3 万亿美元目标）。这是一股独立于加密 beta 的长期交易量顺风。

Sequencing matters. In Phase 2/3 risk-off, BTC trades as high-beta duration first. The sovereign liquidity response — bill purchases, buybacks, facilities, eventual Warsh cuts — is the Phase 3 bid the liquidity-rotation model should be armed for.顺序很重要。在第二／第三阶段 risk-off 中，BTC 会先作为高 beta 久期资产交易。主权流动性回应——国库券购买、回购、工具、最终 Warsh 降息——才是流动性轮动模型应当为之准备的第三阶段买盘。

DXY / gold — dollar-up-gold-up is the regime signatureDXY／黄金——美元涨、黄金涨即是制度特征

The Dollar Milkshake is operative and intentional during the boom phase — hawkish Warsh, global capital suction into dollar AI assets, and the dollar's best day in a year on his first press conference.在繁荣阶段，"美元奶昔"正在运行，而且是有意为之——鹰派的 Warsh、全球资本被吸入美元 AI 资产，以及他首次新闻发布会当天美元创下一年最佳单日表现。

Gold is the anti-bloc instrument. It bids on exactly two things this regime produces: (i) the fiscal-dominance odor from the Accord (the bear-steepening / independence-fear channel), and (ii) accumulation by states outside the AI-dollar loop. Dollar-up-gold-up is not a contradiction here — it is the regime signature.黄金是反阵营工具。它买的正是这个制度会制造的两样东西：（一）来自 Accord 的财政主导气味（熊陡／独立性担忧渠道）；（二）处在 AI–美元循环之外的国家的积累。在这里，美元涨、黄金涨并不矛盾——这正是制度特征。

BABA — urgent, separate workstreamBABA——紧急，需单独开研究流

Flag as urgent: the 1260H designation landed June 8, and the procurement ban is live as of June 30, 2026. This is a structural overhang distinct from the AI-boom thesis and warrants a separate workstream.标记为紧急：1260H 指定于 6 月 8 日落地，采购禁令已于 2026 年 6 月 30 日生效。这是一个独立于 AI 繁荣逻辑的结构性压制，值得单独开一个研究流。

The one-sentence compression一句话压缩

Sources & method资料来源与方法