$10B → $44B in Four Months — All-In on AI's Compounding Year
四个月，从 100 亿到 440 亿——All-In 拆这周的 AI 大事
Anthropic's annualised revenue tripled in Q1 2026, then added another $14B in April. The All-In panel — Chamath, Sacks, Calacanis, with guest Brad Gerstner — uses the number as the entry point to four stories that reshaped AI's center of gravity in a single week: Elon's compute deal, the monopoly question nobody had to ask before, the FDA-for-AI panic, and what the market math actually shows.
Anthropic 的年化收入在 2026 年第一季度直接翻了三倍，四月又往上加了 140 亿美元。All-In 四人围着这串数字，把这一周里搅动 AI 风向的四件事一次说透：Elon 把 Colossus 1 整租给了 Anthropic、那个五个月前还没人问的"AI 垄断"问题、白宫"AI 版 FDA"的恐慌，还有市场账本里真实的那笔数。这一期没有 Friedberg——他病了，由 Brad Gerstner 顶上。
Chapter 1 — Elon's Anthropic Deal
第一章 · Elon 与 Anthropic 的交易
Five days after Chamath suggested on-air that Elon and Dario should "do a deal tomorrow," Elon leased the entirety of Colossus 1 — his first data center — to Anthropic. xAI moved its model training to Colossus 2. The panel agrees the deal does more for both companies than any earnings call could: it solves Anthropic's compute bottleneck and reframes xAI as a hyperscaler.
五天前 Chamath 还在节目里说 Elon 和 Dario "明天就该谈成"。五天后，Elon 把自家第一个数据中心 Colossus 1 整租给了 Anthropic，xAI 自己挪到更新的 Colossus 2 去训练。在场四个人看法一致——这笔交易比两家任何一次财报都管用：Anthropic 的算力瓶颈被一次性拆掉，xAI 也借此把自己升格成超大规模云厂。
Colossus 1, fully leased
Colossus 1 整租给 Anthropic
Anthropic gets 220,000+ Nvidia GPUs and 300+ megawatts of power — Elon's first data center, in one transaction. xAI relocates to Colossus 2, the newer Blackwell-equipped site. Within days of the announcement, Claude doubles its Claude Code rate limits, removes peak-usage caps for paid tiers, and raises API throughput on Opus models. The supply constraint Anthropic spent two years fighting evaporates in a single deal.
一签字，Anthropic 一次性拿到 22 万多颗 Nvidia GPU 和 300 兆瓦以上电力——Elon 第一个数据中心整体出租。xAI 把自家训练挪到更新的 Colossus 2（Blackwell 架构）。消息一出几天内，Claude Code 速率上限翻倍、付费用户的高峰封顶解除、Opus 模型 API 吞吐拉高。Anthropic 啃了两年的算力天花板，被这一单直接推没了。
"Elon Web Services" arrives
"Elon 云"正式登场
Brad calls it EWS — Elon Web Services. With Colossus 1, Elon is now the fourth hyperscaler, competing with AWS, Azure, and GCP. He has three named facilities: Colossus, Macrohard, and Macrohardrr (Brad on-air: "Macro hard, and macro harder") — 1.2 gigawatts on Blackwell alone. Brad estimates the move adds $4–5B of incremental revenue in 2026 on top of analyst mid-$20B forecasts. The strategic upside: it subsidises xAI's training spend without forcing X.AI to deliver near-term revenue.
Brad 给它起了个名字——EWS，Elon Web Services。Colossus 1 一签，Elon 直接成了第四家超大规模云厂，和 AWS、Azure、GCP 同台。三座基地：Colossus、Macrohard、Macrohardrr（Brad 节目里说成"Macro hard, and macro harder"），光 Blackwell 那部分就已有 1.2 吉瓦。Brad 算的账：2026 年这一单能在分析师预估的二十几亿美元收入之上再叠加 40 到 50 亿，正好用现金流养着 xAI 训练那边的烧钱，不必逼 X.AI 短期内出业绩。
"Nobody better at converting electrons to tokens"
"把电变成 token 这事，地球上没人做得过他"
Most data-center operators are good at compute or good at power, not both. Elon's existing competencies — Tesla's factory build-out and Solar City's battery/solar stack — collapse the two into one operating company. Brad: "There's nobody better on planet Earth than Elon at converting electrons to tokens." The activist protests stalling ~50% of this year's 9 GW of new US capacity make the people who already built ahead of the wave even more valuable.
绝大多数搞数据中心的，要么会做算力，要么懂电力，很少两边都行。Elon 不一样：特斯拉那套建工厂的本事 + Solar City 留下来的电池和太阳能堆栈，正好把两件事压缩成一家公司能同时做。Brad 一句话定性："把电变成 token 这事，地球上没人做得过他。"美国今年本来要上 9 吉瓦新产能，但有一半被环保活动人士卡住——这种行情下，谁早早把基础设施搭好了，谁就更值钱。
Blunting the SpaceX bear case
把 SpaceX IPO 的看空说辞拆掉
The biggest red-team on SpaceX's road-show valuation has been the on-the-come value attributed to orbital data centers. Chamath's read: by landing terrestrial capacity now, Elon de-risks the orbital thesis — even if space slips two quarters, the structural hyperscaler revenue is real. Brad expects the SpaceX IPO to trade at 40–50× revenue, against Mag 5 trading at ~25× earnings. The premium is paid for one thing the rest of the Mag 7 stopped delivering: future-pipeline innovation that markets believe.
SpaceX 路演里被空头戳得最狠的一点，就是估值里那块"在轨数据中心"是先记账后兑现。Chamath 的解读是：Elon 现在把地面算力坐实了，等于把这个故事的风险垫了一层——就算太空那部分晚两个季度，超大规模云这块的真金白银先到账了。Brad 据此判断 SpaceX 上市会按市销率 40–50 倍定价，对比 Mag 5 只有 25 倍左右的市盈率。这溢价买的不是当下利润，而是 Mag 7 其他几家最近交不出的那样东西——市场愿意相信的未来管线。
Distributed compute, all the way home
分布式算力，一路压到家里
Chamath's projection: the race goes from factories to data centers to homes. Tesla's Powerwall could ship with compute inside it. Cars are already mobile compute platforms. Starlink turns every connected home into a node. A separately announced deal between PY Homes and SPAN puts Nvidia GPU clusters next to new construction. The end state — and the part "where nobody can complain" — is compute in orbit, an idea the Elon-Anthropic press release also acknowledged.
Chamath 给出下一步：竞速会从工厂打到数据中心，再打进千家万户。特斯拉的 Powerwall 里完全可以塞算力，车本来就是移动的算力台，Starlink 把每个联网的家都变成节点。前两天宣布的 PY Homes × SPAN 合作，干脆把 Nvidia GPU 集群直接装到新房旁边。终点是哪儿？是"谁也没法抗议"的那片地方——在轨数据中心。这次 Elon 和 Anthropic 的联合声明里，也悄悄提到了太空。
Chapter 2 — The Next AI Monopoly?
第二章 · 下一个"史上最大垄断"？
The deal solves Anthropic's compute problem. That makes the monopoly question — one Silicon Valley wasn't seriously asking five months ago — unavoidable. Anthropic grew from $10B ARR on January 1 to $30B on March 31 to $44B by the end of April. Sacks's projection if the trajectory holds: $100B exiting 2026, $1 trillion ARR in 2027. The rest of the Mag 7 — Apple, Microsoft, Alphabet, Amazon, Nvidia, Meta, Tesla — combined for roughly $2.3T of 2025 GAAP revenue. The contest is over whether the panel is watching the most extreme monopoly formation in commercial history or jumping the gun.
交易解决了 Anthropic 的算力问题，但同时把另一个问题推到桌面上：垄断。五个月前没人正经问，现在躲不过。Anthropic 的年化收入：1 月 1 号 100 亿，3 月底 300 亿，4 月底 440 亿。Sacks 顺着这条线推算：2026 年底冲到 1000 亿，2027 年破万亿。再看 2025 年 Mag 7 的合计收入——苹果、微软、Alphabet、亚马逊、Nvidia、Meta、特斯拉加起来大约 2.3 万亿美元。问题随之而来：这是商业史上最极端的一次垄断形成，还是大家太早下结论？四个人在这一段吵得最激烈。
The exponential nobody priced in
没人提前给它定过价的指数曲线
Anthropic's ARR went from $10B to $44B in four months. From January to March it tripled. In April alone it added another $14B. "Nobody in Silicon Valley has ever seen anything like it," Sacks says, "and all we do in Silicon Valley is deal with exponentials." The conventional wisdom entering 2026 was that 10x annual growth couldn't be sustained at this scale. Four months later the only debate left is whether they hit a trillion in 2027 or in 2028.
从 100 亿做到 440 亿，Anthropic 只用了四个月。一季度翻三倍，光是 4 月一个月又往上加了 140 亿。Sacks 一句话："硅谷从没见过这个量级的东西，而我们硅谷天天就是在处理指数曲线。" 进 2026 年的时候，业内共识是这种规模下不可能再延续年化十倍的增速，结果四个月之后，剩下的争论只是"它是 2027 年破万亿，还是 2028 年破万亿"。
"The biggest monopoly in human history"
"人类史上最大的一次垄断"
Sacks's frame is monopoly. Dario's is AGI. "Dario calls it AGI. I call it the biggest monopoly in human history." The math: if Anthropic hits $1T ARR while still growing exponentially, it surpasses the combined revenue of every other tech major. Google does about $100B per quarter. Anthropic — by Sacks's projection — would be doing roughly the same amount per quarter within 18 months, but growing at 1,000% YoY instead of 20%.
同一件事，Dario 用的词是 AGI，Sacks 用的词是垄断。"Dario 叫它 AGI，我叫它人类史上最大的一次垄断。"账面账：如果 Anthropic 真按指数往一万亿年化跑，那它的体量会超过其他所有科技巨头加起来。Google 现在大概一个季度 1000 亿收入；按 Sacks 的推算，再过 18 个月，Anthropic 一个季度也能做到这个数——只不过 Google 同比增速 20%，Anthropic 是 1000%。
Brad's counter: starting line, not finish line
Brad 的反驳：刚起跑，谈不上终点
Brad pushes back from the seat Friedberg usually occupies: "It's ridiculous to think of this as a monopoly. We haven't even left the starting gate of AI." GAAP monthly revenue in March for Anthropic and OpenAI was roughly comparable. Google, Amazon, and Microsoft are producing $100B in free cash flow per year each to justify incremental investment. The last thing Brad wants: "DC trying to preemptively get in the game of picking winners and losers at the starting line."
Brad 坐着 Friedberg 平时的位置，给 Sacks 兜头一盆冷水："AI 才刚出发，就谈垄断？荒唐。" 3 月份按 GAAP 算月度收入，Anthropic 和 OpenAI 其实差不多。Google、亚马逊、微软每家每年都还在用 1000 亿美元级别的自由现金流来证明这笔投入划得来。Brad 最怕的就是另一件事："华盛顿在起跑线上抢着替这场比赛挑赢家输家。"
Imagine John D. Rockefeller in PR class
假如洛克菲勒上过公关课
Sacks's parable runs ~90 seconds. Picture Rockefeller, but better at public relations. He calls his company Safe Oil instead of Standard Oil. Kerosene is dangerous — it can light a house or burn one down — so Rockefeller publicly calls for the creation of a federal agency to regulate kerosene safety. The country debates wick thickness and licensing standards. "People might even have called Rockefeller an effective altruist because of course he was so concerned about the safety of his product." While they argue, he builds the richest, most powerful monopoly of all time. The geometry, Sacks argues, matches what's happening now.
Sacks 这一段讲了大约九十秒。设想一下：洛克菲勒不变，但公关能力突然好起来。他把公司从"标准石油"改名叫"安全石油"。煤油本来就危险——能点亮一座房，也能烧毁一座城——所以洛克菲勒主动呼吁国会成立一个联邦机构，专门监管煤油安全。全国上下开始热烈讨论灯芯该多粗、许可证怎么发。"说不定还会有人叫他'有效利他主义者'呢——人家不是这么关心产品安全嘛。"大家忙着吵法规的时候，他不动声色地把美国史上最赚钱、最有权的垄断盖起来了。Sacks 说，今天 AI 安全那一套，几何形状是一样的。
The $1T coding TAM
一万亿美元的"写代码"市场
Why the projection isn't laughable. Roughly $1 trillion a year is spent globally on software developers and adjacent tooling — the addressable market Anthropic's coding tokens are eating into. Sacks: "I'm not saying they eat that entire market, but I can easily see the market for software doubling." Coding tokens 10× or 100× the value of the underlying developer-hour. If Anthropic captures even a meaningful fraction at exponential growth, a trillion ARR is on the table. It's not the TAM that constrains them.
为什么万亿这个数字并不离谱：全球花在软件工程师及周边工具上的开销，一年大约就是 1 万亿美元——而 Anthropic 的代码 token 正在啃的，就是这块市场。Sacks 说："我不是说他们全吃下来，但我完全能想象软件这个市场会翻一倍。"因为代码 token 的产出，能把一个开发者小时的价值放大十倍、一百倍。Anthropic 只要按指数曲线啃下其中一块像样的份额，万亿年化就在桌面上。卡他们的从来不是 TAM。
The anti-competitive flag
那个值得警惕的小动作
The data point that turned Sacks's monopoly thesis sharper: Anthropic banning OpenAI from using its models. "What conceivable reason did Anthropic have for banning OpenAI using its models? That is anti-competitive, is it not?" Brad's response: "I would double-click on it." Tech markets historically consolidate into monopolies or duopolies; the move to lock competitors out at the API layer, while small in dollar terms, fits the pattern. Sacks's claim is not that Anthropic is a monopoly today — it's that the trajectory and the tactics together make the next 18 months the moment to decide.
让 Sacks 的"垄断论"更锋利的一根刺：Anthropic 把 OpenAI 拉黑了，不许它用自家模型。"Anthropic 凭什么禁止 OpenAI 使用自己的模型？这不就是反竞争行为吗？"Brad 听完只回了一句："这事我会再多看一眼。"科技行业的历史路径就是这样——要么走向垄断，要么走向寡头。从 API 层把对手挡在外面，金额可能不大，但姿势对得上模板。Sacks 的论点不是"Anthropic 今天就是垄断"，而是趋势加打法这两件事叠在一起，接下来的 18 个月就是决定时刻。
Chapter 3 — "FDA for AI" Panic
第三章 · "AI 版 FDA"恐慌
The same week the deal landed, the New York Times reported the White House was drafting an executive order to create an AI working group that would review new models for safety — analogised on-air, by NEC Director Kevin Hassett, to FDA drug approval. The catalyst, per the report: Anthropic's Mythos model, which spooked the administration with its cyber capabilities. Within 24 hours, Susie Wiles, the White House Chief of Staff, issued a statement that the panel reads as a walk-back. The story is messier than the headline.
交易落地那一周，《纽约时报》同时爆料：白宫正在起草行政令，要成立一个 AI 工作组，对新模型逐一做安全审核——美国国家经济委员会主任 Kevin Hassett 当天上 Fox 节目时把这套类比成 FDA 审药。报道说，导火索是 Anthropic 那款让政府不安的"Mythos"模型，因为其网络攻防能力。结果不到 24 小时，白宫幕僚长 Susie Wiles 发了一份声明，四个人都读出"收回"的意思。原始故事其实比标题复杂得多。
What the NYT actually reported
《纽约时报》到底写了什么
The Times said Trump was considering an executive order to create an AI working group of tech executives and government officials to "examine potential oversight procedures, including a review process for new AI models." Kevin Hassett confirmed on Fox Business that the administration was "studying possibly an executive order ... so that future AIs that potentially create vulnerabilities should go through a process so that they're released to the wild after they've been proven safe, just like an FDA drug." That single phrase — "just like an FDA drug" — is what set Silicon Valley on fire.
《纽约时报》写的是：特朗普正在考虑签一份行政令，组建由科技高管和政府官员构成的 AI 工作组，"研究可能的监管程序，其中包括对新 AI 模型的审核流程"。Kevin Hassett 当晚在 Fox Business 印证："我们在研究一个行政令，让未来这些可能制造漏洞的 AI 模型走个流程，'就像 FDA 审药一样'，被证明安全了再放出来。"——就是 "just like an FDA drug" 这一句话，把整个硅谷点炸了。
What Brad and Sacks heard from Hassett
Brad 和 Sacks 私下问到的版本
Both Brad and Sacks called Hassett directly after the clip ran. Their report: nobody senior in the administration supports an FDA-style approval regime. Hassett's framing — Brad's paraphrase — was about coordinating with labs to harden government systems, not pre-approving releases. Susie Wiles published a statement the same night that the panel reads as a clear walk-back. The pre-release-approval reading, Sacks argues, was "fake news" — driven less by White House intent than by Andrew Ross Sorkin's editorial gloss carrying the story into a viral framing.
那段录像一出，Brad 和 Sacks 各自直接拨给了 Hassett。两人转述的口径一致：白宫高层里没人真支持"FDA 式审批"。Hassett 自己的意思（Brad 转的话）是和实验室对接、加固政府系统，不是发布前先盖章。当晚 Susie Wiles 发的那份声明，四个人都解读为明确的"收回"。Sacks 干脆说，"发布前审批"那一版完全是"假新闻"——背后推手与其说是白宫意图，不如说是 Andrew Ross Sorkin 的评论框架先把故事推成了热搜款。
The real issue: cyber, not safety theater
真正该担心的是网络攻防，不是安全大戏
The legitimate concern hiding under the FDA noise. Models like Anthropic's Mythos, OpenAI's offering on Spud / GPT 5.5, and (within 3–6 months) all major frontier labs and Chinese open-source models will be cyber-capable — meaning they can scan code bases, find vulnerabilities, and chain them. Sacks: "That is a certainty, because the same capabilities you use for cyber defense can also be used for cyber offense." The solution isn't a model approval office in DC. It's getting these tools into US cyber-security companies — CrowdStrike, Palo Alto Networks, AI-powered pen-test startups — fast enough to harden systems before attackers catch up.
FDA 这层噪音底下，藏着一桩真问题：模型已经会做网络攻防了。Anthropic 的Mythos、OpenAI 跑在Spud / GPT 5.5 上的版本，再过三到六个月，几乎所有前沿大厂以及中国开源模型，都会具备这能力——会扫代码库、找漏洞、把漏洞串起来。Sacks 一句话定性："这事板上钉钉，因为防御和攻击用的是同一套能力。"解药不是华盛顿开个新审批办公室，而是赶紧把这些工具塞到 CrowdStrike、Palo Alto Networks，以及一批主打 AI 渗透测试的初创手里，赶在攻击方反应过来之前把系统加固掉。
KYC and logging, not pre-approval
实名 + 日志，而不是事前批准
The panel converges on a workable middle ground for the preview period of frontier models: know-your-customer at the API layer, plus suspicious-use logging shared with government. "You should not give that to a company or an actor you don't know who they are." In practice, Sacks says, the labs are already doing it — every frontier lab tracks API usage for anti-distillation reasons and coordinates with the government on suspicious patterns. Formalising the coordination during preview windows is reasonable. Pre-release approval is not.
围绕前沿模型的预览期，四个人能谈拢的方案其实很具体：API 层做实名 / KYC，加上把可疑使用的日志共享给政府。Chamath 一句话："你不能把这个能力给一家你都不知道是谁的公司或人。"Sacks 补一刀——这事其实早在做了，每家前沿实验室出于反蒸馏的需要，都在追踪 API 用量，可疑模式也跟政府对得上号。把这套预览期的对接流程做成制度，合情合理。但"发布前要先批"这事，不行。
Why this is happening now
为什么偏偏现在出事
Chamath's diagnosis: there's a real vibe shift against tech oligarchs on Main Street, and it's seeping into Washington. Regulations are coming — "worse under a Democratic regime, but some form of oversight will exist under a Republican one." He gives the tech-leadership community a D-minus, trending to F on broad-based-benefit messaging. The doomers haven't won the argument; the tech leaders have lost it by default. The Sanders moratorium on data centers is the floor case. The KYC-plus-logging coordination is the ceiling case. The middle is where this lands.
Chamath 给的诊断：美国普通人对"科技寡头"这套的反感是真的有风向变了，而这股风已经开始往华盛顿渗。监管一定会来——"民主党执政会更狠，但即便是共和党执政，某种形式的监管也跑不掉。"他给科技领袖在"广泛受益叙事"这门考试上打了D−，正往 F 滑。AI 末日派并没有真把这一仗打赢，是科技圈自己默默把它输掉了。下限是 Bernie Sanders 那种数据中心禁令，上限是 KYC + 日志这种协调方案。结果会落在中间。
Chapter 4 — Trading the AI Boom
第四章 · 怎么交易这场 AI 上行
What the cloud earnings actually show, what the multiples actually say, and where Chamath thinks the reckoning hits. The S&P 500 is at all-time highs and up only 8% year-to-date — not by itself a bubble signal. But the underlying question — has AI capex earned its return yet — is the one Chamath and Sacks disagree about most sharply on the record.
把云厂财报、估值倍数、以及 Chamath 心里那个"算账日"都摊开看。S&P 500 创了新高，但年初至今只涨了 8%——单看这一项，谈不上泡沫。真正分歧的是底下那个问题：这一轮 AI 资本开支，到底有没有赚回来？Chamath 和 Sacks 在节目里这一段顶得最厉害。
Hyperscaler quarter, in three numbers
三大云厂的一个季度，三个数字
AWS at a $150B run rate, growing 28% YoY on a number that size. Azure at $108B, growing 39%. Google Cloud at $80B, growing 63% — a number Brad calls "stunning everybody." Mag 5 combined headcount up only ~3% over three years while revenue accelerates. That mismatch — top-line acceleration against flat headcount — is the operating-leverage signal that pulls the AI-ROI debate forward.
AWS 年化 1500 亿美元，规模这么大，同比还有 28%；Azure 1080 亿，39%；Google Cloud 800 亿，63%——Brad 说这数字"让所有人傻眼"。再看人头：Mag 5 这三年合计人员增加大约只有 3%，收入却在加速。营收冲、人头不冲——这种错位才是真正撬动 AI 投入产出之争的运营杠杆信号。
Margin expansion: signal or financial engineering?
利润率走高：真信号，还是财务手段？
S&P 500 operating margins moved from ~11% in Q1 2023 to ~13% in the current print — a 200-basis-point lift. Brad reads it as AI-driven operating leverage. Chamath flags the alternative: it could be the same fitness-era cost discipline that started in 2022–23. The consensus forecast says margins keep expanding for two more years. Whether the cause is AI or the financial-engineering playbook is, by Chamath's framing, the entire question of whether AI capex pays off.
S&P 500 的经营利润率，从 2023 年 Q1 的 11% 涨到现在大约 13%——整整 200 个基点。Brad 把这个读成 AI 带来的运营杠杆。Chamath 立刻给另一种解释：这同样可能是 2022–23 那波"裁员、收紧"的余波，跟"AI 红利"无关。市场一致预期是这波利润率扩张还能再走两年。这扩张到底是 AI 给的，还是财务手段挤的——按 Chamath 的话说，整个"AI 投入回报问题"就压在这一道单选题上。
75% of GDP, and a 29-of-39 issue
75% 的 GDP 增量来自它，民意榜上却排第 29
Sacks's macro view in two numbers. AI is ~75% of Q1 2026 US GDP growth by his count — a deflationary force on cost-of-living, a blue-collar construction boom (25–30% wage gains in build zones), and the engine of the current expansion. And yet in voter-priority polling AI ranks 29 of 39 issues. Cost-of-living is #1, the economy is #2. The implication: the effects of AI are popular even when AI itself is not — "if the media would honestly report what was happening."
Sacks 用两个数字描述宏观面：按他算，AI 贡献了 2026 年 Q1 美国 GDP 增量的大约 75%——压低生活成本、催出蓝领建筑热（数据中心施工区工资涨 25–30%）、是整个扩张周期的发动机。但有意思的是，民意调查里"选民最关心的 39 件事"中，AI 排到了第 29 位。第一名是生活成本，第二名是经济。换句话说：AI 本身不讨喜，但 AI 带来的结果其实是讨喜的——"前提是媒体如实报道"。
Not a bubble — yet
至少现在还不像泡沫
Brad's defense against bubble framing, told in P/Es: Meta at 17×, Nvidia 19×, Microsoft 20×, Google 24× fully-taxed GAAP earnings. The memory stocks the panel is bullish on: SK Hynix at 5×, Samsung 6×, Micron 7×. S&P 500 up only 8% YTD. "This is not the stuff that bubbles are made of." The bull case requires those multiples to be defensible; the bear case requires the underlying ROI to show up in the operating margins they're supposedly producing.
Brad 反驳"泡沫论"的方式很简单，就是把市盈率摆出来：Meta 17 倍，Nvidia 19 倍，微软 20 倍，Google 24 倍，全是完税后的 GAAP 利润。他重仓的存储芯片更便宜：SK 海力士 5 倍，三星 6 倍，美光 7 倍。S&P 500 年初至今只涨 8%。一句话："这不是泡沫该有的样子。"多头需要这些倍数撑得住，空头则只能赌：底下那块所谓"AI 带来的利润率"，迟早得真到账。
Chamath's fork in the road, ~500 days out
Chamath 留下的岔路口：大约 500 天后见
The unresolved question the episode hands to the reader. "You can't will profits to go up. Anheuser-Busch eventually has to sell more beer. Nike has to sell more shoes." In roughly 500 days, the panel's spend-X-receive-Y equation has to balance. Two paths: opex shrinks while revenues stay flat (margins expand by headcount contraction) — or revenues grow while opex stays flat (margins expand through scale). The first is a reckoning for workforces. The second is the boom continuing to compound. "It doesn't mean it's not coming. It just means it hasn't shown up yet." Five hundred days. Then it has to.
这一期最后留下的问题，是 Chamath 抛给读者的："利润不是你想涨它就涨。百威最终得多卖啤酒，耐克最终得多卖鞋。"大约还剩 500 天，"花了 X 该收回 Y"这道方程必须解开。两条路：要么营业成本砍下去、收入不动（靠裁员撑出利润率扩张）；要么收入往上跑、营业成本不动（靠规模撑出利润率扩张）。前一条路是工作岗位的清算日；后一条路是这场上行继续复利下去。Chamath 自己留的话："这不代表它不会来，只是说，它现在还没到。"500 天之后，它必须到。