Decoding Anthropic's S-1: The Capital Magic Behind an $80.7 Billion Loss

Decoding Anthropic's S-1: The Capital Magic Behind an $80.7 Billion Loss

Anthropic has officially filed its S-1 with the U.S. SEC, preparing to list on Nasdaq under the ticker ANT, with a valuation range surging to $1 trillion to $2 trillion—making it one of the largest IPOs in human history. But what really set the internet ablaze is another number: revenue of $16.286 billion in the first half of this year, yet a net loss of $80.748 billion.

Many people's first reaction is "losing this much and still daring to go public." But after reading through this prospectus, you'll find that most people have misread it—this $80.7 billion loss is not about burning cash to the point of bankruptcy; it's the AI industry's first great capital magic trick.

Of the $80.7 Billion Loss, $72.4 Billion Is Just Paper Numbers

Let's break down the most terrifying figure first. Of the $80.748 billion net loss, $72.4 billion comes from non-cash accounting adjustments—purely a bookkeeping game.

The logic isn't complicated: Anthropic has issued quite a few convertible bonds and preferred stock financings over the past few years. Now that the company's valuation has skyrocketed, the fair value of these financial instruments has risen accordingly. Under U.S. GAAP accounting standards, this appreciation must be recorded as a loss. To use an analogy: you borrow $1 million from a friend to open a shop, agreeing that the friend can convert the debt into equity if the shares rise in the future. Now the shop's valuation has increased tenfold, so this debt grows on the books—accounting-wise it's a loss, but not a single cent of cash in your pocket has changed.

Deducting this bookkeeping adjustment, Anthropic's real operating loss in the first half was less than $700 million, and after adjustments, it even showed an operating profit of $1.6 billion. In other words, this prospectus doesn't present a company on the brink of bankruptcy, but a high-growth enterprise whose books have been "made to look unprofitable" by financial instruments.

GAAP and non-cash accounting adjustments: GAAP (U.S. Generally Accepted Accounting Principles) requires companies to measure certain financial liabilities at "fair value," especially embedded derivatives such as conversion rights and redemption rights in convertible bonds and preferred stock. When a company's valuation rises, the fair value of these instruments increases, and accounting rules require recognizing this as a current-period loss—even though the company has not paid out any cash. Such losses are added back in the cash flow statement, so when analyzing operating performance, one should focus on "non-cash adjustment" items rather than looking directly at net income. Investors typically refer to adjusted EBITDA or non-GAAP operating profit to strip out this noise and restore the true picture of profitability.

Growth Data: Revenue Up 13x Year-over-Year, Retention Rate at 900%

Setting aside the accounting noise, Anthropic's business data is indeed fierce. First-half revenue rose 13x year-over-year, Q3 grew another 45% quarter-over-quarter, and the net revenue retention rate hit 900%—meaning existing customers spent 9x more this year than last year.

The customer structure is also rapidly moving upmarket: there are 11,900 enterprise customers spending over $100,000 annually, 1,900 at the million-dollar level, and 250 at the ten-million-dollar level—the customer base has grown more than tenfold in a year. This kind of growth rate is rare even in the history of enterprise software, and it explains why capital markets are willing to give it a trillion-dollar valuation.

Net Revenue Retention (NDR): Net revenue retention measures the change in revenue generated by the same cohort of customers over a year. The formula is (beginning cohort's current-period spend − churned customer revenue − downgrade revenue + upgrade revenue) ÷ beginning cohort's prior-period spend. An NDR above 100% means existing customers are expanding or upgrading; reaching 900% means this cohort's spending ballooned to 9x the original within a year—typically seen in API or compute-consumption models where usage grows exponentially with business expansion, rather than in traditional subscription software. The higher this metric, the deeper the product is embedded in customer workflows, but one should also watch for distortion driven by a small base or a handful of large customers.

Five Share Classes: Buying Stock Doesn't Mean Having a Say

What most overturns conventional wisdom in the prospectus is its equity structure. Many people think buying stock makes them a shareholder who can dictate company decisions—at Anthropic, that simply doesn't exist.

The company has designed five classes of shares, with the core being Class F shares: held by the seven founders, each share carries a standalone 50.1% voting power. That is, no matter how many shares public investors buy, how the company's strategy unfolds and how money is spent is entirely up to the founders.

There is also a Class T stock, placed in a long-term benefit trust, specifically responsible for electing the majority of directors. Its core mission is to safeguard AI safety bottom lines—even if it means making less money or no money, AGI must not be allowed to spiral out of control.

Put simply, what public investors buy with real money is essentially dividend rights; company control has nothing to do with it. It's like pooling money with partners to open a restaurant: profits and losses are shared proportionally, but how the menu is set, who gets hired as chef, and whether to open branches are all decided by the boss alone.

Dual-class/multi-class equity structures and voting rights: U.S. capital markets allow companies to set up share classes with different voting rights—commonly Class A with one vote per share and Class B with ten votes per share—so founders can control the company with a smaller ownership stake by holding high-voting shares. Anthropic's Class F shares carry 50.1% voting power per share, an extreme multi-class equity design ensuring the founder team retains firm control over strategic decisions after going public. Such structures are not uncommon among tech companies (e.g., Google, Meta), but regulators and index providers typically restrict their inclusion in certain indices. What public investors buy is economic rights, not equivalent control rights.

Prisoner's Dilemma with Cloud Providers: $75 Billion Compute Commitment

Digging deeper, Anthropic's relationship with cloud providers is a classic prisoner's dilemma. Amazon and Google are both major shareholders and the largest compute suppliers. The prospectus discloses that Anthropic has signed massive non-cancellable cloud computing contracts, with future compute commitments reaching as high as $75 billion.

This creates a perfect internal loop: the money Anthropic earns from enterprise customers flows right back into shareholders' pockets in the form of compute payments. On the surface, it's an industry star; in reality, it's more like a front-facing white glove for cloud providers—helping them monetize compute while bearing all the technology and brand risk.

Jevons Paradox: Unit Costs Fall, Total Spending Rises

There's also a classic economic phenomenon here—the Jevons Paradox. Unit compute costs are indeed falling, with per-token costs down over 13%, but usage is growing even faster. Companies are now building AI agents and complex reasoning, and per-user compute consumption has risen nearly 60%. The result: the cheaper the unit, the faster total usage grows, and total compute spending actually keeps increasing.

It's like gasoline prices dropping, but people drive more frequently, so the monthly fuel bill goes up instead. So Anthropic is not a light-asset software company at all—it is essentially a heavy-asset compute utility company. What it sells isn't code, but electricity and chips packaged by algorithms.

Jevons Paradox: In the 19th century, economist William Jevons observed that improved steam engine efficiency reduced coal consumption per unit, yet total coal consumption rose—because cheaper energy spawned more application scenarios. In the AI compute field, falling per-token costs stimulate companies to deploy more complex models, longer contexts, and more agents, causing total compute demand to rise rather than fall. This paradox means efficiency gains may not reduce infrastructure spending but could accelerate the capex race, giving AI companies heavy-asset characteristics similar to electricity or telecommunications utilities rather than light-asset software firms.

OpenAI's Snipe and Anthropic's Response

Right at the critical moment of the IPO, OpenAI precisely landed a blow: offering enterprise customers migration discounts of 20% to 40%, clearly aimed at the IPO. Wall Street cares most about gross margin and customer retention, and OpenAI is targeting exactly these two points.

Anthropic didn't respond with direct price cuts, but switched to smarter tactics:

• Changing the calculation: Instead of comparing per-token prices, compare the total cost of completing a task. For example, in coding, Claude has a high first-pass success rate and doesn't require repeated debugging—seemingly 30% more expensive per unit, but cheaper for the entire project.
• Prompt caching: Reduces costs by 60% in long-document scenarios, effectively giving existing customers a discount while keeping list prices unchanged, preserving gross margin.
• Doubling down on coding agents: Deeply embedding into enterprise development workflows, making the cost of switching models far higher than the discount difference.
• Cloud service credits: Amazon and Google offset each other with cloud service credits, effectively having cloud providers subsidize the price cuts on their behalf.

Implications for China's AI Industry

Anthropic's IPO formally declares that the large model industry has moved from the light-asset, high-margin internet myth into a heavy-asset, compute-intensive utility era. The window for pure algorithm storytelling has closed; what comes next is competition on compute costs, supply chain control, and vertical scenario implementation capability.

In China, general large model vendors like MiniMax and Zhipu are also pursuing similar commercialization paths, but the binding between domestic cloud providers and large models will be deeper, and a landscape of cloud provider equity stakes plus compute binding is likely to emerge. For entrepreneurs, opportunities in general large models are dwindling, but there is still ample space in vertical industry agents, industry solutions, and large model toolchains.

Anthropic's IPO is not an ordinary tech company listing—it's the first time the AI industry stands before capital markets as an adult. It tears off the disguise of light assets and high margins, laying out the compute arms race, prisoner's dilemma, and commercial games right on the table. Buying its stock is, in a sense, buying a lottery ticket betting on whether humanity can tame superintelligence.