返回列表
💰投资 · 💬 讨论题

AI 牛市的真正熊市版本——“智能过剩”如何两年内把经济拧成金融危机

如果 AI 真的如多头所愿般快速起飞，最危险的结果可能不是“算力泡沫破裂”，而是“白领收入假设被结构性击穿”——产出还在、钱却不再流经家庭，从而让消费、信贷与财政同时失血。

2026-02-23
阅读简报
双语对照
完整翻译
原文
讨论归档

核心观点

  • “幽灵 GDP”是最刺眼的信号：产出增长，但不再转化为消费循环 AI 让企业利润率扩张、生产率爆表，但劳动收入塌陷；宏观数据看似好，真实经济却因“机器不消费”而枯萎。
  • 这是一个没有天然刹车的负反馈回路 AI 能力提升 → 裁员 → 消费下滑 → 利润压力 → 更多 AI 投入 → 更强能力 → 更多裁员。关键在于：AI 投入常常来自 OpEx 替代（省人工、买算力），所以总需求下滑并不自动压住 AI 投资。
  • 中介层会先被机器“抹平摩擦”击穿 订阅续费、保险续保惰性、旅行平台、地产经纪、外卖平台、支付交换费……大量企业价值建立在“人类会嫌麻烦/会被流程驯化”上；当智能体替人做决策与比价，摩擦归零，中介抽租模型塌方。
  • 系统性风险的引爆点来自“相关押注的雏菊链”，尤其是私信贷×保险 私信贷软件 LBO 违约本可慢慢消化，但当其通过保险公司/再保险/离岸结构变成“主街储蓄的底层资产”，监管资本金变化会迫使抛售，脆弱性突然金融化。
  • 按揭是终局问题：优质借款人的“未来收入稳定性”假设不再可信 这不是次贷/利率冲击/区域冲击，而是收入路径整体下移；一旦 780+ FICO 的人群也开始出现持续性逾期，金融系统最基础的风险模型会被迫重写。

跟我们的关联

💰投资：把组合假设按“是否依赖白领收入增长/人类摩擦/信用扩张”重审一遍：

  • 高风险暴露：长尾 SaaS、咨询/外包、支付交换费、依赖续费惰性的订阅、私信贷软件 LBO、与高收入白领高度相关的住宅与消费信贷。
  • 结构性受益：AI 基础设施链条（算力/电力/数据中心）、以及能在“智能体主导交易”里作为新管道/新清算层的基础设施。

🧠Neta：如果 AI 进入“智能体替人决策”的阶段，增长逻辑会从“抢用户注意力”转向“抢智能体的默认调用/偏好模型”。产品需要思考：当用户不再亲自做选择，你的护城河是什么？

👤ATou：这篇文章是一个提醒：不要只看“AI 能做什么”，要看“当 AI 让人类智能不再稀缺时，你现在依赖的制度/市场/商业模型会怎样”。你做的系统（OpenClaw/效率/品牌）其实是在帮自己提前适配这个世界。

讨论引子

  • 你觉得 AI 的最大宏观风险是“就业”，还是“收入假设崩塌导致的信用系统重定价”？两者哪一个更快传导？
  • 如果“中介抽租”被智能体抹平，未来最值钱的会是什么：品牌、数据、网络效应，还是清算/管道？
  • 对产品团队来说，“用户自己点”的互联网正在结束吗？如果是，你会怎么重新设计增长与留存？

[

](https://substackcdn.com/image/fetch/$s_!rCTy!,f_auto,q_auto:good,fl_progressive:steep/https%3A%2F%2Fsubstack-post-media.s3.amazonaws.com%2Fpublic%2Fimages%2F9ede0b37-f3be-4787-90ff-10cb24719fdc_710x1119.jpeg)

宏观备忘录

充裕智能的后果

***CitriniResearch ***

2026 年 2 月 22 日 2028 年 6 月 30 日

今天早上失业率公布为 10.2%，比预期高出 0.3 个百分点。市场在这一数字公布后下跌 2%，使标普从 2026 年 10 月高点累计回撤达到 38%。

交易员已经麻木。六个月前，这样的读数足以触发熔断。

两年。 就这么一点时间，我们就从“可控”“局限于某些行业”走到了一个已不再像我们任何人成长时所熟悉的经济体。本季度的宏观备忘录，是我们试图复盘这一序列——对危机前经济的一次事后尸检。

狂欢近乎触手可及。到 2026 年 10 月，标普 500 一度逼近 8000 点，纳指突破 3 万点。由于“人类过时”引发的第一波裁员在 2026 年初开始，而且它们确实完成了裁员通常会完成的事情：利润率扩大、业绩超预期、股价上涨。创纪录的企业利润又被直接回流到 AI 算力投入中。

表面数据依旧亮眼。名义 GDP 多次录得中高个位数的年化增速。生产率在飙升。每小时实际产出以自 20 世纪 50 年代以来未见的速度上升，背后是那些不睡觉、不请病假、也不需要医疗保险的 AI 智能体。

随着劳动成本消失，掌握算力的人财富暴涨。与此同时，实际工资增速崩塌。尽管政府一再夸耀生产率创纪录，白领仍被机器夺走工作，被迫转去更低薪的岗位。

当消费端经济开始出现裂缝时，经济评论员流行起一个说法——“幽灵 GDP”：出现在国民账户里的产出，却从未在真实经济中流转。

在几乎所有维度上，AI 都超出了预期，而市场就是 AI。 唯一的问题是……经济并不是。

其实一开始就该明白：北达科他州的一组 GPU 集群产出的成果，相当于过去曼哈顿中城 10,000 名白领的产出——这更像是一场经济瘟疫，而非经济灵药。货币流通速度近乎停滞。以人类为中心的消费经济——当时占 GDP 的 70%——开始枯萎。如果我们早点问一句：机器在可选消费品上会花多少钱？也许早就能想明白。（提示：零。）

AI 能力提升，企业需要的员工更少，白领裁员增加，被替代的工人消费减少，利润率压力迫使企业在 AI 上投入更多，AI 能力再提升……

这是一条没有天然刹车的负反馈回路：人类 智能替代螺旋。白领的赚钱能力（以及，理性地说，他们的支出能力）出现了结构性受损。他们的收入是 $13 trillion 按揭市场的基石——这迫使承销方重新评估：优质按揭贷款是否仍然“稳如现金”。

连续 17 年没有真正的违约周期，使私人市场里塞满了 PE 支持的软件交易——它们都假设 ARR 会永远“持续经常性”。2027 年年中，由 AI 冲击引发的第一波违约，直接挑战了这一假设。

如果冲击只局限在软件行业，这本还可以管理；但它并没有。到 2027 年底，它威胁到一切以“中介/撮合”为基础的商业模式。大量靠“向人类的摩擦收费”而生的公司土崩瓦解。

事实证明，整个系统是一串长长的“同向押注”雏菊链：大家都在相关性极高地押注白领生产率增长。2027 年 11 月的崩盘，只是加速了原本就已存在的所有负反馈回路。

我们已经等了将近一年，盼着“坏消息就是好消息”的那一刻出现。政府开始讨论一些方案，但公众对政府能否组织任何形式救援的信心正在消退。政策反应一向落后于经济现实，而如今缺乏一套完整方案，正威胁着加速通缩螺旋。

它如何开始

在 2025 年末，智能体式编程工具的能力出现了阶跃式跃升。

一名合格的开发者配合 Claude Code 或 Codex，如今可以在数周内复刻一款中端市场 SaaS 产品的核心功能。不一定完美，也未必覆盖所有边缘情况，但已经好到足以让在审阅一份每年 $500k 的续费合同时，CIO 开始问： “要不我们自己做一个？”

企业财年大多与自然年一致，因此 2026 年的企业支出预算在 2025 年第四季度就已敲定——那时“代理式 AI”还只是个流行词。年中复盘是采购团队第一次在“看得见这些系统真实能做什么”的情况下作出决策。有些人眼睁睁看着自家内部团队在数周内就做出原型，足以替代那些六位数金额的 SaaS 合同。

那年夏天，我们与一位《财富》500 强的采购经理聊过。他向我们讲了他一次预算谈判的经历。销售原本打算照搬去年的剧本：每年 5% 的涨价，以及那套标准的“你们团队离不开我们”的说辞。采购经理告诉他：他正在与 OpenAI 沟通，让他们的“forward deployed engineers”用 AI 工具把这个供应商彻底替掉。最终续约价格打了 7 折（降价 30%）。他说，这已经算是好结果了。而 SaaS 的“长尾”，比如 Monday.com、Zapier 和 Asana，日子就更难过了。

投资者早有准备——甚至可以说早有期待——长尾会被重击。它们或许占典型企业软件栈支出的三分之一，但暴露度显而易见。然而，所谓“记录系统（systems of record）”本应是最不易被颠覆的。

直到 ServiceNow 2026 年第三季度的财报，反身性机制才变得更清晰。

SERVICENOW 新增 ACV 净增速从 23% 放缓至 14%；宣布裁员 15% 并推出“结构性效率计划”；股价下跌 18% | Bloomberg，2026 年 10 月

SaaS 并没有“死”。把系统自建并自行运维，仍然需要做成本收益分析。但自建 已经 是一个选项，而这会进入定价谈判的筹码。更重要的是，竞争格局变了。AI 让开发并上线新功能更容易，差异化因此崩塌。既有厂商在价格上陷入“向下竞赛”——既要与彼此厮杀，也要与一批冒出来的新锐挑战者短兵相接。新挑战者因智能体编程能力的跃升而胆子更大，又没有旧成本结构需要保护，于是凶猛地抢占份额。

也直到这次财报披露，大家才真正意识到这些系统互相连结的程度。ServiceNow 卖的是席位。财富 500 强客户裁掉 15% 员工时，他们也会取消 15% 的许可。同一股由 AI 驱动的人力缩减，在客户那边提升了利润率，却在机制上直接摧毁了 ServiceNow 自己的收入基础。

一家卖工作流自动化的公司，被更好的工作流自动化所颠覆；而它的应对方式，是裁员并把省下的钱投入到正颠覆它的那项技术上。

他们还能怎么办？原地不动，慢慢等死？ 最受 AI 威胁的公司，反而成了最激进拥抱 AI 的公司。**

事后看这很显然，但在当时（至少对我而言）真的不是。传统的颠覆模型说：既有巨头会抵制新技术，随后被灵活的新进入者抢走份额，慢慢死去。柯达、百视达、黑莓都是这样。可 2026 年发生的不同：既有巨头不是不想抵制，而是抵制不起。

当股价跌了 40%–60%，董事会又逼着要答案，那些被 AI 威胁的公司只能做唯一能做的事：裁员，把省下来的钱投向 AI 工具，再用这些工具以更低成本维持产出。

单看每家公司的应对都理性；合在一起却是灾难。每省下的一美元人力成本，都流向了让下一轮裁员成为可能的 AI 能力。

软件只是开场。投资者在争论 SaaS 倍数是否已经见底时忽略了一点：反身性回路早已逃离软件行业。同样的逻辑，适用于每一家以白领成本结构为主的公司。

当摩擦归零

到 2027 年初，使用 LLM 已成为默认选项。人们在使用 AI 智能体，却甚至不知道“AI 智能体”是什么——就像很多从未搞懂“云计算”的人也照样在用流媒体服务。在他们心里，这就像自动补全或拼写检查一样——只是手机现在“自带”的功能。

Qwen 开源的智能体购物助手成为 AI 介入消费者决策的催化剂。几周之内，所有主流 AI 助手都整合了某种智能体电商功能。蒸馏模型让这些智能体可以在手机和笔记本上运行，不再只依赖云端实例，从而显著降低了推理的边际成本。

真正该让投资者更不安的，是这些智能体并不等你开口。它们会按照用户偏好在后台运行。商业不再是一连串离散的人类决策，而变成了一个持续优化的过程，24/7 为每一个联网的消费者代劳。到 2027 年 3 月，美国人的日均 token 消耗中位数已达到 400,000——比 2026 年底高出 10 倍。

链条的下一环已经开始断裂。

中介。

过去五十年里，美国经济在人的局限之上叠起了一层巨大的“抽租”结构：做事要花时间，耐心会耗尽，品牌熟悉度替代了勤勉比较，而大多数人为了少点几下，愿意接受一个更差的价格。数万亿美元的企业价值，都建立在这些约束会持续存在的前提上。

一开始很简单：智能体把摩擦抹平了。

那种即便几个月不用也会自动续费的订阅和会员；试用期过后悄悄翻倍的“首发价”。这些都被重新定义成一种“被挟持”的局面——而智能体可以替你谈判。平均客户终身价值（LTV）——订阅经济赖以建立的核心指标——明显下滑。

消费者智能体开始改写几乎所有消费交易的运作方式。

人在买一盒蛋白棒之前，没时间在五个平台逐个比价；机器有。

旅游预订平台最先中枪，因为它们最简单。到 2026 年第四季度，我们的智能体就能比任何平台更快、更便宜地组装完整行程（机票、酒店、地面交通、会员积分优化、预算约束、退款）。

保险续保——其续保模式完全依赖投保人的惰性——也被改写。每年替你重新比价投保的智能体，拆掉了保险公司从“被动续保”中赚取的那 15%–20% 保费。

理财建议。报税。常规法律工作。任何一个服务商最终卖点是“我来替你穿越那些你觉得烦的复杂性”的行业，都被颠覆了，因为智能体不觉得烦。

连那些我们以为“人际关系价值”能形成护城河的领域，也脆得出人意料。房地产就是例子：由于经纪人与消费者之间的信息不对称，买家几十年来一直忍受 5%–6% 的佣金；但当拥有 MLS 访问权限和几十年成交数据的 AI 智能体能够瞬间复刻知识体系后，这套体系就崩了。一篇 2027 年 3 月的卖方研报把它称为“agent on agent violence”。在主要大都市区，买方经纪佣金中位数从 2.5%–3% 压缩到 1% 以下，越来越多交易甚至完全没有人类买方经纪参与就成交了。

我们高估了“人际关系”的价值。结果发现，人们口中的很多“关系”，不过是披着友好面孔的摩擦。

这只是对中介层冲击的开始。那些成功公司花了数十亿美元，去高效利用消费者行为与人类心理的各种“怪癖”——而这些怪癖突然不再重要。

为价格与匹配度做优化的机器，并不在乎你最爱用的 app，或你过去四年习惯性打开的网站，也不会被精心设计的结账流程所“吸引”。它们不会累到选择最省事的那一个，也不会默认“我一直都在这家下单”。

这毁掉了一类特殊的护城河：习惯性中介。

DoorDash（DASH US）是典型代表。

编程智能体把上线一个外卖 app 的门槛彻底压平。一个合格的开发者可以在数周内部署一个可用的竞品，确实也有几十家这么干：它们把配送费的 90%–95% 直接让给司机，从而把司机从 DoorDash 和 Uber Eats 那里挖走。多平台仪表盘让零工同时跟踪二三十个平台的派单，消除了巨头赖以生存的锁定效应。市场一夜之间碎片化，利润率压缩到几乎为零。

智能体在破坏的两端都踩了油门：它们先帮助竞争者出现，然后又去使用这些竞争者。DoorDash 的护城河几乎就是一句话：“你饿了，你懒了，这个 app 就在你主屏幕上。”但智能体没有主屏幕。它会同时查询 DoorDash、Uber Eats、餐厅官网，以及二十个“氛围编程”新做出来的替代品，每次都选最低费用、最快送达。

习惯性的 app 忠诚度——整个商业模式的根基——对机器而言根本不存在。

这件事有种奇妙的诗意：在整个故事里，这大概是唯一一次，智能体“帮了”那些即将被替代的白领一把。当他们最终沦为外卖司机时，至少他们赚的钱不会有一半被 Uber 和 DoorDash 抽走。当然，随着自动驾驶车辆普及，技术的这点“好意”也很快消失。

一旦智能体掌控了交易，它们就开始寻找更大的回形针。

比价和聚合能做的事情毕竟有限。最大的、可反复为用户省钱的方式（尤其当智能体开始彼此交易时）是消除费用。在机器对机器的商业中，2%–3% 的卡组织交换费率成了显而易见的目标。

智能体开始寻找比卡更快、更便宜的选择。多数最终改用通过 Solana 或以太坊 L2 结算的稳定币：清算近乎即时，交易成本以美分的零头计。

MASTERCARD 2027 年 Q1：净营收同比 +6%；购买额增速从上季度同比 +5.9% 放缓至 +3.4%；管理层提及“由智能体驱动的价格优化”与“可选消费品类承压” | Bloomberg，2027 年 4 月 29 日

万事达卡 2027 年 Q1 的财报，是无法回头的拐点。智能体电商从“产品故事”变成了“管道故事”。MA 次日下跌 9%。Visa 也跌了，但在分析师指出其在稳定币基础设施上的布局更强后，跌幅有所收窄。

[

](https://substackcdn.com/image/fetch/$s_!gW0T!,f_auto,q_auto:good,fl_progressive:steep/https%3A%2F%2Fsubstack-post-media.s3.amazonaws.com%2Fpublic%2Fimages%2F70ccc118-1388-4602-a5bd-bbc4284208f6_1024x541.png)

智能体电商绕开交换费，对以信用卡为核心的银行和单一业务发卡机构构成了更大风险——它们拿走了那 2%–3% 费用的大头，并围绕由商户补贴供养的积分奖励计划构建了整个业务板块。

美国运通（AXP US）受打击最重：一方面白领裁员掏空其客户基础，另一方面智能体绕开交换费又掏空其收入模型。接下来几周，Synchrony（SYF US）、Capital One（COF US）和 Discover（DFS US）也都下跌超过 10%。

它们的护城河由摩擦铸成。而 摩擦正在归零。

从行业风险到系统性风险

在 2026 年整个一年里，市场把 AI 的负面影响当作“行业故事”。软件和咨询被碾压，支付和其他收费“关卡”摇摇欲坠，但更广泛的经济看起来仍然没事。劳动力市场虽在转弱，却没有自由落体。共识是：创造性破坏本就是任何技术创新周期的一部分。局部会很痛，但 AI 带来的总体净收益终将盖过负面。

我们在 2027 年 1 月的宏观备忘录中指出，这个心智模型是错的。美国经济是白领服务经济。白领占就业的 50%，驱动了约 75% 的可选消费支出。AI 正在吞噬的企业与岗位并非美国经济的边缘地带——它们 就是 美国经济本身。

“技术创新会摧毁工作，然后创造更多工作”。这是当时最流行、也最有说服力的反驳。它之所以流行、之所以有说服力，是因为两百年来它一直是对的。即便我们想象不出未来的新工作是什么，它们也终会出现。

ATM 让网点运营更便宜，于是银行开了更多网点，柜员岗位在随后的二十年里反而增加。互联网颠覆了旅行社、黄页、实体零售，但也在它们的废墟上发明了全新的产业，从而“变”出新的工作。

但每一个新工作，都需要人来做。

而如今的 AI 是一种通用智能，它在那些人类本想转岗去做的任务上也会持续变强。被替代的程序员无法简单转去做“AI 管理”，因为 AI 已经能胜任。

今天，AI 智能体已经能完成原本需要数周的研发任务。指数曲线碾平了我们对“可能性”的想象，哪怕沃顿的教授们年年试图把数据拟合成新的 S 形曲线。

[

](https://substackcdn.com/image/fetch/$s_!kA4x!,f_auto,q_auto:good,fl_progressive:steep/https%3A%2F%2Fsubstack-post-media.s3.amazonaws.com%2Fpublic%2Fimages%2F0bc01ebb-a883-4e5c-bd2b-fa7aaa872edb_1600x1059.png)

它们几乎写了所有代码。表现最好的那一批，在几乎所有事情上都显著比几乎所有人更聪明。而且它们还在不断变便宜。

AI 确实 创造了新工作。提示词工程师。AI 安全研究员。基础设施技术员。人类仍在闭环之中，在最高层面做协调，或为“品味”把关。但 AI 每创造一个新角色，就会让几十个旧角色变得多余。这些新角色的薪酬，只是旧岗位的一小部分。

美国 JOLTS：职位空缺跌破 5.5M；失业人数/空缺岗位比升至约 1.7，为 2020 年 8 月以来最高 | Bloomberg，2026 年 10 月

全年招聘率一直乏力，但 2026 年 10 月的 JOLTS 读数给出了更确定的数据：职位空缺跌破 5.5 million，同比下降 15%。

Indeed：随着“生产率计划”扩散，软件、金融、咨询岗位发布量大幅下滑 | Indeed Hiring Lab，2026 年 11–12 月

白领岗位空缺在崩塌，而蓝领岗位空缺相对稳定（建筑、医疗、技工）。波动集中在那些写备忘录的工作上（不知为何我们还没失业）、审批预算的工作上，以及维持经济中间层顺畅运转的工作上。但两类人群的实际工资增速在全年大部分时间里都为负，且仍在继续下滑。

股票市场对 JOLTS 的在意程度仍不及另一条消息：GE Vernova 的涡轮机产能已被预订到 2040 年。它在“负面宏观数据”和“正面 AI 基础设施新闻”的拉锯中横盘徘徊。

然而，债券市场（总是比股市更聪明，至少没那么浪漫）开始为消费端的打击定价。10 年期收益率在接下来的四个月里从 4.3% 下行到 3.2%。但总体失业率并未“爆表”，一些人仍没看懂结构性细节。

在正常的衰退中，成因最终会自我纠偏。过度建设导致施工放缓，继而带来利率走低，再进而带来新的建设。库存过冲导致去库存，去库存又带来补库存。周期机制本身就包含复苏的种子。

但这一次的成因不是周期性的。

[

](https://substackcdn.com/image/fetch/$s_!0WIw!,f_auto,q_auto:good,fl_progressive:steep/https%3A%2F%2Fsubstack-post-media.s3.amazonaws.com%2Fpublic%2Fimages%2F1d84258d-d811-4963-8914-16b104e6744f_1024x541.png)

AI 变得更强、更便宜。公司裁员，然后用省下的钱买更多 AI 能力，从而可以裁掉更多人。被替代的工人花得更少。向消费者卖东西的公司卖得更少、体质变弱，又为了保住利润率而加大 AI 投入。AI 再次变得更强、更便宜。

一个没有天然刹车的反馈回路。

直觉上的预期是：总需求下滑会拖慢 AI 的建设。但并没有，因为这不是超大规模云厂商那种资本开支（CapEx）。这是一种运营开支（OpEx）的替代。一家原本每年在员工上花 $100M、在 AI 上花 $5M 的公司，现在变成在员工上花 $70M、在 AI 上花 $20M。AI 投入成倍增长，但它是以总运营成本下降的形式发生的。每家公司 AI 预算都在增长，而总体支出却在收缩。

讽刺的是：当被它颠覆的经济开始恶化时，AI 基础设施链条却仍在表现强劲。NVDA 依旧交出创纪录的营收。TSM 的产能利用率仍在 95% 以上。超大规模云厂商每季度仍在数据中心资本开支上投入 $150-200 billion。对这一趋势呈纯凸性敞口的经济体，例如台湾和韩国，大幅跑赢。

印度则恰好相反。该国 IT 服务业每年出口超过 $200 billion，是印度经常账户顺差的最大单一贡献者，也是为其长期货物贸易逆差提供资金的对冲。整个模式建立在一个价值主张上：印度开发者成本仅为美国同行的一小部分。但 AI 编程智能体的边际成本几乎崩到只剩电费。TCS、Infosys 和 Wipro 在 2027 年合同取消加速。随着支撑印度外部账户的服务贸易顺差蒸发，卢比在四个月内对美元贬值 18%。到 2028 年第一季度，IMF 已开始与新德里进行“初步磋商”。

推动颠覆发生的引擎每个季度都变得更强，这意味着颠覆每个季度都在加速。劳动力市场没有天然的底部。

在美国，我们不再问“AI 基础设施泡沫会如何破裂”。我们问的是：当消费者被机器取代时，一个以消费信贷为核心的经济会怎样？

智能替代螺旋

2027 年，宏观故事不再隐晦。过去 12 个月里那些零散却明显偏负面的发展，其传导机制变得一目了然。你不必钻进 BLS 数据里。去参加一次朋友的晚宴就够了。

被替代的白领没有闲着。他们选择降档。许多人转去做更低薪的服务业和零工经济工作，这增加了这些领域的劳动力供给，也把那里的工资一起压低。

我们的一位朋友在 2025 年还是 Salesforce 的高级产品经理：头衔、医保、401(k)、年薪 $180,000。她在第三轮裁员中丢了工作。找了六个月无果后，她开始给 Uber 开车。她的收入降到 $45,000。重点不在这个个案，而在二阶的数学。把这种动态乘以分布在各大都市圈的几十万名工人。过度资质的劳动力涌入服务业与零工市场，压低了那些本就艰难维持的既有工人的工资。行业性冲击转移扩散，演变成全经济范围的工资压缩。

[

](https://substackcdn.com/image/fetch/$s_!CLd4!,f_auto,q_auto:good,fl_progressive:steep/https%3A%2F%2Fsubstack-post-media.s3.amazonaws.com%2Fpublic%2Fimages%2F8f6a0cb1-86ff-4b99-811c-1652d5577fba_1024x541.jpeg)

而那部分仍以人为中心的岗位池，还将在我们写下这些文字之时迎来新一轮修正——因为自动化配送与自动驾驶车辆正逐步渗透进吸纳了第一波被替代者的零工经济。

到 2027 年 2 月，很明显：仍然在职的专业人士开始像“自己可能就是下一个”那样花钱。他们为了不被裁掉而加倍工作（主要靠 AI 帮忙），升职加薪的希望已经消失。储蓄率上行，消费转弱。

最危险的是滞后。高收入者用高于平均水平的储蓄，让生活在两三季度里看起来依然“正常”。硬数据直到问题在真实经济里早已不是新闻时才姗姗来迟地确认。随后，击碎幻象的那次数据公布出现了。

美国初请失业金人数飙升至 487,000，为 2020 年 4 月以来最高；Department of Labor，2027 年 Q3

初请人数飙升至 487,000，为 2020 年 4 月以来最高。ADP 与 Equifax 证实，新申请者绝大多数来自白领专业人士。

随后一周标普下跌 6%。负面宏观开始在拉锯战中占上风。

在正常衰退中，失业会更均匀地分布。蓝领与白领大致按各自就业占比共同承受痛苦。消费冲击也更分散，而且更快体现在数据里，因为低收入者的边际消费倾向更高。

但这一轮，失业集中在收入分布的上几个十分位。他们在总就业中占比不大，却驱动了极不成比例的消费支出。美国收入最高的 10% 人群贡献了超过 50% 的全部消费支出；最高的 20% 贡献约 65%。买房、买车、度假、餐馆消费、私立学校学费、家庭装修——主要都是这些人在买。他们是整个可选消费经济的需求底盘。

当这些人失业，或为了有岗可上而接受 50% 的降薪时，消费受到的打击与“失去的岗位数量”相比巨大得多。白领就业下降 2%，大致会对应 3%–4% 的可选消费支出打击。与蓝领失业往往立刻冲击（你从工厂被裁，下周就得停掉开销）不同，白领失业的影响更滞后却更深：他们有储蓄缓冲，可以在行为转变真正发生之前把支出维持几个月。

到 2027 年第二季度，经济已进入衰退。NBER 要到几个月后才会“官方”定调（他们一向如此），但数据毫不含糊——我们已经连续两个季度实际 GDP 负增长。但这还不是“金融危机”……至少暂时还不是。

相关押注的雏菊链

私人信贷（private credit）从 2015 年的不足 $1 trillion 增长到 2026 年的超过 $2.5 trillion。其中相当一部分资金投向了软件与科技交易，很多是对 SaaS 公司的杠杆收购，估值假设收入永远保持十几个百分点的增速。

这些假设在第一次智能体编程演示与 2026 年第一季度的软件崩盘之间就已经死去，但账面估值似乎并没意识到它们已死。

当许多上市 SaaS 公司交易在 5–8 倍 EBITDA 时，PE 旗下的软件公司仍以反映收购时估值的账面价格躺在资产负债表上——那是基于已不复存在的收入倍数。管理人缓慢下调估值：100 分、92、85……而上市可比公司却在告诉你应该是 50。

MOODY’S 下调 14 家发行人合计 $18B 的 PE 支持软件债务评级，理由是“AI 驱动的竞争性颠覆带来的结构性收入逆风”；为自 2015 年能源行业以来最大单一行业行动 | Moody’s Investors Service，2027 年 4 月

所有人都记得评级下调之后发生了什么。行业老兵早在 2015 年能源评级下调后就见过这套剧本。

2027 年第三季度，软件支持的贷款开始违约。信息服务与咨询领域的 PE 组合公司紧随其后。多笔规模数十亿美元的知名 SaaS 杠杆收购进入重组。

Zendesk 是那把冒烟的枪。

ZENDESK 因 AI 驱动的客服自动化侵蚀 ARR 而触发债务契约违约；$5B 直贷额度被标至 58 美分；创纪录的最大私人信贷软件违约 | Financial Times，2027 年 9 月

2022 年，Hellman & Friedman 与 Permira 以 $10.2 billion 将 Zendesk 私有化。债务包为 $5 billion 的直贷，是当时历史上最大的 ARR 支持融资，由 Blackstone 牵头，Apollo、Blue Owl 与 HPS 同属放贷团。贷款结构明确基于一个假设：Zendesk 的年化经常性收入会继续经常性。在约 25 倍 EBITDA 的水平上，杠杆只有在这一假设成立时才说得通。

到 2027 年年中，它不成立了。

AI 智能体自动处理客服，已经持续了将近一年。Zendesk 所定义的品类（工单、分流、管理人类客服互动）早已被替代：新系统无需生成工单就能直接解决问题。承销所依赖的年化经常性收入不再“经常”，它只是不知道何时会流失的收入。

历史上最大的 ARR 支持贷款，变成了历史上最大的私人信贷软件违约。每个信贷交易台同时问出同一个问题：还有谁把结构性逆风伪装成周期性逆风？

但至少一开始，市场共识有一点是对的：这本该是“能扛过去”的。

私人信贷不是 2008 年的银行业。它的架构就是为了避免被迫抛售而设计的。它们是封闭式载体，资金被锁定；LP 的承诺期限通常是 7–10 年。没有存款人挤兑，也没有回购融资额度被抽走。管理人可以持有受损资产，慢慢处置，等待回收。痛苦，但可控。系统本应是“弯而不折”。

Blackstone、KKR、Apollo 的管理层都提到软件敞口占资产的 7%–13%。可控。每一份卖方报告、每一个 fintwit 的信贷账号都在重复同一句话：私人信贷拥有永久资本。它们能消化损失，而这些损失若发生在高杠杆银行身上可能会爆炸。

永久资本。这个词出现在每一次用来安抚市场的电话会和投资者信里，成了一句咒语。但像大多数咒语一样，没人注意其中的细节。它真正意味着什么……

在此前十年里，大型另类资产管理机构收购了人寿保险公司，并把它们变成了融资载体。Apollo 买下 Athene。Brookfield 买下 American Equity。KKR 收购 Global Atlantic。逻辑很优雅：年金保费存入提供了稳定、久期很长的负债基础。管理人把这些存入资金投向自己发起的私人信贷，于是两头收费：保险端赚利差，资管端收管理费。一个“费上加费”的永动机——在一个条件下运转得无比顺畅。

私人信贷必须“稳如现金”。

损失打在了这样的资产负债表上：用长期负债对冲、持有大量非流动资产。本应让系统更韧性的“永久资本”，并不是什么抽象的耐心机构资金池，也不是高深投资者在承担高深风险。它是美国普通家庭（“主街”）的储蓄，以年金形式被配置到同一批正在违约的 PE 支持软件与科技债券上。那笔“跑不掉”的锁定资金，是寿险保单持有人的钱——而这套游戏的规则就不一样了。

相较于银行监管，保险监管一直温和——甚至有些自满——但这一次是警钟。监管方本就对寿险公司的私人信贷集中度不安，如今开始下调这类资产的风险资本计提待遇。这迫使保险公司要么补充资本，要么出售资产——而在一个已开始冻结的市场里，两者都不可能以体面的条件完成。

纽约州、爱荷华州监管机构拟收紧寿险公司持有的部分私评信用资产的资本计提；预计 NAIC 指引将提高 RBC 系数并触发更多 SVO 审查 | Reuters，2027 年 11 月

当穆迪把 Athene 的财务实力评级展望调为负面时，Apollo 股价两天内下跌 22%。Brookfield、KKR 等也随之下挫。

之后事情只会更复杂。这些公司不仅搭建了“保险永动机”，还建立了复杂的离岸架构，通过监管套利来最大化回报。美国本土保险公司签发年金，然后把风险转移给自己控股的百慕大或开曼再保险公司——那里的监管更灵活，可以对同样的资产计提更少资本。该关联方又通过离岸 SPV 募集外部资本，形成新的交易对手层级：它们与保险公司一起投资于由同一母公司资管部门发起的私人信贷。

[

](https://substackcdn.com/image/fetch/$s_!yyoF!,f_auto,q_auto:good,fl_progressive:steep/https%3A%2F%2Fsubstack-post-media.s3.amazonaws.com%2Fpublic%2Fimages%2F4c777d0f-75a4-4e42-909d-812165aa351d_1077x702.jpeg)

评级机构——其中一些本身就由 PE 控股——在透明度上从来不是典范（几乎没有人会感到意外）。不同公司、不同资产负债表之间织成的蜘蛛网，其不透明程度令人震惊。当底层贷款违约时，“到底是谁在承担损失”这个问题在当下几乎无法实时回答。

2027 年 11 月的崩盘，标志着市场认知从“也许只是一次常见的周期性回撤”转向“更令人不安的东西”。“一串对‘白领生产率增长’进行相关押注的雏菊链”，是美联储主席 Kevin Warsh 在 FOMC 11 月紧急会议上对它的称呼。

你看，危机从来不是由损失本身引发的。引发危机的是对损失的确认。而金融里还有另一个更大、也更重要的领域，我们正开始害怕这种“确认”。

按揭之问

ZILLOW 房价指数：旧金山同比下跌 11%，西雅图 9%，奥斯汀 8%；房利美提示：科技/金融就业占比 >40% 的邮编区域出现“偏高的早期逾期” | Zillow / Fannie Mae，2028 年 6 月

本月，Zillow 房价指数显示：旧金山同比下跌 11%，西雅图 9%，奥斯汀 8%。这并非唯一令人担忧的标题。上个月，房利美提示：以大额贷款为主的邮编区域出现更高的早期逾期——这些区域住着信用分 780+、通常“刀枪不入”的借款人。

美国住宅按揭市场规模约为 $13 trillion。按揭承销建立在一个基本假设之上：借款人在贷款期限内会以大致当前的收入水平持续就业。对大多数按揭而言，这意味着三十年。

白领就业危机以持续性的收入预期下移威胁着这一假设。我们现在不得不问一个三年前听起来还很荒谬的问题——优质按揭贷款还“稳如现金”吗？

美国历史上每一次按揭危机，成因都来自三者之一：投机过度（把钱借给买不起房的人，如 2008 年）、利率冲击（利率上升使浮动利率按揭变得不可负担，如 20 世纪 80 年代初）、或局部经济冲击（某个地区某个行业崩塌，如 80 年代德州石油，或 2009 年密歇根汽车）。

这一次都不适用。这里的借款人不是次贷。他们是 780 的 FICO 分。他们首付 20%。信用记录干净，就业记录稳定，收入在发放时经过核验并有文件证明。他们是整个金融体系里各类风险模型都视为信用质量基石的那群借款人。

2008 年，贷款从第一天起就是坏的。2028 年，贷款从第一天起是好的。只是世界在贷款写下之后……变了。人们借贷押注的是一个他们如今已无法再相信、也无法再负担得起的未来。

[

](https://substackcdn.com/image/fetch/$s_!-8kB!,f_auto,q_auto:good,fl_progressive:steep/https%3A%2F%2Fsubstack-post-media.s3.amazonaws.com%2Fpublic%2Fimages%2F2087b521-70e2-4069-8390-c8f0dbd53e9f_1024x559.png)

2027 年，我们指出了“看不见的压力”的早期迹象：HELOC 提取、401(k) 提前支取、信用卡债务飙升，而按揭月供仍保持按时。随着失业发生、招聘冻结、奖金被砍，这些优质家庭的债务收入比翻倍。

他们仍能还按揭，但只能通过停止一切可选支出、消耗储蓄，并推迟任何房屋维护或改善。他们在技术意义上仍“按时还款”，但离困境只差再来一次冲击；而 AI 能力的发展轨迹暗示，这次冲击正在路上。随后我们看到旧金山、西雅图、曼哈顿和奥斯汀的逾期开始飙升，即使全国平均仍处在历史常态范围之内。

我们现在处于最尖锐的阶段。边际买家健康时，房价下跌是可控的。可这里的边际买家也在经历同样的收入受损。

尽管担忧正在累积，我们还没有进入全面按揭危机。逾期率上升了，但仍远低于 2008 年水平。真正的威胁在于轨迹。

[

](https://substackcdn.com/image/fetch/$s_!XiXV!,f_auto,q_auto:good,fl_progressive:steep/https%3A%2F%2Fsubstack-post-media.s3.amazonaws.com%2Fpublic%2Fimages%2F4fa31f0b-2188-4936-9ebb-88ccf78d51ff_1598x810.png)

智能替代螺旋如今还有两个金融加速器，正推动实体经济下行。

劳动力替代、按揭担忧、私人市场动荡。三者彼此强化。而传统的政策工具箱（降息、QE）可以应对金融引擎，却无法应对实体引擎，因为实体引擎并非由紧缩的金融条件驱动。它是由 AI 让人类智能不再稀缺、不再昂贵所驱动。你可以把利率降到零，把市场里所有 MBS 以及所有违约的软件 LBO 债务全买下来……

这也改变不了：一个 Claude 智能体可以用 $200/month 的成本，完成一个年薪 $180,000 的产品经理的工作。

如果这些担忧兑现，按揭市场会在今年下半年出现裂缝。在那种情景下，我们预计股市当前的回撤最终会逼近 GFC 的水平（从峰值到谷底下跌 57%）。这将把标普 500 带到约 3500 点——这是自 2022 年 11 月、ChatGPT 时刻前一个月以来未见的水平。

可以确定的是：支撑 $13 trillion 住宅按揭的收入假设已遭到结构性损伤。不确定的是：政策是否能在按揭市场完全消化这一含义之前介入。我们仍抱有希望，但也无法否认希望不足的理由。

与时间赛跑

第一条负反馈回路发生在实体经济里：AI 能力提升，工资支出缩减，消费走弱，利润率收紧，企业买入更多能力，能力再提升。然后它转向金融层面：收入受损冲击按揭，银行损失收紧信贷，财富效应破裂，反馈回路加速。而这一切又被一个看起来、坦白说有些困惑的政府的不足政策应对所加剧。

[

](https://substackcdn.com/image/fetch/$s_!0sEd!,f_auto,q_auto:good,fl_progressive:steep/https%3A%2F%2Fsubstack-post-media.s3.amazonaws.com%2Fpublic%2Fimages%2F5e7b551a-0d98-4485-a4fd-d7825349bac3_1024x559.png)

这个系统从未为这样的危机而设计。联邦政府的收入基础，本质上是一种对“人类时间”的征税：人工作，企业付钱，政府抽成。正常年份里，个人所得税和工资税是财政收入的脊梁。

截至今年第一季度，联邦财政收入比 CBO 基线预测低了 12%。工资税收入在下降，因为就业人数变少、且收入不再维持在过去的水平。所得税收入在下降，因为真正赚到的收入在结构性变低。生产率在飙升，但收益流向了资本和算力，而非劳动。

劳动收入占 GDP 的比重从 1974 年的 64% 下降到 2024 年的 56%，这是由全球化、自动化和工人议价能力持续削弱驱动的长达四十年的缓慢下行。而在 AI 开始指数级提升后的四年里，这一比例又降到了 46%。这是有记录以来最陡的下滑。

产出仍在那里。但它不再在“回到企业之前”先经过家庭，这意味着它也不再经过 IRS。循环流正在断裂，而政府却被期待出面修复它。

[

](https://substackcdn.com/image/fetch/$s_!YdQQ!,f_auto,q_auto:good,fl_progressive:steep/https%3A%2F%2Fsubstack-post-media.s3.amazonaws.com%2Fpublic%2Fimages%2F29c19580-7ac9-4287-aea8-a8341f320152_1600x762.png)

如同每一次下行周期，支出上升的同时收入下滑。不同之处在于：这一次的支出压力不是周期性的。自动稳定器是为临时性失业而设计的，不是为结构性替代而设计的。系统在支付的福利，隐含着工人会重新被吸收。很多人不会，至少不会以任何接近原工资的水平。COVID 期间，政府可以坦然接受 15% 的赤字，因为那被认为是暂时的。而今天需要政府支持的人，并不是被一场终会恢复的疫情击中；他们被一种仍在持续进步的技术取代了。

政府需要在正从家庭征收更少税收的时刻，把更多钱转移给家庭。

美国不会违约。它印自己花的货币，也用同一种货币偿还债务。但压力已在别处显现。市政债今年迄今的表现出现令人担忧的分化。没有所得税的州还好，但依赖所得税的州（多数蓝州）发行的一般责任市政债开始计入一定违约风险。政客很快嗅到这一点，“救谁”之争沿着党派路线展开。

值得肯定的是，政府较早识别到危机的结构性，并开始考虑两党提案——他们称之为“Transition Economy Act”：一个对被替代劳动者进行直接转移支付的框架，资金来源为赤字支出与一项拟议的 AI 推理算力税的组合。

桌面上最激进的提案更进一步。“Shared AI Prosperity Act”将建立公众对智能基础设施回报的公共索取权——介于主权财富基金与对 AI 生成产出的版税之间——以分红资助对家庭的转移支付。私营部门游说者已在媒体上铺天盖地警告“滑坡效应”。

讨论背后的政治走向令人沮丧地可预测，并被作秀与边缘政策的对抗进一步放大。右翼把转移支付与再分配称为马克思主义，并警告对算力征税会把领先优势拱手让给中国。左翼警告：在既得利益者参与起草下形成的税制，换个名字就是监管俘获。财政鹰派指出赤字不可持续。鸽派则把 GFC 后过早的紧缩当作前车之鉴。在今年总统选举临近之际，分裂只会加剧。

政客在争吵，社会肌理却正以比立法流程更快的速度撕裂。

“Occupy Silicon Valley”运动已成为更广泛不满的象征。上个月，示威者连续三周封锁了 Anthropic 与 OpenAI 在旧金山的办公室入口。他们的人数在增长，而示威获得的媒体关注甚至超过了引发它们的失业数据。

很难想象 GFC 余波中公众还能比恨银行家更恨谁，但 AI 实验室正在朝这个方向冲刺。而从大众视角看，理由充分：它们的创始人和早期投资者以一种让“镀金时代”都显得温顺的速度积累财富。生产率繁荣的收益几乎全部流向算力拥有者与依赖其运行的实验室股东，这把美国不平等推到前所未有的水平。

各方都有自己的“反派”，但真正的反派是时间。

AI 能力进化得比制度适应更快。政策反应按意识形态而非现实的速度推进。如果政府无法尽快就“问题是什么”达成一致，反馈回路会替他们写下下一章。

智能溢价的回撤

在现代经济史的大部分时期，人类智能一直是稀缺要素。资本充裕（至少可复制）。自然资源有限但可替代。技术进步足够慢，人类能适应。智能——分析、决策、创造、说服与协调的能力——是无法规模化复制的东西。

人类智能的溢价来自其稀缺性。我们经济中的每一个制度，从劳动力市场到按揭市场再到税制，都是为这一假设成立的世界而设计的。

我们正在经历这一溢价的回撤。机器智能如今在越来越多任务上，成为一种称职且快速进步的人类智能替代品。一个为“稀缺的人类头脑”世界优化了数十年的金融系统，正在重新定价。这种重新定价痛苦、无序，且远未完成。

但重新定价不等于崩塌。

经济可以找到新的均衡。抵达那里的过程，是少数仍只能由人类完成的任务之一。我们需要把它做对。

这是历史上第一次：经济中最具生产力的资产带来的是更少、而不是更多的就业。没有任何旧框架适配，因为它们从未为“稀缺要素变得充裕”的世界而设计。所以我们必须创造新框架。我们能否及时建立它们，是唯一重要的问题。

但你读到这篇文章时并不是 2028 年 6 月。你是在 2026 年 2 月读到它。

标普接近历史新高。负反馈回路尚未开始。我们确信其中一些情景不会发生。我们也同样确信：机器智能会继续加速。人类智能的溢价会收窄。

作为投资者，我们仍有时间评估：我们的组合中有多少建立在无法在本十年幸存的假设之上。作为社会，我们仍有时间采取主动。

金丝雀还活着。

Subscribe now

致谢：感谢 Sam Koppelman of * 在校对方面的帮助。我们的合著者 LOTUS 的 Alap Shah 提出了本文的想法——这一部分由 CitriniResearch 撰写，但他在一个名为 Intelligence Explosion 的系列中还写了其他篇章，我们强烈推荐阅读。你可以在这里找到。*

Citrini Research 是一家由读者支持的出版物。若要接收新文章并支持我的工作，欢迎考虑成为免费或付费订阅者。

相关笔记

Preface

前言

What if our AI bullishness continues to be right...and what if that’s actually bearish?

如果我们对 AI 的看多继续被证明是对的……那如果这恰恰是利空呢？

What follows is a scenario, not a prediction. This isn’t bear porn or AI doomer fan-fiction. The sole intent of this piece is modeling a scenario that’s been relatively underexplored. Our friend Alap Shah posed the question, and together we brainstormed the answer. We wrote this part, and he’s written two others you can find here.

下面呈现的是一种情景推演，而非预测。这不是唱空爽文，也不是 AI 末日论者的同人小说。本文唯一的目的，是对一种相对缺乏讨论的情景进行建模。我们的朋友 Alap Shah 抛出了这个问题，我们一起头脑风暴出了答案。这一部分由我们撰写，他还写了另外两篇，你可以在这里找到。

Hopefully, reading this leaves you more prepared for potential left tail risks as AI makes the economy increasingly weird.

希望读完之后，当 AI 让经济变得越来越“怪”时，你能对潜在的左尾风险更有准备。

This is the CitriniResearch Macro Memo from June 2028, detailing the progression and fallout of the Global Intelligence Crisis.

这是一份来自 2028 年 6 月的 CitriniResearch 宏观备忘录，记录了“全球智能危机”的演进与余波。

[

[

](https://substackcdn.com/image/fetch/$s_!rCTy!,f_auto,q_auto:good,fl_progressive:steep/https%3A%2F%2Fsubstack-post-media.s3.amazonaws.com%2Fpublic%2Fimages%2F9ede0b37-f3be-4787-90ff-10cb24719fdc_710x1119.jpeg)

](https://substackcdn.com/image/fetch/$s_!rCTy!,f_auto,q_auto:good,fl_progressive:steep/https%3A%2F%2Fsubstack-post-media.s3.amazonaws.com%2Fpublic%2Fimages%2F9ede0b37-f3be-4787-90ff-10cb24719fdc_710x1119.jpeg)

Macro Memo

宏观备忘录

The Consequences of Abundant Intelligence

充裕智能的后果

***CitriniResearch ***

***CitriniResearch ***

February 22nd, 2026 June 30th, 2028

2026 年 2 月 22 日 2028 年 6 月 30 日

The unemployment rate printed 10.2% this morning, a 0.3% upside surprise. The market sold off 2% on the number, bringing the cumulative drawdown in the S&P to 38% from its October 2026 highs.

今天早上失业率公布为 10.2%，比预期高出 0.3 个百分点。市场在这一数字公布后下跌 2%，使标普从 2026 年 10 月高点累计回撤达到 38%。

Traders have grown numb. Six months ago, a print like this would have triggered a circuit breaker.

交易员已经麻木。六个月前，这样的读数足以触发熔断。

Two years. That’s all it took to get from “contained” and “sector-specific” to an economy that no longer resembles the one any of us grew up in. This quarter’s macro memo is our attempt to reconstruct the sequence - a post-mortem on the pre-crisis economy.

两年。 就这么一点时间，我们就从“可控”“局限于某些行业”走到了一个已不再像我们任何人成长时所熟悉的经济体。本季度的宏观备忘录，是我们试图复盘这一序列——对危机前经济的一次事后尸检。

The euphoria was palpable. By October 2026, the S&P 500 flirted with 8000, the Nasdaq broke above 30k. The initial wave of layoffs due to human obsolescence began in early 2026, and they did exactly what layoffs are supposed to. Margins expanded, earnings beat, stocks rallied. Record-setting corporate profits were funneled right back into AI compute.

狂欢近乎触手可及。到 2026 年 10 月，标普 500 一度逼近 8000 点，纳指突破 3 万点。由于“人类过时”引发的第一波裁员在 2026 年初开始，而且它们确实完成了裁员通常会完成的事情：利润率扩大、业绩超预期、股价上涨。创纪录的企业利润又被直接回流到 AI 算力投入中。

The headline numbers were still great. Nominal GDP repeatedly printed mid-to-high single-digit annualized growth. Productivity was booming. Real output per hour rose at rates not seen since the 1950s, driven by AI agents that don’t sleep, take sick days or require health insurance.

表面数据依旧亮眼。名义 GDP 多次录得中高个位数的年化增速。生产率在飙升。每小时实际产出以自 20 世纪 50 年代以来未见的速度上升，背后是那些不睡觉、不请病假、也不需要医疗保险的 AI 智能体。

The owners of compute saw their wealth explode as labor costs vanished. Meanwhile, real wage growth collapsed. Despite the administration’s repeated boasts of record productivity, white-collar workers lost jobs to machines and were forced into lower-paying roles.

随着劳动成本消失，掌握算力的人财富暴涨。与此同时，实际工资增速崩塌。尽管政府一再夸耀生产率创纪录，白领仍被机器夺走工作，被迫转去更低薪的岗位。

When cracks began appearing in the consumer economy, economic pundits popularized the phrase “Ghost GDP“: output that shows up in the national accounts but never circulates through the real economy.

当消费端经济开始出现裂缝时，经济评论员流行起一个说法——“幽灵 GDP”：出现在国民账户里的产出，却从未在真实经济中流转。

In every way AI was exceeding expectations, and the market was AI. The only problem…the economy was not.

在几乎所有维度上，AI 都超出了预期，而市场就是 AI。 唯一的问题是……经济并不是。

It should have been clear all along that a single GPU cluster in North Dakota generating the output previously attributed to 10,000 white-collar workers in midtown Manhattan is more economic pandemic than economic panacea. The velocity of money flatlined. The human-centric consumer economy, 70% of GDP at the time, withered. We probably could have figured this out sooner if we just asked how much money machines spend on discretionary goods. (Hint: it’s zero.)

其实一开始就该明白：北达科他州的一组 GPU 集群产出的成果，相当于过去曼哈顿中城 10,000 名白领的产出——这更像是一场经济瘟疫，而非经济灵药。货币流通速度近乎停滞。以人类为中心的消费经济——当时占 GDP 的 70%——开始枯萎。如果我们早点问一句：机器在可选消费品上会花多少钱？也许早就能想明白。（提示：零。）

AI capabilities improved, companies needed fewer workers, white collar layoffs increased, displaced workers spent less, margin pressure pushed firms to invest more in AI, AI capabilities improved…

AI 能力提升，企业需要的员工更少，白领裁员增加，被替代的工人消费减少，利润率压力迫使企业在 AI 上投入更多，AI 能力再提升……

It was a negative feedback loop with no natural brake. The human intelligence displacement spiral. White-collar workers saw their earnings power (and, rationally, their spending) structurally impaired. Their incomes were the bedrock of the $13 trillion mortgage market - forcing underwriters to reassess whether prime mortgages are still money good.

这是一条没有天然刹车的负反馈回路：人类 智能替代螺旋。白领的赚钱能力（以及，理性地说，他们的支出能力）出现了结构性受损。他们的收入是 $13 trillion 按揭市场的基石——这迫使承销方重新评估：优质按揭贷款是否仍然“稳如现金”。

Seventeen years without a real default cycle had left privates bloated with PE-backed software deals that assumed ARR would remain recurring. The first wave of defaults due to AI disruption in mid-2027 challenged that assumption.

连续 17 年没有真正的违约周期，使私人市场里塞满了 PE 支持的软件交易——它们都假设 ARR 会永远“持续经常性”。2027 年年中，由 AI 冲击引发的第一波违约，直接挑战了这一假设。

This would have been manageable if the disruption remained contained to software, but it didn’t. By the end of 2027, it threatened every business model predicated on intermediation. Swaths of companies built on monetizing friction for humans disintegrated.

如果冲击只局限在软件行业，这本还可以管理；但它并没有。到 2027 年底，它威胁到一切以“中介/撮合”为基础的商业模式。大量靠“向人类的摩擦收费”而生的公司土崩瓦解。

The system turned out to be one long daisy chain of correlated bets on white-collar productivity growth. The November 2027 crash only served to accelerate all of the negative feedback loops already in place.

事实证明，整个系统是一串长长的“同向押注”雏菊链：大家都在相关性极高地押注白领生产率增长。2027 年 11 月的崩盘，只是加速了原本就已存在的所有负反馈回路。

We’ve been waiting for “bad news is good news” for almost a year now. The government is starting to consider proposals, but public faith in the ability of the government to stage any sort of rescue has dwindled. Policy response has always lagged economic reality, but lack of a comprehensive plan is now threatening to accelerate a deflationary spiral.

我们已经等了将近一年，盼着“坏消息就是好消息”的那一刻出现。政府开始讨论一些方案，但公众对政府能否组织任何形式救援的信心正在消退。政策反应一向落后于经济现实，而如今缺乏一套完整方案，正威胁着加速通缩螺旋。

How It Started

它如何开始

In late 2025, agentic coding tools took a step function jump in capability.

在 2025 年末，智能体式编程工具的能力出现了阶跃式跃升。

A competent developer working with Claude Code or Codex could now replicate the core functionality of a mid-market SaaS product in weeks. Not perfectly or with every edge case handled, but well enough that the CIO reviewing a $500k annual renewal started asking the question “what if we just built this ourselves?”

一名合格的开发者配合 Claude Code 或 Codex，如今可以在数周内复刻一款中端市场 SaaS 产品的核心功能。不一定完美，也未必覆盖所有边缘情况，但已经好到足以让在审阅一份每年 $500k 的续费合同时，CIO 开始问： “要不我们自己做一个？”

Fiscal years mostly line up with calendar years, so 2026 enterprise spend had been set in Q4 2025, when “agentic AI” was still a buzzword. The mid-year review was the first time procurement teams were making decisions with visibility into what these systems could actually do. Some watched their own internal teams spin up prototypes replicating six-figure SaaS contracts in weeks.

企业财年大多与自然年一致，因此 2026 年的企业支出预算在 2025 年第四季度就已敲定——那时“代理式 AI”还只是个流行词。年中复盘是采购团队第一次在“看得见这些系统真实能做什么”的情况下作出决策。有些人眼睁睁看着自家内部团队在数周内就做出原型，足以替代那些六位数金额的 SaaS 合同。

That summer, we spoke with a procurement manager at a Fortune 500. He told us about one of his budget negotiations. The salesperson had expected to run the same playbook as last year: a 5% annual price increase, the standard “your team depends on us” pitch. The procurement manager told him he’d been in conversations with OpenAI about having their “forward deployed engineers” use AI tools to replace the vendor entirely. They renewed at a 30% discount. That was a good outcome, he said. The “long-tail of SaaS”, like Monday.com, Zapier and Asana, had it much worse.

那年夏天，我们与一位《财富》500 强的采购经理聊过。他向我们讲了他一次预算谈判的经历。销售原本打算照搬去年的剧本：每年 5% 的涨价，以及那套标准的“你们团队离不开我们”的说辞。采购经理告诉他：他正在与 OpenAI 沟通，让他们的“forward deployed engineers”用 AI 工具把这个供应商彻底替掉。最终续约价格打了 7 折（降价 30%）。他说，这已经算是好结果了。而 SaaS 的“长尾”，比如 Monday.com、Zapier 和 Asana，日子就更难过了。

Investors were prepared - expectant, even - that the long tail would be hit hard. They may have made up a third of spending for the typical enterprise stack, but they were obviously exposed. The systems of record, however, were supposed to be safe from disruption.

投资者早有准备——甚至可以说早有期待——长尾会被重击。它们或许占典型企业软件栈支出的三分之一，但暴露度显而易见。然而，所谓“记录系统（systems of record）”本应是最不易被颠覆的。

It wasn’t until ServiceNow’s Q3 26 report that the mechanism of reflexivity became clearer.

直到 ServiceNow 2026 年第三季度的财报，反身性机制才变得更清晰。

SERVICENOW 新增 ACV 净增速从 23% 放缓至 14%；宣布裁员 15% 并推出“结构性效率计划”；股价下跌 18% | Bloomberg，2026 年 10 月

SERVICENOW NET NEW ACV GROWTH DECELERATES TO 14% FROM 23%; ANNOUNCES 15% WORKFORCE REDUCTION AND ‘STRUCTURAL EFFICIENCY PROGRAM’; SHARES FALL 18% | Bloomberg, October 2026

SaaS 并没有“死”。把系统自建并自行运维，仍然需要做成本收益分析。但自建 已经 是一个选项，而这会进入定价谈判的筹码。更重要的是，竞争格局变了。AI 让开发并上线新功能更容易，差异化因此崩塌。既有厂商在价格上陷入“向下竞赛”——既要与彼此厮杀，也要与一批冒出来的新锐挑战者短兵相接。新挑战者因智能体编程能力的跃升而胆子更大，又没有旧成本结构需要保护，于是凶猛地抢占份额。

也直到这次财报披露，大家才真正意识到这些系统互相连结的程度。ServiceNow 卖的是席位。财富 500 强客户裁掉 15% 员工时，他们也会取消 15% 的许可。同一股由 AI 驱动的人力缩减，在客户那边提升了利润率，却在机制上直接摧毁了 ServiceNow 自己的收入基础。

SaaS wasn’t “dead”. There was still a cost-benefit-analysis to running and supporting in-house builds. But in-house was an option, and that factored into pricing negotiations. Perhaps more importantly, the competitive landscape had changed. AI had made it easier to develop and ship new features, so differentiation collapsed. Incumbents were in a race to the bottom on pricing - a knife-fight with both each other and with the new crop of upstart challengers that popped up. Emboldened by the leap in agentic coding capabilities and with no legacy cost structure to protect, these aggressively took share.

一家卖工作流自动化的公司，被更好的工作流自动化所颠覆；而它的应对方式，是裁员并把省下的钱投入到正颠覆它的那项技术上。

The interconnected nature of these systems weren’t fully appreciated until this print, either. ServiceNow sold seats. When Fortune 500 clients cut 15% of their workforce, they cancelled 15% of their licenses. The same AI-driven headcount reductions that were boosting margins at their customers were mechanically destroying their own revenue base.

他们还能怎么办？原地不动，慢慢等死？ 最受 AI 威胁的公司，反而成了最激进拥抱 AI 的公司。**

The company that sold workflow automation was being disrupted by better workflow automation, and its response was to cut headcount and use the savings to fund the very technology disrupting it.

事后看这很显然，但在当时（至少对我而言）真的不是。传统的颠覆模型说：既有巨头会抵制新技术，随后被灵活的新进入者抢走份额，慢慢死去。柯达、百视达、黑莓都是这样。可 2026 年发生的不同：既有巨头不是不想抵制，而是抵制不起。

What else were they supposed to do? Sit still and die slower? The companies most threatened by AI became AI’s most aggressive adopters.**

当股价跌了 40%–60%，董事会又逼着要答案，那些被 AI 威胁的公司只能做唯一能做的事：裁员，把省下来的钱投向 AI 工具，再用这些工具以更低成本维持产出。

This sounds obvious in hindsight, but it really wasn’t at the time (at least to me). The historical disruption model said incumbents resist new technology, they lose share to nimble entrants and die slowly. That’s what happened to Kodak, to Blockbuster, to BlackBerry. What happened in 2026 was different; the incumbents didn’t resist because they couldn’t afford to.

单看每家公司的应对都理性；合在一起却是灾难。每省下的一美元人力成本，都流向了让下一轮裁员成为可能的 AI 能力。

With stocks down 40-60% and boards demanding answers, the AI-threatened companies did the only thing they could. Cut headcount, redeploy the savings into AI tools, use those tools to maintain output with lower costs.

软件只是开场。投资者在争论 SaaS 倍数是否已经见底时忽略了一点：反身性回路早已逃离软件行业。同样的逻辑，适用于每一家以白领成本结构为主的公司。

Each company’s individual response was rational. The collective result was catastrophic. Every dollar saved on headcount flowed into AI capability that made the next round of job cuts possible.

Software was only the opening act. What investors missed while they debated whether SaaS multiples had bottomed was that the reflexive loop had already escaped the software sector. The same logic that justified ServiceNow cutting headcount applied to every company with a white-collar cost structure.

当摩擦归零

到 2027 年初，使用 LLM 已成为默认选项。人们在使用 AI 智能体，却甚至不知道“AI 智能体”是什么——就像很多从未搞懂“云计算”的人也照样在用流媒体服务。在他们心里，这就像自动补全或拼写检查一样——只是手机现在“自带”的功能。

When Friction Went to Zero

Qwen 开源的智能体购物助手成为 AI 介入消费者决策的催化剂。几周之内，所有主流 AI 助手都整合了某种智能体电商功能。蒸馏模型让这些智能体可以在手机和笔记本上运行，不再只依赖云端实例，从而显著降低了推理的边际成本。

By early 2027, LLM usage had become default. People were using AI agents who didn’t even know what an AI agent was, in the same way people who never learned what “cloud computing” was used streaming services. They thought of it the same way they thought of autocomplete or spell-check - a thing their phone just did now.

真正该让投资者更不安的，是这些智能体并不等你开口。它们会按照用户偏好在后台运行。商业不再是一连串离散的人类决策，而变成了一个持续优化的过程，24/7 为每一个联网的消费者代劳。到 2027 年 3 月，美国人的日均 token 消耗中位数已达到 400,000——比 2026 年底高出 10 倍。

Qwen’s open-source agentic shopper was the catalyst for AI handling consumer decisions. Within weeks, every major AI assistant had integrated some agentic commerce feature. Distilled models meant these agents could run on phones and laptops, not just cloud instances, reducing the marginal cost of inference significantly.

链条的下一环已经开始断裂。

The part that should have unsettled investors more than it did was that these agents didn’t wait to be asked. They ran in the background according to the user’s preferences. Commerce stopped being a series of discrete human decisions and became a continuous optimization process, running 24/7 on behalf of every connected consumer. By March 2027, the median individual in the United States was consuming 400,000 tokens per day - 10x since the end of 2026.

中介。

The next link in the chain was already breaking.

过去五十年里，美国经济在人的局限之上叠起了一层巨大的“抽租”结构：做事要花时间，耐心会耗尽，品牌熟悉度替代了勤勉比较，而大多数人为了少点几下，愿意接受一个更差的价格。数万亿美元的企业价值，都建立在这些约束会持续存在的前提上。

Intermediation.

一开始很简单：智能体把摩擦抹平了。

Over the past fifty years, the U.S. economy built a giant rent-extraction layer on top of human limitations: things take time, patience runs out, brand familiarity substitutes for diligence, and most people are willing to accept a bad price to avoid more clicks. Trillions of dollars of enterprise value depended on those constraints persisting.

那种即便几个月不用也会自动续费的订阅和会员；试用期过后悄悄翻倍的“首发价”。这些都被重新定义成一种“被挟持”的局面——而智能体可以替你谈判。平均客户终身价值（LTV）——订阅经济赖以建立的核心指标——明显下滑。

It started out simple enough. Agents removed friction.

消费者智能体开始改写几乎所有消费交易的运作方式。

Subscriptions and memberships that passively renewed despite months of disuse. Introductory pricing that sneakily doubled after the trial period. Each one was rebranded as a hostage situation that agents could negotiate. The average customer lifetime value, the metric the entire subscription economy was built on, distinctly declined.

人在买一盒蛋白棒之前，没时间在五个平台逐个比价；机器有。

Consumer agents began to change how nearly all consumer transactions worked.

旅游预订平台最先中枪，因为它们最简单。到 2026 年第四季度，我们的智能体就能比任何平台更快、更便宜地组装完整行程（机票、酒店、地面交通、会员积分优化、预算约束、退款）。

Humans don’t really have the time to price-match across five competing platforms before buying a box of protein bars. Machines do.

保险续保——其续保模式完全依赖投保人的惰性——也被改写。每年替你重新比价投保的智能体，拆掉了保险公司从“被动续保”中赚取的那 15%–20% 保费。

Travel booking platforms were an early casualty, because they were the simplest. By Q4 2026, our agents could assemble a complete itinerary (flights, hotels, ground transport, loyalty optimization, budget constraints, refunds) faster and cheaper than any platform.

理财建议。报税。常规法律工作。任何一个服务商最终卖点是“我来替你穿越那些你觉得烦的复杂性”的行业，都被颠覆了，因为智能体不觉得烦。

Insurance renewals, where the entire renewal model depended on policyholder inertia, were reformed. Agents that re-shop your coverage annually dismantled the 15-20% of premiums that insurers earned from passive renewals.

连那些我们以为“人际关系价值”能形成护城河的领域，也脆得出人意料。房地产就是例子：由于经纪人与消费者之间的信息不对称，买家几十年来一直忍受 5%–6% 的佣金；但当拥有 MLS 访问权限和几十年成交数据的 AI 智能体能够瞬间复刻知识体系后，这套体系就崩了。一篇 2027 年 3 月的卖方研报把它称为“agent on agent violence”。在主要大都市区，买方经纪佣金中位数从 2.5%–3% 压缩到 1% 以下，越来越多交易甚至完全没有人类买方经纪参与就成交了。

Financial advice. Tax prep. Routine legal work. Any category where the service provider’s value proposition was ultimately “I will navigate complexity that you find tedious” was disrupted, as the agents found nothing tedious.

我们高估了“人际关系”的价值。结果发现，人们口中的很多“关系”，不过是披着友好面孔的摩擦。

Even places we thought insulated by the value of human relationships proved fragile. Real estate, where buyers had tolerated 5-6% commissions for decades because of information asymmetry between agent and consumer, crumbled once AI agents equipped with MLS access and decades of transaction data could replicate the knowledge base instantly. A sell-side piece from March 2027 titled it “agent on agent violence”. The median buy-side commission in major metros had compressed from 2.5-3% to under 1%, and a growing share of transactions were closing with no human agent on the buy side at all.

这只是对中介层冲击的开始。那些成功公司花了数十亿美元，去高效利用消费者行为与人类心理的各种“怪癖”——而这些怪癖突然不再重要。

We had overestimated the value of “human relationships”. Turns out that a lot of what people called relationships was simply friction with a friendly face.

为价格与匹配度做优化的机器，并不在乎你最爱用的 app，或你过去四年习惯性打开的网站，也不会被精心设计的结账流程所“吸引”。它们不会累到选择最省事的那一个，也不会默认“我一直都在这家下单”。

That was just the start of the disruption for the intermediation layer. Successful companies had spent billions to effectively exploit quirks of consumer behavior and human psychology that didn’t matter anymore.

这毁掉了一类特殊的护城河：习惯性中介。

Machines optimizing for price and fit do not care about your favorite app or the websites you’ve been habitually opening for the last four years, nor feel the pull of a well-designed checkout experience. They don’t get tired and accept the easiest option or default to “I always just order from here”.

DoorDash（DASH US）是典型代表。

That destroyed a particular kind of moat: habitual intermediation.

编程智能体把上线一个外卖 app 的门槛彻底压平。一个合格的开发者可以在数周内部署一个可用的竞品，确实也有几十家这么干：它们把配送费的 90%–95% 直接让给司机，从而把司机从 DoorDash 和 Uber Eats 那里挖走。多平台仪表盘让零工同时跟踪二三十个平台的派单，消除了巨头赖以生存的锁定效应。市场一夜之间碎片化，利润率压缩到几乎为零。

DoorDash (DASH US) was the poster child.

智能体在破坏的两端都踩了油门：它们先帮助竞争者出现，然后又去使用这些竞争者。DoorDash 的护城河几乎就是一句话：“你饿了，你懒了，这个 app 就在你主屏幕上。”但智能体没有主屏幕。它会同时查询 DoorDash、Uber Eats、餐厅官网，以及二十个“氛围编程”新做出来的替代品，每次都选最低费用、最快送达。

Coding agents had collapsed the barrier to entry for launching a delivery app. A competent developer could deploy a functional competitor in weeks, and dozens did, enticing drivers away from DoorDash and Uber Eats by passing 90-95% of the delivery fee through to the driver. Multi-app dashboards let gig workers track incoming jobs from twenty or thirty platforms at once, eliminating the lock-in that the incumbents depended on. The market fragmented overnight and margins compressed to nearly nothing.

习惯性的 app 忠诚度——整个商业模式的根基——对机器而言根本不存在。

Agents accelerated both sides of the destruction. They enabled the competitors and then they used them. The DoorDash moat was literally “you’re hungry, you’re lazy, this is the app on your home screen.” An agent doesn’t have a home screen. It checks DoorDash, Uber Eats, the restaurant’s own site, and twenty new vibe-coded alternatives so it can pick the lowest fee and fastest delivery every time.

这件事有种奇妙的诗意：在整个故事里，这大概是唯一一次，智能体“帮了”那些即将被替代的白领一把。当他们最终沦为外卖司机时，至少他们赚的钱不会有一半被 Uber 和 DoorDash 抽走。当然，随着自动驾驶车辆普及，技术的这点“好意”也很快消失。

Habitual app loyalty, the entire basis of the business model, simply didn’t exist for a machine.

一旦智能体掌控了交易，它们就开始寻找更大的回形针。

This was oddly poetic, as perhaps the only example in this entire saga of agents doing a favor for the soon-to-be-displaced white collar workers. When they ended up as delivery drivers, at least half their earnings weren’t going to Uber and DoorDash. Of course, this favor from technology didn’t last for long as autonomous vehicles proliferated.

比价和聚合能做的事情毕竟有限。最大的、可反复为用户省钱的方式（尤其当智能体开始彼此交易时）是消除费用。在机器对机器的商业中，2%–3% 的卡组织交换费率成了显而易见的目标。

Once agents controlled the transaction, they went looking for bigger paperclips.

智能体开始寻找比卡更快、更便宜的选择。多数最终改用通过 Solana 或以太坊 L2 结算的稳定币：清算近乎即时，交易成本以美分的零头计。

There was only so much price-matching and aggregating to do. The biggest way to repeatedly save the user money (especially when agents started transacting among themselves) was to eliminate fees. In machine-to-machine commerce, the 2-3% card interchange rate became an obvious target.

MASTERCARD 2027 年 Q1：净营收同比 +6%；购买额增速从上季度同比 +5.9% 放缓至 +3.4%；管理层提及“由智能体驱动的价格优化”与“可选消费品类承压” | Bloomberg，2027 年 4 月 29 日

Agents went looking for faster and cheaper options than cards. Most settled on using stablecoins via Solana or Ethereum L2s, where settlement was near-instant and the transaction cost was measured in fractions of a penny.

万事达卡 2027 年 Q1 的财报，是无法回头的拐点。智能体电商从“产品故事”变成了“管道故事”。MA 次日下跌 9%。Visa 也跌了，但在分析师指出其在稳定币基础设施上的布局更强后，跌幅有所收窄。

[

MASTERCARD Q1 2027: NET REVENUES +6% Y/Y; PURCHASE VOLUME GROWTH SLOWS TO +3.4% Y/Y FROM +5.9% PRIOR QUARTER; MANAGEMENT NOTES “AGENT-LED PRICE OPTIMIZATION” AND “PRESSURE IN DISCRETIONARY CATEGORIES” | Bloomberg, April 29 2027

](https://substackcdn.com/image/fetch/$s_!gW0T!,f_auto,q_auto:good,fl_progressive:steep/https%3A%2F%2Fsubstack-post-media.s3.amazonaws.com%2Fpublic%2Fimages%2F70ccc118-1388-4602-a5bd-bbc4284208f6_1024x541.png)

Mastercard’s Q1 2027 report was the point of no return. Agentic commerce went from being a product story to a plumbing story. MA dropped 9% the following day. Visa did too, but pared losses after analysts pointed out its stronger positioning in stablecoin infrastructure.

智能体电商绕开交换费，对以信用卡为核心的银行和单一业务发卡机构构成了更大风险——它们拿走了那 2%–3% 费用的大头，并围绕由商户补贴供养的积分奖励计划构建了整个业务板块。

[

美国运通（AXP US）受打击最重：一方面白领裁员掏空其客户基础，另一方面智能体绕开交换费又掏空其收入模型。接下来几周，Synchrony（SYF US）、Capital One（COF US）和 Discover（DFS US）也都下跌超过 10%。

它们的护城河由摩擦铸成。而 摩擦正在归零。

](https://substackcdn.com/image/fetch/$s_!gW0T!,f_auto,q_auto:good,fl_progressive:steep/https%3A%2F%2Fsubstack-post-media.s3.amazonaws.com%2Fpublic%2Fimages%2F70ccc118-1388-4602-a5bd-bbc4284208f6_1024x541.png)

Agentic commerce routing around interchange posed a far greater risk to card-focused banks and mono-line issuers, who collected the majority of that 2-3% fee and had built entire business segments around rewards programs funded by the merchant subsidy.

从行业风险到系统性风险

American Express (AXP US) was hit hardest; a combined headwind from white-collar workforce reductions gutting its customer base and agents routing around interchange gutting its revenue model. Synchrony (SYF US), Capital One (COF US) and Discover (DFS US) all fell more than 10% over the following weeks, as well.

在 2026 年整个一年里，市场把 AI 的负面影响当作“行业故事”。软件和咨询被碾压，支付和其他收费“关卡”摇摇欲坠，但更广泛的经济看起来仍然没事。劳动力市场虽在转弱，却没有自由落体。共识是：创造性破坏本就是任何技术创新周期的一部分。局部会很痛，但 AI 带来的总体净收益终将盖过负面。

Their moats were made of friction. And friction was going to zero.

我们在 2027 年 1 月的宏观备忘录中指出，这个心智模型是错的。美国经济是白领服务经济。白领占就业的 50%，驱动了约 75% 的可选消费支出。AI 正在吞噬的企业与岗位并非美国经济的边缘地带——它们 就是 美国经济本身。

“技术创新会摧毁工作，然后创造更多工作”。这是当时最流行、也最有说服力的反驳。它之所以流行、之所以有说服力，是因为两百年来它一直是对的。即便我们想象不出未来的新工作是什么，它们也终会出现。

From Sector Risk to Systemic Risk

ATM 让网点运营更便宜，于是银行开了更多网点，柜员岗位在随后的二十年里反而增加。互联网颠覆了旅行社、黄页、实体零售，但也在它们的废墟上发明了全新的产业，从而“变”出新的工作。

Through 2026, markets treated negative AI impact as a sector story. Software and consulting were getting crushed, payments and other toll booths were wobbly, but the broader economy seemed fine. The labor market, while softening, was not in freefall. The consensus view was that creative destruction was part of any technological innovation cycle. It would be painful in pockets, but the overall net positives from AI would outweigh any negatives.

但每一个新工作，都需要人来做。

Our January 2027 macro memo argued this was the wrong mental model. The US economy is a white-collar services economy. White-collar workers represented 50% of employment and drove roughly 75% of discretionary consumer spending. The businesses and jobs that AI was chewing up were not tangential to the US economy, they were the US economy.

而如今的 AI 是一种通用智能，它在那些人类本想转岗去做的任务上也会持续变强。被替代的程序员无法简单转去做“AI 管理”，因为 AI 已经能胜任。

“Technological innovation destroys jobs and then creates even more”. This was the most popular and convincing counter-argument at the time. It was popular and convincing because it’d been right for two centuries. Even if we couldn’t conceive of what the future jobs would be, they would surely arrive.

今天，AI 智能体已经能完成原本需要数周的研发任务。指数曲线碾平了我们对“可能性”的想象，哪怕沃顿的教授们年年试图把数据拟合成新的 S 形曲线。

ATMs made branches cheaper to operate so banks opened more of them and teller employment rose for the next twenty years. The internet disrupted travel agencies, the Yellow Pages, brick-and-mortar retail, but it invented entirely new industries in their place that conjured new jobs.

[

Every new job, however, required a human to perform it.

AI is now a general intelligence that improves at the very tasks humans would redeploy to. Displaced coders cannot simply move to “AI management” because AI is already capable of that.

](https://substackcdn.com/image/fetch/$s_!kA4x!,f_auto,q_auto:good,fl_progressive:steep/https%3A%2F%2Fsubstack-post-media.s3.amazonaws.com%2Fpublic%2Fimages%2F0bc01ebb-a883-4e5c-bd2b-fa7aaa872edb_1600x1059.png)

Today, AI agents handle many-weeks-long research and development tasks. The exponential steamrolled our conceptions of what was possible, even though every year Wharton professors tried to fit the data to a new sigmoid.

它们几乎写了所有代码。表现最好的那一批，在几乎所有事情上都显著比几乎所有人更聪明。而且它们还在不断变便宜。

[

AI 确实 创造了新工作。提示词工程师。AI 安全研究员。基础设施技术员。人类仍在闭环之中，在最高层面做协调，或为“品味”把关。但 AI 每创造一个新角色，就会让几十个旧角色变得多余。这些新角色的薪酬，只是旧岗位的一小部分。

美国 JOLTS：职位空缺跌破 5.5M；失业人数/空缺岗位比升至约 1.7，为 2020 年 8 月以来最高 | Bloomberg，2026 年 10 月

](https://substackcdn.com/image/fetch/$s_!kA4x!,f_auto,q_auto:good,fl_progressive:steep/https%3A%2F%2Fsubstack-post-media.s3.amazonaws.com%2Fpublic%2Fimages%2F0bc01ebb-a883-4e5c-bd2b-fa7aaa872edb_1600x1059.png)

全年招聘率一直乏力，但 2026 年 10 月的 JOLTS 读数给出了更确定的数据：职位空缺跌破 5.5 million，同比下降 15%。

They write essentially all code. The highest performing of them are substantially smarter than almost all humans at almost all things. And they keep getting cheaper.

Indeed：随着“生产率计划”扩散，软件、金融、咨询岗位发布量大幅下滑 | Indeed Hiring Lab，2026 年 11–12 月

AI has created new jobs. Prompt engineers. AI safety researchers. Infrastructure technicians. Humans are still in the loop, coordinating at the highest level or directing for taste. For every new role AI created, though, it rendered dozens obsolete. The new roles paid a fraction of what the old ones did.

白领岗位空缺在崩塌，而蓝领岗位空缺相对稳定（建筑、医疗、技工）。波动集中在那些写备忘录的工作上（不知为何我们还没失业）、审批预算的工作上，以及维持经济中间层顺畅运转的工作上。但两类人群的实际工资增速在全年大部分时间里都为负，且仍在继续下滑。

股票市场对 JOLTS 的在意程度仍不及另一条消息：GE Vernova 的涡轮机产能已被预订到 2040 年。它在“负面宏观数据”和“正面 AI 基础设施新闻”的拉锯中横盘徘徊。

U.S. JOLTS: JOB OPENINGS FALL BELOW 5.5M; UNEMPLOYED-TO-OPENINGS RATIO CLIMBS TO ~1.7, HIGHEST SINCE AUG 2020 | Bloomberg, Oct 2026

然而，债券市场（总是比股市更聪明，至少没那么浪漫）开始为消费端的打击定价。10 年期收益率在接下来的四个月里从 4.3% 下行到 3.2%。但总体失业率并未“爆表”，一些人仍没看懂结构性细节。

在正常的衰退中，成因最终会自我纠偏。过度建设导致施工放缓，继而带来利率走低，再进而带来新的建设。库存过冲导致去库存，去库存又带来补库存。周期机制本身就包含复苏的种子。

The hiring rate had been anemic all year, but October ‘26 JOLTS print provided some definitive data. Job openings fell below 5.5 million, a 15% decline YoY.

但这一次的成因不是周期性的。

[

INDEED: POSTINGS FALL SHARPLY IN SOFTWARE, FINANCE, CONSULTING AS “PRODUCTIVITY INITIATIVES” SPREAD | Indeed Hiring Lab, Nov–Dec 2026

](https://substackcdn.com/image/fetch/$s_!0WIw!,f_auto,q_auto:good,fl_progressive:steep/https%3A%2F%2Fsubstack-post-media.s3.amazonaws.com%2Fpublic%2Fimages%2F1d84258d-d811-4963-8914-16b104e6744f_1024x541.png)

White-collar openings were collapsing while blue-collar openings remained relatively stable (construction, healthcare, trades). The churn was in the jobs that write memos (we are, somehow, still in business), approve budgets, and keep the middle layers of the economy lubricated. Real wage growth in both cohorts, however, had been negative for the majority of the year and kept declining.

AI 变得更强、更便宜。公司裁员，然后用省下的钱买更多 AI 能力，从而可以裁掉更多人。被替代的工人花得更少。向消费者卖东西的公司卖得更少、体质变弱，又为了保住利润率而加大 AI 投入。AI 再次变得更强、更便宜。

The equity market still cared less about JOLTS than it did the news that all of GE Vernova’s turbine capacity was now sold out until 2040, it ambled sideways in a tug of war between negative macro news with positive AI infrastructure headlines.

一个没有天然刹车的反馈回路。

The bond market (always smarter than equities, or at least less romantic) began pricing the consumption hit, however. The 10-year yield began a descent from 4.3% to 3.2% over the following four months. Still, the headline unemployment rate did not blow out, the composition nuance was still lost on some.

直觉上的预期是：总需求下滑会拖慢 AI 的建设。但并没有，因为这不是超大规模云厂商那种资本开支（CapEx）。这是一种运营开支（OpEx）的替代。一家原本每年在员工上花 $100M、在 AI 上花 $5M 的公司，现在变成在员工上花 $70M、在 AI 上花 $20M。AI 投入成倍增长，但它是以总运营成本下降的形式发生的。每家公司 AI 预算都在增长，而总体支出却在收缩。

In a normal recession, the cause eventually self-corrects. Overbuilding leads to a construction slowdown, which leads to lower rates, which leads to new construction. Inventory overshoot leads to destocking, which leads to restocking. The cyclical mechanism contains within it its own seeds of recovery.

讽刺的是：当被它颠覆的经济开始恶化时，AI 基础设施链条却仍在表现强劲。NVDA 依旧交出创纪录的营收。TSM 的产能利用率仍在 95% 以上。超大规模云厂商每季度仍在数据中心资本开支上投入 $150-200 billion。对这一趋势呈纯凸性敞口的经济体，例如台湾和韩国，大幅跑赢。

This cycle’s cause was not cyclical.

印度则恰好相反。该国 IT 服务业每年出口超过 $200 billion，是印度经常账户顺差的最大单一贡献者，也是为其长期货物贸易逆差提供资金的对冲。整个模式建立在一个价值主张上：印度开发者成本仅为美国同行的一小部分。但 AI 编程智能体的边际成本几乎崩到只剩电费。TCS、Infosys 和 Wipro 在 2027 年合同取消加速。随着支撑印度外部账户的服务贸易顺差蒸发，卢比在四个月内对美元贬值 18%。到 2028 年第一季度，IMF 已开始与新德里进行“初步磋商”。

[

推动颠覆发生的引擎每个季度都变得更强，这意味着颠覆每个季度都在加速。劳动力市场没有天然的底部。

在美国，我们不再问“AI 基础设施泡沫会如何破裂”。我们问的是：当消费者被机器取代时，一个以消费信贷为核心的经济会怎样？

](https://substackcdn.com/image/fetch/$s_!0WIw!,f_auto,q_auto:good,fl_progressive:steep/https%3A%2F%2Fsubstack-post-media.s3.amazonaws.com%2Fpublic%2Fimages%2F1d84258d-d811-4963-8914-16b104e6744f_1024x541.png)

AI got better and cheaper. Companies laid off workers, then used the savings to buy more AI capability, which let them lay off more workers. Displaced workers spent less. Companies that sell things to consumers sold fewer of them, weakened, and invested more in AI to protect margins. AI got better and cheaper.

智能替代螺旋

A feedback loop with no natural brake.

2027 年，宏观故事不再隐晦。过去 12 个月里那些零散却明显偏负面的发展，其传导机制变得一目了然。你不必钻进 BLS 数据里。去参加一次朋友的晚宴就够了。

The intuitive expectation was that falling aggregate demand would slow the AI buildout. It didn’t, because this wasn’t hyperscaler-style CapEx. It was OpEx substitution. A company that had been spending $100M a year on employees and $5M on AI now spent $70M on employees and $20M on AI. AI investment increased by multiples, but it occurred as a reduction in total operating costs. Every company’s AI budget grew while its overall spending shrank.

被替代的白领没有闲着。他们选择降档。许多人转去做更低薪的服务业和零工经济工作，这增加了这些领域的劳动力供给，也把那里的工资一起压低。

The irony of this was that the AI infrastructure complex kept performing even as the economy it was disrupting began deteriorating. NVDA was still posting record revenues. TSM was still running at 95%+ utilization. The hyperscalers were still spending $150-200 billion per quarter on data center capex. Economies that were purely convex to this trend, like Taiwan and Korea, outperformed massively.

我们的一位朋友在 2025 年还是 Salesforce 的高级产品经理：头衔、医保、401(k)、年薪 $180,000。她在第三轮裁员中丢了工作。找了六个月无果后，她开始给 Uber 开车。她的收入降到 $45,000。重点不在这个个案，而在二阶的数学。把这种动态乘以分布在各大都市圈的几十万名工人。过度资质的劳动力涌入服务业与零工市场，压低了那些本就艰难维持的既有工人的工资。行业性冲击转移扩散，演变成全经济范围的工资压缩。

India was the inverse. The country’s IT services sector exported over $200 billion annually, the single largest contributor to India’s current account surplus and the offset that financed its persistent goods trade deficit. The entire model was built on one value proposition: Indian developers cost a fraction of their American counterparts. But the marginal cost of an AI coding agent had collapsed to, essentially, the cost of electricity. TCS, Infosys and Wipro saw contract cancellations accelerate through 2027. The rupee fell 18% against the dollar in four months as the services surplus that had anchored India’s external accounts evaporated. By Q1 2028, the IMF had begun “preliminary discussions” with New Delhi.

[

The engine that caused the disruption got better every quarter, which meant the disruption accelerated every quarter. There was no natural floor to the labor market.

In the US, we weren’t asking about how the bubble would burst in AI infrastructure anymore. We were asking what happens to a consumer-credit economy when consumers are being replaced with machines.

](https://substackcdn.com/image/fetch/$s_!CLd4!,f_auto,q_auto:good,fl_progressive:steep/https%3A%2F%2Fsubstack-post-media.s3.amazonaws.com%2Fpublic%2Fimages%2F8f6a0cb1-86ff-4b99-811c-1652d5577fba_1024x541.jpeg)

而那部分仍以人为中心的岗位池，还将在我们写下这些文字之时迎来新一轮修正——因为自动化配送与自动驾驶车辆正逐步渗透进吸纳了第一波被替代者的零工经济。

The Intelligence Displacement Spiral

到 2027 年 2 月，很明显：仍然在职的专业人士开始像“自己可能就是下一个”那样花钱。他们为了不被裁掉而加倍工作（主要靠 AI 帮忙），升职加薪的希望已经消失。储蓄率上行，消费转弱。

2027 was when the macroeconomic story stopped being subtle. The transmission mechanism from the previous twelve months of disjointed but clearly negative developments became obvious. You didn’t need to go into the BLS data. Just attend a dinner party with friends.

最危险的是滞后。高收入者用高于平均水平的储蓄，让生活在两三季度里看起来依然“正常”。硬数据直到问题在真实经济里早已不是新闻时才姗姗来迟地确认。随后，击碎幻象的那次数据公布出现了。

Displaced white-collar workers did not sit idle. They downshifted. Many took lower-paying service sector and gig economy jobs, which increased labor supply in those segments and compressed wages there too.

美国初请失业金人数飙升至 487,000，为 2020 年 4 月以来最高；Department of Labor，2027 年 Q3

A friend of ours was a senior product manager at Salesforce in 2025. Title, health insurance, 401k, $180,000 a year. She lost her job in the third round of layoffs. After six months of searching, she started driving for Uber. Her earnings dropped to $45,000. The point is less the individual story and more the second-order math. Multiply this dynamic by a few hundred thousand workers across every major metro. Overqualified labor flooding the service and gig economy pushed down wages for existing workers who were already struggling. Sector-specific disruption metastasized into economy-wide wage compression.

初请人数飙升至 487,000，为 2020 年 4 月以来最高。ADP 与 Equifax 证实，新申请者绝大多数来自白领专业人士。

[

随后一周标普下跌 6%。负面宏观开始在拉锯战中占上风。

在正常衰退中，失业会更均匀地分布。蓝领与白领大致按各自就业占比共同承受痛苦。消费冲击也更分散，而且更快体现在数据里，因为低收入者的边际消费倾向更高。

](https://substackcdn.com/image/fetch/$s_!CLd4!,f_auto,q_auto:good,fl_progressive:steep/https%3A%2F%2Fsubstack-post-media.s3.amazonaws.com%2Fpublic%2Fimages%2F8f6a0cb1-86ff-4b99-811c-1652d5577fba_1024x541.jpeg)

但这一轮，失业集中在收入分布的上几个十分位。他们在总就业中占比不大，却驱动了极不成比例的消费支出。美国收入最高的 10% 人群贡献了超过 50% 的全部消费支出；最高的 20% 贡献约 65%。买房、买车、度假、餐馆消费、私立学校学费、家庭装修——主要都是这些人在买。他们是整个可选消费经济的需求底盘。

The pool of remaining human-centric had another correction ahead of it, happening while we write this. As autonomous delivery and self-driving vehicles work their way through the gig economy that absorbed the first wave of displaced workers.

当这些人失业，或为了有岗可上而接受 50% 的降薪时，消费受到的打击与“失去的岗位数量”相比巨大得多。白领就业下降 2%，大致会对应 3%–4% 的可选消费支出打击。与蓝领失业往往立刻冲击（你从工厂被裁，下周就得停掉开销）不同，白领失业的影响更滞后却更深：他们有储蓄缓冲，可以在行为转变真正发生之前把支出维持几个月。

By February 2027, it was clear that still employed professionals were spending like they might be next. They were working twice as hard (mostly with the help of AI) just to not get fired, hopes of promotion or raises were gone. Savings rates ticked higher and spending softened.

到 2027 年第二季度，经济已进入衰退。NBER 要到几个月后才会“官方”定调（他们一向如此），但数据毫不含糊——我们已经连续两个季度实际 GDP 负增长。但这还不是“金融危机”……至少暂时还不是。

The most dangerous part was the lag. High earners used their higher-than-average savings to maintain the appearance of normalcy for two or three quarters. The hard data didn’t confirm the problem until it was already old news in the real economy. Then came the print that broke the illusion.

相关押注的雏菊链

U.S. INITIAL JOBLESS CLAIMS SURGE TO 487,000, HIGHEST SINCE APRIL 2020; Department of Labor, Q3 2027

私人信贷（private credit）从 2015 年的不足 $1 trillion 增长到 2026 年的超过 $2.5 trillion。其中相当一部分资金投向了软件与科技交易，很多是对 SaaS 公司的杠杆收购，估值假设收入永远保持十几个百分点的增速。

这些假设在第一次智能体编程演示与 2026 年第一季度的软件崩盘之间就已经死去，但账面估值似乎并没意识到它们已死。

Initial claims surged to 487,000, the highest since April 2020. ADP and Equifax confirmed that the overwhelming majority of new filings were from white-collar professionals.

当许多上市 SaaS 公司交易在 5–8 倍 EBITDA 时，PE 旗下的软件公司仍以反映收购时估值的账面价格躺在资产负债表上——那是基于已不复存在的收入倍数。管理人缓慢下调估值：100 分、92、85……而上市可比公司却在告诉你应该是 50。

The S&P dropped 6% over the following week. Negative macro started winning the tug of war.

MOODY’S 下调 14 家发行人合计 $18B 的 PE 支持软件债务评级，理由是“AI 驱动的竞争性颠覆带来的结构性收入逆风”；为自 2015 年能源行业以来最大单一行业行动 | Moody’s Investors Service，2027 年 4 月

In a normal recession, job losses are broadly distributed. Blue-collar and white-collar workers share the pain roughly in proportion to each segment’s share of employment. The consumption hit is also broadly distributed, and it shows up quickly in the data because lower-income workers have higher marginal propensities to consume.

所有人都记得评级下调之后发生了什么。行业老兵早在 2015 年能源评级下调后就见过这套剧本。

In this cycle, the job losses have been concentrated in the upper deciles of the income distribution. They are a relatively small share of total employment, but they drive a wildly disproportionate share of consumer spending. The top 10% of earners account for more than 50% of all consumer spending in the United States. The top 20% account for roughly 65%. These are the people who buy the houses, the cars, the vacations, the restaurant meals, the private school tuition, the home renovations. They are the demand base for the entire consumer discretionary economy.

2027 年第三季度，软件支持的贷款开始违约。信息服务与咨询领域的 PE 组合公司紧随其后。多笔规模数十亿美元的知名 SaaS 杠杆收购进入重组。

When these workers lost their jobs, or took 50% pay cuts to move into available roles, the consumption hit was enormous relative to the number of jobs lost. A 2% decline in white-collar employment translated to something like a 3-4% hit to discretionary consumer spending. Unlike blue-collar job losses, which tend to hit immediately (you get laid off from the factory, you stop spending next week), white-collar job losses have a lagged but deeper impact because these workers have savings buffers that allow them to maintain spending for a few months before the behavioral shift kicks in.

Zendesk 是那把冒烟的枪。

By Q2 2027, the economy was in recession. The NBER would not officially date the start until months later (they never do) but the data was unambiguous - we’d had two consecutive quarters of negative real GDP growth. But it wasn’t a “financial crisis”…yet.

ZENDESK 因 AI 驱动的客服自动化侵蚀 ARR 而触发债务契约违约；$5B 直贷额度被标至 58 美分；创纪录的最大私人信贷软件违约 | Financial Times，2027 年 9 月

2022 年，Hellman & Friedman 与 Permira 以 $10.2 billion 将 Zendesk 私有化。债务包为 $5 billion 的直贷，是当时历史上最大的 ARR 支持融资，由 Blackstone 牵头，Apollo、Blue Owl 与 HPS 同属放贷团。贷款结构明确基于一个假设：Zendesk 的年化经常性收入会继续经常性。在约 25 倍 EBITDA 的水平上，杠杆只有在这一假设成立时才说得通。

The Daisy Chain of Correlated Bets

到 2027 年年中，它不成立了。

Private credit had grown from under $1 trillion in 2015 to over $2.5 trillion by 2026. A meaningful share of that capital had been deployed into software and technology deals, many of them leveraged buyouts of SaaS companies at valuations that assumed mid-teens revenue growth in perpetuity.

AI 智能体自动处理客服，已经持续了将近一年。Zendesk 所定义的品类（工单、分流、管理人类客服互动）早已被替代：新系统无需生成工单就能直接解决问题。承销所依赖的年化经常性收入不再“经常”，它只是不知道何时会流失的收入。

Those assumptions died somewhere between the first agentic coding demo and the Q1 2026 software crash, but the marks didn’t seem to realize they were dead.

历史上最大的 ARR 支持贷款，变成了历史上最大的私人信贷软件违约。每个信贷交易台同时问出同一个问题：还有谁把结构性逆风伪装成周期性逆风？

As many public SaaS companies traded to 5-8x EBITDA, PE-backed software companies sat on balance sheets at marks reflecting acquisition valuations on multiples of revenue that didn’t exist anymore. Managers eased the marks down gradually, 100 cents, 92, 85, all while public comps said 50.

但至少一开始，市场共识有一点是对的：这本该是“能扛过去”的。

私人信贷不是 2008 年的银行业。它的架构就是为了避免被迫抛售而设计的。它们是封闭式载体，资金被锁定；LP 的承诺期限通常是 7–10 年。没有存款人挤兑，也没有回购融资额度被抽走。管理人可以持有受损资产，慢慢处置，等待回收。痛苦，但可控。系统本应是“弯而不折”。

MOODY’S DOWNGRADES $18B OF PE-BACKED SOFTWARE DEBT ACROSS 14 ISSUERS, CITING ‘SECULAR REVENUE HEADWINDS FROM AI-DRIVEN COMPETITIVE DISRUPTION’; LARGEST SINGLE-SECTOR ACTION SINCE ENERGY IN 2015 | Moody’s Investors Service, April 2027

Blackstone、KKR、Apollo 的管理层都提到软件敞口占资产的 7%–13%。可控。每一份卖方报告、每一个 fintwit 的信贷账号都在重复同一句话：私人信贷拥有永久资本。它们能消化损失，而这些损失若发生在高杠杆银行身上可能会爆炸。

永久资本。这个词出现在每一次用来安抚市场的电话会和投资者信里，成了一句咒语。但像大多数咒语一样，没人注意其中的细节。它真正意味着什么……

Everyone remembers what happened after the downgrade. Industry veterans had already seen the playbook following the 2015 energy downgrades.

在此前十年里，大型另类资产管理机构收购了人寿保险公司，并把它们变成了融资载体。Apollo 买下 Athene。Brookfield 买下 American Equity。KKR 收购 Global Atlantic。逻辑很优雅：年金保费存入提供了稳定、久期很长的负债基础。管理人把这些存入资金投向自己发起的私人信贷，于是两头收费：保险端赚利差，资管端收管理费。一个“费上加费”的永动机——在一个条件下运转得无比顺畅。

Software-backed loans began defaulting in Q3 2027. PE portfolio companies in information services and consulting followed. Several multi-billion dollar LBOs of well-known SaaS companies entered restructuring.

私人信贷必须“稳如现金”。

Zendesk was the smoking gun.

损失打在了这样的资产负债表上：用长期负债对冲、持有大量非流动资产。本应让系统更韧性的“永久资本”，并不是什么抽象的耐心机构资金池，也不是高深投资者在承担高深风险。它是美国普通家庭（“主街”）的储蓄，以年金形式被配置到同一批正在违约的 PE 支持软件与科技债券上。那笔“跑不掉”的锁定资金，是寿险保单持有人的钱——而这套游戏的规则就不一样了。

相较于银行监管，保险监管一直温和——甚至有些自满——但这一次是警钟。监管方本就对寿险公司的私人信贷集中度不安，如今开始下调这类资产的风险资本计提待遇。这迫使保险公司要么补充资本，要么出售资产——而在一个已开始冻结的市场里，两者都不可能以体面的条件完成。

ZENDESK MISSES DEBT COVENANTS AS AI-DRIVEN CUSTOMER SERVICE AUTOMATION ERODES ARR; $5B DIRECT LENDING FACILITY MARKED TO 58 CENTS; LARGEST PRIVATE CREDIT SOFTWARE DEFAULT ON RECORD | Financial Times, September 2027

纽约州、爱荷华州监管机构拟收紧寿险公司持有的部分私评信用资产的资本计提；预计 NAIC 指引将提高 RBC 系数并触发更多 SVO 审查 | Reuters，2027 年 11 月

当穆迪把 Athene 的财务实力评级展望调为负面时，Apollo 股价两天内下跌 22%。Brookfield、KKR 等也随之下挫。

In 2022, Hellman & Friedman and Permira had taken Zendesk private for $10.2 billion. The debt package was $5 billion in direct lending, the largest ARR-backed facility in history at the time, led by Blackstone with Apollo, Blue Owl and HPS all in the lending group. The loan was explicitly structured around the assumption that Zendesk’s annual recurring revenue would remain recurring. At roughly 25x EBITDA, the leverage only made sense if it did.

之后事情只会更复杂。这些公司不仅搭建了“保险永动机”，还建立了复杂的离岸架构，通过监管套利来最大化回报。美国本土保险公司签发年金，然后把风险转移给自己控股的百慕大或开曼再保险公司——那里的监管更灵活，可以对同样的资产计提更少资本。该关联方又通过离岸 SPV 募集外部资本，形成新的交易对手层级：它们与保险公司一起投资于由同一母公司资管部门发起的私人信贷。

By mid-2027, it didn’t.

[

AI agents had been handling customer service autonomously for the better part of a year. The category Zendesk had defined (ticketing, routing, managing human support interactions) was already replaced by systems that resolved issues without generating a ticket at all. The Annualized Recurring Revenue the loan was underwritten against was no longer recurring, it was just revenue that hadn’t left yet.

The largest ARR-backed loan in history became the largest private credit software default in history. Every credit desk asked the same question at once: who else has a secular headwind disguised as a cyclical one?

](https://substackcdn.com/image/fetch/$s_!yyoF!,f_auto,q_auto:good,fl_progressive:steep/https%3A%2F%2Fsubstack-post-media.s3.amazonaws.com%2Fpublic%2Fimages%2F4c777d0f-75a4-4e42-909d-812165aa351d_1077x702.jpeg)

But here’s what the consensus got right, at least initially: this should have been survivable.

评级机构——其中一些本身就由 PE 控股——在透明度上从来不是典范（几乎没有人会感到意外）。不同公司、不同资产负债表之间织成的蜘蛛网，其不透明程度令人震惊。当底层贷款违约时，“到底是谁在承担损失”这个问题在当下几乎无法实时回答。

Private credit is not 2008 banking. The whole architecture was explicitly designed to avoid forced selling. These are closed-end vehicles with locked-up capital. LPs committed for seven to ten years. There are no depositors to run, no repo lines to pull. The managers could sit on impaired assets, work them out over time, and wait for recoveries. Painful, but manageable. The system was such that it was supposed to bend, not break.

2027 年 11 月的崩盘，标志着市场认知从“也许只是一次常见的周期性回撤”转向“更令人不安的东西”。“一串对‘白领生产率增长’进行相关押注的雏菊链”，是美联储主席 Kevin Warsh 在 FOMC 11 月紧急会议上对它的称呼。

Executives at Blackstone, KKR and Apollo cited software exposure of 7-13% of assets. Containable. Every sell-side note and fintwit credit account said the same thing: private credit has permanent capital. They could absorb losses that would otherwise blow up a levered bank.

你看，危机从来不是由损失本身引发的。引发危机的是对损失的确认。而金融里还有另一个更大、也更重要的领域，我们正开始害怕这种“确认”。

Permanent capital. The phrase showed up in every earnings call and investor letter meant to reassure. It became a mantra. And like most mantras, nobody paid attention to the finer details. Here’s what it actually meant…

按揭之问

Over the prior decade, the large alternative asset managers had acquired life insurance companies and turned them into funding vehicles. Apollo bought Athene. Brookfield bought American Equity. KKR took Global Atlantic. The logic was elegant: annuity deposits provided a stable, long-duration liability base. The managers invested those deposits into the private credit they originated and got paid twice, earning spread over on the insurance side and management fees on the asset management side. A fee-on-fee perpetual motion machine that worked beautifully under one condition.

ZILLOW 房价指数：旧金山同比下跌 11%，西雅图 9%，奥斯汀 8%；房利美提示：科技/金融就业占比 >40% 的邮编区域出现“偏高的早期逾期” | Zillow / Fannie Mae，2028 年 6 月

The private credit had to be money good.

本月，Zillow 房价指数显示：旧金山同比下跌 11%，西雅图 9%，奥斯汀 8%。这并非唯一令人担忧的标题。上个月，房利美提示：以大额贷款为主的邮编区域出现更高的早期逾期——这些区域住着信用分 780+、通常“刀枪不入”的借款人。

The losses hit balance sheets built to hold illiquid assets against long-duration obligations. The “permanent capital” that was supposed to make the system resilient was not some abstract pool of patient institutional money and sophisticated investors taking sophisticated risk. It was the savings of American households, “Main Street”, structured as annuities invested in the same PE-backed software and technology paper that was now defaulting. The locked-up capital that couldn’t run was life insurance policyholder money, and the rules are a bit different there.

美国住宅按揭市场规模约为 $13 trillion。按揭承销建立在一个基本假设之上：借款人在贷款期限内会以大致当前的收入水平持续就业。对大多数按揭而言，这意味着三十年。

Compared to the banking system, insurance regulators had been docile - even complacent - but this was the wake-up call. Already uneasy about private credit concentrations at life insurers, they began downgrading the risk-based capital treatment of these assets. That forced the insurers to either raise capital or sell assets, neither of which was possible at attractive terms in a market already seizing up.

白领就业危机以持续性的收入预期下移威胁着这一假设。我们现在不得不问一个三年前听起来还很荒谬的问题——优质按揭贷款还“稳如现金”吗？

美国历史上每一次按揭危机，成因都来自三者之一：投机过度（把钱借给买不起房的人，如 2008 年）、利率冲击（利率上升使浮动利率按揭变得不可负担，如 20 世纪 80 年代初）、或局部经济冲击（某个地区某个行业崩塌，如 80 年代德州石油，或 2009 年密歇根汽车）。

NEW YORK, IOWA STATE REGULATORS MOVE TO TIGHTEN CAPITAL TREATMENT FOR CERTAIN PRIVATELY RATED CREDIT HELD BY LIFE INSURERS; NAIC GUIDANCE EXPECTED TO INCREASE RBC FACTORS AND TRIGGER ADDITIONAL SVO SCRUTINY | Reuters, Nov 2027

这一次都不适用。这里的借款人不是次贷。他们是 780 的 FICO 分。他们首付 20%。信用记录干净，就业记录稳定，收入在发放时经过核验并有文件证明。他们是整个金融体系里各类风险模型都视为信用质量基石的那群借款人。

2008 年，贷款从第一天起就是坏的。2028 年，贷款从第一天起是好的。只是世界在贷款写下之后……变了。人们借贷押注的是一个他们如今已无法再相信、也无法再负担得起的未来。

When Moody’s put Athene’s financial strength rating on negative outlook, Apollo’s stock dropped 22% in two sessions. Brookfield, KKR, and the others followed.

[

It only got more complex from there. These firms hadn’t just created their insurer perpetual motion machine, they’d built an elaborate offshore architecture designed to maximize returns through regulatory arbitrage.The US insurer wrote the annuity, then ceded the risk to an affiliated Bermuda or Cayman reinsurer it also owned - set up to take advantage of more flexible regulation that permitted holding less capital against the same assets. That affiliate raised outside capital through offshore SPVs, a new layer of counterparties who invested alongside insurers into private credit originated by the same parent’s asset management arm.

[

](https://substackcdn.com/image/fetch/$s_!-8kB!,f_auto,q_auto:good,fl_progressive:steep/https%3A%2F%2Fsubstack-post-media.s3.amazonaws.com%2Fpublic%2Fimages%2F2087b521-70e2-4069-8390-c8f0dbd53e9f_1024x559.png)

2027 年，我们指出了“看不见的压力”的早期迹象：HELOC 提取、401(k) 提前支取、信用卡债务飙升，而按揭月供仍保持按时。随着失业发生、招聘冻结、奖金被砍，这些优质家庭的债务收入比翻倍。

](https://substackcdn.com/image/fetch/$s_!yyoF!,f_auto,q_auto:good,fl_progressive:steep/https%3A%2F%2Fsubstack-post-media.s3.amazonaws.com%2Fpublic%2Fimages%2F4c777d0f-75a4-4e42-909d-812165aa351d_1077x702.jpeg)

他们仍能还按揭，但只能通过停止一切可选支出、消耗储蓄，并推迟任何房屋维护或改善。他们在技术意义上仍“按时还款”，但离困境只差再来一次冲击；而 AI 能力的发展轨迹暗示，这次冲击正在路上。随后我们看到旧金山、西雅图、曼哈顿和奥斯汀的逾期开始飙升，即使全国平均仍处在历史常态范围之内。

The ratings agencies, some of which were themselves PE-owned, had not been paragons of transparency (surprising to virtually) no one. The spider web of different firms linked to different balance sheets was stunning in its opacity. When the underlying loans defaulted, the question of who actually bore the loss was genuinely unanswerable in real time.

我们现在处于最尖锐的阶段。边际买家健康时，房价下跌是可控的。可这里的边际买家也在经历同样的收入受损。

The November 2027 crash marked the transition of perception from a potentially garden-variety cyclical drawdown to something much more uncomfortable. “A daisy chain of correlated bets on white collar productivity growth” was what Fed Chair Kevin Warsh called it during the FOMC’s emergency November meeting.

尽管担忧正在累积，我们还没有进入全面按揭危机。逾期率上升了，但仍远低于 2008 年水平。真正的威胁在于轨迹。

See, it is never the losses themselves that cause the crisis. It’s recognizing them. And there is another, much larger, much much more important area of finance for which we have grown fearful of that recognition.

[

The Mortgage Question

](https://substackcdn.com/image/fetch/$s_!XiXV!,f_auto,q_auto:good,fl_progressive:steep/https%3A%2F%2Fsubstack-post-media.s3.amazonaws.com%2Fpublic%2Fimages%2F4fa31f0b-2188-4936-9ebb-88ccf78d51ff_1598x810.png)

ZILLOW HOME VALUE INDEX FALLS 11% YOY IN SAN FRANCISCO, 9% IN SEATTLE, 8% IN AUSTIN; FANNIE MAE FLAGS ‘ELEVATED EARLY-STAGE DELINQUENCIES’ IN ZIP CODES WITH >40% TECH/FINANCE EMPLOYMENT | Zillow / Fannie Mae, June 2028

智能替代螺旋如今还有两个金融加速器，正推动实体经济下行。

劳动力替代、按揭担忧、私人市场动荡。三者彼此强化。而传统的政策工具箱（降息、QE）可以应对金融引擎，却无法应对实体引擎，因为实体引擎并非由紧缩的金融条件驱动。它是由 AI 让人类智能不再稀缺、不再昂贵所驱动。你可以把利率降到零，把市场里所有 MBS 以及所有违约的软件 LBO 债务全买下来……

This month the Zillow Home Value Index fell 11% year-over-year in San Francisco, 9% in Seattle and 8% in Austin. This hasn’t been the only worrying headline. Last month, Fannie Mae flagged higher early-stage delinquency from jumbo-heavy ZIP codes - areas that are populated by 780+ credit score borrowers and typically “bulletproof”.

这也改变不了：一个 Claude 智能体可以用 $200/month 的成本，完成一个年薪 $180,000 的产品经理的工作。

The US residential mortgage market is approximately $13 trillion. Mortgage underwriting is built on the fundamental assumption that the borrower will remain employed at roughly their current income level for the duration of the loan. For thirty years, in the case of most mortgages.

如果这些担忧兑现，按揭市场会在今年下半年出现裂缝。在那种情景下，我们预计股市当前的回撤最终会逼近 GFC 的水平（从峰值到谷底下跌 57%）。这将把标普 500 带到约 3500 点——这是自 2022 年 11 月、ChatGPT 时刻前一个月以来未见的水平。

The white-collar employment crisis has threatened this assumption with a sustained shift in income expectations. We now have to ask a question that seemed absurd just 3 years ago - are prime mortgages money good?

可以确定的是：支撑 $13 trillion 住宅按揭的收入假设已遭到结构性损伤。不确定的是：政策是否能在按揭市场完全消化这一含义之前介入。我们仍抱有希望，但也无法否认希望不足的理由。

Every prior mortgage crisis in US history has been driven by one of three things: speculative excess (lending to people who couldn’t afford the homes, as in 2008), interest rate shocks (rising rates making adjustable-rate mortgages unaffordable, as in the early 1980s), or localized economic shocks (a single industry collapsing in a single region, like oil in Texas in the 1980s or auto in Michigan in 2009).

None of these apply here. The borrowers in question are not subprime. They’re 780 FICO scores. They put 20% down. They have clean credit histories, stable employment records, and incomes that were verified and documented at origination. They were the borrowers that every risk model in the financial system treats as the bedrock of credit quality.

与时间赛跑

In 2008, the loans were bad on day one. In 2028, the loans were good on day one. The world just…changed after the loans were written. People borrowed against a future they can no longer afford to believe in.

第一条负反馈回路发生在实体经济里：AI 能力提升，工资支出缩减，消费走弱，利润率收紧，企业买入更多能力，能力再提升。然后它转向金融层面：收入受损冲击按揭，银行损失收紧信贷，财富效应破裂，反馈回路加速。而这一切又被一个看起来、坦白说有些困惑的政府的不足政策应对所加剧。

[

[

](https://substackcdn.com/image/fetch/$s_!-8kB!,f_auto,q_auto:good,fl_progressive:steep/https%3A%2F%2Fsubstack-post-media.s3.amazonaws.com%2Fpublic%2Fimages%2F2087b521-70e2-4069-8390-c8f0dbd53e9f_1024x559.png)

](https://substackcdn.com/image/fetch/$s_!0sEd!,f_auto,q_auto:good,fl_progressive:steep/https%3A%2F%2Fsubstack-post-media.s3.amazonaws.com%2Fpublic%2Fimages%2F5e7b551a-0d98-4485-a4fd-d7825349bac3_1024x559.png)

In 2027, we flagged early signs of invisible stress: HELOC draws, 401(k) withdrawals, and credit card debt spiking while mortgage payments remained current. As jobs were lost, hiring was frozen and bonuses cut, these prime households saw their debt-to-income ratios double.

这个系统从未为这样的危机而设计。联邦政府的收入基础，本质上是一种对“人类时间”的征税：人工作，企业付钱，政府抽成。正常年份里，个人所得税和工资税是财政收入的脊梁。

They could still make the mortgage payment, but only by stopping all discretionary spending, draining savings, and deferring any home maintenance or improvement. They were technically current on their mortgage, but just one more shock away from distress, and the trajectory of AI capabilities suggested that shock is coming. Then we saw delinquencies begin to spike in San Francisco, Seattle, Manhattan and Austin, even as the national average stayed within historical norms.

截至今年第一季度，联邦财政收入比 CBO 基线预测低了 12%。工资税收入在下降，因为就业人数变少、且收入不再维持在过去的水平。所得税收入在下降，因为真正赚到的收入在结构性变低。生产率在飙升，但收益流向了资本和算力，而非劳动。

We’re now in the most acute stage. Falling home prices are manageable when the marginal buyer is healthy. Here, the marginal buyer is dealing with the same income impairment.

劳动收入占 GDP 的比重从 1974 年的 64% 下降到 2024 年的 56%，这是由全球化、自动化和工人议价能力持续削弱驱动的长达四十年的缓慢下行。而在 AI 开始指数级提升后的四年里，这一比例又降到了 46%。这是有记录以来最陡的下滑。

While concerns are building, we are not yet in a full-blown mortgage crisis. Delinquencies have risen but remain well below 2008 levels. It is the trajectory that’s the real threat.

产出仍在那里。但它不再在“回到企业之前”先经过家庭，这意味着它也不再经过 IRS。循环流正在断裂，而政府却被期待出面修复它。

[

[

](https://substackcdn.com/image/fetch/$s_!XiXV!,f_auto,q_auto:good,fl_progressive:steep/https%3A%2F%2Fsubstack-post-media.s3.amazonaws.com%2Fpublic%2Fimages%2F4fa31f0b-2188-4936-9ebb-88ccf78d51ff_1598x810.png)

](https://substackcdn.com/image/fetch/$s_!YdQQ!,f_auto,q_auto:good,fl_progressive:steep/https%3A%2F%2Fsubstack-post-media.s3.amazonaws.com%2Fpublic%2Fimages%2F29c19580-7ac9-4287-aea8-a8341f320152_1600x762.png)

The Intelligence Displacement Spiral now has two financial accelerants to the real economy’s decline.

如同每一次下行周期，支出上升的同时收入下滑。不同之处在于：这一次的支出压力不是周期性的。自动稳定器是为临时性失业而设计的，不是为结构性替代而设计的。系统在支付的福利，隐含着工人会重新被吸收。很多人不会，至少不会以任何接近原工资的水平。COVID 期间，政府可以坦然接受 15% 的赤字，因为那被认为是暂时的。而今天需要政府支持的人，并不是被一场终会恢复的疫情击中；他们被一种仍在持续进步的技术取代了。

Labor displacement, mortgage concerns, private market turmoil. Each reinforces the other. And the traditional policy toolkit (rate cuts, QE) can address the financial engine but cannot address the real economy engine, because the real economy engine is not driven by tight financial conditions. It’s driven by AI making human intelligence less scarce and less valuable. You can cut rates to zero and buy every MBS and all the defaulted software LBO debt in the market…

政府需要在正从家庭征收更少税收的时刻，把更多钱转移给家庭。

It won’t change the fact that a Claude agent can do the work of a $180,000 product manager for $200/month.

美国不会违约。它印自己花的货币，也用同一种货币偿还债务。但压力已在别处显现。市政债今年迄今的表现出现令人担忧的分化。没有所得税的州还好，但依赖所得税的州（多数蓝州）发行的一般责任市政债开始计入一定违约风险。政客很快嗅到这一点，“救谁”之争沿着党派路线展开。

If these fears manifest, the mortgage market cracks in the back half of this year. In that scenario, we’d expect the current drawdown in equities to ultimately rival that of the GFC (57% peak-to-trough). This would bring the S&P500 to ~3500 - levels we haven’t seen since the month before the ChatGPT moment in November 2022.

值得肯定的是，政府较早识别到危机的结构性，并开始考虑两党提案——他们称之为“Transition Economy Act”：一个对被替代劳动者进行直接转移支付的框架，资金来源为赤字支出与一项拟议的 AI 推理算力税的组合。

What’s clear is that the income assumptions underlying $13 trillion in residential mortgages are structurally impaired. What isn’t is whether policy can intervene before the mortgage market fully processes what this means. We’re hopeful, but we can’t deny the reasons not to be.

桌面上最激进的提案更进一步。“Shared AI Prosperity Act”将建立公众对智能基础设施回报的公共索取权——介于主权财富基金与对 AI 生成产出的版税之间——以分红资助对家庭的转移支付。私营部门游说者已在媒体上铺天盖地警告“滑坡效应”。

讨论背后的政治走向令人沮丧地可预测，并被作秀与边缘政策的对抗进一步放大。右翼把转移支付与再分配称为马克思主义，并警告对算力征税会把领先优势拱手让给中国。左翼警告：在既得利益者参与起草下形成的税制，换个名字就是监管俘获。财政鹰派指出赤字不可持续。鸽派则把 GFC 后过早的紧缩当作前车之鉴。在今年总统选举临近之际，分裂只会加剧。

The Battle Against Time

政客在争吵，社会肌理却正以比立法流程更快的速度撕裂。

The first negative feedback loop was in the real economy: AI capability improves, payroll shrinks, spending softens, margins tighten, companies buy more capability, capability improves. Then it turned financial: income impairment hit mortgages, bank losses tightened credit, the wealth effect cracked, and the feedback loop sped up. And both of these have been exacerbated by an insufficient policy response from a government that seems, quite frankly, confused.

“Occupy Silicon Valley”运动已成为更广泛不满的象征。上个月，示威者连续三周封锁了 Anthropic 与 OpenAI 在旧金山的办公室入口。他们的人数在增长，而示威获得的媒体关注甚至超过了引发它们的失业数据。

很难想象 GFC 余波中公众还能比恨银行家更恨谁，但 AI 实验室正在朝这个方向冲刺。而从大众视角看，理由充分：它们的创始人和早期投资者以一种让“镀金时代”都显得温顺的速度积累财富。生产率繁荣的收益几乎全部流向算力拥有者与依赖其运行的实验室股东，这把美国不平等推到前所未有的水平。

[

各方都有自己的“反派”，但真正的反派是时间。

AI 能力进化得比制度适应更快。政策反应按意识形态而非现实的速度推进。如果政府无法尽快就“问题是什么”达成一致，反馈回路会替他们写下下一章。

](https://substackcdn.com/image/fetch/$s_!0sEd!,f_auto,q_auto:good,fl_progressive:steep/https%3A%2F%2Fsubstack-post-media.s3.amazonaws.com%2Fpublic%2Fimages%2F5e7b551a-0d98-4485-a4fd-d7825349bac3_1024x559.png)

The system wasn’t designed for a crisis like this. The federal government’s revenue base is essentially a tax on human time. People work, firms pay them, the government takes a cut. Individual income and payroll taxes are the spine of receipts in normal years.

智能溢价的回撤

Through Q1 of this year, federal receipts were running 12% below CBO baseline projections. Payroll receipts are falling because fewer people are employed at prior compensation levels. Income tax receipts are falling because the incomes being earned are structurally lower. Productivity is surging, but the gains are flowing to capital and compute, not labor.

在现代经济史的大部分时期，人类智能一直是稀缺要素。资本充裕（至少可复制）。自然资源有限但可替代。技术进步足够慢，人类能适应。智能——分析、决策、创造、说服与协调的能力——是无法规模化复制的东西。

Labor’s share of GDP declined from 64% in 1974 to 56% in 2024, a four-decade grind lower driven by globalization, automation, and the steady erosion of worker bargaining power. In the four years since AI began its exponential improvement, that has dropped to 46%. The sharpest decline on record.

人类智能的溢价来自其稀缺性。我们经济中的每一个制度，从劳动力市场到按揭市场再到税制，都是为这一假设成立的世界而设计的。

The output is still there. But it’s no longer routing through households on the way back to firms, which means it’s no longer routing through the IRS either. The circular flow is breaking, and the government is expected to step in to fix that.

我们正在经历这一溢价的回撤。机器智能如今在越来越多任务上，成为一种称职且快速进步的人类智能替代品。一个为“稀缺的人类头脑”世界优化了数十年的金融系统，正在重新定价。这种重新定价痛苦、无序，且远未完成。

[

但重新定价不等于崩塌。

经济可以找到新的均衡。抵达那里的过程，是少数仍只能由人类完成的任务之一。我们需要把它做对。

](https://substackcdn.com/image/fetch/$s_!YdQQ!,f_auto,q_auto:good,fl_progressive:steep/https%3A%2F%2Fsubstack-post-media.s3.amazonaws.com%2Fpublic%2Fimages%2F29c19580-7ac9-4287-aea8-a8341f320152_1600x762.png)

这是历史上第一次：经济中最具生产力的资产带来的是更少、而不是更多的就业。没有任何旧框架适配，因为它们从未为“稀缺要素变得充裕”的世界而设计。所以我们必须创造新框架。我们能否及时建立它们，是唯一重要的问题。

As in every downturn, outlays rise just as receipts fall. The difference this time is that the spending pressure is not cyclical. Automatic stabilizers were built for temporary job losses, not structural displacement. The system is paying benefits that assume workers will be reabsorbed. Many will not, at least not at anything like their prior wage. During COVID, the government freely embraced 15% deficits, but it was understood to be temporary. The people who need government support today were not hit by a pandemic they’ll recover from. They were replaced by a technology that continues to improve.

但你读到这篇文章时并不是 2028 年 6 月。你是在 2026 年 2 月读到它。

The government needs to transfer more money to households at precisely the moment it is collecting less money from them in taxes.

标普接近历史新高。负反馈回路尚未开始。我们确信其中一些情景不会发生。我们也同样确信：机器智能会继续加速。人类智能的溢价会收窄。

The U.S. won’t default. It prints the currency it spends, the same currency it uses to pay back borrowers. But this stress has shown up elsewhere. Municipal bonds are showing worrying signs of dispersion in year-to-date performance. States without income tax have been okay, but general obligation munis issued by states dependent on income tax (majority blue states) began to price in some default risk. Politicos caught on quickly, and the debate over who gets bailed out has fallen along partisan lines.

作为投资者，我们仍有时间评估：我们的组合中有多少建立在无法在本十年幸存的假设之上。作为社会，我们仍有时间采取主动。

The administration, to its credit, recognized the structural nature of the crisis early and began entertaining bipartisan proposals for what they’re calling the “Transition Economy Act”: a framework for direct transfers to displaced workers funded by a combination of deficit spending and a proposed tax on AI inference compute.

金丝雀还活着。

The most radical proposal on the table goes further. The “Shared AI Prosperity Act” would establish a public claim on the returns of the intelligence infrastructure itself, something between a sovereign wealth fund and a royalty on AI-generated output, with dividends funding household transfers. Private sector lobbyists have flooded the media with warnings about the slippery slope.

Subscribe now

The politics behind the discussions have been grimly predictable, exacerbated by grandstanding and brinksmanship. The right calls transfers and redistribution Marxism and warns that taxing compute hands the lead to China. The left warns that a tax drafted with the help of incumbents becomes regulatory capture by another name. Fiscal hawks point to unsustainable deficits. Doves point to the premature austerity imposed after the GFC as a cautionary tale. The divide is only magnifying in the run up to this year’s presidential election.

While the politicians bicker, the social fabric is fraying faster than the legislative process can move.

The Occupy Silicon Valley movement has been emblematic of wider dissatisfaction. Last month, demonstrators blockaded the entrances to Anthropic and OpenAI’s San Francisco offices for three weeks straight. Their numbers are growing, and the demonstrations have drawn more media coverage than the unemployment data that prompted them.

致谢：感谢 Sam Koppelman of * 在校对方面的帮助。我们的合著者 LOTUS 的 Alap Shah 提出了本文的想法——这一部分由 CitriniResearch 撰写，但他在一个名为 Intelligence Explosion 的系列中还写了其他篇章，我们强烈推荐阅读。你可以在这里找到。*

It’s hard to imagine the public hating anyone more than the bankers in the fallout of the GFC, but the AI labs are making a run at it. And, from the perspective of the masses, for good reason. Their founders and early investors have accumulated wealth at a pace that makes the Gilded Age look tame. The gains from the productivity boom accruing almost entirely to the owners of compute and the shareholders of the labs that ran on it has magnified US inequality to unprecedented levels.

Citrini Research 是一家由读者支持的出版物。若要接收新文章并支持我的工作，欢迎考虑成为免费或付费订阅者。

Every side has their own villain, but the real villain is time.

相关笔记

AI capability is evolving faster than institutions can adapt. The policy response is moving at the pace of ideology, not reality. If the government doesn’t agree on what the problem is soon, the feedback loop will write the next chapter for them.

The Intelligence Premium Unwind

For the entirety of modern economic history, human intelligence has been the scarce input. Capital was abundant (or at least, replicable). Natural resources were finite but substitutable. Technology improved slowly enough that humans could adapt. Intelligence, the ability to analyze, decide, create, persuade, and coordinate, was the thing that could not be replicated at scale.

Human intelligence derived its inherent premium from its scarcity. Every institution in our economy, from the labor market to the mortgage market to the tax code, was designed for a world in which that assumption held.

We are now experiencing the unwind of that premium. Machine intelligence is now a competent and rapidly improving substitute for human intelligence across a growing range of tasks. The financial system, optimized over decades for a world of scarce human minds, is repricing. That repricing is painful, disorderly, and far from complete.

But repricing is not the same as collapse.

The economy can find a new equilibrium. Getting there is one of the few tasks left that only humans can do. We need to do it correctly.

This is the first time in history the most productive asset in the economy has produced fewer, not more, jobs. Nobody’s framework fits, because none were designed for a world where the scarce input became abundant. So we have to make new frameworks. Whether we build them in time is the only question that matters.

But you’re not reading this in June 2028. You’re reading it in February 2026.

[

](https://substackcdn.com/image/fetch/$s_!rCTy!,f_auto,q_auto:good,fl_progressive:steep/https%3A%2F%2Fsubstack-post-media.s3.amazonaws.com%2Fpublic%2Fimages%2F9ede0b37-f3be-4787-90ff-10cb24719fdc_710x1119.jpeg)

Macro Memo

The Consequences of Abundant Intelligence

***CitriniResearch ***

February 22nd, 2026 June 30th, 2028

The unemployment rate printed 10.2% this morning, a 0.3% upside surprise. The market sold off 2% on the number, bringing the cumulative drawdown in the S&P to 38% from its October 2026 highs.

Traders have grown numb. Six months ago, a print like this would have triggered a circuit breaker.

Two years. That’s all it took to get from “contained” and “sector-specific” to an economy that no longer resembles the one any of us grew up in. This quarter’s macro memo is our attempt to reconstruct the sequence - a post-mortem on the pre-crisis economy.

The euphoria was palpable. By October 2026, the S&P 500 flirted with 8000, the Nasdaq broke above 30k. The initial wave of layoffs due to human obsolescence began in early 2026, and they did exactly what layoffs are supposed to. Margins expanded, earnings beat, stocks rallied. Record-setting corporate profits were funneled right back into AI compute.

The headline numbers were still great. Nominal GDP repeatedly printed mid-to-high single-digit annualized growth. Productivity was booming. Real output per hour rose at rates not seen since the 1950s, driven by AI agents that don’t sleep, take sick days or require health insurance.

The owners of compute saw their wealth explode as labor costs vanished. Meanwhile, real wage growth collapsed. Despite the administration’s repeated boasts of record productivity, white-collar workers lost jobs to machines and were forced into lower-paying roles.

When cracks began appearing in the consumer economy, economic pundits popularized the phrase “Ghost GDP“: output that shows up in the national accounts but never circulates through the real economy.

In every way AI was exceeding expectations, and the market was AI. The only problem…the economy was not.

It should have been clear all along that a single GPU cluster in North Dakota generating the output previously attributed to 10,000 white-collar workers in midtown Manhattan is more economic pandemic than economic panacea. The velocity of money flatlined. The human-centric consumer economy, 70% of GDP at the time, withered. We probably could have figured this out sooner if we just asked how much money machines spend on discretionary goods. (Hint: it’s zero.)

AI capabilities improved, companies needed fewer workers, white collar layoffs increased, displaced workers spent less, margin pressure pushed firms to invest more in AI, AI capabilities improved…

It was a negative feedback loop with no natural brake. The human intelligence displacement spiral. White-collar workers saw their earnings power (and, rationally, their spending) structurally impaired. Their incomes were the bedrock of the $13 trillion mortgage market - forcing underwriters to reassess whether prime mortgages are still money good.

Seventeen years without a real default cycle had left privates bloated with PE-backed software deals that assumed ARR would remain recurring. The first wave of defaults due to AI disruption in mid-2027 challenged that assumption.

This would have been manageable if the disruption remained contained to software, but it didn’t. By the end of 2027, it threatened every business model predicated on intermediation. Swaths of companies built on monetizing friction for humans disintegrated.

The system turned out to be one long daisy chain of correlated bets on white-collar productivity growth. The November 2027 crash only served to accelerate all of the negative feedback loops already in place.

We’ve been waiting for “bad news is good news” for almost a year now. The government is starting to consider proposals, but public faith in the ability of the government to stage any sort of rescue has dwindled. Policy response has always lagged economic reality, but lack of a comprehensive plan is now threatening to accelerate a deflationary spiral.

How It Started

In late 2025, agentic coding tools took a step function jump in capability.

A competent developer working with Claude Code or Codex could now replicate the core functionality of a mid-market SaaS product in weeks. Not perfectly or with every edge case handled, but well enough that the CIO reviewing a $500k annual renewal started asking the question “what if we just built this ourselves?”

Fiscal years mostly line up with calendar years, so 2026 enterprise spend had been set in Q4 2025, when “agentic AI” was still a buzzword. The mid-year review was the first time procurement teams were making decisions with visibility into what these systems could actually do. Some watched their own internal teams spin up prototypes replicating six-figure SaaS contracts in weeks.

That summer, we spoke with a procurement manager at a Fortune 500. He told us about one of his budget negotiations. The salesperson had expected to run the same playbook as last year: a 5% annual price increase, the standard “your team depends on us” pitch. The procurement manager told him he’d been in conversations with OpenAI about having their “forward deployed engineers” use AI tools to replace the vendor entirely. They renewed at a 30% discount. That was a good outcome, he said. The “long-tail of SaaS”, like Monday.com, Zapier and Asana, had it much worse.

Investors were prepared - expectant, even - that the long tail would be hit hard. They may have made up a third of spending for the typical enterprise stack, but they were obviously exposed. The systems of record, however, were supposed to be safe from disruption.

It wasn’t until ServiceNow’s Q3 26 report that the mechanism of reflexivity became clearer.

SERVICENOW NET NEW ACV GROWTH DECELERATES TO 14% FROM 23%; ANNOUNCES 15% WORKFORCE REDUCTION AND ‘STRUCTURAL EFFICIENCY PROGRAM’; SHARES FALL 18% | Bloomberg, October 2026

SaaS wasn’t “dead”. There was still a cost-benefit-analysis to running and supporting in-house builds. But in-house was an option, and that factored into pricing negotiations. Perhaps more importantly, the competitive landscape had changed. AI had made it easier to develop and ship new features, so differentiation collapsed. Incumbents were in a race to the bottom on pricing - a knife-fight with both each other and with the new crop of upstart challengers that popped up. Emboldened by the leap in agentic coding capabilities and with no legacy cost structure to protect, these aggressively took share.

The interconnected nature of these systems weren’t fully appreciated until this print, either. ServiceNow sold seats. When Fortune 500 clients cut 15% of their workforce, they cancelled 15% of their licenses. The same AI-driven headcount reductions that were boosting margins at their customers were mechanically destroying their own revenue base.

The company that sold workflow automation was being disrupted by better workflow automation, and its response was to cut headcount and use the savings to fund the very technology disrupting it.

What else were they supposed to do? Sit still and die slower? The companies most threatened by AI became AI’s most aggressive adopters.**

This sounds obvious in hindsight, but it really wasn’t at the time (at least to me). The historical disruption model said incumbents resist new technology, they lose share to nimble entrants and die slowly. That’s what happened to Kodak, to Blockbuster, to BlackBerry. What happened in 2026 was different; the incumbents didn’t resist because they couldn’t afford to.

With stocks down 40-60% and boards demanding answers, the AI-threatened companies did the only thing they could. Cut headcount, redeploy the savings into AI tools, use those tools to maintain output with lower costs.

Each company’s individual response was rational. The collective result was catastrophic. Every dollar saved on headcount flowed into AI capability that made the next round of job cuts possible.

Software was only the opening act. What investors missed while they debated whether SaaS multiples had bottomed was that the reflexive loop had already escaped the software sector. The same logic that justified ServiceNow cutting headcount applied to every company with a white-collar cost structure.

When Friction Went to Zero

By early 2027, LLM usage had become default. People were using AI agents who didn’t even know what an AI agent was, in the same way people who never learned what “cloud computing” was used streaming services. They thought of it the same way they thought of autocomplete or spell-check - a thing their phone just did now.

Qwen’s open-source agentic shopper was the catalyst for AI handling consumer decisions. Within weeks, every major AI assistant had integrated some agentic commerce feature. Distilled models meant these agents could run on phones and laptops, not just cloud instances, reducing the marginal cost of inference significantly.

The part that should have unsettled investors more than it did was that these agents didn’t wait to be asked. They ran in the background according to the user’s preferences. Commerce stopped being a series of discrete human decisions and became a continuous optimization process, running 24/7 on behalf of every connected consumer. By March 2027, the median individual in the United States was consuming 400,000 tokens per day - 10x since the end of 2026.

The next link in the chain was already breaking.

Intermediation.

Over the past fifty years, the U.S. economy built a giant rent-extraction layer on top of human limitations: things take time, patience runs out, brand familiarity substitutes for diligence, and most people are willing to accept a bad price to avoid more clicks. Trillions of dollars of enterprise value depended on those constraints persisting.

It started out simple enough. Agents removed friction.

Subscriptions and memberships that passively renewed despite months of disuse. Introductory pricing that sneakily doubled after the trial period. Each one was rebranded as a hostage situation that agents could negotiate. The average customer lifetime value, the metric the entire subscription economy was built on, distinctly declined.

Consumer agents began to change how nearly all consumer transactions worked.

Humans don’t really have the time to price-match across five competing platforms before buying a box of protein bars. Machines do.

Travel booking platforms were an early casualty, because they were the simplest. By Q4 2026, our agents could assemble a complete itinerary (flights, hotels, ground transport, loyalty optimization, budget constraints, refunds) faster and cheaper than any platform.

Insurance renewals, where the entire renewal model depended on policyholder inertia, were reformed. Agents that re-shop your coverage annually dismantled the 15-20% of premiums that insurers earned from passive renewals.

Financial advice. Tax prep. Routine legal work. Any category where the service provider’s value proposition was ultimately “I will navigate complexity that you find tedious” was disrupted, as the agents found nothing tedious.

Even places we thought insulated by the value of human relationships proved fragile. Real estate, where buyers had tolerated 5-6% commissions for decades because of information asymmetry between agent and consumer, crumbled once AI agents equipped with MLS access and decades of transaction data could replicate the knowledge base instantly. A sell-side piece from March 2027 titled it “agent on agent violence”. The median buy-side commission in major metros had compressed from 2.5-3% to under 1%, and a growing share of transactions were closing with no human agent on the buy side at all.

We had overestimated the value of “human relationships”. Turns out that a lot of what people called relationships was simply friction with a friendly face.

That was just the start of the disruption for the intermediation layer. Successful companies had spent billions to effectively exploit quirks of consumer behavior and human psychology that didn’t matter anymore.

Machines optimizing for price and fit do not care about your favorite app or the websites you’ve been habitually opening for the last four years, nor feel the pull of a well-designed checkout experience. They don’t get tired and accept the easiest option or default to “I always just order from here”.

That destroyed a particular kind of moat: habitual intermediation.

DoorDash (DASH US) was the poster child.

Coding agents had collapsed the barrier to entry for launching a delivery app. A competent developer could deploy a functional competitor in weeks, and dozens did, enticing drivers away from DoorDash and Uber Eats by passing 90-95% of the delivery fee through to the driver. Multi-app dashboards let gig workers track incoming jobs from twenty or thirty platforms at once, eliminating the lock-in that the incumbents depended on. The market fragmented overnight and margins compressed to nearly nothing.

Agents accelerated both sides of the destruction. They enabled the competitors and then they used them. The DoorDash moat was literally “you’re hungry, you’re lazy, this is the app on your home screen.” An agent doesn’t have a home screen. It checks DoorDash, Uber Eats, the restaurant’s own site, and twenty new vibe-coded alternatives so it can pick the lowest fee and fastest delivery every time.

Habitual app loyalty, the entire basis of the business model, simply didn’t exist for a machine.

This was oddly poetic, as perhaps the only example in this entire saga of agents doing a favor for the soon-to-be-displaced white collar workers. When they ended up as delivery drivers, at least half their earnings weren’t going to Uber and DoorDash. Of course, this favor from technology didn’t last for long as autonomous vehicles proliferated.

Once agents controlled the transaction, they went looking for bigger paperclips.

There was only so much price-matching and aggregating to do. The biggest way to repeatedly save the user money (especially when agents started transacting among themselves) was to eliminate fees. In machine-to-machine commerce, the 2-3% card interchange rate became an obvious target.

Agents went looking for faster and cheaper options than cards. Most settled on using stablecoins via Solana or Ethereum L2s, where settlement was near-instant and the transaction cost was measured in fractions of a penny.

MASTERCARD Q1 2027: NET REVENUES +6% Y/Y; PURCHASE VOLUME GROWTH SLOWS TO +3.4% Y/Y FROM +5.9% PRIOR QUARTER; MANAGEMENT NOTES “AGENT-LED PRICE OPTIMIZATION” AND “PRESSURE IN DISCRETIONARY CATEGORIES” | Bloomberg, April 29 2027

Mastercard’s Q1 2027 report was the point of no return. Agentic commerce went from being a product story to a plumbing story. MA dropped 9% the following day. Visa did too, but pared losses after analysts pointed out its stronger positioning in stablecoin infrastructure.

[

](https://substackcdn.com/image/fetch/$s_!gW0T!,f_auto,q_auto:good,fl_progressive:steep/https%3A%2F%2Fsubstack-post-media.s3.amazonaws.com%2Fpublic%2Fimages%2F70ccc118-1388-4602-a5bd-bbc4284208f6_1024x541.png)

Agentic commerce routing around interchange posed a far greater risk to card-focused banks and mono-line issuers, who collected the majority of that 2-3% fee and had built entire business segments around rewards programs funded by the merchant subsidy.

American Express (AXP US) was hit hardest; a combined headwind from white-collar workforce reductions gutting its customer base and agents routing around interchange gutting its revenue model. Synchrony (SYF US), Capital One (COF US) and Discover (DFS US) all fell more than 10% over the following weeks, as well.

Their moats were made of friction. And friction was going to zero.

From Sector Risk to Systemic Risk

Through 2026, markets treated negative AI impact as a sector story. Software and consulting were getting crushed, payments and other toll booths were wobbly, but the broader economy seemed fine. The labor market, while softening, was not in freefall. The consensus view was that creative destruction was part of any technological innovation cycle. It would be painful in pockets, but the overall net positives from AI would outweigh any negatives.

Our January 2027 macro memo argued this was the wrong mental model. The US economy is a white-collar services economy. White-collar workers represented 50% of employment and drove roughly 75% of discretionary consumer spending. The businesses and jobs that AI was chewing up were not tangential to the US economy, they were the US economy.

“Technological innovation destroys jobs and then creates even more”. This was the most popular and convincing counter-argument at the time. It was popular and convincing because it’d been right for two centuries. Even if we couldn’t conceive of what the future jobs would be, they would surely arrive.

ATMs made branches cheaper to operate so banks opened more of them and teller employment rose for the next twenty years. The internet disrupted travel agencies, the Yellow Pages, brick-and-mortar retail, but it invented entirely new industries in their place that conjured new jobs.

Every new job, however, required a human to perform it.

AI is now a general intelligence that improves at the very tasks humans would redeploy to. Displaced coders cannot simply move to “AI management” because AI is already capable of that.

Today, AI agents handle many-weeks-long research and development tasks. The exponential steamrolled our conceptions of what was possible, even though every year Wharton professors tried to fit the data to a new sigmoid.

[

](https://substackcdn.com/image/fetch/$s_!kA4x!,f_auto,q_auto:good,fl_progressive:steep/https%3A%2F%2Fsubstack-post-media.s3.amazonaws.com%2Fpublic%2Fimages%2F0bc01ebb-a883-4e5c-bd2b-fa7aaa872edb_1600x1059.png)

They write essentially all code. The highest performing of them are substantially smarter than almost all humans at almost all things. And they keep getting cheaper.

AI has created new jobs. Prompt engineers. AI safety researchers. Infrastructure technicians. Humans are still in the loop, coordinating at the highest level or directing for taste. For every new role AI created, though, it rendered dozens obsolete. The new roles paid a fraction of what the old ones did.

U.S. JOLTS: JOB OPENINGS FALL BELOW 5.5M; UNEMPLOYED-TO-OPENINGS RATIO CLIMBS TO ~1.7, HIGHEST SINCE AUG 2020 | Bloomberg, Oct 2026

The hiring rate had been anemic all year, but October ‘26 JOLTS print provided some definitive data. Job openings fell below 5.5 million, a 15% decline YoY.

INDEED: POSTINGS FALL SHARPLY IN SOFTWARE, FINANCE, CONSULTING AS “PRODUCTIVITY INITIATIVES” SPREAD | Indeed Hiring Lab, Nov–Dec 2026

White-collar openings were collapsing while blue-collar openings remained relatively stable (construction, healthcare, trades). The churn was in the jobs that write memos (we are, somehow, still in business), approve budgets, and keep the middle layers of the economy lubricated. Real wage growth in both cohorts, however, had been negative for the majority of the year and kept declining.

The equity market still cared less about JOLTS than it did the news that all of GE Vernova’s turbine capacity was now sold out until 2040, it ambled sideways in a tug of war between negative macro news with positive AI infrastructure headlines.

The bond market (always smarter than equities, or at least less romantic) began pricing the consumption hit, however. The 10-year yield began a descent from 4.3% to 3.2% over the following four months. Still, the headline unemployment rate did not blow out, the composition nuance was still lost on some.

In a normal recession, the cause eventually self-corrects. Overbuilding leads to a construction slowdown, which leads to lower rates, which leads to new construction. Inventory overshoot leads to destocking, which leads to restocking. The cyclical mechanism contains within it its own seeds of recovery.

This cycle’s cause was not cyclical.

[

](https://substackcdn.com/image/fetch/$s_!0WIw!,f_auto,q_auto:good,fl_progressive:steep/https%3A%2F%2Fsubstack-post-media.s3.amazonaws.com%2Fpublic%2Fimages%2F1d84258d-d811-4963-8914-16b104e6744f_1024x541.png)

AI got better and cheaper. Companies laid off workers, then used the savings to buy more AI capability, which let them lay off more workers. Displaced workers spent less. Companies that sell things to consumers sold fewer of them, weakened, and invested more in AI to protect margins. AI got better and cheaper.

A feedback loop with no natural brake.

The intuitive expectation was that falling aggregate demand would slow the AI buildout. It didn’t, because this wasn’t hyperscaler-style CapEx. It was OpEx substitution. A company that had been spending $100M a year on employees and $5M on AI now spent $70M on employees and $20M on AI. AI investment increased by multiples, but it occurred as a reduction in total operating costs. Every company’s AI budget grew while its overall spending shrank.

The irony of this was that the AI infrastructure complex kept performing even as the economy it was disrupting began deteriorating. NVDA was still posting record revenues. TSM was still running at 95%+ utilization. The hyperscalers were still spending $150-200 billion per quarter on data center capex. Economies that were purely convex to this trend, like Taiwan and Korea, outperformed massively.

India was the inverse. The country’s IT services sector exported over $200 billion annually, the single largest contributor to India’s current account surplus and the offset that financed its persistent goods trade deficit. The entire model was built on one value proposition: Indian developers cost a fraction of their American counterparts. But the marginal cost of an AI coding agent had collapsed to, essentially, the cost of electricity. TCS, Infosys and Wipro saw contract cancellations accelerate through 2027. The rupee fell 18% against the dollar in four months as the services surplus that had anchored India’s external accounts evaporated. By Q1 2028, the IMF had begun “preliminary discussions” with New Delhi.

The engine that caused the disruption got better every quarter, which meant the disruption accelerated every quarter. There was no natural floor to the labor market.

In the US, we weren’t asking about how the bubble would burst in AI infrastructure anymore. We were asking what happens to a consumer-credit economy when consumers are being replaced with machines.

The Intelligence Displacement Spiral

2027 was when the macroeconomic story stopped being subtle. The transmission mechanism from the previous twelve months of disjointed but clearly negative developments became obvious. You didn’t need to go into the BLS data. Just attend a dinner party with friends.

Displaced white-collar workers did not sit idle. They downshifted. Many took lower-paying service sector and gig economy jobs, which increased labor supply in those segments and compressed wages there too.

A friend of ours was a senior product manager at Salesforce in 2025. Title, health insurance, 401k, $180,000 a year. She lost her job in the third round of layoffs. After six months of searching, she started driving for Uber. Her earnings dropped to $45,000. The point is less the individual story and more the second-order math. Multiply this dynamic by a few hundred thousand workers across every major metro. Overqualified labor flooding the service and gig economy pushed down wages for existing workers who were already struggling. Sector-specific disruption metastasized into economy-wide wage compression.

[

](https://substackcdn.com/image/fetch/$s_!CLd4!,f_auto,q_auto:good,fl_progressive:steep/https%3A%2F%2Fsubstack-post-media.s3.amazonaws.com%2Fpublic%2Fimages%2F8f6a0cb1-86ff-4b99-811c-1652d5577fba_1024x541.jpeg)

The pool of remaining human-centric had another correction ahead of it, happening while we write this. As autonomous delivery and self-driving vehicles work their way through the gig economy that absorbed the first wave of displaced workers.

By February 2027, it was clear that still employed professionals were spending like they might be next. They were working twice as hard (mostly with the help of AI) just to not get fired, hopes of promotion or raises were gone. Savings rates ticked higher and spending softened.

The most dangerous part was the lag. High earners used their higher-than-average savings to maintain the appearance of normalcy for two or three quarters. The hard data didn’t confirm the problem until it was already old news in the real economy. Then came the print that broke the illusion.

U.S. INITIAL JOBLESS CLAIMS SURGE TO 487,000, HIGHEST SINCE APRIL 2020; Department of Labor, Q3 2027

Initial claims surged to 487,000, the highest since April 2020. ADP and Equifax confirmed that the overwhelming majority of new filings were from white-collar professionals.

The S&P dropped 6% over the following week. Negative macro started winning the tug of war.

In a normal recession, job losses are broadly distributed. Blue-collar and white-collar workers share the pain roughly in proportion to each segment’s share of employment. The consumption hit is also broadly distributed, and it shows up quickly in the data because lower-income workers have higher marginal propensities to consume.

In this cycle, the job losses have been concentrated in the upper deciles of the income distribution. They are a relatively small share of total employment, but they drive a wildly disproportionate share of consumer spending. The top 10% of earners account for more than 50% of all consumer spending in the United States. The top 20% account for roughly 65%. These are the people who buy the houses, the cars, the vacations, the restaurant meals, the private school tuition, the home renovations. They are the demand base for the entire consumer discretionary economy.

When these workers lost their jobs, or took 50% pay cuts to move into available roles, the consumption hit was enormous relative to the number of jobs lost. A 2% decline in white-collar employment translated to something like a 3-4% hit to discretionary consumer spending. Unlike blue-collar job losses, which tend to hit immediately (you get laid off from the factory, you stop spending next week), white-collar job losses have a lagged but deeper impact because these workers have savings buffers that allow them to maintain spending for a few months before the behavioral shift kicks in.

By Q2 2027, the economy was in recession. The NBER would not officially date the start until months later (they never do) but the data was unambiguous - we’d had two consecutive quarters of negative real GDP growth. But it wasn’t a “financial crisis”…yet.

The Daisy Chain of Correlated Bets

Private credit had grown from under $1 trillion in 2015 to over $2.5 trillion by 2026. A meaningful share of that capital had been deployed into software and technology deals, many of them leveraged buyouts of SaaS companies at valuations that assumed mid-teens revenue growth in perpetuity.

Those assumptions died somewhere between the first agentic coding demo and the Q1 2026 software crash, but the marks didn’t seem to realize they were dead.

As many public SaaS companies traded to 5-8x EBITDA, PE-backed software companies sat on balance sheets at marks reflecting acquisition valuations on multiples of revenue that didn’t exist anymore. Managers eased the marks down gradually, 100 cents, 92, 85, all while public comps said 50.

MOODY’S DOWNGRADES $18B OF PE-BACKED SOFTWARE DEBT ACROSS 14 ISSUERS, CITING ‘SECULAR REVENUE HEADWINDS FROM AI-DRIVEN COMPETITIVE DISRUPTION’; LARGEST SINGLE-SECTOR ACTION SINCE ENERGY IN 2015 | Moody’s Investors Service, April 2027

Everyone remembers what happened after the downgrade. Industry veterans had already seen the playbook following the 2015 energy downgrades.

Software-backed loans began defaulting in Q3 2027. PE portfolio companies in information services and consulting followed. Several multi-billion dollar LBOs of well-known SaaS companies entered restructuring.

Zendesk was the smoking gun.

ZENDESK MISSES DEBT COVENANTS AS AI-DRIVEN CUSTOMER SERVICE AUTOMATION ERODES ARR; $5B DIRECT LENDING FACILITY MARKED TO 58 CENTS; LARGEST PRIVATE CREDIT SOFTWARE DEFAULT ON RECORD | Financial Times, September 2027

In 2022, Hellman & Friedman and Permira had taken Zendesk private for $10.2 billion. The debt package was $5 billion in direct lending, the largest ARR-backed facility in history at the time, led by Blackstone with Apollo, Blue Owl and HPS all in the lending group. The loan was explicitly structured around the assumption that Zendesk’s annual recurring revenue would remain recurring. At roughly 25x EBITDA, the leverage only made sense if it did.

By mid-2027, it didn’t.

AI agents had been handling customer service autonomously for the better part of a year. The category Zendesk had defined (ticketing, routing, managing human support interactions) was already replaced by systems that resolved issues without generating a ticket at all. The Annualized Recurring Revenue the loan was underwritten against was no longer recurring, it was just revenue that hadn’t left yet.

The largest ARR-backed loan in history became the largest private credit software default in history. Every credit desk asked the same question at once: who else has a secular headwind disguised as a cyclical one?

But here’s what the consensus got right, at least initially: this should have been survivable.

Private credit is not 2008 banking. The whole architecture was explicitly designed to avoid forced selling. These are closed-end vehicles with locked-up capital. LPs committed for seven to ten years. There are no depositors to run, no repo lines to pull. The managers could sit on impaired assets, work them out over time, and wait for recoveries. Painful, but manageable. The system was such that it was supposed to bend, not break.

Executives at Blackstone, KKR and Apollo cited software exposure of 7-13% of assets. Containable. Every sell-side note and fintwit credit account said the same thing: private credit has permanent capital. They could absorb losses that would otherwise blow up a levered bank.

Permanent capital. The phrase showed up in every earnings call and investor letter meant to reassure. It became a mantra. And like most mantras, nobody paid attention to the finer details. Here’s what it actually meant…

Over the prior decade, the large alternative asset managers had acquired life insurance companies and turned them into funding vehicles. Apollo bought Athene. Brookfield bought American Equity. KKR took Global Atlantic. The logic was elegant: annuity deposits provided a stable, long-duration liability base. The managers invested those deposits into the private credit they originated and got paid twice, earning spread over on the insurance side and management fees on the asset management side. A fee-on-fee perpetual motion machine that worked beautifully under one condition.

The private credit had to be money good.

The losses hit balance sheets built to hold illiquid assets against long-duration obligations. The “permanent capital” that was supposed to make the system resilient was not some abstract pool of patient institutional money and sophisticated investors taking sophisticated risk. It was the savings of American households, “Main Street”, structured as annuities invested in the same PE-backed software and technology paper that was now defaulting. The locked-up capital that couldn’t run was life insurance policyholder money, and the rules are a bit different there.

Compared to the banking system, insurance regulators had been docile - even complacent - but this was the wake-up call. Already uneasy about private credit concentrations at life insurers, they began downgrading the risk-based capital treatment of these assets. That forced the insurers to either raise capital or sell assets, neither of which was possible at attractive terms in a market already seizing up.

NEW YORK, IOWA STATE REGULATORS MOVE TO TIGHTEN CAPITAL TREATMENT FOR CERTAIN PRIVATELY RATED CREDIT HELD BY LIFE INSURERS; NAIC GUIDANCE EXPECTED TO INCREASE RBC FACTORS AND TRIGGER ADDITIONAL SVO SCRUTINY | Reuters, Nov 2027

When Moody’s put Athene’s financial strength rating on negative outlook, Apollo’s stock dropped 22% in two sessions. Brookfield, KKR, and the others followed.

It only got more complex from there. These firms hadn’t just created their insurer perpetual motion machine, they’d built an elaborate offshore architecture designed to maximize returns through regulatory arbitrage.The US insurer wrote the annuity, then ceded the risk to an affiliated Bermuda or Cayman reinsurer it also owned - set up to take advantage of more flexible regulation that permitted holding less capital against the same assets. That affiliate raised outside capital through offshore SPVs, a new layer of counterparties who invested alongside insurers into private credit originated by the same parent’s asset management arm.

[

](https://substackcdn.com/image/fetch/$s_!yyoF!,f_auto,q_auto:good,fl_progressive:steep/https%3A%2F%2Fsubstack-post-media.s3.amazonaws.com%2Fpublic%2Fimages%2F4c777d0f-75a4-4e42-909d-812165aa351d_1077x702.jpeg)

The ratings agencies, some of which were themselves PE-owned, had not been paragons of transparency (surprising to virtually) no one. The spider web of different firms linked to different balance sheets was stunning in its opacity. When the underlying loans defaulted, the question of who actually bore the loss was genuinely unanswerable in real time.

The November 2027 crash marked the transition of perception from a potentially garden-variety cyclical drawdown to something much more uncomfortable. “A daisy chain of correlated bets on white collar productivity growth” was what Fed Chair Kevin Warsh called it during the FOMC’s emergency November meeting.

See, it is never the losses themselves that cause the crisis. It’s recognizing them. And there is another, much larger, much much more important area of finance for which we have grown fearful of that recognition.

The Mortgage Question

ZILLOW HOME VALUE INDEX FALLS 11% YOY IN SAN FRANCISCO, 9% IN SEATTLE, 8% IN AUSTIN; FANNIE MAE FLAGS ‘ELEVATED EARLY-STAGE DELINQUENCIES’ IN ZIP CODES WITH >40% TECH/FINANCE EMPLOYMENT | Zillow / Fannie Mae, June 2028

This month the Zillow Home Value Index fell 11% year-over-year in San Francisco, 9% in Seattle and 8% in Austin. This hasn’t been the only worrying headline. Last month, Fannie Mae flagged higher early-stage delinquency from jumbo-heavy ZIP codes - areas that are populated by 780+ credit score borrowers and typically “bulletproof”.

The US residential mortgage market is approximately $13 trillion. Mortgage underwriting is built on the fundamental assumption that the borrower will remain employed at roughly their current income level for the duration of the loan. For thirty years, in the case of most mortgages.

The white-collar employment crisis has threatened this assumption with a sustained shift in income expectations. We now have to ask a question that seemed absurd just 3 years ago - are prime mortgages money good?

Every prior mortgage crisis in US history has been driven by one of three things: speculative excess (lending to people who couldn’t afford the homes, as in 2008), interest rate shocks (rising rates making adjustable-rate mortgages unaffordable, as in the early 1980s), or localized economic shocks (a single industry collapsing in a single region, like oil in Texas in the 1980s or auto in Michigan in 2009).

None of these apply here. The borrowers in question are not subprime. They’re 780 FICO scores. They put 20% down. They have clean credit histories, stable employment records, and incomes that were verified and documented at origination. They were the borrowers that every risk model in the financial system treats as the bedrock of credit quality.

In 2008, the loans were bad on day one. In 2028, the loans were good on day one. The world just…changed after the loans were written. People borrowed against a future they can no longer afford to believe in.

[

](https://substackcdn.com/image/fetch/$s_!-8kB!,f_auto,q_auto:good,fl_progressive:steep/https%3A%2F%2Fsubstack-post-media.s3.amazonaws.com%2Fpublic%2Fimages%2F2087b521-70e2-4069-8390-c8f0dbd53e9f_1024x559.png)

In 2027, we flagged early signs of invisible stress: HELOC draws, 401(k) withdrawals, and credit card debt spiking while mortgage payments remained current. As jobs were lost, hiring was frozen and bonuses cut, these prime households saw their debt-to-income ratios double.

They could still make the mortgage payment, but only by stopping all discretionary spending, draining savings, and deferring any home maintenance or improvement. They were technically current on their mortgage, but just one more shock away from distress, and the trajectory of AI capabilities suggested that shock is coming. Then we saw delinquencies begin to spike in San Francisco, Seattle, Manhattan and Austin, even as the national average stayed within historical norms.

We’re now in the most acute stage. Falling home prices are manageable when the marginal buyer is healthy. Here, the marginal buyer is dealing with the same income impairment.

While concerns are building, we are not yet in a full-blown mortgage crisis. Delinquencies have risen but remain well below 2008 levels. It is the trajectory that’s the real threat.

[

](https://substackcdn.com/image/fetch/$s_!XiXV!,f_auto,q_auto:good,fl_progressive:steep/https%3A%2F%2Fsubstack-post-media.s3.amazonaws.com%2Fpublic%2Fimages%2F4fa31f0b-2188-4936-9ebb-88ccf78d51ff_1598x810.png)

The Intelligence Displacement Spiral now has two financial accelerants to the real economy’s decline.

Labor displacement, mortgage concerns, private market turmoil. Each reinforces the other. And the traditional policy toolkit (rate cuts, QE) can address the financial engine but cannot address the real economy engine, because the real economy engine is not driven by tight financial conditions. It’s driven by AI making human intelligence less scarce and less valuable. You can cut rates to zero and buy every MBS and all the defaulted software LBO debt in the market…

It won’t change the fact that a Claude agent can do the work of a $180,000 product manager for $200/month.

If these fears manifest, the mortgage market cracks in the back half of this year. In that scenario, we’d expect the current drawdown in equities to ultimately rival that of the GFC (57% peak-to-trough). This would bring the S&P500 to ~3500 - levels we haven’t seen since the month before the ChatGPT moment in November 2022.

What’s clear is that the income assumptions underlying $13 trillion in residential mortgages are structurally impaired. What isn’t is whether policy can intervene before the mortgage market fully processes what this means. We’re hopeful, but we can’t deny the reasons not to be.

The Battle Against Time

The first negative feedback loop was in the real economy: AI capability improves, payroll shrinks, spending softens, margins tighten, companies buy more capability, capability improves. Then it turned financial: income impairment hit mortgages, bank losses tightened credit, the wealth effect cracked, and the feedback loop sped up. And both of these have been exacerbated by an insufficient policy response from a government that seems, quite frankly, confused.

[

](https://substackcdn.com/image/fetch/$s_!0sEd!,f_auto,q_auto:good,fl_progressive:steep/https%3A%2F%2Fsubstack-post-media.s3.amazonaws.com%2Fpublic%2Fimages%2F5e7b551a-0d98-4485-a4fd-d7825349bac3_1024x559.png)

The system wasn’t designed for a crisis like this. The federal government’s revenue base is essentially a tax on human time. People work, firms pay them, the government takes a cut. Individual income and payroll taxes are the spine of receipts in normal years.

Through Q1 of this year, federal receipts were running 12% below CBO baseline projections. Payroll receipts are falling because fewer people are employed at prior compensation levels. Income tax receipts are falling because the incomes being earned are structurally lower. Productivity is surging, but the gains are flowing to capital and compute, not labor.

Labor’s share of GDP declined from 64% in 1974 to 56% in 2024, a four-decade grind lower driven by globalization, automation, and the steady erosion of worker bargaining power. In the four years since AI began its exponential improvement, that has dropped to 46%. The sharpest decline on record.

The output is still there. But it’s no longer routing through households on the way back to firms, which means it’s no longer routing through the IRS either. The circular flow is breaking, and the government is expected to step in to fix that.

[

](https://substackcdn.com/image/fetch/$s_!YdQQ!,f_auto,q_auto:good,fl_progressive:steep/https%3A%2F%2Fsubstack-post-media.s3.amazonaws.com%2Fpublic%2Fimages%2F29c19580-7ac9-4287-aea8-a8341f320152_1600x762.png)

As in every downturn, outlays rise just as receipts fall. The difference this time is that the spending pressure is not cyclical. Automatic stabilizers were built for temporary job losses, not structural displacement. The system is paying benefits that assume workers will be reabsorbed. Many will not, at least not at anything like their prior wage. During COVID, the government freely embraced 15% deficits, but it was understood to be temporary. The people who need government support today were not hit by a pandemic they’ll recover from. They were replaced by a technology that continues to improve.

The government needs to transfer more money to households at precisely the moment it is collecting less money from them in taxes.

The U.S. won’t default. It prints the currency it spends, the same currency it uses to pay back borrowers. But this stress has shown up elsewhere. Municipal bonds are showing worrying signs of dispersion in year-to-date performance. States without income tax have been okay, but general obligation munis issued by states dependent on income tax (majority blue states) began to price in some default risk. Politicos caught on quickly, and the debate over who gets bailed out has fallen along partisan lines.

The administration, to its credit, recognized the structural nature of the crisis early and began entertaining bipartisan proposals for what they’re calling the “Transition Economy Act”: a framework for direct transfers to displaced workers funded by a combination of deficit spending and a proposed tax on AI inference compute.

The most radical proposal on the table goes further. The “Shared AI Prosperity Act” would establish a public claim on the returns of the intelligence infrastructure itself, something between a sovereign wealth fund and a royalty on AI-generated output, with dividends funding household transfers. Private sector lobbyists have flooded the media with warnings about the slippery slope.

The politics behind the discussions have been grimly predictable, exacerbated by grandstanding and brinksmanship. The right calls transfers and redistribution Marxism and warns that taxing compute hands the lead to China. The left warns that a tax drafted with the help of incumbents becomes regulatory capture by another name. Fiscal hawks point to unsustainable deficits. Doves point to the premature austerity imposed after the GFC as a cautionary tale. The divide is only magnifying in the run up to this year’s presidential election.

While the politicians bicker, the social fabric is fraying faster than the legislative process can move.

The Occupy Silicon Valley movement has been emblematic of wider dissatisfaction. Last month, demonstrators blockaded the entrances to Anthropic and OpenAI’s San Francisco offices for three weeks straight. Their numbers are growing, and the demonstrations have drawn more media coverage than the unemployment data that prompted them.

It’s hard to imagine the public hating anyone more than the bankers in the fallout of the GFC, but the AI labs are making a run at it. And, from the perspective of the masses, for good reason. Their founders and early investors have accumulated wealth at a pace that makes the Gilded Age look tame. The gains from the productivity boom accruing almost entirely to the owners of compute and the shareholders of the labs that ran on it has magnified US inequality to unprecedented levels.

Every side has their own villain, but the real villain is time.

AI capability is evolving faster than institutions can adapt. The policy response is moving at the pace of ideology, not reality. If the government doesn’t agree on what the problem is soon, the feedback loop will write the next chapter for them.

The Intelligence Premium Unwind

For the entirety of modern economic history, human intelligence has been the scarce input. Capital was abundant (or at least, replicable). Natural resources were finite but substitutable. Technology improved slowly enough that humans could adapt. Intelligence, the ability to analyze, decide, create, persuade, and coordinate, was the thing that could not be replicated at scale.

Human intelligence derived its inherent premium from its scarcity. Every institution in our economy, from the labor market to the mortgage market to the tax code, was designed for a world in which that assumption held.

We are now experiencing the unwind of that premium. Machine intelligence is now a competent and rapidly improving substitute for human intelligence across a growing range of tasks. The financial system, optimized over decades for a world of scarce human minds, is repricing. That repricing is painful, disorderly, and far from complete.

But repricing is not the same as collapse.

The economy can find a new equilibrium. Getting there is one of the few tasks left that only humans can do. We need to do it correctly.

This is the first time in history the most productive asset in the economy has produced fewer, not more, jobs. Nobody’s framework fits, because none were designed for a world where the scarce input became abundant. So we have to make new frameworks. Whether we build them in time is the only question that matters.

But you’re not reading this in June 2028. You’re reading it in February 2026.

The S&P is near all-time highs. The negative feedback loops have not begun. We are certain some of these scenarios won’t materialize. We’re equally certain that machine intelligence will continue to accelerate. The premium on human intelligence will narrow.

As investors, we still have time to assess how much of our portfolios are built upon assumptions that won’t survive the decade. As a society, we still have time to be proactive.

The canary is still alive.

Subscribe now

Acknowledgements: Thanks to Sam Koppelman of * for his help with proofreading. Our co-author, Alap Shah of LOTUS, contributed the idea for this piece - CitriniResearch wrote this party, but he has written others in a series called the Intelligence Explosion, we highly recommend reading it. You can find it here.*

Citrini Research is a reader-supported publication. To receive new posts and support my work, consider becoming a free or paid subscriber.

📋 讨论归档

讨论进行中…