Anthropic files S-1 to pursue a $2 trillion IPO, with $650 billion in compute commitments far exceeding its cash reserves

[[**Anthropic publicly files S-1, seeking to hit the largest IPO in history at a valuation of up to $2 trillion; the $80.7 billion net loss in the first half is mainly mark-to-market revaluation, and the real pressure is over $650 billion in compute commitments far exceeding cash reserves.**

**Key points**

**1. The $80.7 billion net loss is mark-to-market revaluation; operations are improving**

$72.388 billion comes from revaluation of rising valuations; second-quarter operating profit was $974 million, a margin of 8%, and compute costs fell from 201% of revenue to 55%.

**2. $650 billion in compute commitments far exceeds cash**

Commitments over the next decade exceed $650 billion, while cash and short-term investments are about $131.7 billion, a gap of nearly $520 billion; some contracts are non-cancelable or require making up the difference if usage falls short.

**3. The cloud vendors' circular network inflates revenue, making real demand hard to separate**

Amazon, Google, Microsoft and others are simultaneously investors, customers, distribution partners and competitors; nearly half of first-half revenue was sold with cloud vendors' help.

**4. The 900% retention rate stems from usage-based billing**

About 80% of revenue is usage-based; enterprises buy call counts rather than seats; heavy Claude Code users can spend up to 20 times as much as ordinary enterprise-edition users, but if usage stagnates, growth stops.

**5. The valuation bets that compute costs will keep falling as a share of revenue**

Compute costs are locked in 12 to 18 months ahead while revenue lags; the ratio fell from 2.5 times in the first quarter to below 0.5 in the second quarter; if it flattens, it becomes a meter for the boom. 2026-10-27 17:12 The largest IPO in history, and also the largest IOU in history: Anthropic's $2 trillion gamble Future Turing Quick Look

This article is from the WeChat public account: Future Turing, author: Zhang Fengjing

On the evening of October 26, a crowd of people waited at the SEC website for a document, and crashed the site.

Anthropic publicly filed its S-1, ticker "ANT", on the Nasdaq Select Market. "ANT" in Chinese means ant. Many people's first reaction was: was Ant's ticker taken? In June this year it only filed confidentially; this time it is the public version, 335 pages, downloadable by anyone. The valuation is expected to reach $1.5 trillion to $2 trillion, aiming for the largest IPO in history.

After opening it, most people were struck by two numbers: first-half revenue of $16.286 billion, versus $1.163 billion a year earlier; net loss for the same period of $80.748 billion.

But what really caught my attention was another detail: of these 335 pages, only 64 pages discuss the business, while 94 pages discuss risk. The prospectus states in black and white that advanced AI could cause "catastrophic harm," severe enough that "a single insurer, or even the entire insurance industry, may be unable to provide sufficient coverage."

A company tells you on one hand that "what I build could destroy the world," while on the other hand reaching out to ask you for $2 trillion. This kind of narrative tension is unique in the entire history of IPOs. Its old rival OpenAI has already made clear this September that it will not proceed with an IPO. Anthropic got ahead of it.

The $80.7 billion loss is accounting magic; what is truly valuable is that circular network

First, break down the loss.

Of the $80.7 billion, $72.388 billion is mark-to-market revaluation: convertible notes, derivative liabilities and SAFEs became more expensive as the company's valuation rose, and accounting rules require this increase in value to be recorded as a loss. Not a single cent of cash went out; it is purely a byproduct of the soaring valuation.

The actual business is improving instead. Second-quarter operating profit was $974 million, an operating margin of 8%, after fully including stock-based compensation. In the first half, compute and infrastructure cost $8.945 billion, falling from 201% of revenue a year earlier to 55%. In the first quarter of 2025, it spent 2.5 times its revenue on compute; by the second quarter of this year, that multiple had fallen below 0.5.

But the loss on the income statement is only accounting magic. What truly cannot be escaped is that IOU off the balance sheet: over the next decade, Anthropic's signed compute procurement commitments roughly exceed $650 billion: at least $114.5 billion from Google, at least $110 billion from Amazon, $44.9 billion from Microsoft, $161.2 billion for Broadcom equipment leasing, and up to $120.5 billion from SpaceX.AI. A considerable share of these are "non-cancelable" or "pay whether you use it or not." Compare the resources: as of the end of June, cash plus short-term investments were about $131.7 billion. $650 billion versus $131.7 billion leaves a hole of nearly $520 billion. This is the hardest account in the prospectus.

And the other side of this IOU is a circular network hidden in the books of four companies. Amazon invested in Anthropic, partly through convertible notes; Anthropic used the money to buy compute, and the supplier list includes Amazon itself, Google, Microsoft, Broadcom and SpaceX.AI. And these four are also its sales channels. Nearly half of first-half revenue was sold with the help of cloud vendors; they are also its shareholders; and they are also building their own large models. The prospectus itself describes this relationship with four words: investor, customer, distribution partner, competitor, and adds that these partners "are able to understand our pricing and commercial terms."

Money starts from Amazon's balance sheet, is recorded as Anthropic's cost, turns into Google's revenue, and then returns to the cloud vendors' income statements as channel sharing. The same sum of money leaves a mark in four places.

The disagreement lies here too. The bulls say every link in the chain is growing and demand is real; the bears say the internal back-and-forth within the chain gets double-counted, and no one can disentangle how much is genuine end demand. This document cannot answer that question, but it makes the question publicly discussable for the first time.

A 900% retention rate means the unit of billing has changed

The most easily misread number is the net revenue retention rate, about 900%.

Many people read it as "extremely high customer loyalty." The prospectus's definition is: for the same group of enterprise customers that were already paying a year ago, the money they spent this year is nine times last year's; individual subscriptions are not included. This number holds up not because customers became more loyal, but because the billing method changed.

In the first half, $13.1 billion, about 80% of revenue, came from usage-based billing, and only about 20% from subscriptions. Enterprises are not buying seats; they are buying calls. Buying software used to mean signing for headcount; buying intelligence now means signing for workload.

This explains a detail that seems odd at first glance: among enterprise users who do not use Claude Code, average monthly spending is under $75; among those who do use Claude Code, the top 5% of spenders pay 20 times that per month. The same product, two user groups 20 times apart: one is just asking questions in a chat box, the other has already stuffed the model into the pipeline for writing code, running tests and fixing bugs.

The old ceiling for SaaS was "how many employees in the enterprise are willing to pay to use it"; the ceiling for this kind of product is business volume. This is a real commercial change, and the only reason a high-multiple valuation can be uttered out loud.

The other side of the coin is equally clear: revenue follows call volume, and once customer usage stagnates, growth stops. The prospectus is restrained in its wording: many large customers have no long-term contracts, or have shorter contract terms, and spending levels may fluctuate from period to period.

Pricing sets the curve, not the company

There is only one number in the entire prospectus worth reading over and over: compute and infrastructure costs as a share of revenue.

That curve falling from 2.5 times to below 0.5 is the engine of the entire valuation narrative. The price Wall Street is giving is not for today's profit, but for this line to keep going down. The cheaper intelligence gets, the more widely usage spreads.

The engine has a physical constraint: compute must be ordered 12 to 18 months in advance, costs are locked at that moment, and revenue only slowly grows afterward. The gap in between is filled entirely by growth.

How hard are the commitments? Both the Google and Amazon contracts state that shortfalls must be made up, and the Microsoft one says you pay regardless of how much you use. But not everything is welded shut: the SpaceX.AI contract leaves an exit, allowing termination with 90 days' notice after the initial lock-up period ends. This is a rare live opening in the whole web of contracts.

So "betting ten years" needs a discount. The real question is not how long it signed for, but how many contracts can still be let go on the day growth decelerates.

The curve is about how money is spent; one other thing is about how power is divided.

The seven founders kept a card in reserve. At the helm is CEO Dario Amodei, who left OpenAI with six colleagues in 2021 and five years later brought the company to the doorstep of a $2 trillion valuation.

The founders use a class of non-economic Class F shares to lock in 50.1% of voting power; a long-term interest trust uses a Class T share to hold the right to elect a majority of board seats. You can buy the right to dividends, but not the steering wheel. This structure would have been cursed a decade ago; now it is written up as "preventing short-term capital from kidnapping long-term goals."

The bet is placed on the table first

The weight of this IPO lies not in how much it raises, nor in whether the $2 trillion holds. For the first time, it puts a price on "the marginal cost of intelligence." If that curve falling from 2.5 times to below 0.5 can keep going down, Anthropic is the next generation of infrastructure; if it flattens halfway, these 335 pages become the most precise meter of this boom.

As said at the start, of these 335 pages only 64 discuss the business, while 94 discuss risk. After reading, you realize what is truly worth seeing is not those 94 pages of warnings, but the few contract tables in the middle. Writing a lot about risk is an attitude; signing contracts hard is the business.

As for who pays for the bet, the last line of the announcement is still blank: the number of shares offered and the price range await updates in subsequent filings.

A prospectus never answers the future. It merely places the bet on the table first. Channel: Finance and Business. This content comes from the internet, original link; the views represent only the author and do not represent Huxiu's position. For copyright issues, please contact hezuo@huxiu.com, and we will verify and handle it promptly. Everyone who is changing the world and wants to change the world is in the Huxiu APP [ ]