Anthropic's $1.48 trillion valuation rests entirely on sustained growth

1. Growth with a missing quarter — ▶ 4:00

   Revenue grew more than 1,300% year over year and reached $16 billion-plus in the last reported quarter. The customer base is 11,900, of which about 1,900 spend over $1 million, and the annualized run rate could reach $100 billion by year-end. The bear reading starts with an omission. Third-quarter numbers are absent, and Anthropic's investor meetings suggest growth stalled. Candidate causes are a compute constraint, the impossibility of 100% quarterly growth at scale, and OpenAI competing on price. The top five customers still account for 25% of revenue.

2. Gross versus net revenue — ▶ 6:24

   Anthropic reports gross revenue, while OpenAI has shifted to net. Partnerships with big tech companies send 15-20% of revenue back to the partner, and Anthropic books that as a marketing expense instead of netting it out. The speaker says this is not deceptive, but it makes the two companies' revenue figures non-comparable. He adjusts for it by moving the partner share into cost of goods sold, which is why his gross margin of roughly 80% is a little below outside estimates.

3. Margins built on fixed costs — ▶ 7:11

   Compute costs fell from 250% of revenue six quarters ago to under 50%, and Anthropic reported its first operating profit in the latest quarter. The speaker says this is partly economies of scale and partly a reclassification of compute from a variable to a fixed cost. The trade-off is that operating profit now requires continued growth. Stock-based compensation, large and rising more slowly than revenue, becomes a drain if growth fades. Positive margins are therefore a bet on growth, not evidence of durable profitability.

4. Reinvestment hidden in commitments — ▶ 9:53

   Capex and invested capital look small, but the real investment sits in contractual compute commitments of more than $450 billion, or over $600 billion collectively, with Broadcom the largest beneficiary. Bulls see locked-in supply in a business where compute separates winners from losers, plus chips that may get cheaper and some flexibility in the contracts. Bears see a young, loss-making company with debt-like obligations that cannot be met without growth, and note that data centers are getting costlier to build, so compute may not get cheaper.

5. Prospectus as kindergarten report card — ▶ 16:30

   A prospectus describes the past, and valuation concerns the future. The speaker compares it to a report card for a precocious kindergartener: too early to plan for college. Almost no big question in AI can be answered from it, though it frames those questions. He also advises avoiding conspiracy theories about a cabal of AI companies. He reads circularity as self-interest, and the $72.6 billion convertible expense as a conversion option kicking in, which he calls a feature of paying with options.

6. Market size drives the revenue target — ▶ 20:27

   Total addressable market depends on four questions: whether AI is a tool or a replacement for employees, whether it targets high-paid or all workers, whether it reaches some or all industries, and whether it is US-centric or global. The answers yield less than $2 trillion or more than $10 trillion. The speaker holds $5-10 trillion and sees Anthropic leading the premium segment, citing the 80% of revenue from usage rather than subscriptions and from enterprises. He assumes $1 trillion in revenue in 2036, more than any company earns today.

7. Margin mix and capital intensity — ▶ 24:30

   He splits the market into a commoditized mass tier, with roughly 10% margins like business services, and a premium tier near 40%, like enterprise software. A 75/25 premium-to-mass mix gives a 32.5% operating margin, which he cuts by 5 points for regulatory costs. Because chips age quickly and power, real estate and water are also needed, compute remains a continuing drain even if easier to get. He assumes $3 of revenue per $1 of capital in steady state, and says scarce compute would lower that ratio.

8. Three layers of risk — ▶ 28:08

   Risk enters in three places. The cost of capital is 9.74%, near the third quartile for US companies (9.88%) against a median of about 9.1%, and it converges toward the median over time. A 10% probability of failure is attached, with a distressed sale at half value. Regulation is treated as a mild margin hit, as Facebook's compliance hiring raised costs without weakening its competitive position. Offsetting factors are roughly $200 billion in post-IPO cash and a perception that Anthropic is too big to fail. Its bond rating is BB+.

9. Commitments treated as debt — ▶ 32:56

   Operating assets are worth about $1.7 trillion. Adding about $220 billion in cash, including roughly $100 billion of IPO proceeds, and subtracting $473 billion of debt, the present value of compute commitments, gives $1.48 trillion of equity after failure-risk adjustment, about $447 per share on a provisional share count. The speaker is unconcerned that accounting does not treat the commitments as debt. He notes that he capitalized operating leases starting in 1992, while accountants waited until 2019, and says a real contractual burden belongs in the valuation.

10. Scenario range spans bankruptcy — ▶ 36:19

    The valuation outcomes span negative to $4 trillion. In the worst case AI becomes a commoditized tool, revenue stays low and the cost of capital is 9.88%, leaving equity at minus $373 billion, which means bankruptcy. In the best case AI replaces workers across industries, revenue reaches $1.5 trillion, margins hit 40%, compute gets cheap and the cost of capital falls to the first quartile, yielding $4 trillion. The speaker uses the spread to argue that investors need their own story, tested for whether it is possible, plausible and probable.

11. Five share classes, one winner — ▶ 38:17

    The speaker defines governance as shareholders' power to force change. Anthropic's structure gives public Class A shares one vote each, while Alphabet and Amazon hold non-voting Class N shares. A benefit trust holding Class T shares elects the board majority, and seven co-founders hold a single Class F share with 50.1% of votes. As a Delaware public benefit corporation it can make decisions that hurt shareholders without being sued. He compares this to Google's 2004 dual-class precedent and to ESG, and says it demands a much bigger discount.

Anthropic's prospectus does not settle whether it is a good business. It shows a company whose profitability depends on growth continuing, and that is a different thing. The speaker's method is to read each metric twice, once as a bull and once as a bear. Revenue rose from almost nothing to more than $16 billion in the last reported quarter, with growth above 1,300% year over year. The customer base has reached 11,900, of which about 1,900 spend more than $1 million. A bull sees sticky, expanding spend. A bear sees the same figures and asks how long they can keep compounding.

The strongest bear evidence is something the prospectus leaves out. Third-quarter numbers are absent, yet Anthropic's investor meetings suggest growth stalled. The company blames a compute constraint. The speaker lists two less comfortable explanations: 100% quarterly growth cannot persist at scale, and OpenAI is competing on price. Concentration adds to the worry, since the top five customers still make up 25% of revenue. Revenue is also reported gross, with the 15-20% partner share booked as a marketing expense, whereas OpenAI has moved to net. That does not mislead, but it blurs comparisons between the two.

The profitability story turns on a classification choice. Compute fell from 250% of revenue six quarters ago to under 50%, and the company posted an operating profit in the latest quarter. The speaker credits economies of scale but also notes that Anthropic has effectively moved compute from a variable cost to a fixed one. Fixed costs improve unit economics when volume is rising and punish a company when it is not. Stock-based compensation compounds the exposure. Positive margins here are a statement about the trajectory, not a settled fact about the business.

The same logic explains why reinvestment looks small. Capex and invested capital are modest because the real commitment sits in contracts with big tech companies: more than $450 billion on a contractual basis and over $600 billion collectively. Bulls read this as secured supply in a market where compute separates winners from losers, and hope chips get cheaper. Bears read it as debt-like. The speaker sides with the bears on accounting. His valuation treats the present value of those commitments, $473 billion, as debt, defending the choice with his own history of capitalizing operating leases in 1992, long before accountants did so in 2019.

With a prospectus that cannot answer forward-looking questions, the valuation becomes an exercise in stated assumptions. The speaker assumes a market of $5-10 trillion, Anthropic leading the premium segment, and revenue of $1 trillion in 2036, a figure no company reaches today. He blends 40% premium margins with 10% mass-market margins at 75/25, lowers the result by 5 points for regulation, and assumes $3 of revenue per $1 of capital. He uses a 9.74% cost of capital and a 10% failure probability, with failure meaning a sale at half value. These steps produce $1.7 trillion in operating value and $1.48 trillion in equity.

What the analysis shows is how little the mechanics matter next to the inputs. The scenarios run from minus $373 billion, in which the company goes bankrupt, to $4 trillion. A model with outputs that wide cannot confirm anyone's view. It can only translate stories into values. The speaker concedes he may be massively wrong, and he asks readers to enter their own assumptions and test them for whether they are possible, plausible and probable. Editorially, that makes the 10% failure probability and the $1 trillion target the points to watch, since both are judgments rather than findings from the filing.

Governance is the one subject on which the speaker is categorical. Anthropic will have five classes of shares. Public holders get one vote per share, Alphabet and Amazon hold non-voting shares with a claim on profits, and a benefit trust elects the board majority. Seven co-founders hold a single share carrying 50.1% of the votes. Anthropic is also a Delaware public benefit corporation, which, as the speaker describes it, lets it make decisions that hurt shareholders without legal recourse. He sees this as the Google dual-class precedent in a more extreme form, and likens it to ESG, where power was given to a few people on the promise of doing what is right.

That objection changes how he would act, not how he values the company. He says governance requires a much bigger discount before he would buy. If the IPO prices above his $1.48 trillion, he stays out, and a higher price makes him less interested. He dismisses the lock-up structure, in which non-founder insiders can sell up to 10% on day one and all insiders are free by day 180. Markets, he argues, know when unlocks are coming. The speaker adds that prices swing on news and he will revalue as it arrives, recalling that he eventually bought Tesla after valuing it from 2013 and calling it overvalued in 2018.

Behind everything is a question the prospectus cannot answer. Is Anthropic's slowdown a matter of capacity, which new data centers could fix, or of demand and competition? If it is capacity, commitments of more than $450 billion look prescient. If it is demand, they look like a liability that growth cannot cover. The two readings imply very different values, and no figure in the filing, nor in the speaker's model, can yet distinguish between them.