Market

Anthropic's IPO: the economic strategy behind the S-1

Revenue, a mostly accounting loss, hundreds of billions committed to compute, founders in control: Anthropic's strategy as read in its S-1 filing.

By ·

On October 26, Anthropic made public its Form S-1, the document a company files with the SEC, the US securities regulator, before going public. It plans to list on the Nasdaq under the ticker "ANT" (Anthropic's S-1). The share price and the amount to be raised are still blank (TechCrunch). So is third-quarter revenue: the filing leaves that amount empty, as it does the total of its compute commitments.

What follows separates three things: facts from the S-1, estimates published by the press and analysts (each attributed), and our own reading, flagged as such. Unless stated otherwise, figures are in US dollars and come from the S-1.

Growth driven by enterprises and consumption

Revenue went from $41 million in 2023 to $386 million in 2024, then $4.59 billion in 2025. In the first half of 2026 it reached $16.29 billion, up from $1.16 billion a year earlier, a 1,300% increase according to the S-1. These are reported revenues, not an annualised run-rate.

The model rests first on what Anthropic calls "pay-for-intelligence": customers pay for the tokens they use, at the price of the model they pick. That accounted for $13.1 billion, or 80% of revenue in the half. Subscriptions (Claude Pro, Max, enterprise plans) brought in $3.2 billion, or 20%. Anthropic says it expects consumption to remain "the substantial majority" of its revenue.

In $B
█ revenue
░ compute and infrastructure costs

Q1 2025 revenue  0.38 │█
        compute  0.94 │░
Q2 2025 revenue  0.79 │█
        compute  1.39 │░░
Q3 2025 revenue  1.38 │██
        compute  2.05 │░░░
Q4 2025 revenue  2.04 │███
        compute  2.95 │░░░░░
Q1 2026 revenue  4.73 │███████
        compute  3.39 │░░░░░
Q2 2026 revenue 11.56 │██████████████████
        compute  5.56 │░░░░░░░░░

Chart 1. Anthropic's revenue and compute costs by quarter, in billions of dollars. Source: Anthropic's S-1, unaudited quarterly results.

The chart shows the tipping point. Until the end of 2025, every quarter cost more in compute than it brought in: 251% of revenue in the first quarter of 2025. In the second quarter of 2026, that ratio fell to 48%. Year over year, revenue for that quarter grew 1,368% while compute costs grew 300%. Anthropic explains that it buys capacity 12 to 18 months ahead, so these costs are set in advance and do not follow demand in the moment.

The customer base follows the same curve. The S-1 counts about 20 customers with more than $100,000 of trailing twelve-month revenue at the end of 2023, 200 at the end of 2024, 1,500 at the end of 2025, and more than 11,900 on September 30, 2026. On that date, more than 1,900 customers were above $1 million and about 250 above $10 million. Net revenue retention is about 900%: customers who were there a year ago now spend, together, about nine times as much as they did then (Pro and Max subscribers excluded from the calculation).

Customers above $100,000

Customers above $100,000
DateCustomers above $100,000
Dec 31, 2023about 20
Dec 31, 2024about 200
Dec 31, 2025about 1,500
Sep 30, 2026more than 11,900

These figures extend the trend measured by Menlo Ventures, which at the end of 2025 estimated Anthropic's share of enterprise spending on model APIs at 40%, against 27% for OpenAI (Menlo Ventures, 2025). Menlo is an Anthropic investor, which it discloses itself.

The flip side is concentration. The top five customers accounted for about 25% of revenue in the first nine months of 2026, although none was above 10%. In the third quarter, 79% of revenue came from customers outside that top five, up from 64% in the first quarter. Another sign of dependence: model versions released in the previous six months made up more than 80% of third-quarter revenue.

An $80.7 billion loss, mostly on paper

The net loss for the half reached $80.75 billion. The number is striking, but it mostly reflects the rise in Anthropic's valuation.

H1 2026, $B
(negative = expense)

Operating loss    −0.69 │▒
Interest income   +1.11 │░
Convertibles FV  −72.39 │█████████████
Other expense     −8.76 │██
Net loss         −80.75 │▓▓▓▓▓▓▓▓▓▓▓▓▓▓

Chart 2. From operating result to net loss in the first half of 2026, in billions of dollars "Convertibles FV": change in fair value of convertible notes, derivatives and SAFEs; $21 million of income taxes not shown. Source: Anthropic's S-1.

Most of it, $72.39 billion, comes from remeasuring convertible notes, derivatives and SAFEs: when the company's valuation rises, these instruments, which will convert into shares, are worth more, and the difference is booked as a non-cash expense. On top of that come $8.76 billion of "other expense", tied according to the S-1 to the value of its note purchase agreements. The S-1 says these changes will continue until the instruments convert.

On the operating side, the loss was only $693 million for the half, against $2.1 billion a year earlier. The second quarter of 2026 is the first profitable one in the published series, with $974 million of operating income. On an adjusted basis (excluding stock-based compensation in particular), Anthropic reports $1.62 billion of operating income, a 10% margin. Operating cash flow was a positive $3.93 billion. The S-1 still warns that the company "may not achieve or sustain profitability".

The real bet: compute

This is where the strategy plays out. The S-1 describes a single "pool" of compute spread across AWS Trainium, Google TPUs, NVIDIA GPUs and AMD GPUs, which Anthropic continuously assigns to research, customers or internal use. It also announces a shift: on top of cloud contracts, Anthropic now leases chips directly, hosted in dedicated facilities, an approach it calls "Anthropic Direct".

Amounts in the S-1, $B

Broadcom (rent)  161.2 │████████████████
SpaceXAI (up to) 120.5 │████████████
Google (min.)    114.5 │███████████
Amazon (min.)    110.0 │███████████
Microsoft (min.)  44.9 │████
Nscale (up to)    44.6 │████
Lambda            41.3 │████
AMD (est.)        20.0 │██

Chart 3. Main compute and infrastructure commitments described in the S-1, in billions of dollars. Guaranteed minimums (Google, Amazon, Microsoft), caps (SpaceXAI, Nscale), an approximate amount (Lambda) or an estimate (AMD, more than $20B). Source: Anthropic's S-1.

The detail is dizzying. The S-1 lists at least $114.5 billion with Google through 2033, plus an estimated $125 billion to $150 billion for more TPUs delivered between 2028 and 2031. It also lists at least $110 billion with Amazon through 2036, at least $44.9 billion with Microsoft, up to $120.5 billion with SpaceXAI, formerly xAI, and about $161.2 billion of rent on Broadcom equipment leased for five years. Adding up these amounts, which are of different kinds, gives more than $650 billion: that is our calculation, an order of magnitude and not a published total. On the day of the filing, Bank of America, Blackstone, Citi and Morgan Stanley announced $63 billion of financing to buy Broadcom accelerators to be leased to Anthropic, with deliveries starting in 2027 (Broadcom press release).

Most of these contracts must be paid "regardless" of revenue or utilisation, the S-1 notes. Several suppliers, such as Amazon, Google and Microsoft, are also Anthropic's investors, customers and competitors. On June 30, Anthropic had $131.7 billion in cash and short-term investments, after raising $141.1 billion in total. In the first half, the Series H alone brought in $65.2 billion.

A governance that keeps control

Anthropic remains a public benefit corporation: its board must balance shareholders' interests, those of people affected by its business, and its mission, "the responsible development and maintenance of advanced AI". The S-1 stacks several share classes:

The listing will therefore be that of a "controlled company" under Nasdaq rules. Anthropic has never paid a dividend and does not plan to. The money raised will go to "general corporate purposes": working capital, compute and infrastructure, hiring, possible acquisitions. Anthropic has also given 5% of its fully diluted shares to a public benefit fund. It plans up to 2% more each time the share price rises 50%, and the co-founders have pledged to give away 80% of their shares.

What markets and analysts say

What follows comes from outside estimates, not from the S-1. According to Reuters, Anthropic is seeking to raise up to $100 billion, and some investors value it at about $1.5 trillion, down from the $2 trillion previously expected. Its last funding round, in May, valued it at $965 billion (Reuters via Investing.com, NPR). Also according to Reuters, the listing is expected in mid-November. It would be the largest IPO in history, ahead of SpaceX's $75 billion in June.

Our reading

The strategy that comes out of the S-1 is a loop: sell intelligence by the token to enterprises, reinvest almost everything in compute, and rely on more efficient models to lower the cost per task. Anthropic says, for example, that Opus 5.5 cut the cost per task by 40% compared with Opus 5 without hurting its margins. The IPO is not there to cover operating losses, which have almost disappeared, but to fund hundreds of billions of compute commitments without relying on its cloud partners alone.

Three things are worth watching. The first is third-quarter revenue, left blank: the S-1 itself says growth will slow. The second is the pace of compute payments, which fall due whether customers follow or not. The third is the price, which will show how much investors are willing to pay for a company where the founders and a trust keep control.

Sources

Spotted an error in this article? Report it · Our corrections policy

Rekha icon: a house drawn in a single stroke

From the same author Self-promotion

An Eulerian path, drawn with your finger

Rekha, the iPhone game by OrigoTech's author: connect every dot in a single stroke without tracing the same link twice, across 500 drawings that are all solvable.

The OrigoTech newsletter

A sourced weekly digest of AI tool news.

Read next