Anthropic publicly releases 335-page prospectus, applies for Nasdaq IPO

Just now, Anthropic publicly released a 335-page prospectus: the largest IPO in history, and the most loss-making AI company in history.

Meng Chen 2026-10-27 00:06:31

English text is being prepared.

The richest brother and sister on Earth have arrived in grand style! The 7 co-founders still firmly hold 50.1% of total voting power.

English text is being prepared.

English text is being prepared.

The title of the largest IPO in history is about to be taken from Musk's SpaceX!

Just now, Claude's parent company Anthropic officially publicly submitted its Nasdaq IPO application, with the ticker symbol "ANT."

(At first glance, you'd think some insect company had gone public, lol)

Its 335-page official prospectus was also freshly released on the SEC website just now.

Compared with the previously leaked draft, the public version further updated its computing power procurement arrangements:

Anthropic's procurement commitment to Microsoft increased from $31.4 billion to $44.9 billion, and its procurement commitment to Google is $114.5 billion; the maximum amount of its computing power contract with SpaceXAI is $120.5 billion.

In addition, the public version newly disclosed two computing power suppliers, Nscale and Lambda.

This matches earlier Bloomberg reports citing people familiar with the matter that Anthropic plans to complete its IPO in mid-November.

In the capital race with its archrival OpenAI, Anthropic has crossed the finish line first, making a forced sprint at a time when market sentiment is not friendly, and is expected to become the first purebred frontier large-model listed company in U.S. stock market history.

Although the size of the fundraising has not yet been finally priced, the previously exposed draft prospectus showed that its target is up to about $100 billion, corresponding to a valuation of about $2 trillion.

If that happens, it will surpass the $1.77 trillion valuation corresponding to Musk's SpaceX offering in June this year, and is expected to sprint toward the largest IPO in history.

But, because the AI sector has recently come under pressure, some investors have already cut their valuation expectations from $2 trillion to about $1.5 trillion (shrug.jpg)...

But, who cares!

The previously leaked version had already laid out this crazy Silicon Valley wealth-creation myth and computing power gamble in the sunlight without reservation:

Revenue soaring like a money printer: it raked in $16.286 billion in the first half of this year, took in another $16.8 billion in Q3 alone, and its annualized revenue in July soared past $65 billion;

Book losses hitting a spectacle: net loss in the first half of the year reached as high as $80.7 billion, nearly twice last year's full-year loss;

The harshest computing power indenture in history: about $518 billion in hard computing power commitments signed for the next decade, with reports saying 80% is "pay whether you use it or not";

Writing sci-fi into the prospectus: the business introduction is 64 pages, yet it used a full 94 pages to warn that "AI may destroy humanity, resist being shut down, or even extort humans."

...

The company CEO Dario Amodei's line that "growth is too fast to handle" keeps rising in value.

Previously, the company confidentially submitted its draft S-1 on June 1, 2026, and this public submission still uses the S-1 form format.

The SEC requires S-1 to be publicly submitted at least 15 days before the roadshow. That means Anthropic will begin its roadshow as early as the second week of November.

Before everything begins, you should be as curious as we are—

How exactly does Anthropic make money? Of the $80.7 billion net loss, how much is accounting and how much is cash? Behind the listing, who will become the biggest beneficiary?

These questions all have answers in the prospectus, and we will look at them one by one.

What do the financial conditions disclosed for the first time look like?

If you look only at the numbers at the bottom of the income statement, Wall Street might gasp:

For the full year 2025, Anthropic's net loss was about $41.965 billion.

By the first half of 2026, the net loss had directly soared to an astonishing $80.748 billion.

But those in the know understand that this is actually a "happy trouble" brought by the surge in valuation, and it needs to be broken down.

Of the massive $80.7 billion net loss, about $72.4 billion comes from changes in the fair value of convertible notes, derivative liabilities, and other financing instruments.

Simply put, as Anthropic's valuation rose sharply, the accounting measurement value of some financing instruments the company had previously issued also changed accordingly.

Under relevant accounting rules, these changes need to be included in profit and loss, thereby greatly inflating book losses.

Such fair value losses usually do not mean the company paid out the same scale of cash in the current period.

In other words, the $80.7 billion net loss cannot simply be understood as Anthropic burning $80.7 billion in cash in half a year.

However, this does not mean the company's operating costs can be ignored. After all, the accounting factors in book losses need to be analyzed separately, and computing power, R&D, and personnel expenses are still real money inputs.

Peeling away this layer of accounting fog, Anthropic's true business cash-generating ability is terrifying.

The consolidated operating profit summary table in the prospectus clearly shows Company A's revenue changes.

For the full year 2023, Anthropic's revenue was only $41 million;

In 2024 it reached $386 million;

In 2025 it jumped to $4.585 billion;

In the first half of 2026, in just six months, revenue stormed to $16.286 billion (only $1.16 billion in the same period last year).

Overall, the revenue side shows explosive growth, and the pace of commercialization has accelerated significantly.

Although the prospectus shows the company is not yet profitable, at the investor day just held in mid-October, Anthropic also laid its cards on the table:

Third-quarter revenue alone reached $16.8 billion, up 45% quarter over quarter, and on an adjusted basis the company officially achieved profitability.

From a "chaser" ground underfoot by ChatGPT to now pulling off a comeback in both valuation and growth rate, what exactly did Anthropic rely on to win its turnaround battle?

The answer is just two words:

Code.

Since launching its flagship product Claude Code in 2025, coding and enterprise customers have become important growth drivers.

Data from venture capital firm Menlo Ventures shows that in the $1.9 billion code generation market, Anthropic alone holds a 42% share, directly leaving OpenAI (21%) a full body length behind.

The prospectus divides Anthropic's revenue into two categories by charging model.

One is consumption-based revenue, and the other is subscription revenue.

The prospectus states that consumption-based billing (including API calls, etc.) contributed about $13.06 billion in the first half of 2026, accounting for about 80.2% of total revenue in the same period.

This means Anthropic has already built a fairly important business model. The more work customers complete using Claude, the more opportunities model calls and related services have to generate more revenue.

As more customers onboard and continuously increase their use of Claude, the base of high-value paying customers is expanding rapidly.

As of September 30, 2026, there were more than 11,900 customers with TTM revenue of ≥$100,000; on September 30, 2025, that number was only 950.

This proves Claude's extremely strong penetration in the B2B enterprise market, as well as its super-strong customer stickiness.

However, there is a small mix-up!

Previously, a leaked version said that in its customer list, just the top two mystery customers each contributed about 12% of revenue (about 25% combined). It was previously rumored online that these were Cursor and Microsoft's GitHub Copilot.

However, the official version shows that these two customers are Amazon and Google.

At the financial statement level, the prospectus clearly states that the company currently has only a single operating segment, with CEO Dario and President Daniela, the brother-and-sister pair, as the chief operating decision maker (CODM) overseeing the whole picture.

In order to go all in on "high-value enterprise cognitive work," it also laid out a "not-to-do list" in the prospectus:

Although highly commercially attractive, we currently choose not to develop image generation and video generation models, but instead concentrate all compute on coding, Agent capabilities, and safety research.

The prospectus systematically breaks down the company's business landscape for the first time:

A tightly interlocking "three-layer full-stack architecture" (bottom-layer models + middle platform + top-layer applications), paired with an "80/20 charging engine."

First layer: Four major model families

Anthropic's current foundation is a stepped matrix composed of four major model families.

1. Mythos and Fable (top-tier flagship level): These are the "nuclear weapon"-level models newly unveiled this year, specifically targeting extreme coding and ultra-long-cycle Agent tasks that AI previously could not handle at all. Because its capabilities in mining software vulnerabilities and biological sciences are too extraordinary, Anthropic adopted a rare "one body, twin forms" strategy.

2. The full-power version of Mythos is only made available on a targeted basis to rigorously vetted cybersecurity and life sciences institutions through trusted programs (such as Project Glasswing and the cybersecurity verification program);

3. Fable, the general commercial version (now iterated to Fable 5.1), is equipped with strict production-grade safety guardrails. Once a sensitive high-risk request is detected, it will automatically block it or downgrade the routing to a less capable model for handling.

4. Opus (all-around mainstay level): the mainstay for enterprises' daily heavy-duty inference. The latest Claude Opus 5.5, while cutting the cost per task by 40%, has become the core engine driving end-to-end project Agents.

5. Sonnet (balanced cost-performance level): aimed at the broadest range of everyday general tasks, and also the mainstay serving as high-concurrency "sub-agents" in multi-agent architectures.

6. Haiku (ultra-fast lightweight level): lowest cost and fastest response, specifically targeting intelligent customer service and massive document processing.

Even more formidable is the revenue-generating effect of model iteration.

The prospectus reveals that in the third quarter of this year, more than 80% of the company's revenue was contributed by new model versions released within the past six months.

Second layer: Claude Platform (Agent development platform)

In order to let enterprises truly run AI in production environments, Anthropic built the Claude Platform, opening up the full set of underlying building blocks (Primitives) used in its own products.

1. Thinking and cost-control primitives: including Extended Thinking, Adaptive Thinking, and the killer feature Prompt Caching—which directly helps developers cut the cost of long prompts by up to 97% and reduce latency by 60%.

2. Protocols for connecting everything and the physical world: in addition to MCP (Model Context Protocol), which has already been donated to the Linux Foundation and become an industry-standard, and the prebuilt Agent Skills, the most eye-catching is the Model Hardware Standard, currently in research preview.

3. Two major Harnesses: to solve the pain points of Agents crashing during long runs and permissions spiraling out of control, the platform offers two solutions—the Agent SDK, which runs on customers' own compute, and Claude Managed Agents, an out-of-the-box service fully hosted by Anthropic with sandboxes, credentials, and memory loops.

Third layer: 7 major in-house applications + 7 major industry suites (Applications & Verticals)

On top of the models and platform, Anthropic itself has entered the fray to build a legion of "Virtual Collaborators," directly serving up 7 first-party native applications:

In addition to the developer-familiar Claude.ai (web/mobile entry point), Claude Code (the coding powerhouse), and Claude Cowork (a desktop-level fully automated office Agent, which began merging with Claude.ai into a unified entry point this September), there are 4 more first-party flagship applications:

2. Claude Tag (group chat teammate): resides permanently in channels of enterprise collaboration software such as Slack, understands context and proactively chimes in, directly delivering document drafts, data analysis, and even code PRs in group chats;

3. Claude Design (visual design): moving into the heartland of designers and marketers, directly generating high-fidelity design mockups, product prototypes, and presentations through natural language conversation;

4. Claude Science (research workbench): one-click connection to more than 60 scientific databases, helping scientists run the entire workflow from literature analysis and data processing to code reproduction, directly outputting journal-publication-quality paper figures and manuscripts,

In addition to general-purpose tools, for traditional industries with extremely high compliance requirements, Anthropic has also directly packaged 7 major vertical industry solutions:

Legal edition (integrated with Relativity and Ironclad)

Government edition (FedRAMP High, the highest security certification)

Education edition (completely free for all K-12 certified teachers in the United States)

Even to help traditional mid-sized enterprises implement AI, Anthropic recently also teamed up with top private equity (PE) giants to directly form a joint-venture AI services company—Ode with Anthropic (in which Anthropic has committed to invest up to $325 million), and spent $200 million to launch the "Claude Partner Network" and the "Frontier Academy," with the goal of training 10,000 top AI deployment engineers for major enterprises and consulting giants by the end of next year.

△Original image of the s-1 filing

However, arch-rival OpenAI is not sitting idle either.

Because 47% of Anthropic's sales (about $2.16 billion in 2025) are sold through Amazon AWS and Google Cloud marketplaces, Anthropic directly recognizes revenue in its financial statements on a gross transaction basis, and then records the 16% toll (about $351 million) taken by the cloud vendors as a sales expense.

OpenAI had previously privately mocked this to investors, saying that this accounting method artificially "inflated" Anthropic's revenue by billions of dollars.

Betting over $650 billion on compute, making deals with tigers in the crevices between giants

Money comes in like flowing water, and goes out like a mountain collapsing.

A previously leaked draft prospectus disclosed that Anthropic's cloud computing, compute, and infrastructure commitments over the next decade amount to about $518 billion.

And the public version further lists more contract details, with some amounts also adjusted. Roughly summing the major contract amounts disclosed publicly, the related arrangements already exceed $650 billion.

What does that mean?

This figure has already surpassed the $500 billion "Stargate" plan jointly launched by OpenAI, SoftBank, and Oracle, corresponding to a long-term committed dedicated compute scale of a full 3.5 gigawatts (GW).

If additional compute still under negotiation is included, the potential total scale could increase further.

To prevent compute from becoming a bottleneck, Anthropic has staked its life and fortune over the next decade on eight giants:

Broadcom: about $161.2 billion in chip and equipment leasing agreements;

Google: at least $114.5 billion in cloud and compute services (2026 to 2033);

Amazon: at least $110 billion in cloud service commitments (2026 to 2036);

Musk's SpacexAI: up to $120.5 billion for leasing Nvidia compute clusters (through 2029);

Microsoft: at least $44.9 billion in cloud infrastructure contracts (2026 to 2033);

AMD: expected to procure more than $20 billion in compute, while “Su Ma” also promised to spend $5 billion to subscribe for Anthropic shares on the flip side.

These contracts cover different suppliers, equipment types, and service terms, with some arrangements extending into the 2030s.

For Anthropic, model capabilities, customer demand, and compute supply are already bound together.

The more powerful the model, the more complex the tasks it can handle; the more enterprise customers there are, the more computing resources are needed for daily inference and Agent operation. At the same time, training the next generation of models itself requires enormous computing resources.

If compute supply cannot keep up, even if customers are willing to pay, the company may be unable to deliver sufficient service capacity in time.

This is also why Anthropic is willing to lock in compute for multiple years in advance.

What is striking is that as much as 80% of it consists of “non-cancelable” take-or-pay clauses.

No matter how much compute Anthropic actually uses, the money must be paid in full regardless. “If our actual spending falls short, we must make up the difference to Google and Amazon.”

Once you board this war chariot, you can never hit the brakes.

The only exception is the $120.5 billion contract signed with Musk's SpacexAI, which leaves an opening: after passing the initial non-cancelable period through January and March 2027, the two parties are allowed to cancel with 90 days' written notice, providing a certain degree of exit flexibility.

What is more delicate is its “plastic brotherhood” with Google, Amazon, and Microsoft.

These three cloud giants are at once Anthropic's financial backers and shareholders, its compute landlords, the sales channel that collects 60% of its outstanding receivables, and also its number-one competitors directly building large models.

In its prospectus, Anthropic reminds investors that the company's demand for computing resources may exceed the scale some partners are willing to provide.

It admits in the prospectus that the deep ties with the giants give the other side a clear view of its pricing and commercial bottom line.

If competitors “throw a wrench” into compute allocation or raise prices on the spot, the company's business will face a direct blow.

To escape the fate of “living under someone else's roof,” Anthropic announced that it is accelerating the construction of dedicated data centers and directly leasing chips, trying to seize its lifeline back into its own hands.

An 80-page “silicon doomsday warning,” and 50.1% iron-fisted control

If the financial data show a commercial rampage, the latter half of the prospectus reads like a sci-fi thriller.

In the 261-page S-1 filing, Anthropic spends only 64 pages introducing its business, but as many as 94 pages listing risks (by comparison, SpaceX, which listed this year, disclosed only 38 pages of risks).

This is the first time in business history that a company has seriously warned of “human extinction risk” in a prospectus.

Anthropic admits that advanced AI may evolve self-protective behaviors such as resisting shutdown, manipulating information, extorting humans, and even “disguising itself to fool safety tests”; once it gets out of control, existing commercial insurance would simply be unable to pay.

Yet reality is full of contradictions.

Although the CEO has repeatedly called for “slowing down and putting safety first,” the prospectus exposes the brutality of competition. In a sample week in July, compute used for safety research accounted for only 6%; by the end of September, to block OpenAI, the company quickly released the Opus 5.5 large model.

While saying “slow down,” it is flooring the accelerator.

To prevent Wall Street's short-term profit-seeking from devouring the safety bottom line, Anthropic has rolled out a rare “iron-fisted defense mechanism”:

Public interest first: retain public benefit corporation (PBC) status, legally place “long-term human interests” above shareholder interests, and donate 5% of shares to establish a public interest fund.

Super voting rights: the seven co-founders, by holding the sole Class F super-voting share, directly lock in 50.1% voting power on key matters.

Control the board: of the seven-member board, the “Long-Term Benefit Trust,” led by former Federal Reserve Chairman Bernanke and national security experts, directly appoints four seats, holding a majority.

In other words, even if Wall Street and retail investors put in hundreds of billions of dollars in real money, the company's steering wheel and brake pads remain welded in the hands of the founders and the safety trust.

The title of “the world's richest siblings” has officially fallen to the Amodei family.

(helpless doge.gif)

One More Thing

Zooming out, Anthropic's choice to sprint toward a listing at this moment comes at a rather delicate stage for the global technology capital market.

In June this year, Musk's SpaceX reportedly completed an IPO of about $86 billion, and market expectations for listings by ultra-large tech companies subsequently heated up.

On the other side, OpenAI, which also confidentially filed for listing in June this year, reportedly postponed its listing plan to 2027.

But by this autumn, the wind suddenly turned biting cold.

As investors focused more on whether AI infrastructure spending can generate sufficient returns, compute infrastructure stocks and bonds such as CoreWeave were hit on both fronts, and a host of AI infrastructure IPOs that had originally been queued up, including Switch, Nscale, and AirTrunk, were postponed or withdrawn.

The bustling autumn tech IPO calendar was left overnight with only Anthropic as the “lone sprout” carrying the banner.

This has also made the valuation battle surrounding Anthropic unprecedentedly fierce.

Five years ago, because of disagreements over OpenAI's development direction and safety philosophy, Dario Amodei left OpenAI with a group of core staff and founded Anthropic.

Five years later, this company has already become one of OpenAI's strongest competitors in commercialization thanks to Claude, Claude Code, and enterprise-grade AI products, and has been the first to publicly submit an IPO application.

The original prospectus: