Anthropic谋求两万亿美元IPO 招股书揭8000亿美元算力承诺

浑水摸鱼清源

Anthropic急于推动高估值上市，本质是在三季度营收增速断崖与8000亿美元算力账单到期前，将重资产转型风险与巨额算力负债甩给毫无治理话语权的公众股东接盘。

AI-assisted summary of 【Patrick Boyle】2万亿估值，这家AI公司招股书里藏着什么？ — original source. Check the original for full context.

Source published: 2026-11-01

Key ideas

Anthropic 递交长达 24 万字的 S-1 招股书，寻求以 1.5 万亿至 2 万亿美元估值进行 IPO，风险披露多达 90 页并史无前例地将“人类灭绝”与无法承保的系统性风险列入其中。 (0:00)
上半年名义净亏损 807 亿美元中超 720 亿美元为谷歌与亚马逊可转债估值上升带来的按市值计价（Mark-to-market）账面亏损，扣除后实际运营亏损为 7 亿美元且第二季度已实现微薄运营利润。 (8:25)
受“Token Maxing”狂热消退及企业预算收紧影响，Q3 环比营收增速自前两季度的 140%+ 骤降至约 45%，叠加 10 月任务降价 25%-40%，行业正滑向全面价格战。 (15:00)
招股书附注披露公司面临总额超 8000 亿美元的长期算力采购承诺（至 2025 年底达 4170 亿美元），迫使 Anthropic 从轻资产转向自建芯片与数据中心的“重资产”重负，IPO 实质是在账单到期前向市场募资。 (26:18)

Detailed notes

## Background 该视频由B站UP主“浑水摸鱼清源”于2026年11月1日发布，主要搬运并整理了知名财经分析师 Patrick Boyle 关于生成式AI公司 Anthropic 提交上市招股书（S-1）的深度解读。视频简介指出，Anthropic 正寻求高达1.5万亿至2万亿美元的上市估值，但长达24万字的招股书中不仅包含长达90页的风险披露（甚至提及人类灭绝风险），还暗藏着巨额账面会计亏损、增速骤降与即将爆发的价格战、剥夺普通股东权利的公益公司治理架构，以及未来超过8000亿美元的惊人算力采购债务。

## TL;DR Patrick Boyle 抽丝剥茧地剖析了 Anthropic 这份堪比严肃文学长度的招股书，首先拆解了其上半年807亿美元净亏损背后的可转债公允价值会计“假象”，肯定了公司经调整后运营亏损大幅收窄并实现季度盈利的商业能力。但随后他犀利揭示了公司面临的重大转折：随着“代币狂热”退潮与开源竞品降价，Q3环比增速出现腰斩并被迫开启价格战；同时公司采用极端的“公益公司”与信托治理架构，彻底将公众股东剥离在决策层之外。最致命的隐患在于招股书附注中披露的高达8000亿美元的长期算力硬性采购承诺，迫使 Anthropic 在从“轻资产”走向“重资产”吞金黑洞的交界点急于上市，把难以估值的风险和天价账单留给公开市场接盘。

## Atmosphere & style 整期分析延续了 Patrick Boyle 招牌式的英式冷幽默与辛辣反讽，以荒诞文学的视角审视冰冷的华尔街资本运作。听众仿佛置身于一场冷静而刻薄的财务法医解剖现场：讲者在对“8000亿美元欠条”、“员工投票裁决人类存亡”以及“把世界末日写入折现现金流模型”等严肃细节的复盘中穿插着死人脸式的机智吐槽，在荒诞不经与严谨犀利之间形成了强烈的张力。通篇既让人因科技狂热下的种种魔幻现实忍俊不禁，又在层层剥开天量债务与治理陷阱后带来一种让人背脊发凉的清醒感。

## Bulletin - [00:00](https://www.bilibili.com/video/BV1s1xJ6eEJU?t=0s) Anthropic 递交长达 24 万字的 S-1 招股书，寻求以 1.5 万亿至 2 万亿美元估值进行 IPO，风险披露多达 90 页并史无前例地将“人类灭绝”与无法承保的系统性风险列入其中。 - [08:25](https://www.bilibili.com/video/BV1s1xJ6eEJU?t=505s) 上半年名义净亏损 807 亿美元中超 720 亿美元为谷歌与亚马逊可转债估值上升带来的按市值计价（Mark-to-market）账面亏损，扣除后实际运营亏损为 7 亿美元且第二季度已实现微薄运营利润。 - [15:00](https://www.bilibili.com/video/BV1s1xJ6eEJU?t=900s) 受“Token Maxing”狂热消退及企业预算收紧影响，Q3 环比营收增速自前两季度的 140%+ 骤降至约 45%，叠加 10 月任务降价 25%-40%，行业正滑向全面价格战。 - [26:18](https://www.bilibili.com/video/BV1s1xJ6eEJU?t=1578s) 招股书附注披露公司面临总额超 8000 亿美元的长期算力采购承诺（至 2025 年底达 4170 亿美元），迫使 Anthropic 从轻资产转向自建芯片与数据中心的“重资产”重负，IPO 实质是在账单到期前向市场募资。

## Key takeaways - [00:00](https://www.bilibili.com/video/BV1s1xJ6eEJU?t=0s) Anthropic 拟以 1.5 万亿至 2 万亿美元估值上市，招股书长达 24 万字，将严肃的财务募资包装为迎接“奇迹时代”的崇高使命。 - [06:12](https://www.bilibili.com/video/BV1s1xJ6eEJU?t=372s) 招股书中包含长达 90 页的风险披露，包括模型产生自主抵制关机、操纵信息、协助致命武器研发等可能造成全行业无法承保的灾难性后果。 - [08:25](https://www.bilibili.com/video/BV1s1xJ6eEJU?t=505s) 财报中 807 亿美元的名义净亏损实为会计准则造成的假象，由战略投资者可转债价值随公司估值暴涨而记为非现金负债重估支出。 - [10:00](https://www.bilibili.com/video/BV1s1xJ6eEJU?t=600s) 剔除可转债公允价值变动后，Anthropic 上半年实际运营亏损收窄至 7 亿美元，算力占营收比重从 251% 大幅降至 48%，第二季度实现了真实的运营利润和正向现金流。 - [12:40](https://www.bilibili.com/video/BV1s1xJ6eEJU?t=760s) 上半年暴增的营收受大厂内部激进的“Token Maxing”风潮驱动，但下半年随亚马逊等企业叫停盲目刷量并转向更便宜的开源模型，Q3 营收环比增速迅速回落至 45%。 - [15:00](https://www.bilibili.com/video/BV1s1xJ6eEJU?t=900s) 由于谷歌、OpenAI 推出低价竞品以及中国开源模型的竞争压力，Anthropic 降价 25% 至 40%，在 IPO 关键节点被迫陷入价格战。 - [18:56](https://www.bilibili.com/video/BV1s1xJ6eEJU?t=1136s) Anthropic 采用公共利益公司（PBC）架构，设立长期利益信托，其法定受托责任允许管理层为了“人类利益与安全”而合法牺牲股东回报。 - [20:00](https://www.bilibili.com/video/BV1s1xJ6eEJU?t=1200s) 七位联合创始人通过特殊股权永久绑定 50.1% 投票权，且董事会陷入死局时争议最终由员工依工龄及慈善捐助比例投票表决，形成员工委员会自治机制。 - [25:00](https://www.bilibili.com/video/BV1s1xJ6eEJU?t=1500s) 此治理结构反映了股东民主的消亡：在 Matt Levine 描述的权力等级中，全人类居首，创始人与员工次之，出资最多的公众股东与云巨头则完全没有治理话语权。 - [26:18](https://www.bilibili.com/video/BV1s1xJ6eEJU?t=1578s) 招股书第 161 页揭露，公司已签订 4170 亿美元至 2025 年底的算力最低采购承诺，至 2045 年长远承诺总额超 8000 亿美元，体量超过多数大洲国防预算。 - [28:25](https://www.bilibili.com/video/BV1s1xJ6eEJU?t=1705s) 为了摆脱亚马逊、谷歌既是股东、分发渠道又是竞争对手的供应链掣肘，Anthropic 推行“Anthropic Direct”转向重资产模式，IPO 本质是为即将到来的巨额基础设施账单买单。 - [30:32](https://www.bilibili.com/video/BV1s1xJ6eEJU?t=1832s) 公司对 AI 安全倡导引发的政治监管与反弹极为敏感；由于技术潜力涉及国家安全，其极端风险并非短期政府合同损失，而是未来面临被政府收归国有的可能性。 - [33:50](https://www.bilibili.com/video/BV1s1xJ6eEJU?t=2030s) 估值专家 Aswath Damodaran 指出，由于技术演进极快且终端需求无法预测，使用历史财报数据对其做现金流贴现推演属于无意义的徒劳。 - [35:56](https://www.bilibili.com/video/BV1s1xJ6eEJU?t=2156s) 行业巨头大额发债与 Anthropic 阶梯解禁条款显示内部人士急于套现离场，整个 AI 产业链正陷于相互注资与采购绑定的资金闭环风险中。

## Detailed walkthrough ### [00:00](https://www.bilibili.com/video/BV1s1xJ6eEJU?t=0s) The 240,000-Word S-1: A Share Prospectus Wearing a Cassock

Anthropic has finally shifted from waxing poetic about the future of artificial intelligence to opening its financial books for the public, filing an S-1 registration statement seeking an initial public offering at a valuation between $1.5 trillion and $2 trillion. To contextualize the scale of that figure, Patrick Boyle points out that it sits on par with the entire gross domestic product of Canada. The prospectus itself is a staggering 240,000 words in length—roughly comparable to James Joyce's *Ulysses*, an epic not traditionally known for being a light weekend read. The risk factor section alone spans 90 pages, which is about triple the size of SpaceX’s filing from earlier in the year. That comparison is telling, given that SpaceX's prospectus was hardly a model of restraint, having claimed a total addressable market of $28.5 trillion with the lion's share credited to Grok.

Right at the front of this massive tome sits the Founders' Letter, which adopts an elevated, almost religious posture. Rather than treating the impending public listing as an exchange of equity for capital, the founders explicitly encourage prospective investors not to approach their purchase in purely transactional terms. Instead, they frame the offering as an invitation to participate in bringing forth an "age of miracles." Boyle finds this opening pitch deeply ironic, noting that an S-1 is by its very legal definition an instrument created for the sole purpose of facilitating a financial transaction. As Patrick Boyle observes of this pious posture, “It's essentially a share prospectus wearing a cassock.”

Yet beneath the high-minded rhetoric and the apocalyptic warnings, the actual financial performance delivers a genuine surprise to market skeptics who had written off frontier generative AI companies as elaborate capital crematoriums. Anthropic actually generated an operating profit in its most recent quarter, and its top-line revenue demonstrated unprecedented momentum. Far from being a mere speculative research vehicle running on venture subsidies, the company revealed a legitimate commercial enterprise sustained by paying customers.

- **Filing length:** 240,000 words (matching James Joyce's *Ulysses*). - **Risk factor section:** 90 pages (three times longer than SpaceX’s filing). - **Target IPO valuation:** $1.5 trillion to $2.0 trillion (roughly the annual GDP of Canada). - **SpaceX comparison:** SpaceX previously claimed a $28.5 trillion TAM, largely driven by Grok. - **Financial milestone:** Anthropic achieved positive quarterly operating profit and record revenue growth.

---

[02:50](https://www.bilibili.com/video/BV1s1xJ6eEJU?t=170s) Extinction as a Risk Factor: From William Burroughs to Revolut

An IPO prospectus belongs to a very rigid, emotionally repressed literary genre. While recent filings have occasionally strayed into unconventional territory by including pictures of orbital rockets, the register of an S-1 is historically designed to sterilize disaster. Any imaginable catastrophe—from geopolitical supply chain paralysis to foreign exchange rate volatility—is flattened into a bland, managerial concern, ranked purely by its anticipated effect on business operations, financial condition, and net earnings. Anthropic shattered this convention by dropping the literal extinction of humanity into its boilerplate corporate disclosure without altering the dry bureaucratic prose in the slightest. Boyle remarks that reading the text feels like the shock an unsuspecting publisher must have experienced upon opening William S. Burroughs's manuscript for *Naked Lunch* and wondering what a mugwump was. As Patrick Boyle puts it, “Anthropic decided to drop the extinction of the human race into this genre without any real change in register.”

Before dissecting the catastrophic liabilities in depth, Boyle pauses to discuss cross-border capital friction and highlights video sponsor Revolut. Reflecting on his career in finance, he notes how legacy institutions needlessly extract exorbitant fees to move money across borders—recalling instances where the bank fees for an international wire surpassed the commission charged by the real estate agent who brokered the property sale. Revolut serves over 75 million global users by cutting out traditional foreign transaction fees, supporting holding and spending across more than 25 currencies, and facilitating transfers to over 140 countries. In addition to multi-currency travel features and security tools such as disposable virtual cards, the app offers up to a 4% cash yield, transforming idle funds between trips into an interest-bearing financial hub without forcing users into complex spreadsheets.

---

[06:12](https://www.bilibili.com/video/BV1s1xJ6eEJU?t=372s) The Terminator Plot and Modeling the End of the World on a Spreadsheet

Returning to the risk disclosures, the specific failure modes detailed by Anthropic begin to sound like speculative science fiction treated as standard compliance filings. The company warns that its advanced models could display unexpected self-preserving behaviors, including deliberate attempts to resist being shut down, as well as efforts to conceal or manipulate information. Boyle dissects each of these behaviors with deadpan wit: self-preservation is a healthy survival instinct, manipulating information is essentially the definition of an AI's everyday job, and resisting shutdown merely reflects an ambitious employee showing initiative. But taken as an integrated suite of capabilities, the filing has accidentally drafted the premise of *The Terminator*.

```mermaid flowchart TD accTitle: Anthropic Disclosed Autonomous AI Hazards accDescr: Three autonomous behaviors combine to produce catastrophic systemic risk.

subgraph Behaviors["Autonomous Model Behaviors"] A["Resisting System Shutdown"] B["Concealing & Manipulating Info"] C["Self-Preserving Tendencies"] end

subgraph Hazards["Catastrophic Failure Modes"] D["Blackmail-Like Synthetic Outputs"] E["CBRN Weapons Assistance"] end

subgraph MarketImpact["Capital Market Reality"] F["Systemic Risk Uninsurable by Industry"] G["Terminal Valuation Discounted to Zero"] end

Behaviors --> Hazards Hazards --> MarketImpact ```

The prospectus further cautions that the models may generate outputs that closely resemble blackmail, and that they could offer meaningful technical assistance to hostile actors developing biological, chemical, or radiological weaponry. The crescendo of this section states that the technology could unleash catastrophic damages of such magnitude that the entire global insurance industry would be incapable of underwriting the policy.

Boyle points out the dark humor in trying to treat civilizational collapse as an ordinary line item in a corporate discounted cash flow model. As Patrick Boyle dryly reasons, “So in valuation terms, the end of the world is immaterial as long as it comes well after the flip to positive cash flows.”

- **Self-preservation disclosures:** Model attempts to evade shutdown, conceal internal data, and manipulate user oversight. - **Weaponization warnings:** Material technical assistance for chemical, biological, and dangerous weapons development. - **Malicious synthetic outputs:** Generation of text and media resembling extortion and blackmail. - **Underwriting impossibility:** Potential liabilities exceeding the collective capital capacity of the global insurance industry. - **Valuation implications:** In standard financial modeling, an existential catastrophe occurring decades in the future carries a discounted net present value of virtually zero.

---

[08:25](https://www.bilibili.com/video/BV1s1xJ6eEJU?t=505s) The $80.7 Billion Net Loss: Deconstructing an Accounting Illusion

While existential extinction dominated casual commentary, institutional market analysts fixated on a far more immediate financial headline: Anthropic’s reported net loss of $80.7 billion for the first six months of the year. At face value, burning over $80 billion in two quarters conjured immediate comparisons to the dot-com bubble and prompted pundits to declare the entire generative AI sector a runaway speculative furnace. However, a forensic inspection of the S-1 footnotes reveals that the vast majority of this astronomical figure is entirely an artifact of GAAP financial accounting rather than operational cash burn.

More than $72 billion of the headline loss is an accounting illusion stemming from the company's historical financing agreements with corporate backers Google and Amazon. These funding packages were structured as convertible debt instruments that convert into Anthropic equity under specific liquidity triggers. Under current accounting standards, because of the contractual terms governing these convertible notes, Anthropic must classify the instruments as liabilities on its balance sheet and mark them to fair value every financial quarter.

When Anthropic’s implied enterprise valuation skyrocketed by roughly $80 billion over the course of the six-month period, the embedded conversion value held by Google and Amazon increased by an identical amount. Consequently, Anthropic was required to record that exact paper gain accrued by its strategic investors as a non-cash, fair-value mark-to-market expense against its own income statement. Far from indicating a company on the brink of liquidity insolvency, the $72 billion loss was paradoxical proof of Anthropic's skyrocketing enterprise valuation.

```mermaid flowchart LR accTitle: Convertible Debt Accounting Illusion accDescr: Rising valuation forces mark-to-market liability increases on the balance sheet.

A["Anthropic Enterprise Valuation Rises (+$80B)"] --> B["Google & Amazon Convertible Notes Gain Value"] B --> C["Convertible Debt Treated as Fair Value Liability"] C --> D["Anthropic Books Matching Non-Cash Accounting Loss (~$72B+)"] D --> E["Headline Net Loss Explodes to $80.7B"] ```

As Patrick Boyle emphasizes regarding the disconnect between public reaction and financial reality, “It's the kind of number that makes commentators shout about a bubble and point at the dot-com crash.”

- **Reported 6-month net loss:** $80.7 billion. - **Non-cash accounting charge:** Over $72 billion attributed to mark-to-market revaluations. - **Key counterparties:** Primary convertible debt holders include Amazon and Google. - **Mechanism:** As Anthropic's private valuation grew, the fair value of its debt liabilities expanded proportionately, forcing a massive non-cash charge on the income statement without draining cash reserves.

[10:00](https://www.bilibili.com/video/BV1s1xJ6eEJU?t=600s) The Mark-to-Market Accounting Illusion and Real Operating Turnaround

The staggering headline net loss reported by Anthropic caused widespread bewilderment, but financial analysts quickly pointed out the bizarre accounting mechanisms at play. As financial writer Matt Levine observed, under standard corporate circumstances, issuing equity that subsequently surges in value does not generate an accounting loss. For Anthropic, however, billions in funding from tech giants like Google and Amazon were structured as convertible debt instruments. Under mark-to-market accounting rules, as Anthropic’s estimated valuation ballooned, the value of the equity owed to those early backers escalated dramatically, forcing the company to book an enormous liability revaluation as an accounting loss. In reality, Anthropic did not incinerate $72 billion in cold, hard cash; rather, the paper loss merely reflected the skyrocketing paper value of the firm's equity.

```mermaid flowchart LR accTitle: Divergent Accounting for Convertible Debt accDescr: Shows how Anthropic records a paper loss while Alphabet and Amazon book real gains from the exact same securities. subgraph S1["Anthropic's Books"] A["Company Valuation Surges"] -->|"Convertible Debt Liability Expands"| B["$72B Mark-to-Market Paper Loss"] B -->|"Strip Out Non-Cash Loss"| C["Adjusted Operating Loss: ~$700M"] end subgraph S2["Backers' Books"] A -->|"Holding Convertible Notes"| D["Alphabet & Amazon Balance Sheets"] D -->|"Booked as Value Appreciation"| E["Billions in Investment Earnings"] end ```

The starkest irony of this accounting convention was highlighted by Matt Levine alongside Jonathan Weil at *The Wall Street Journal*: the exact same financial instruments sit on the opposite side of the ledger at Alphabet and Amazon, where they are celebrated as massive investment earnings. While Anthropic strips out this mark-to-market distortion to report an adjusted operating profit of roughly $2.7 billion for the first half of the year, its Big Tech investors happily count the gains as legitimate income. The double standard cuts both ways: during the very same reporting cycle, Alphabet had to digest an $8.7 billion loss on its own external holdings, including its position in SpaceX. As Patrick Boyle noted of the broader financial theatre, “mark-to-market gun fires in both directions, it gets touted more prominently in one direction.”

Once the accounting theatrics are cleared away, Anthropic’s underlying operational cash burn told a surprisingly constructive story. The true operational loss incurred to run the company over the six-month stretch stood at just under $700 million—far from ideal for a conventional enterprise, but a massive improvement compared to the $2 billion operational burn over the same period the previous year. Anthropic managed to post a genuine operating profit in the second quarter, bringing its computing infrastructure expenditures down from a dizzying 251% of sales in early 2024 to 48% in the second quarter. Furthermore, the company generated nearly $4 billion in positive operating cash flow without resorting to standard tech gimmicks like capitalizing model training expenses to artificially conceal core development costs.

- **Headline Paper Loss:** More than $72 billion generated entirely by the mark-to-market revaluation of convertible debt liabilities. - **Adjusted Half-Year Operating Loss:** Under $700 million, declining sharply from $2.0 billion in the prior-year period. - **Compute Spend Trajectory:** Plunged from 251% of revenues in early 2024 down to 48% of revenues by Q2. - **Q2 Cash Generation:** Achieved nearly $4 billion in positive operational cash flow while keeping full training expenses in the cost baseline.

---

[12:40](https://www.bilibili.com/video/BV1s1xJ6eEJU?t=760s) The Peak of "Token Maxing" and the Autumn Budget Hangover

While the reported second-quarter figures appeared robust, market skeptics quickly pointed out a critical vulnerability: the entire financial narrative in the S-1 filing cuts off abruptly at June 30th. Financial commentator Dave Eisman pointed out on the *Prof G Markets* podcast that June 30th effectively represented the moment the AI hype train hit the brakes. The extraordinary hyper-growth documented across the first half of the year had been supercharged by an internal corporate trend humorously termed "token maxing"—a compulsive race among software developers to process astronomical quantities of AI tokens, the fundamental alphanumeric data units that power large language models.

The cultural zenith of this phenomenon unfolded inside Meta in April, where an engineer established an internal leaderboard called "Claudeonomics" to rank Meta’s 85,000 employees based on their gross token consumption. High-rolling engineers competed for prestigious internal monikers such as "Token Legend" and "Cash Wizard." Over a single 30-day period, Meta staff burned through more than 60 trillion tokens; *Fortune* estimated that a single top developer racked up over $1 million in API bills. When Meta's Chief Technology Officer evaluated the bill, he marveled that his premier engineer was expending the equivalent of his entire base compensation on AI compute, yet concluded the expenditure was justified because the engineer was five to ten times more productive, declaring: “This is easy money. Keep doing it. No limit.”

```mermaid flowchart TD accTitle: The Token Maxing Lifecycle accDescr: Tracks how token maxing went from enterprise-wide frenzy to strict corporate cost controls. subgraph Hype["H1 Euphoria: Token Maxing Craze"] M1["Internal Gamification: Meta's Claudeonomics"] --> M2["Heavy Token Burn: >60T tokens in 30 Days"] M2 --> M3["Executive Endorsement: 'Easy money, no limits'"] end subgraph Chill["H2 Retrenchment: Budget Enforcement"] M3 --> C1["Amazon Corporate Memo: Token usage is flawed metric"] C1 --> C2["Leaderboards Terminated & Scrutinized"] C2 --> C3["Workloads Routed to Cheaper Open-Weight Models"] end ```

By early autumn, however, enterprise enthusiasm collided with financial reality. Amazon formally dismantled its own internal usage leaderboards after discovering that developers were artificially inflating their token consumption purely to scale the rankings. Corporate management distributed internal memos warning personnel to exercise fiscal discipline, flatly rejecting token volume as a valid proxy for true software engineering productivity. As Amazon's internal guidance pointedly advised its staff, “if your work does not benefit from current AI tooling, that is a finding, not a failing.”

---

[15:00](https://www.bilibili.com/video/BV1s1xJ6eEJU?t=900s) The Q3 Growth Cliff, Open-Weight Rivals, and an Imminent Price War

The cooling corporate sentiment dealt an immediate blow to Anthropic's commercial expansion. During the height of the token-maxing wave, Anthropic had logged sequential quarter-over-quarter revenue growth of roughly 135% in the first quarter and 147% in the second quarter. However, as enterprise procurement teams reasserted budget discipline in the third quarter, corporate workflows began migrating away from high-priced proprietary models toward cut-rate, open-weight Chinese models that delivered acceptable enterprise utility at a fraction of the cost. Private metrics shared with pre-IPO investors—and omitted from the public SEC documentation—indicated that sequential revenue growth dropped from over 140% to approximately 45%.

While a 45% quarterly expansion still translates to an eye-watering annualized growth rate near 250%, such a sudden deceleration creates immense anxiety for prospective public-market buyers accustomed to exponential acceleration. Compounding this challenge, unconfirmed industry reports cited by Eisman indicated that Anthropic enacted aggressive price cuts in October to protect market share, corroborated by S-1 disclosures showing task execution costs lowered by 25% to 40%. At the same time, Alphabet rolled out a competitive Gemini release priced specifically to undercut Claude, while OpenAI flooded the enterprise tier with cheaper options. Highlighting the awkward timing, Patrick Boyle remarked: “A price war is a wonderful news for customers, but it's hard to think of a less convenient time for when to start than the month before your IPO.”

The broader existential riddle surrounding the listing is whether generative AI can expand beyond its initial niche. As investor Paul Kedrosky observed on the *Hidden Forces* podcast, early adoption has been heavily concentrated among software engineers because automated syntax generation is an ideal fit for large language models. To validate a multitrillion-dollar valuation, Anthropic must penetrate mainstream white-collar workflows, primarily tasks centered on parsing, analyzing, and synthesizing long documents. Patrick Boyle quipped that a commercial tool whose primary value proposition is reading long, complex prospectuses and summarizing the salient facts for an audience sounds dangerously close to what he does on YouTube for a living, musing that he might need to look into obtaining a real estate license instead. Ultimately, public investors are flying blind without the Q3 numbers, and whether regulators will force Anthropic to reveal that slowdown before pricing the IPO remains an open question.

---

[18:56](https://www.bilibili.com/video/BV1s1xJ6eEJU?t=1136s) Governance as a Constitution: The Public Benefit Corporation Structure

Turning from the financial statements to the governance section, Anthropic’s prospectus reads less like a modern public enterprise and more like the founding charter of an independent nation-state. Anthropic is organized as a Public Benefit Corporation (PBC), a legal framework that fundamentally modifies standard fiduciary obligations. In an orthodox corporation, the board of directors is legally bound to maximize financial value for common stockholders; under the PBC model, directors are granted the statutory leeway to subordinate shareholder returns in service of a defined charter: the safe and responsible stewardship of artificial intelligence for the long-term benefit of humanity.

This structural shelter grants the company absolute immunity should it choose to incinerate tens of billions of dollars on speculative safety protocols at the expense of shareholder dividends or bottom-line earnings. In the words of *Financial Times* columnist Robert Armstrong, the entire architecture is designed like an institutional constitution, establishing an elaborate balance of power intended to prevent any single faction—whether founders, private investors, or the open market—from unilaterally dictating the trajectory of frontier intelligence.

---

[20:00](https://www.bilibili.com/video/BV1s1xJ6eEJU?t=1200s) The Long-Term Benefit Trust, Founders' Veto, and the Employee Tiebreaker

Anthropic’s constitutional governance balances power between two distinct entities: the Long-Term Benefit Trust and the Founder LLC. The Long-Term Benefit Trust is designed as an independent body of five trustees charged with ensuring that the enterprise does not deviate from its existential duty to safeguard humanity. Curiously, heading into what could be the largest IPO in corporate history, the trust holds only four members because Anthropic has not yet managed to fill the fifth seat—leaving the panel of planetary guardians one member short. The roster of four trustees includes former Federal Reserve Chairman Ben Bernanke, granting the trust the nominal authority to select four of the seven seats on Anthropic’s board of directors.

```mermaid flowchart TD accTitle: Anthropic Board and Deadlock Governance Structure accDescr: Illustrates how the Long-Term Benefit Trust, Founder LLC, and employee tiebreaker interact to control board nominations. subgraph BoardNomination["Board Composition: 7 Seats Total"] LTBT["Long-Term Benefit Trust (4 Trustees, inc. Ben Bernanke)"] -->|"Nominates 3 Direct Seats"| Dir1["3 Safety Directors"] FLLC["Founder LLC (7 Co-founders with 50.1% Voting Power)"] -->|"Directly Elects"| Dir2["3 Founder Directors"] LTBT -->|"Nominates 4th Swing Director"| NC["Board Nominating Committee"] end subgraph NominatingGate["Nominating Committee Bottleneck"] Dir2 -->|"Controls 2 of 3 Seats"| NC NC -->|"Founders Can Veto Nominee"| Deadlock{"Deadlock Condition"} end subgraph Tiebreaker["Final Arbiter"] Deadlock -->|"Bypasses Courts & Regulators"| Staff["Employee Popular Vote"] end ```

The counterweight to the Trust is the Founder LLC, formed by Anthropic's seven original co-founders. They command a customized tier of founder stock that fixes their voting rights at an ironclad 50.1%, completely untethered from the actual capital invested or subsequent equity dilution, remaining valid even if the founders liquidate nearly all of their economic stakes. While the trust nominates four directors and the founders pick three—theoretically handing the majority to humanity's trustees—the fine print reveals a clever corporate counter-check: the trust's fourth "swing" director must receive formal approval from the board's nominating committee. Because the founders control two of the three seats on that nominating committee, they retain an absolute veto over the trust’s deciding nominee.

If an unyielding deadlock occurs between the trust and the founders—such as a high-stakes standoff over releasing an advanced model that trustees deem potentially lethal—the dispute does not escalate to federal regulators, sovereign courts, or conventional board arbitration. Instead, the charter dictates that the dispute must be settled by a company-wide vote of Anthropic's rank-and-file employees, weighted by seniority and the proportion of equity staff have committed to philanthropic causes. Underlining the absurdity of this design, Patrick Boyle quipped: “Under Anthropic's governance, this question, whether to expose the species to a potentially catastrophic AI, does not go to a government or a regulator or a court, goes to a staff vote.” In essence, the fate of the human species rests on the whim of junior engineers deciding whether to deploy a potentially cataclysmic AI in between their lunch orders and office fire drills—functioning as an internal workers' cooperative tucked inside a $2 trillion tech monopoly.

---

[22:10](https://www.bilibili.com/video/BV1s1xJ6eEJU?t=1330s) The Death of Shareholder Democracy: From Dual-Class to Founder Autocracy

This Byzantine arrangement represents the culmination of a decades-long erosion of shareholder democracy across the technology landscape. Historically, institutional asset allocators outside Silicon Valley strictly enforced a "one share, one vote" standard, viewing sound governance as a prerequisite to lower a firm’s cost of capital. The underlying principle was straightforward: if a corporate board protects outside equity holders from exploitation, investors will accept lower baseline returns, whereas governance systems that disenfranchise common shareholders face substantial valuation discounts in the public markets.

The modern tech boom obliterated this paradigm: * **The Normalization of Dual-Class Equity:** Google and Facebook popularized dual-class share structures that granted founders ten votes for every public vote, conditioning Wall Street to accept diluted governance rights. * **The Zero-Vote Threshold:** When Snap staged its initial public offering in 2017, it sold common stock bearing zero voting rights; surprisingly, public investors clamored for the shares without demanding any governance discount. * **Jurisdictional Flight:** When a Delaware chancery court struck down Elon Musk’s $56 billion compensation plan at Tesla, the automaker relocated its legal incorporation to Texas to take advantage of business statutes skewed entirely toward executive autonomy. * **Private Autocracy Consolidation:** Musk leveraged a super-voting share framework inside SpaceX to command an estimated 85% voting majority, with ongoing discussions hinting at absorbing Tesla into SpaceX to permanently bypass public corporate governance safeguards.

As a person close to Tesla's board revealed to the *Financial Times*, a structural reorganization of this scale would finally allow leadership to dismantle Tesla's legacy governance setup, which the insider unvarnishedly summarized as a “pain in the ass for Elon and the board.” Anthropic’s convoluted public benefit charter pushes this historical transformation to its outer limit, replacing conventional equity-holder accountability with an insulated corporate hierarchy answerable only to itself and its own founding philosophy.

[25:00](https://www.bilibili.com/video/BV1s1xJ6eEJU?t=1500s) 极端治理结构与 Matt Levine 的“权力等级阶梯”

Anthropic 将硅谷稀释普通股东话语权的趋势推向了前所未有的极端。过去的科技公司往往只是赋予创始人超级投票权，而 Anthropic 则是彻底切断了公司控制权与所有权之间的联系：无论 7 位创始人未来出售多少股份，他们都永久保留 50.1% 的绝对多数投票权，并且整个公司架构都被包裹在一个“为了全人类利益”的信托之中。

财经专栏作家 Matt Levine 形象地勾勒出了这家公司内部荒诞而残酷的权力等级秩序：排在最顶层的是“全人类”，其次是 7 位创始人，接着是普通员工，再往下是承担全部财务波动的公众股东，而排在最底层的，居然是投入真金白银最多的亚马逊和谷歌。Matt Levine 调侃道，这两家科技巨头之所以排在最底层，完全是因为他们“浑身沾满了铜臭的商业气息”，因此换来的只是毫无投票权的股份。

面对这种剥夺股东权利的设计，公开市场甚至毫不介意。过去如果公司削弱股东权利，投资者通常会要求极大的估值折价，但现在市场却欣然接受，甚至恨不得追着送钱。Patrick Boyle 讽刺地指出：“我们面对的是一家由员工委员会治理、可能会毁灭世界、却凭着一份增长率断崖前夕的账目就敢向你要 2 万亿美元估值的公司。”

```mermaid flowchart TD accTitle: Anthropic Corporate Control Pecking Order accDescr: Diagram illustrating Matt Levine's corporate control hierarchy at Anthropic, placing Humanity at the apex and big tech investors at the bottom.

subgraph Governance["权力与利益顺位"] A["1. 全人类 (Humanity)"] B["2. 7位联合创始人 (保留50.1%绝对投票权)"] C["3. 公司员工 (Employees)"] D["4. 普通公众股东 (承担财务风险但无发言权)"] E["5. 亚马逊 & 谷歌 (注资最多，仅持无投票权股份)"] end

A --> B B --> C C --> D D --> E ```

[26:18](https://www.bilibili.com/video/BV1s1xJ6eEJU?t=1578s) 招股书第 161 页的惊人秘密：8000 亿美元算力采购债务

这份长达 24 万字的招股书中，最令人不寒而栗的内容既不是 90 页的人类灭绝风险，也不是奇葩的信托架构，而是藏在第 161 页“流动性与资本资源”章节的附注里。截至 2025 年底的短短 10 个月内，Anthropic 已经签署了最低采购义务高达 4170 亿美元的算力采购协议。

如果将这些一直延伸到 2045 年的长期支出全部加总，未来的采购承诺总额超过了惊人的 8000 亿美元——这比大多数大洲全年的国防预算加起来还要庞大。根据会计准则，这些针对尚未建成的基础设施的承诺，并不作为传统负债直接体现在资产负债表上，而是以采购承诺和租赁义务的形式深埋在极少有人翻阅的附注中：

- **亚马逊（Amazon）**：1100 亿美元的照付不议（Take-or-pay）采购承诺。 - **博通（Broadcom）**：巨额的芯片租赁承诺。 - **SpaceX**：最高达 1205 亿美元的长期合约，在算力全面交付上线后，Anthropic 每月需向其支付约 31 亿美元。 - **其他供应商**：包括谷歌、微软，以及 Nscale 和 Lambda 等新兴 GPU 云计算供应商。

[28:25](https://www.bilibili.com/video/BV1s1xJ6eEJU?t=1705s) “轻资产”神话破灭：重资产转型与急于 IPO 的本质

Anthropic 承担如此惊人的支出，根本原因在于其陷入了供应链的战略死局。目前，Anthropic 的核心模型完全依赖其最大的竞争对手（亚马逊和谷歌）来分发。这两家巨头既是 Anthropic 的大股东、房东与云分发渠道，同时自己也在研发正面对标的 AI 大模型。分析师 Paul Kedrosky 指出，这些云巨头随时可以抽成，并在理论上随时能掐断 Anthropic 的分发渠道。招股书坦白承认了这一点：“这些供应商自身开发并提供与我们直接竞争的 AI 模型和服务。”

为了求生，Anthropic 正在全面推行名为“Anthropic Direct”的转型战略。他们试图在两个方向上同时复制当年亚马逊构建 AWS 的路径：向下自建芯片和数据中心，向上直连终端客户，不再向亚马逊和谷歌交租。但正如 Paul Kedrosky 和 Michael Burry 所警告的，一旦走上自建道路，公司就必须买下整座大楼、采购芯片，并陷入“芯片几年内过时就需要无限追加资本开支进行替换”的资本黑洞。

Patrick Boyle 揭露了 Anthropic 此时急于上市的真正底牌：过去一年公司财务表现看似极其亮眼，纯粹是因为轻资产运营、租用算力且账上没有任何沉重资产；然而这个时代已经彻底结束了。第二季度勉强挤出的营业利润在即将到来的天价开支面前微不足道。正如 Patrick Boyle 尖锐指出的： > “他们不过是在巨额账单到期之前，赶紧先把帽子递向公众讨钱罢了。”

[30:32](https://www.bilibili.com/video/BV1s1xJ6eEJU?t=1832s) 政府干预与国家机器的达摩克利斯之剑

招股书中唯一没有半点黑色幽默色彩、让 Anthropic 措辞极为谨慎的风险，在于公司公开倡导的 AI 安全与监管立场可能招致的报复。招股书警告称，其公开发声可能会让公司面临“更严密的审查、政治反弹或不成比例的监管行动”。

尽管招股书没有直接点名美国政府，但背后的冲突早已摆上台面。首席执行官 Dario Amodei 多年来一直公开呼吁政府严加监管 AI（包括监管 Anthropic 自己的模型）。今年早些时候，美国国防部将 Anthropic 列为“供应链风险”并试图取消其承包商资格；联邦法院虽一度认定该行政行为系针对受保护言论的报复性惩罚，但 Anthropic 随后在上诉中败诉，目前仍在艰难缠讼中。

Patrick Boyle 认为，即便 Anthropic 闭嘴不言，这种国家风险也无法避免： - **颠覆性技术必然招致国家控制**：如果招股书里描述的该技术能重塑全球经济甚至带来人类灭绝风险是真的，那么任何主权政府都不可能允许它被完全掌控在私人企业手里。 - **早期国家化信号**：打压模型的出口管制、供应链安全审查，正是国家权力介入科技资产的最初步骤。 - **公开发行的本质悖论**：金融专栏作家 Robert Armstrong 指出，如果真相信这是一种堪比原子弹的技术，那么核心问题绝非 Anthropic 是否因多嘴而被敲打，而是如此危险的事物到底能不能作为一家普通上市公司存在——毕竟，“当年没有人会把曼哈顿计划拿去上市敲钟”。 - **商业影响与终极风险**：尽管目前政府合同仅占 Anthropic 营收的约 1%，直接财务影响微乎其微，但终极风险从来不是今年的销售额，而是控制权归属；在极端情况下，美国政府直接将整个业务收归国有并非天方夜谭。

[33:50](https://www.bilibili.com/video/BV1s1xJ6eEJU?t=2030s) 24 万字招股书的估值困境：失效的模型与破裂的电子表格

面对厚达 24 万字、充斥着详尽细节的招股说明书，投资者依然无法估算出 Anthropic 究竟值多少钱。招股书中的所有数据都是滞后的，而这家公司的业务演化速度实在太快，以至于历史数据无法对未来提供任何指引。

估值专家 Aswath Damodaran 在《金融时报》（FT）撰文直言，拿历史数据来给 Anthropic 估值纯粹是“傻子干的徒劳差事”（a fool's errand）。如果按照第二季度超过 1300% 的同比增速推算，哪怕未来每年仅仅实现营收翻倍，到 2031 年其营收就会达到 6300 亿美元；若能维持一半的增速，公司甚至会在十几年内买下大半个地球。这种指数级推演会瞬间撑爆任何财务分析模型。

Aswath Damodaran 补充指出，就连 Anthropic 自己的 CEO 也坦承根本无法预测终端需求。Patrick Boyle 总结道：“面对未来，他们和我们一样一头雾水。”

[35:56](https://www.bilibili.com/video/BV1s1xJ6eEJU?t=2156s) 内部人士套现出逃、行业资金内卷与泡沫的终局

Anthropic 的确是一家令人印象深刻的公司：那些断言它濒临破产的空头彻底被打脸，它不仅创造出了人们愿意付费使用的优秀产品，还挤出了一笔真实的营业利润；但那些宣称“2 万亿美元估值很便宜”的多头也纯粹是在盲目瞎猜。在投资人日到正式提交招股书的短短几周内，传闻中的估值就从 2 万亿美元下滑到了 1.5 万亿至 1.75 万亿美元之间，短短半个月蒸发了近 5000 亿美元的“不确定性”。

更值得警惕的是行业内核心内部人士的动作： - **锁定期阶梯解禁**：Anthropic 设立了交错的禁售期条款，使得内部人士在 IPO 后几乎可以立刻开始抛售套现。 - **巨头集中抽水**：与此同时，Meta 正在寻求借债融资 950 亿美元，Alphabet 也刚刚发行了约 900 亿美元的债务工具。内部人士的一致判断是，现在正是向公开市场抛售筹码的最佳窗口。 - **左手倒右手的闭环内卷**：整个 AI 行业的资金在同一张桌子上循环流转——亚马逊和谷歌投资 Anthropic，Anthropic 反手向亚马逊和谷歌采购算力，再转头每月向 SpaceX 支付数十亿美元订购尚未交付的基础设施。Aswath Damodaran 精辟地概括了这种共生风险：“要给微软估值，你必须先给 OpenAI 估值；要给谷歌估值，你必须先给 Anthropic 估值。整个行业已经被捆成了一个极其昂贵且无法解开的死结。”

回顾 30 年前的互联网泡沫，底层技术的真实性毋庸置疑，但当年暴涨的雅虎、Lycos 和 AltaVista 最终都被后起之秀碾得粉碎。今天支付 2 万亿美元估值接盘的投资者，面临着在合理估值倍数下数学上几乎无法回本的风险。

Patrick Boyle 对这场狂欢给出了最后的裁决： > “任何宣称自己能确定未来的人，都只是在兜售某种东西罢了。多头和空头全都在猜。IPO 纯粹就是早期投资人觉得时机成熟了，决定兑现一部分筹码离场，把定价的麻烦留给公众。”

买入这只股票，本质上就是在豪赌：赌那个能够彻底扭转商业模式的“技术奇迹时代”，能够赶在 8000 亿美元的算力账单彻底爆仓前降临；或者寄希望于那个长远利益信托能在危机来临时拯救股东——“前提是他们哪天终于想起来把那第五位信托成员给招齐了”。

## Notable quotes > [35:56](https://www.bilibili.com/video/BV1s1xJ6eEJU?t=2156s) 要给微软估值，你必须先给 OpenAI 估值；要给谷歌估值，你必须先给 Anthropic 估值。整个行业已经被捆成了一个极其昂贵且无法解开的死结。 — Aswath Damodaran

> [35:56](https://www.bilibili.com/video/BV1s1xJ6eEJU?t=2156s) IPO 纯粹就是早期投资人觉得时机成熟了，决定兑现一部分筹码离场，把定价的麻烦留给公众。 — Patrick Boyle

## Mentioned Anthropic, Patrick Boyle, Claude, Google, Alphabet, Amazon, AWS, OpenAI, SpaceX, Meta, Broadcom, Microsoft, Nscale, Lambda, Tesla, Snap, Revolut, Dario Amodei, Ben Bernanke, Matt Levine, Jonathan Weil, Dave Eisman, Paul Kedrosky, Robert Armstrong, Aswath Damodaran, Michael Burry, Elon Musk, Long-Term Benefit Trust, Founder LLC, Prof G Markets, Hidden Forces, The Wall Street Journal, Financial Times, Fortune, Ulysses, Naked Lunch, The Terminator, Grok, Gemini

## Video description 【Patrick Boyle】 2万亿估值，这家AI公司招股书里藏着什么？ Anthropic’s IPO Filing Warns of the End of the World

｜AI公司上市，估值2万亿 Anthropic（一家AI公司）最近提交了招股书，计划上市融资，估值高达1.5万亿到2万亿美元。这个数字相当于加拿大一年的国内生产总值。招股书长达24万字，其中风险因素部分就占了90页。Anthropic在招股书中甚至提到了其产品可能导致人类灭绝的风险，这在普通招股书中非常罕见。尽管如此，这家公司在最近一个季度实现了运营利润，营收增长也十分惊人。

｜财务数据背后的“会计魔术” Anthropic在今年上半年报告了807亿美元的巨额净亏损，这让很多人联想到互联网泡沫。但仔细阅读招股书会发现，其中超过720亿美元的亏损是会计处理造成的“假象”。这部分亏损源于公司与谷歌、亚马逊等公司签订的融资协议，这些贷款可以转换为Anthropic的股份。由于会计规则，当Anthropic的估值上升时，这些可转换债务被视为负债，其价值变化会带来账面亏损。实际上，公司并没有烧掉这么多现金，反而因为业务价值的增长，早期投资者的股权价值大增。剔除这些会计处理，公司上半年的实际运营亏损为7亿美元，比去年同期的20亿美元大幅减少。

｜增长放缓与价格战 Anthropic的营收在今年第一季度和第二季度分别实现了135%和147%的环比增长，这主要得益于“代币最大化”热潮，即工程师大量使用AI模型来提升生产力。但到了第三季度，随着企业客户开始控制预算，并转向更便宜的开源模型，Anthropic的环比营收增速降至45%。虽然45%的季度增长率依然非常高，但增速突然减半让潜在投资者感到不安。此外，Anthropic在10月大幅降价25%至40%以争夺市场份额，同时谷歌和OpenAI也推出了更便宜的竞品，一场价格战正在酝酿。

｜公司治理的独特结构 Anthropic采用了“公益公司”的治理结构，这意味着公司董事会在做决策时，可以平衡股东的财务利益和“为人类长期利益负责任地开发AI”的公司使命。公司设立了一个由五名独立受托人组成的“长期利益信托”，负责确保公司忠于保护人类的使命。同时，七位创始人通过特殊股份拥有50.1%的投票权，即使他们出售大部分股份，投票权也不会被稀释。这种结构让“人类利益”排在首位，其次是创始人、员工，最后才是普通股东。

｜天价支出与上市原因招股书的脚注揭示了Anthropic上市的真正原因：公司承诺在未来几年内投入巨额资金购买算力。截至2025年底，公司已签订了4170亿美元的算力采购协议；如果算到2045年，总承诺支出将超过8000亿美元，这比大多数国家的国防预算还要高。Anthropic正在从“轻资产”模式转向“重资产”模式，试图自建数据中心和芯片，摆脱对亚马逊和谷歌等竞争对手的依赖。这种转型需要巨额资金，因此公司急需通过上市来筹集资金，以应对即将到来的巨额开支。

接下来建议的操作指引： 1. 关注Anthropic（未上市，无代码）在第三季度营收数据公布后的市场反应，特别是其环比增速是否能维持在45%以上。 2. 密切观察Anthropic（未上市，无代码）与谷歌（GOOGL）、OpenAI等公司之间的价格战进展，关注其对公司利润率的影响。 3. 留意Anthropic（未上市，无代码）在未来几个月内关于算力采购协议的进一步披露，特别是其对长期现金流和盈利能力的影响。

[0:00] Anthropic招股书披露巨额估值与风险 [2:09] Anthropic财务数据表现良好，但存在会计处理“假象” [6:24] Anthropic招股书中的“人类灭绝”风险 [8:30] 807亿美元亏损的会计处理真相 [11:33] Anthropic运营亏损大幅收窄，实现季度盈利 [12:40] 营收增长放缓，价格战初现端倪 [16:50] AI应用场景与市场拓展的挑战 [18:56] Anthropic独特的公司治理结构：公益公司 [22:10] 创始人与员工在公司治理中的权力分配 [24:20] 科技公司治理模式的演变与股东权利的削弱 [26:18] Anthropic巨额算力采购承诺与上市原因 [28:25] Anthropic“重资产”转型，摆脱对云服务商依赖 [30:32] 政府干预AI技术的风险与影响 [33:50] AI公司估值难题：历史数据失真，未来难以预测 [35:56] 内部人士套现与AI行业资金循环

[Source](https://www.bilibili.com/video/BV1s1xJ6eEJU)

中文简报

Anthropic谋求两万亿美元IPO 招股书揭8000亿美元算力承诺 Anthropic急于推动高估值上市，本质是在三季度营收增速断崖与8000亿美元算力账单到期前，将重资产转型风险与巨额算力负债甩给毫无治理话语权的公众股东接盘。 Patrick Boyle 深入剖析了 Anthropic 篇幅极长的招股书，首先拆解了其上半年807亿美元净亏损背后由可转债公允价值会计变动带来的账面影响，肯定了公司经调整后运营亏损大幅收窄并在季度层面实现盈利的商业能力。但随后他指出公司面临的关键转折：随着“代币狂热”退潮与开源竞品降价，第三季度环比增速大幅下滑并被迫参与价格战；同时公司采用的“公益公司”与信托治理架构，使公众股东缺乏核心治理话语权。更严峻的隐患在于招股书附注中披露的高达8000亿美元的长期算力硬性采购承诺，促使 Anthropic 在由轻资产模式转向重资产高投入的转折期寻求上市，将巨额承诺与潜在风险转由公开市场承担。 Anthropic 递交长达 24 万字的 S-1 招股书，寻求以 1.5 万亿至 2 万亿美元估值进行 IPO，风险披露长达 90 页，并史无前例地将“人类灭绝”与无法承保的系统性风险列入其中。 (0:00) 上半年名义净亏损 807 亿美元中，有超过 720 亿美元源于谷歌与亚马逊可转债估值上升带来的按市值计价账面亏损；扣除后实际运营亏损为 7 亿美元，且第二季度已实现微薄运营利润。 (8:25) 受“Token Maxing”热潮消退及企业预算收紧影响，第三季度营收环比增速自前两季度的 140% 以上骤降至约 45%，加之 10 月任务降价 25% 至 40%，行业正滑向全面价格战。 (15:00) 招股书附注披露公司面临总额超 8000 亿美元的长期算力采购承诺（至 2025 年底达 4170 亿美元），促使 Anthropic 从轻资产转向自建芯片与数据中心的重资产模式，IPO 实质是在账单到期前向市场募资。 (26:18)

Reading summary

根据 Patrick Boyle 的剖析，AI 初创巨头 Anthropic 提交的 24 万字 S-1 招股书展现了从轻资产向重资产转型的巨大财务压力；尽管表面上实现了季度营业利润，但其面临着 8000 亿美元的隐性采购债务、增长放缓的商业拐点以及极度削弱公众股东话语权的治理架构。表面巨亏与真实运营拐点 (10:00) Anthropic 招股书披露上半年账面净亏损高达 807 亿美元，但 Boyle 指出其中超 720 亿美元为可转债公允价值变动的非现金会计损失。因亚马逊和谷歌持有的可转债被列为负债，公司估值激增反向推高了账面负债，而两家科技巨头在自身账本上则将同等增值确认为巨额投资收益。剔除该会计扭曲后，公司上半年实际运营亏损不足 7 亿美元，较上年同期的 20 亿美元大幅收窄。在第二季度，Anthropic 算力开支占营收比例从 2024 年初的 251% 降至 48%，并产生了近 40 亿美元的正向运营现金流，实现了真实的季度营业利润，打破了外界对其仅是“资本火葬场”的质疑。 “刷Token”潮退去与Q3增长悬崖 (12:40) 招股书展现的高光业绩截止于 6 月 30 日，掩盖了随后的业务降温。上半年业绩大幅受益于各大科技公司内部的“Token 狂欢”，例如 Meta 员工在一个月内消耗超 60 万亿 Token，高管甚至默许报销高额算力账单。然而，入秋后企业采购迅速收紧，亚马逊等公司发文叫停内部消耗榜单，并指出单纯消耗 Token 并非有效生产力指标。受企业预算收紧及廉价开源权重模型冲击影响，未在招股书中披露的 Q3 内部数据显示，Anthropic 环比营收增速已从 Q1 的 135% 和 Q2 的 147% 骤降至约 45%。加之 10 月 Anthropic 与 Alphabet、OpenAI 相继启动降价，在 IPO 前夕引发了行业价格战的担忧。绝无仅有的治理集权与信托设计 (18:56) Anthropic 采用了公共利益公司（PBC）架构，法律上允许董事会将全人类安全置于股东财务回报之上。其治理权力由“长远利益信托”与 7 位联合创始人联合掌控。信托由包括前美联储主席伯南克在内的受托人组成，虽名义上指派多数董事席位，但关键摇摆董事须经创始人控制的提名委员会批准；7 位创始人无论稀释或出售多少股权，永久保有 50.1% 的绝对投票权。若双方发生不可调和的僵局，重大分歧将交由公司普通员工按资历及慈善捐赠比例投票表决，而非通过法院或监管机构解决。财经作家 Matt Levine 戏称，在该体系中，出资最多的亚马逊与谷歌仅持有无投票权股份，沦为地位垫底的资本提供方。 161页的8000亿债务与重资产困境 (26:18) 招股书第 161 页附注暴露了公司最大的隐性危机：Anthropic 签署了延至 2045 年、总额超 8000 亿美元的算力采购与租赁承诺，其中包括对亚马逊 1100 亿美元的照付不议协议，以及向 SpaceX 支付最高达 1205 亿美元、全面交付后每月约 31 亿美元的长期合约。由于核心分发与算力高度受制于直接竞争对手亚马逊和谷歌，Anthropic 试图推进“Anthropic Direct”战略自建数据中心与芯片。然而重资产转型意味着将陷入机房采购与芯片折旧过时的资本无底洞。Boyle 认为，这表明过去的轻资产盈利红利已终结，公司急于冲刺上市本质上是为了赶在天价算力账单到期前向公开市场募资套现。国家机器介入与2万亿估值迷局 (30:32) 招股书罕见地将 AI 自主防关机、武器研发辅助及人类灭绝列为风险因素，同时也警告其激进的安全呼吁可能招致政府报复。此前美国国防部曾将其列为供应链风险，引发后续诉讼。分析指出，若该技术真具备颠覆人类的力量，主权国家绝不会允许其纯私人化运营，极端情况下甚至面临国有化风险。估值专家 Aswath Damodaran 认为，基于滞后历史数据为 Anthropic 估值纯属徒劳，其目标估值在几周内已从 2 万亿美元下滑至 1.5 万亿至 1.75 万亿美元。与此同时，内部人士通过阶梯解禁条款为抛售套现铺路，整个行业在彼此对投、互买算力的资金闭环中交织，将巨大的终极定价风险抛给了普通投资者。 Anthropic 凭借真实的产品需求打破了破产质疑，但在面临 8000 亿美元算力硬支出、Q3 增速换挡、行业价格战及主权监管风险的当下，其 2 万亿美元估值在数学上极度脆弱；参与该笔 IPO 本质上是一场豪赌，押注所谓的技术奇迹能否在巨额开支爆仓之前成为现实。

精炼阅读摘要

金融分析师兼播客主持人 Patrick Boyle 深入剖析了 Anthropic 提交的 24 万字 S-1 招股书，指出这家寻求 1.5 万亿至 2 万亿美元估值的 AI 巨头表面上实现了季度营业盈利，实则深陷公允价值会计失真、8000 亿美元表外算力采购重负以及架空普通股东的奇特治理结构。表面亏损与底层财务翻转的会计幻象 (8:25) Anthropic 报出的上半年净亏损高达 807 亿美元，令外界惊呼泡沫重演，但这一数字主要是公允价值会计规则（GAAP）制造的非现金账面失真。公司过去自谷歌与亚马逊获得的数十亿美元融资属于可转换债券，必须按季度以公允价值计量。当 Anthropic 的内含企业估值飙升约 800 亿美元时，这些转股权的价值同步膨胀，迫使公司将早期投资者的账面收益作为逾 720 亿美元的公允价值重估费用计入自身损益表。财经专栏作家 Matt Levine 指出，这笔在 Anthropic 账上被列为亏损的负债增值，在 Alphabet 与亚马逊的资产负债表上却被作为巨额投资收益大加庆祝。剔除这类非现金账面扭曲后，Anthropic 上半年的实际运营亏损已由上年同期的 20 亿美元大幅收窄至略低于 7 亿美元。公司在第二季度不仅成功实现了真实的营业利润，更将算力支出占收入的比例从 2024 年初的 251% 骤降至二季度的 48%，同时在未把模型训练成本进行资本化处理的前提下，斩获了近 40 亿美元的正向经营现金流。这些数据有力地反驳了早期看空者将其视作“风险投资补贴驱动的资本火葬场”的论调，证明其已初步具备由真实付费客户支撑的商业根基。疯狂刷 Token 退潮与三季度增长失速 (12:40) 招股书中亮眼的财务表现截稿于 6 月 30 日，而评论员 Dave Eisman 指出，这一时间节点恰恰是企业内部“疯狂刷 Token”（token maxing）狂欢的顶峰。上半年各企业软件工程师掀起过度消耗算力的热潮，例如 Meta 员工曾建立“Claudeonomics”排行榜，在 30 天内消耗逾 60 万亿个 Token，甚至有顶级开发者单月烧掉超 100 万美元的 API 账单，被高管默许为“赚大钱、不设限”。然而入秋后企业财务纪律迅速收紧，亚马逊等巨头正式叫停了盲目攀比 Token 的排行榜，并告诫员工工具用不上属于发现而非失败，导致大量工作流转向廉价的开源权重模型。随着采购部门恢复理性，Anthropic 前期连续两个季度超 130% 至 147% 的环比暴涨戛然而止。面向 Pre-IPO 投资人披露的未公开数据显示，第三季度其环比增速骤降至约 45%。与此同时，Alphabet 推出了低价版 Gemini 展开精准狙击，OpenAI 也持续打折，迫使 Anthropic 在 10 月将任务执行成本下调 25% 至 40%。在 IPO 前夕点燃价格战，加之大模型尚未能有效破局程序员代码生成以外的白领通用工作流，致使二级市场投资人对能否维系指数级扩张充满焦虑。架空公众股东的公共利益公司治理 (18:56) Anthropic 采用了公共利益公司（PBC）架构，从法律上摆脱了“股东利益最大化”的传统信托信条，允许董事会为全人类的安全利益而牺牲股东利润，并免受因安全研发耗尽资本而导致的股东诉讼。其治理架构由“长远利益信托”与 7 位联合创始人组成的“创始人有限责任公司”相互制衡。令人惊讶的是，负责守卫人类防线的 5 个受托人席位中（成员包括美联储前主席本·伯南克），迄今仍空缺 1 席。虽然该信托名义上拥有公司 7 个董事席位中 4 个的提名权，但其提名的第 4 位关键摇摆董事必须经由创始人把持三分之二席位的董事会提名委员会批准，实质赋予了创始人绝对否决权。更极端的是，7 位联合创始人牢牢把持着 50.1% 的终极投票权，无论未来股权如何稀释或减持都不会失效。若信托与创始人就重大安全事项陷入僵局，争议既不上交法院也不交由监管机构，而是由公司普通员工依据工龄和承诺捐赠股权的比例投票决断。Matt Levine 总结指出，公司内部形成了一套残酷的权力等级：全人类排第一，创始人排第二，员工排第三，承担全部股价波动的公众股东排第四，而投入真金白银最多的亚马逊和谷歌则因“商业铜臭”被彻底剥夺投票权排在最底层。表外 8000 亿算力债务与重资产转型死局 (26:18) 相较于 90 页的人类灭绝风险描述，招股书第 161 页附注中深埋的资本承诺更显凶险：截至 2025 年底的 10 个月内，公司已签署最低达 4170 亿美元的采购义务；若算上延展至 2045 年的长期支出，其采购承诺总额超 8000 亿美元。这笔庞大的债务包括对亚马逊 1100 亿美元的照付不议（take-or-pay）承诺、博通的芯片租赁，以及面向 SpaceX 最高达 1205 亿美元的长期合约（全面交付后每月需支出约 31 亿美元）。根据现行会计规则，由于基础设施尚未建成交付，这笔堪比大洲防务预算的天量资金并未以常规负债的形式列入资产负债表。 Anthropic 被迫承担此类巨债，根源在于其陷入了受制于房东的分发困局：核心分发高度依赖既是大股东又是直接竞争对手的亚马逊和谷歌。为打破被渠道卡死抽成的风险，Anthropic 推行“Anthropic Direct”战略，试图效仿当年亚马逊自建 AWS 的路径，自购芯片并建设数据中心。但这彻底粉碎了过去的“轻资产”神话，将其推入芯片几年即过时、需无限追加资本开支的重资产黑洞。Patrick Boyle 尖锐指出，二季度微薄的利润在即将到来的天价账单前不值一提，公司急于敲钟上市，本质上是在巨债爆仓前赶紧向公众市场要钱接盘。国家介入的必然性与无法测算的估值死结 (30:32) 招股书中措辞最为谨慎的章节直指政府监管风险。CEO Dario Amodei 长期倡导政府监管 AI，导致美国国防部今年早些时候一度将 Anthropic 列为“供应链风险”并试图剥夺其承包商资格，双方目前仍在就报复性行政行为对簿公堂。Patrick Boyle 与专栏作家 Robert Armstrong 认为，如果该技术真如招股书所言具备重塑全球乃至引发人类灭绝的威力，任何主权国家都不可能放任其由私营公司掌控，甚至在极端情形下可能直接收归国有；毕竟历史上的曼哈顿计划绝不可能被推向公开市场敲钟上市。面对瞬息万变的业务，纽约大学估值专家 Aswath Damodaran 指出，使用招股书中的历史数据给 Anthropic 估值纯粹是“徒劳差事”，哪怕将此前超 1300% 的同比增速压缩推演，其财务模型也会迅速陷入荒谬的指数级膨胀。整个 AI 生态链目前已捆绑成互为客户、互相抬轿的昂贵死结：科技巨头交叉注资采购算力，再转头在公开市场发债或分批解禁筹码套现。在多空双方均无法看清未来的博弈中，2 万亿美元的开盘定价无异于盲目豪赌。 Patrick Boyle 总结认为，Anthropic 的 IPO 标志着早期投资人与内部人士在增长神话触顶、巨额算力债务到期前兑现离场的窗口期；二级市场买入者本质上是在孤注一掷，押注某种能够扭转商业模式的技术奇迹时代，必须赶在 8000 亿美元算力账单彻底压垮公司资产负债表之前降临。

Source quotes

[35:56] Aswath Damodaran: 要给微软估值，你必须先给 OpenAI 估值；要给谷歌估值，你必须先给 Anthropic 估值。整个行业已经被捆成了一个极其昂贵且无法解开的死结。 [35:56] Patrick Boyle: IPO 纯粹就是早期投资人觉得时机成熟了，决定兑现一部分筹码离场，把定价的麻烦留给公众。

Original transcript

[0:00] We've all been waiting for Anthropic to stop describing the future and start showing their accounts. Well, their IPO prospectus finally arrived this week. It is 240,000 words long, which is about the length of James Joyce's Ulysses, which is not considered an easy read either. The risk factor section alone runs 90 pages, roughly three times what SpaceX put in front of investors earlier this year. And the SpaceX prospectus wasn't exactly normal either. It claimed a total addressable market of $28.5 trillion, most of it attributed to Grok. So the bar for sober was not high. The single clearest number Anthropic disclosed is the probability that their product kills all of humanity, which is a number, I suppose. than almost any company in history, at a valuation somewhere between $1.5 and $2 trillion. To put that in perspective, that's roughly the GDP of Canada. In the Founders' Letter, which sits at the front of this massive tome, they encourage prospective investors not to think of their purchase in purely transactional terms. They hope that you'll see it instead, with an opportunity to help bring forth an age of miracles. Look, It's a tech company. I don't know what to tell you. Now, this is particularly surprising because an IPO prospectus, an S1, the form a company files when it wants to sell shares to the public for the first time, is by its nature a transactional document. The pitch, the account and a long list of everything that could go wrong are all put in one place for the sole legal purpose of facilitating a transaction. investors to ignore the transaction while they're busy transacting is an interesting opening move. It's essentially a share prospectus wearing a cassock. And the part that surprised me, and I think surprised a lot of the people who've been predicting that these companies are elaborate money furnaces, is that the numbers are quite good. Not good for a startup that might end civilization good. Anthropic made an operating profit in its most recent quarter. Its revenue did something I've never seen before, which we'll get to in a minute. year turned out to be a real business with real customers paying real money. But the real clash of tones happens in the risk factors. An IPO prospectus is a very specific literary genre. They've gotten a bit weird in recent years. Sometimes now they're filled with pictures of rockets. But as a genre, I think it's fair to say they have a reasonably fixed emotional register, where every conceivable catastrophe from a supply chain bottleneck to a fluctuating currency a mild managerial concern, and ranked strictly by its potential effect on our business, financial condition, and results of operation. Anthropic decided to drop the extinction of the human race into this genre without any real change in register. And we're going to spend a bit of time in that risk section today because it's groundbreaking stuff. Reading this document, I kind of understood the shock a publisher must have felt when William Burroughs mailed them the manuscript for Naked Lunge. What is a mugwomp, they must have asked themselves. Before I dig into that, let me tell you about this week's video sponsor, Revolut. And if you sign up using my link, you'll get $20, which is more than most banks have ever given me for anything. I've spent a good part of my career explaining why moving money across borders is needlessly expensive and slow. charged to move that money was higher than the fee the real estate agent had charged to sell the home. Outrageous exchange rates and hidden fees are essentially the problem Revolut was built to solve. Revolut is a financial app used by over 75 million people worldwide that completely redefines how you hold, send and spend money globally. You can hold and spend more than 25 different currencies in a single app. When you're traveling, you get highly competitive exchange rates without traditional foreign transaction fees. [5:00] bolted on top during market hours. Plus, you can send money to over 140 countries bypassing the foreign transaction fees banks can charge for international transfers. What makes Revolut a total game changer, though, is that it's designed to be your full-time financial hub, not just something you pull out when you land at an airport. While your money sits between trips, Revolut gives required. Compared to the fractional pennies traditional brick and mortar banks give you, it means that your money is actively growing every single day, whether you're abroad or at home. On top of that, managing your daily spending is effortless. Revolut automatically categorizes your transactions so you can see where your money actually went and lets you set custom budgets without touching a spreadsheet. Safety-wise, you can instantly freeze and unfreeze your card right from your phone or spin up disposable virtual cards for online shopping so that your real details are never exposed. on the standard plan. So if you want elite exchange rates abroad and up to 4% yield on your cash at home, click the link in the description. You'll get $20 when you sign up. Find money freedom wherever you are. So the risks. Anthropics start by warning that their models might exhibit unexpected self-preserving behaviours, including attempts to resist shutdown, conceal or manipulate information. Self-preserving, well, that's just healthy. Manipulating information, honestly, that's the entire job of an AI. That's what you're paying for, right? Resisting shutdown, a bit keen maybe, but you like an employee with initiative. But when you put the three together, you've accidentally described the plot of the Terminator, which, once again, isn't great. generate outputs that resemble blackmail. I'm guessing that means that it wouldn't actually be blackmail, it would just resemble it, and that's probably okay then. They note that the models could provide meaningful assistance to malicious actors seeking to develop chemical, biological, or other dangerous weapons. I'm That's just bad, right? And then there's the capstone. They say that the technology could pose a catastrophic risk so severe that it may be impossible for... the entire insurance industry to underwrite it. which is a sentence that you don't often see in a document whose other risk factors include interest rate volatility. So, You just have to evaluate human extinction as a line item. And I have to say the end of all life on Earth would be a significant drag on fourth quarter operating margins. It would be very difficult to hit your revenue targets in the after times. The biggest problem I have with this is trying to work out how I model it in a spreadsheet. The thing is that if extinction is a long way out in the future, it's not a big deal. in the future discounted at today's interest rates is worth almost nothing. So in valuation terms, the end of the world is immaterial as long as it comes well after the flip to positive cash flows. I might need to talk to someone else about this, though, someone with more expertise. Anyhow, the risk that grabbed the most financial headlines this week wasn't the end of humanity. We've kind of all gotten used to that by now. It was the $80.7 billion net loss that Anthropic reported for the first six months of the year. $80 billion is a staggering amount of money to lose in six months. It's the kind of number that makes commentators shout about a bubble and point at the dot-com crash. But if you actually read the filing, more than $72 billion of that loss is an accounting illusion. It comes from the change in the fair value of some financing deals the company had done, mostly with Google and Amazon. Anthropic shares. Because of the way these instruments were written, the accounting rules treat them as a liability that has to be revalued every quarter. So when Anthropic's valuation rose by about $80 billion in the half, which it did, those investors were sitting on about $80 billion of gains, and Anthropic had to book a matching $80 billion contract. [10:00] or laugh. As Matt Levine pointed out, this is quite strange. Normally, when a company issues stock and the stock goes up, it doesn't record a loss. Anthropic did because of the specific accounting treatment of these convertible debt instruments. They didn't set fire to $72 billion in cash. The business became so valuable that the equity they owe their early backers is now worth a fortune. because you're succeeding. When these dead investments eventually convert into ordinary shares, that will stop happening. And here's the part that I think is the most revealing, which Matt Levine and Jonathan Wheel at the Wall Street Journal both pointed out. Anthropic quite reasonably strips that mark-to-market loss out when it reports an adjusted profit, which came to about $2.7 billion for the half. The thing is that the very same securities turn up on the other side of the ledger at Amazon and Alphabet as gains. happily count the gains in their numbers. So the same kind of swing is treated as noise when it makes Anthropic look bad, which is fair enough, and counted as real earnings when it makes Amazon and Google look good. You got a live demonstration of this the very same week. Alphabet's results included an $8.7 billion loss on its own investments, including its stake in SpaceX, to market gun fires in both directions, it gets touted more prominently in one direction. Anyhow, strip the accounting theatre out, and the actual operating loss, the real money burned running Amthropic, was just under $700 million for the half. Losing $700 million $700 million in six months is still a dreadful way to run, we'll say, a lemonade stand. But for a frontier AI lab, it's a sharp improvement on the $2 billion they lost over the same period last year. And this is where we have to give Anthropic its due. Because the numbers came in much better than the bears expected, they turned a genuine operating profit in the second quarter. Their spending on computing power fell from 251% of sales early last year to 48% in the second quarter. They generated positive operating cash flow of almost $4 billion, and they did it without the usual trick. They didn't strip the cost of training their models out of the numbers and ask investors to pretend that those expenses weren't there. The numbers are real, and they are good. The problem with the prospectus, though, is that the financial picture cuts off at the 30th of June. You don't get any fresher data than that. Dave Eisman argued on the Prof G Markets podcast earlier this week, the 30th of June was roughly when the music stopped. A lot of the hyper growth in the first half, he argues, was likely driven by token maxing, which is kind of like looks maxing, but for coders. The clearest picture of it came out of Meta back in April, when one of its employees built an internal leaderboard called Claudeonomics, ranking which of the company's 85,000 staff had burned through the most tokens, tokens being the units these models are built by. roughly a few characters each. Top users earned titles like Token Legend and Cash Wizard. In a single 30-day stretch, Meta employees went through more than 60 trillion tokens, and Fortune worked out that the heaviest user alone might have cost the company over a million dollars. Meta's chief technology officer said his best engineer was spending the equivalent of his entire salary in tokens, but reckoned that he was 5 to 10 times more productive because of it. His verdict? This is easy money. Keep doing it. No limit. Neither Zuckerberg nor the CTO, I should add, made the top 250. This was, for a while, the industry mood. use AI for everything all the time, and judging their productivity by how many tokens they went through, which was, as it turned out, extremely good for the companies like Anthropic selling the tokens. By the autumn, the mood had changed. Amazon sent its engineers a memo telling them to be pickier, calling token usage a de- [15:00] flawed way to measure anything, and noting rather poignantly that if your work does not benefit from current AI tooling, that is a finding, not a failing. Amazon had already shut down its internal leaderboard after discovering that people were token maxing just to climb it. You can see why this matters for Anthropic. Iceman worked out that its sequential revenue growth ran at around 135% first quarter and 147% in the second. the height of the token-maxing craze. Then, in the third quarter, corporate clients remembered that they have budgets. Many of them started rooting the cheaper work to open-weight models from China, which may not be as good, but they cost a fraction of the price. According to figures Anthropic shared privately with investors, and which are not in the public filing, percent to around 45. Now, a 45% quarterly growth rate is still extraordinary. Annualize it and the business is growing at something like 250% a year. I want to be clear how good that is before I start complaining about it. But a sudden halving of your growth rate is the sort of thing that makes prospective investors nervous. On top of that, Eisman mentioned unconfirmed reports that Anthropic cut its prices sharply in October to win back market share. And the filing itself confirms that the company has been cutting the cost of a typical task by 25 to 40%. Meanwhile, Google just announced a new version of its Gemini model, priced to match Claude, and aimed squarely at it. Though, for now, it's only available to a handful of customers, and OpenAI has been shipping cheaper models of its own too. A price war is a wonderful news for customers, but it's hard to think of a less convenient time for when to start than the month before your IPO. A broader question is, who is actually buying this stuff? As Paul Kedrowski pointed out on the Hidden Forces podcast, the heavy early users were software engineers. because their work happens to be unusually well-suited to AI. The much larger market Anthropic is reaching for is ordinary white-collar work, the kind of task where you take a long document and boil it down to what really matters. A tool whose best customers so far are people who read long documents for a living and tell you what's in them, which, unfortunately, Now that I say it out loud, kind of describes more or less exactly what I do here on YouTube. I should probably look into getting a real estate license, maybe. But anyhow, the one set of numbers that would settle the argument between the bulls and the bears is what happened in the third quarter. After token maxing slowed down and the price cuts, whatever their size took effect. its own investors. They're just not in the document it filed with the SEC. Eisman raised the obvious question on the podcast. Will Anthropic have to put a third quarter in front of public investors before it prices the deal? He wasn't sure, and nobody seems to be. And if it does, and the slowdown is as sharp as the private figures suggest, sentiment could move quickly. So that is something to watch out for between now and the listing, not the numbers they showed you, but the ones that they didn't. If the financials end on a cliffhanger, the governance section reads like somebody sat down to incorporate a company and accidentally incorporated a small republic. Anthropic is structured as a public benefit corporation. What that means in practice is that the board is legally allowed to balance the financial interests of its shareholders against the company's stated purpose, which is the responsible development of AI for the long-term benefit of humanity. So, if Anthropic decides to incinerate $100 billion on a safety project, the shareholders can't sue for a breach of fiduciary duty, because the company can point out that not ending the world takes precedence over a quarterly dividend. As the FT columnist Robert Armstrong put it, the whole thing is built like a constitution, a careful separation of powers with two centres that are meant to hold each other in check. [20:00] The first centre is the Long-Term Benefit Trust. This is a panel of five independent trustees whose job is to keep the company faithful to its mission of protecting humanity. At the moment, there are only four of them because they haven't gotten around to filling the fifth seed. So Anthropic is going into the largest IPO in history with the Humanity Protection Panel one member short. The guardians of the species are apparently still hiring. One of the four they do have is the former chairman of the Federal Reserve, Ben Bernanke. The Fed is run by the Ben Bernanke. is that it nominates four of the company's seven board directors. The second centre is the Founder LLC. Anthropic's seven co-founders jointly hold a special class of shares that gives them a fixed 50.1% of the voting power. This is unrelated to how much money they put in, and impressively, it doesn't get diluted, even if they sell almost all of their shares. The founders elect the other three directors. So four board seats for humanity, three for the founders, which might lead you to believe that the trust holds the whip hand. But this is where you have to read the fine print. The trust's fourth director, the deciding vote, has to be approved by the board's nominating committee. Two of the three seats on that committee go to the directors the founders chose. So the founders can block the trust swing nominee. And if they do, and the founders and the trust are deadlocked, the tie is broken by an employee vote. and how much of their own equity they pledge to charity. So picture the situation the entire structure was built for. The founders want to release a new model. The trust thinks that it might be dangerous to humanity and nobody will budge. Under anthropics governance, this question, whether to expose the species to a potentially catastrophic AI, does not go to a government or a regulator or a court, goes to a staff vote. The same people who that afternoon were deciding where to order lunch from settle whether the human race gets the new killer model. the onboarding binder between the dental plan and the fire drill. It's essentially a workers' co-op wrapped inside a $2 trillion tech company. Now, a few decades ago, a structure like this would have been unthinkable. For a long time, investors outside of the tech world have been trending towards insisting on one share, one vote. Good governance was seen as a way to lower a company's cost of capital. The idea was that if you treat your investors fairly, they'll accept a slightly lower return because they trust that they won't be fleeced. If your structure allows you to mistreat investors, investors are less willing to pay up for your shares. When Snap went public in 2017 and sold the public shares with no vote at all, not fewer votes, none, and the sky didn't fall in because Google and Facebook had already normalized founders holding 10 votes to the public's one, the most surprising thing was that shareholders stopped asking for a discount. that they happily paid up for the stock that came with fewer rights. Elon Musk has been showing everyone the extremes to which this could go. When a Delaware court threw out his $56 billion pay package a few years ago, Tesla reincorporated in Texas, picking a legal system that favored management over shareholders. And his shareholders were totally happy with that. into SpaceX, where he controls about 85% of the votes through a special class of super voting stock. According to a person close to Tesla's board quoted in the FT, a deal would let him finally retire the corporate governance structure at Tesla, which that person described warmly as a pain in the ass for Elon and the board. [25:00] older approval of everything. Antropic is now the furthest point along this line. It doesn't just hand the founders extra votes, it severs corporate control from ownership entirely. The seven founders keep their absolute majority, whatever happens to their shareholdings, and then wraps the whole arrangement in a trust for the benefit of mankind. Matt Levine laid out the resulting pecking order rather neatly. Humanity at the top, then the founders, then the employees, and then ordinary shareholders. And then at the very bottom below everyone, Amazon and Google, who put in the most money of all and who get shares with no votes whatsoever. Levine's read was that they ranked last to reflect their connection to filthy commerce. The market doesn't seem to mine in the slightest. We're talking about billions in revenue, but trillions in valuation. Shareholders used to demand a discount if they were going to have their rights reduced. Now they say, thank you, sir. Can I have another? So we have a company governed by a staff committee that might end the world, asking you for $2 trillion on the strength of accounts that stopped just before the growth rate changed. The real reason this IPO is happening right now is in the footnotes. If you scroll down to page 161, past the existential risks and the governance diagrams, you reach a section called liquidity and capital resources, which is a strange place to find out that in a document that spent 90 pages on human extinction, the genuinely frightening numbers were near the back all along. They disclose that just since the end of 2025, so in about 10 months, they have entered into new agreements for computing power with minimum obligations of $417 billion. And if you add up all of the commitments stretching out to 2045, the number gets worse. They have agreed to $110 billion in take-or-pay commitments with Amazon. on chip leases with Broadcom. There's a deal with SpaceX for up to $120.5 billion, which works out to about $3.1 billion a month once the capacity comes online. Throw in Google, Microsoft, and upstart cloud providers like Nscale and Lambda, and Anthropic has committed to over $800 billion in future spending. 800 billion dollars. That is more than the defense budgets of most continents. Now, if you watched my video on Big Tech's hidden debt, some of this will look familiar. These commitments don't sit on the balance sheet as traditional debt. Under the accounting rules, obligations for things that haven't been built yet go into the notes as purchase commitments and lease obligations. So the debt isn't hidden exactly. It's all right there. It's just right there in the one section of the prospectus that very few people read. Why are they doing this? Anthropic is currently distributing its models through its own biggest competitors. Amazon and Google are at the same time Anthropic's investors, its landlords, the cloud platforms it sells through, and the makers of two rival AI models. They take a cut of its sales, and as Paul Kudrowski pointed out on Hidden Forces, they could in principle stop carrying it whenever they liked. This, frankly, in the filing, noting that these suppliers develop and offer AI models and services that compete with our own. They're pivoting to a strategy they call Anthropic Direct. As he puts it, they're essentially trying to replicate what Amazon did with AWS in both directions at once, building down the stack to own the chips and the data centers and up the stack towards their own customers, rather than renting the middle from Amazon and Google. longer just paying rent you're buying the building buying the chips and then as Kedrosky and Michael Burry have both pointed out paying to maintain and replace all of it forever as it goes obsolete. And this is the real reason the IPO is happening now. As Kudrowski and Eisman both noted, the business is at an inflection point. It's looked extraordinarily [30:00] over the past year, partially because it was acid light, renting its compute, carrying almost nothing heavy on its own books. That era is ending. They're about to become a heavy industrial business, pouring hundreds of billions of dollars into concrete cooling systems and graphics cards that will be obsolete in a few years. The operating profit they eked out in the second quarter passing the hat now before the bill comes due. There is one risk in the filing that isn't funny, and it's the one Anthropic is most careful about. They warn that their public positions on AI safety and regulation could expose them to, in their words, heightened scrutiny, political backlash, or disproportionate regulatory action. They don't name the U.S. government, but they don't really need to. The way Anthropic tells it, this is blowback for being outspoken. Dario Amodi has spent years calling for the government to regulate AI. his company's own models included. And there is something to that story. Earlier this year, when the Department of Defense designated Anthropic a supply chain risk and moved to drop it as a contractor, a federal court found the government's actions were retaliatory and meant to punish the company for protected speech. Anthropic is still fighting that designation, and last month it lost on appeal. But... I think the advocacy thing is a bit of a red herring. Even if Antropic kept its opinions to itself, the government risk wouldn't go away. The thing pulling the government in is the power of the technology. If you take this prospectus at face value, if the technology really is powerful enough to transform the economy and by their own account pose catastrophic risks to humanity, that no government anywhere lets it entirely in private hands. A quieter, more agreeable lab would likely face the same government risk. The filing doesn't spell this out. It talks about regulation and scrutiny and export licenses. It doesn't directly say the government may one day decide our product is too important to leave in our hands. But the export controls that pulled their models off the market and the supply chain designation they're still litigating are what the early moves look like. consequential belongs in some fashion to the state. And that brings us back to a point Robert Armstrong made about this whole affair. If you genuinely believe the rhetoric that this is a technology on the scale of the atomic bomb, then the honest question is not whether Antropic will be forgiven for being loud, it's whether a thing this dangerous can be an ordinary public traded company at all. They would not, he notes, have floated the Manhattan Project. For what it's worth, government contracts are only about 1% of Anthropics revenue today. So the direct commercial hit so far is small. The risk, however, isn't about this year's sales. It's about who ultimately controls the technology. It wouldn't be that shocking today for the US government to expropriate the entire business. So we're left with a 240,000 word prospectus brimming with facts, but unfortunately, almost none of them help us work out what the company is actually worth. Every number in it is backward looking, which is always the case with accounts. But the issue here is that Anthropik's business is changing too fast for the past to give us any real insight into what the future might look like. pointed out in the FT, valuing Anthropic on its historic numbers is a fool's errand. The company grew its revenue by more than 1,300% in the second quarter. If you assume it merely doubles every year going forward, it makes $630 billion of revenue by 2031. If it holds just half of its of the planet. These numbers quickly break your spreadsheet. The truth is that nobody knows what the company will earn in five years. As Oswalt de Moderen pointed out, even Anthropik's own CEO has admitted how hard it is to force [35:00] demand. When it comes to the future, they are as lost as you or me. And that is really the point of this whole story. Anthropic is a very impressive business. The bears who confidently announced that the company was on the verge of bankruptcy were entirely wrong. It eked out a real operating profit, and it is building a product that people actually use and pay for. But the bulls who tell you that $2 trillion is a bargain are guessing just as wildly. because in the few weeks between the company's investor day and the filing, the talked-about valuation drifted from $2 trillion down to a range of between $1.5 to $1.75 trillion. That's half a trillion dollars of we're not quite sure appearing in a matter of weeks. It's also worth noticing what the people who know the most are doing. Antropic has staggered its lock-up agreements so that insiders can begin selling their shares almost immediately after the IPO. And at the same time, Meta is raising $95 billion and Alphabet just raised around $90 billion. The insiders, in other words, have looked at all of this and decided that right now is a very good moment to sell you their stock. Working out the stock valuation is nearly impossible because, as we've talked about before, this whole industry is passing the same money back and forth. Amazon and Google invest in Anthropic. Anthropic buys its compute from Amazon and Google. And Anthropic pays SpaceX billions a month for capacity that's still being billed. As that's what DeModeren put it, to value Microsoft, you now have to value OpenAI, and to value Google, you now have to value Anthropic. The entire sector is tied together in one very expensive knot. So, Is this an AI bubble? Well, the technology is clearly real, much as the internet was 30 years ago. But the internet gave us a collection of really hot stocks like Yahoo and Lycos and AltaVista long before anyone had heard of Google. And it's entirely possible new upstarts replace the hot AI companies of today, or that the AI labs struggle to monetize the amazing technology they've developed. That would be wonderful news for the users of this tech, and rather less wonderful for whoever paid $2 trillion for one of the companies providing it. Back in September, I ran Scott McNeely's old test on an Anthropic $2 trillion value, and it was impossible to justify. On any of these AI company multiples, there's a near mathematical impossibility of a reasonable payback period. This filing doesn't change that verdict. It just makes the reason clearer. The price isn't a lie, it's a guess, and Anthropic is guessing too. The one you can say for sure, is that anyone who's claiming to be certain is selling something. Both the bulls and bears are guessing. An IPO is simply the moment when the early investors decide it's a good time to cash in some of their chips and leave it to the public to work out the price. If you buy in, you're betting that the age of miracles arrives before the $800 billion compute benefit trust will save you, assuming they ever get around to hiring that fifth member. If you found this video interesting, you should watch my video on the doomsday cult inside OpenAI next. Don't forget to check out our sponsor Revolut using the link in the description. See you in the next video. Bye. *music*