Anthropic Files Prospectus: First-Half Revenue of $16.286 Billion, Seven Founders Hold 50.1% Voting Power

[[**Anthropic files IPO prospectus, planning to list on Nasdaq with high growth, heavy compute investment and absolute founder control; behind the revenue surge, a hundred-billion-dollar compute bill and channel dependence will test its AI business loop.**

**Key points**

**1. Revenue and per-capita output soar**

Revenue in the first half of 2026 was $16.286 billion, up about 1,300% year on year; its roughly 5,300 employees generated $6.6 million in revenue per capita, 90% of code is written by AI, and more than 80% of Q3 revenue came from models launched within the past six months.

**2. Compute and channel costs run high**

Compute and infrastructure spending for the half-year was $8.945 billion, 55% of revenue; minimum purchase commitments to cloud giants exceed $260 billion, and the Broadcom lease is $161.2 billion; CSP channel revenue accounts for 48%, and sales expenses rose 2,540% year on year.

**3. Losses include large non-cash items, operations already turned positive**

Of the $80.748 billion net loss, $72.388 billion was non-cash losses, while the operating loss narrowed 67% to $693 million; Q2 operating profit was $974 million, and cash and equivalents stood at $112.4 billion.

**4. Absolute founder control locked in**

The seven founders hold 50.1% of voting power through 1 share of Class F stock, which is not diluted by new issuance, and the LTBT holds Class T stock that can elect a majority of the board; Class A public shareholders will find it hard to change strategy, while the PBC and RSP provide protection for safety decisions. Original 2026-10-27 07:49The secrets inside Anthropic's prospectus: the compute devouring and control defenses behind $2 trillionLiang Ka'erQuick Take

When Dario Amodei and his six co-founders decided to leave OpenAI in 2021, they may not have anticipated that just five years later, a company called "Anthropic" would file a prospectus that could set a Wall Street record.

When they left, the commercialization race for large models was just beginning, and this group's anxiety about technical safety and the pace of commercialization became the direct reason they started over. Yet five years later, the company they founded has a valuation exceeding OpenAI's, at one point discussed at $2 trillion.

At 21:00 Beijing time on the evening of October 26, the U.S. Securities and Exchange Commission website disclosed this hundreds-of-pages-long S-1 filing. The prospectus opens with a letter from the founders, and the seven co-founders state their position directly: "We believe artificial intelligence is one of the most profoundly impactful technologies in human history."

After that, however, the document presents a sense of fragmentation. On one hand, it shows that the company's revenue in the first half of 2026 soared to $16.286 billion, up about 1,300% year on year, shedding the predicament of three years ago when it relied on a single major customer; on the other hand, it tells the outside world that it carries a net loss of up to $80.748 billion and infrastructure spending commitments in the hundreds of billions of dollars.

More noteworthy is the unique voting-rights mechanism design, which gives the company a governance architecture that strictly isolates public-market capital from the core of corporate power. As a result, this prospectus is more like an "declaration of independence" drafted by seven technological idealists, carrying an enormous bill.

Anthropic plans to list on the Nasdaq exchange under the ticker "ANT". Next, it will pitch to investors the shares it plans to sell. More than 20 institutions are involved, with Goldman Sachs and Morgan Stanley serving as joint lead underwriters, and people familiar with the matter expect a listing in mid-to-late November. Media cited informed sources saying the fundraising size is about $100 billion.

The prospectus shows that the main purpose of this offering is to increase our capital strength and financial flexibility, facilitate the orderly distribution of shares by selling shareholders, and open access to the public equity markets. Anthropic said the net proceeds will include investment in compute and infrastructure, as well as the recruitment and retention of highly skilled talent, and may also use a portion for acquisitions or investments.

When compute devours manpower

Traditional enterprise software companies expand by piling up sales and R&D headcount, but Anthropic has demonstrated the terrifying leverage of the AI era.

The prospectus shows that as of September 30, 2026, the company had about 5,300 employees. Over the past 12 months (TTM), this team generated $6.6 million in revenue per employee, up about 371.43% year on year, nearly 3 times Apple's figure for the same period and 6 times Microsoft's.

Anthropic had only 178 employees in 2023 and only about 940 at the end of 2024. Considering this, the per-capita figures above are a fairly conservative calculation. As for the reason for the improvement, Anthropic disclosed that 90% of its internal code is now written by AI. Compared with 2024, in the third quarter of 2026, the average amount of code written by engineers per day has increased more than 45-fold.

This large-model beast is beginning to show the compounding effect of using AI to improve AI. Growth is highly sensitive to the pace of model updates. The prospectus shows that new model versions released in the six months ended September 2026 contributed more than 80% of revenue in the third quarter of that year. Opus 5.5 reduced the cost per task by about 40% compared with Opus 5, and lower task costs did not depress total revenue but instead unlocked more usage.

Anthropic is not a typical subscription-model SaaS (software as a service) company. In the first half of 2026, its usage-based consumption revenue reached $13.1 billion, accounting for 80% of total revenue, while subscription revenue was $3.2 billion, accounting for 20%. Customers mainly pay based on models and token usage; although some enterprise contracts carry annual or multi-year minimum spending commitments, what is ultimately recognized is still mainly usage revenue.

However, this does not come without a cost: supporting the high revenue is heavy capital expenditure.

In the first half of 2026 alone, Anthropic's compute and infrastructure spending reached as high as $8.945 billion, exceeding the full-year 2025 figure of $7.326 billion and accounting for 55% of half-year revenue. In addition, the prospectus also discloses enormous long-term commitments: minimum spending commitments to the three cloud giants Amazon, Google, and Microsoft exceeding $260 billion, as well as a five-year finance lease agreement with Broadcom worth as much as $161.2 billion.

At the same time, because its large models need to be distributed through cloud channels and it recognizes revenue on a gross basis, the channel fees Anthropic pays to cloud providers are recorded as "sales and marketing expenses." This caused that expense to soar to $3.168 billion in the first half of 2026, an increase of 2,540% year over year. This shows that although the revenue base is expanding, a considerable portion of the profit margin is offset by channel costs.

Channel dependence is more alarming than expense growth. The share of revenue generated through CSP cloud marketplaces has risen from 11% in 2023 to 48% in the first half of 2026; as of the end of June 2026, related net accounts receivable accounted for 58% of total accounts receivable. The three major cloud providers not only provide Anthropic with compute power, but also control the distribution, invoicing, collection, and cash recovery of nearly half of its revenue.

However, one piece of good news is that the net loss of $80.748 billion was largely affected by changes in the fair value of convertible notes, derivative liabilities, and SAFEs, with related non-cash losses alone reaching $72.388 billion. By comparison, the operating loss, which better reflects the state of the main business, was $693 million, narrowing by about 67% year over year.

An even more important signal is that in the second quarter, its operating profit turned positive for the first time in history, reaching $974 million. This stands in sharp contrast to OpenAI, whose losses are widening.

Image source: prospectus

Before the IPO, Anthropic also had an extremely thick cash buffer. As of June 30, its cash and cash equivalents were $112.4 billion, accounting for 80% of the $141.1 billion it had raised over the past six years. In addition, Anthropic also held $19.3 billion in short-term securities investments, covering money market funds and various other securities.

A control defense line amid the capital frenzy

Seeing such financial data, if you think you can gain a say through investment, you would be badly mistaken. This company has built one of the most extreme control defense lines in Silicon Valley history, excluding ordinary shareholders from most of the company's decision-making.

The prospectus shows that the five different classes of common stock issued by Anthropic (Class A, N, T, F, and E) are basically equal in economic rights, but are deeply divided in voting mechanisms and transfer restrictions. Through this extremely complex equity matrix, combined with the Long-Term Benefit Trust (LTBT), the company ensures the absolute control of seven founders over the company.

The seven founders established a private company called Founder LLC. Through this company, they issued only one share of Class F stock, yet it carries 50.1% of the absolute voting power and is not diluted by subsequent issuances. Economically, they have achieved an extremely rare absolute equality, with each holding about 240 million shares, that is, equal voting rights.

At the same time, the board also includes an LTBT composed of five experts who hold no economic interest in the company. By holding Class T stock, the trust has the legal right to elect and remove a majority of board seats, constituting the ultimate safety lock. This means that even if it faces commercial pressure from external capital in the future, the trust can directly remove decision-makers deemed to be catering to capital by re-electing directors.

In addition, Anthropic's listing entity is a PBC (public benefit corporation). Under Delaware law, this model gives management the legal right to balance shareholder interests with the public interest. If management refuses to release a product because it believes it threatens human safety, shareholders could theoretically sue, but the safe harbor provisions would protect that decision, making litigation very difficult to succeed.

As for the Class N stock, it serves as a public benefit and matching donation and has no voting rights. The Class E stock granted to employees only has a "one-vote check" on certain specific restricted matters and no right to participate in ordinary governance. Class A stock is the core of this IPO issuance, held by public investors in the secondary market and early external financial investors. Although it is formally "one share, one vote," actual power has long been hollowed out by Class F and Class T.

To clear the obstacles for the IPO, Anthropic also initiated a share conversion operation. The prospectus discloses that although Anthropic has historically issued a very large number of preferred stock financing rounds, extending from Series A all the way to Series H, with different sub-series under each round, all of its preferred stock was divided into two categories for automatic conversion: those with a "-1" suffix and Series A were converted into Class A, while the others were converted into Class N.

Even the recruited CFO Krishna Rao holds only 9.49 million Class A shares. Among the seven founders, apart from company chief architect Sam McCandlish, who has slightly fewer at 231 million Class A shares, the others are all highly concentrated at around 240 million shares. Krishna Rao does not even hold a fraction of that, which thoroughly draws the line between "building the empire" and "the IPO operator."

"We never set out to start a company"

History always rhymes.

In 1957, Robert Noyce, Gordon Moore and six other young scientists, dissatisfied with their boss's dictatorship, left en masse to found Fairchild Semiconductor, which subsequently spawned semiconductor giants such as Intel and AMD and nurtured what later became Silicon Valley.

Nearly 70 years later, Anthropic's seven co-founders, in addition to the two mentioned above, also include Dario Amodei's sister, Anthropic President Daniela Amodei, Chief Compute Officer Tom Brown, Jack Clark, who leads the research institute, Chief Science Officer Jared Kaplan, and Christopher Olah, who leads interpretability research.

Image source: Screenshot from a video released by Anthropic

They defected from OpenAI en masse for much the same reason: dissatisfaction with the frantic commercialization drive at the time and extreme anxiety about safety. In an interview, they admitted frankly, "Not one of us initially wanted to start a company." What they initially wanted to establish was a nonprofit safety organization. Founding Anthropic was purely out of a sense of responsibility in the face of an AGI explosion and the reality of having no choice.

For Silicon Valley's frontier AI circles, this was tantamount to a powerful earthquake. From GPT-2 to GPT-3, these founders and first creators of Scaling Laws may have acted out of idealistic faith, but the brutal commercial and technological reality soon confronted them: frontier AI R&D requires computing power and engineering investment in the tens of billions to hundreds of billions of dollars.

To avoid becoming just another commercial machine driven by capital, they concretized their safety commitment into Anthropic's core technical and governance mechanism, the RSP (Responsible Scaling Policy), and regarded it as the company's "constitution." The prospectus also explains the source of the RSP's "legal binding force."

Simply put, if a model is detected in red-team testing as having breached a specific risk threshold, no matter how great the commercial market pressure or how severe the PR damage, once the model crosses a specific risk threshold in evaluations, the RSP requires the company to pause or restrict deployment and further scaling until the corresponding safety measures meet the requirements.

This underlying logic explains the core contradiction in the prospectus: Anthropic wants to be both a high-speed commercial machine and a safety barrier against AI expansion. "The majority of our revenue comes from sales to enterprise customers, and our product and investment decisions are guided to a significant extent by our commitment to AI safety" appears as a core statement when introducing the business model and positioning.

From this perspective, only a founder team with absolute voting control could achieve this.

Anthropic's prospectus is more like an experiment in the "form of the future super commercial entity." It clearly conveys a signal to secondary-market investors: you can buy the company's economic dividends and growth upside, but you cannot change the company's management or major strategic direction through voting, still less buy the power to decide humanity's future.

But from the perspective of a listed company, the test is already on the schedule. Starting in 2027, hundreds of billions of dollars in compute rent and performance obligations will arrive in quick succession, and the market will scrutinize this "seven-person co-governance" company with a critical eye: can per-capita revenue of $6.6 million and quarterly operating profit of $974 million cover hundreds of billions of dollars in compute procurement, leasing and performance commitments and keep the flywheel running smoothly?

If it can make this work, Anthropic will provide a brand-new business model for high-capital-intensity tech companies. But if it fails, those compute bills in the hundreds of billions may become the blade that bursts the AI bubble.

Article title: The secrets in Anthropic's prospectus: the compute devouring and the control defense line behind $2 trillion

Article link: https://www.huxiu.com/article/4900650.html Read the original: The secrets in Anthropic's prospectus: the compute devouring and the control defense line behind $2 trillion_Huxiu

Published 78 articles on Huxiu. This content may not be reproduced without permission. For licensing matters, please contact hezuo@huxiu.com. Everyone who is changing the world and wants to change it is on the Huxiu APP[ ]