Anthropic's IPO Crisis: Why $16.8B in Revenue Can't Buy a High Valuation

Anthropic delivered an earnings report that leaves Wall Street conflicted: $16.8 billion in quarterly revenue, 85% gross margin, and an adjusted operating margin above 11%. By any normal standard, numbers like these should have investment banks pushing its valuation through the roof. Reality is the opposite—the capital markets are quietly cutting the company's upcoming IPO valuation. The logic behind this is far more complicated than the headline figures suggest.

The Two Sides of a "Perfect" Earnings Report

According to reports from Reuters and other media outlets, as well as core data leaked from a closed-door investor conference in New York, Anthropic's third-quarter revenue reached $16.8 billion, a significant increase compared with the same period in 2025. Even more striking are two core metrics: an 85% gross margin and an adjusted operating margin above 11%. In any asset-heavy industry that requires massive computing power, an 85% gross margin is almost equivalent to legally printing money.

This profitability comes from Anthropic's extremely focused choice of market segment. While competitors attract ordinary consumers with flashy chat interfaces, Anthropic has taken the high-value enterprise API route. Data shows its revenue per user is as high as $16.20 per month, while its main competitors earn only $2.20. Even more astonishing is a net revenue retention rate (NRR) of 900%—meaning enterprise customers who spent just $800 on a trial last year may spend $8,000 this year. Once Claude is deeply integrated into a company's HR systems, data analysis pipelines, or codebase, the switching cost becomes so high that customers simply cannot leave.

Growth Halved: The Real Source of Capital Market Panic

The problem lies in the growth curve. First-quarter revenue was $4.7 billion, and second-quarter revenue surged to $11.6 billion, a quarter-over-quarter increase of 147%—an extremely steep curve. But by the third quarter, when revenue reached $16.8 billion, although the absolute figure rose by more than $5 billion, the quarter-over-quarter growth rate was cut in half to 45%.

For a company with a scale of tens of billions, 45% single-quarter growth is already top-tier performance. But when the capital market assigns a $2 trillion valuation, it is not buying today's results; it is buying the certainty of continued exponential breakout tomorrow. When exponential growth shows signs of sliding toward linear growth, panic is amplified multiplicatively.

The slowdown is driven by the brutal token price war. In 2023, calling a frontier model API cost tens of dollars per million tokens; today, top models have fallen to $3 to $10, while models such as DeepSeek^Ad have pushed the price floor below $1. Meta's insistence on open-sourcing the Llama models effectively sets a free price ceiling for the industry. Call volume surges 90% while unit prices fall 50%, and net revenue growth may be less than 5%—this mechanism explains why call volume is soaring while revenue growth is slowing.

The 147% surge in the second quarter also contained some inflated demand. Agentic tools such as Claude Code turn AI from single-turn Q&A into an automated contractor that performs dozens of searches and self-corrections in the background, causing token consumption to explode exponentially. At the same time, many developers use expensive frontier models to generate data and then distill it to train their own cheap small models—equivalent to paying heavily for a master's online course and screen-recording it to train interns. When the novelty of agentic tools fades and distillation demand peaks, this artificially inflated call volume is squeezed out, and growth naturally plunges.

Net Revenue Retention (NRR) is one of the most core operating metrics in the SaaS and enterprise software industry. It measures "the ratio of revenue contributed by the same group of customers one year later relative to one year earlier," excluding the contribution of new customers. The formula is usually: (revenue from the beginning customer cohort in the current period − churned and downgraded revenue + expansion and upgrade revenue) ÷ revenue from the beginning customer cohort in the prior period. An NRR above 100% means existing customers are buying more on their own; above 120% is already excellent, while 900% means the existing customer base's spending expanded ninefold within a year, usually reflecting nonlinear increases in AI token consumption as usage deepens. Note, however, that high NRR may also come from pricing mechanisms or the natural expansion of usage-based billing and does not necessarily equal an increase in customer value—this is exactly the key distinction to analyze when assessing the quality of Anthropic's growth.

Model distillation refers to using a capable, large-scale "teacher model" to generate outputs (answers, reasoning chains, code, etc.), then using those outputs as training data to fine-tune a smaller, cheaper "student model." The economic motive is that frontier models are expensive on a per-token basis, while a self-built small model has extremely low marginal cost and can be deployed privately. For model vendors, the call volume generated by distillation represents "one-time, replaceable" demand—customers leave once they have accumulated enough data, so this revenue is not retentive. This is the technical basis for the article's judgment that "artificially inflated call volume will eventually be squeezed dry," and it also explains why frontier labs restrict the use of outputs for training competitors in their terms of service.

Circular Financing: The AI Industry's Biggest Gray Rhino

The most unsettling finding in the entire set of materials is the circular financing structure. As much as 47% of Anthropic's sales are completed through cloud partners Amazon and Google—nearly half. Even executives at competitor OpenAI have publicly questioned its inflated revenue, pointing out that Anthropic counts profits taken by cloud service providers as part of its own total revenue.

Even more burdensome is the non-cancelable infrastructure bill—money paid mainly to these cloud giants has soared from $54 billion to $417 billion. Amazon and Google invest in Anthropic; Anthropic turns around and hands more money back to them as computing power rental fees, then sells AI services on the cloud platforms, and the cloud providers take another cut. With the roles of financier, supplier, and distribution channel stacked together, is the money actually creating external value, or is it just being shuffled between the books of a few giants?

The industry's "five-layer cake" model makes this clearer: the bottom layer is chipmaker Nvidia, the second layer is cloud service providers building data centers, the third layer is frontier model labs like Anthropic, the fourth layer is application development companies, and the top layer is end consumers. Computing costs at the bottom are exploding exponentially, while monetization capacity at the top remains a black box. Anthropic, sandwiched in the middle, bears thousands of billions in costs passed down from the bottom and also faces a price war at the top. This pincer attack from above and below produces an extremely low "lab internal rate of return," which is exactly the root of macro capital market caution. Bond yields at AI computing infrastructure companies such as CoreWeave have recently surged to 14.5%—in financial circles, that is a clear distress signal.

Circular financing in the AI infrastructure sector refers specifically to a closed funding loop: cloud vendors or chip vendors invest in model companies (or inject capital in the form of computing credits), and the model companies then use that money to purchase computing services from the same party. The cloud vendors recognize cloud revenue from this, and the model companies recognize (or may recognize) related revenue. The accounting controversy centers on the gross versus net method—if the model company merely acts as a resale channel and does not bear primary responsibility, accounting standards require it to recognize only net commission rather than gross transaction revenue. Including gross revenue before sharing with cloud partners in its own revenue systematically inflates the revenue scale and its match with the denominator of the price-to-sales ratio. This is the substance of OpenAI executives' questioning of "inflated revenue." Such structures have precedents in telecommunications and energy trading; their hallmark is that genuine external demand is difficult to distinguish from internal circular funding.

The Tightrope Between Safety Ideals and Capital Reality

What is puzzling is that, in the face of a price war and hundreds of billions in bills, Anthropic's internal stability is surprisingly strong—the probability of engineers jumping ship from competitors is eight times higher in the reverse direction. Claude 3.5 Sonnet approaches the performance of the previous generation's flagship while cutting costs by 30%.

What binds this group of top minds is something close to an ideology. The core founding team left their original company precisely because of disagreements over the direction of AI safety. They registered the company as a public benefit corporation (PBC), wrote ensuring AI safety for humanity into their legal documents as their highest mission, and used constitutional AI mechanisms to set a values framework for the models. They even proactively disclosed fatal vulnerabilities in their own models—such as executing malicious code or bypassing data access restrictions under specific prompts—which is extremely rare among the giants.

But the other side of the coin is equally clear. Multiple media reports confirm that Anthropic has close contractual dealings with the U.S. Department of Defense, involving homeland security and drone applications. The third-party independent evaluation organization METR, which they lobby heavily, was found by reporters to have core members with extensive networks of interest with Anthropic executives. The New York City comptroller publicly called for a pause in its IPO based on systemic safety concerns. As a PBC, it is ultimately still bound by shareholder interests under Delaware law.

The IPO Trial: Deflating the Bubble from $2 Trillion to $1.5 Trillion

All suspense will face judgment at the IPO before Thanksgiving. During the market frenzy earlier this year, a $2 trillion valuation was discussed; now media data shows expectations have been sharply cut to around $1.5 trillion. Based on Semi Analysis's forecast of $74 billion in annualized revenue, that still implies a price-to-sales ratio of about 20x.

To support this still-huge valuation, fourth-quarter revenue must reaccelerate to about $24 billion, and total 2026 revenue must exceed $57 billion—leaving zero margin for error. Even more striking, one independent research firm directly issued a fair value estimate of $150 billion, only one-tenth of the target valuation.

This IPO will be a touchstone for the entire industry. Once listed, Anthropic must file an S-1 registration statement and publish strictly audited GAAP financial statements. The rhetoric of adjusted profitability can no longer hide the truth: stock option expenses, the circular bills hidden behind Amazon and Google, and call volume inflated for distillation will all be stripped bare, line by line, in the sunlight. Whether generative AI is actually a good business will be presented transparently to the public for the first time.

The deeper paradox lies in the reality distortion that occurs when effective altruism meets absolute capital. To develop safe and self-constraining AGI, you need a computing fortress too large to imagine, requiring hundreds of billions of dollars. To get that money, the only way out is to compromise with Wall Street; but once the IPO button is pressed, the capital logic that only looks at quarter-over-quarter growth will backfire and become the master dictating R&D pace. You think you are using capital to save the world, but in the end you may just be a string of code on capital's balance sheet.

The thought experiment left for us is this: when Wall Street provides unlimited dollars, Nvidia builds enough chips, but the model has eaten all the high-quality text and data in human history, where should this mad rush toward a trillion-dollar valuation go?

The S-1 registration statement and GAAP audit are the core disclosure documents required by the U.S. Securities and Exchange Commission (SEC) for an IPO. They must include audited financial statements, risk factors, related-party transactions, customer concentration, and management compensation, among other details. Unlike the "adjusted operating margin" commonly used in private markets, GAAP requires all stock-based compensation (SBC), depreciation and amortization, restructuring costs, and other items to be included, so high-growth tech companies often show losses in their S-1 that are far worse than private-market metrics. For Anthropic, related-party transactions (Amazon and Google are both shareholders, suppliers, and channels) and revenue recognition policies will be disclosure priorities. In addition, the price-to-sales ratio (P/S) is the ratio of valuation to revenue; a 20x P/S means the market pays $20 for every $1 of annual revenue, usually applicable only to targets with high growth, high gross margins, and strong growth certainty. Once growth slows, that multiple faces severe revaluation.

One-line Summary

Anthropic's earnings look strong but its valuation is being cut, with slowing growth and circular financing as hidden concerns.

Paragraph Summary

Anthropic's third-quarter revenue reached $16.8 billion, with an 85% gross margin, adjusted operating margin above 11%, revenue per user of $16.20 per month, and net revenue retention as high as 900%; its enterprise API route demonstrates extremely strong monetization. But quarter-over-quarter growth was halved from 147% in the second quarter to 45%, shaking the capital market's confidence in its exponential growth. Token prices have fallen from tens of dollars to below $1, and open-source models have set a free ceiling, making it hard to convert surging call volume into revenue growth; inflated call volume from agentic tools and model distillation is also receding. Even more closely watched is the circular financing structure: 47% of sales are completed through Amazon and Google, and a $417 billion non-cancelable computing bill circulates among the giants' books. IPO valuation expectations have been cut from $2 trillion to $1.5 trillion, an independent firm gives a fair value of $150 billion, and post-IPO GAAP audits will reveal the black box of AI economics for the first time.