Anthropic Prospectus Deep Dive: The World’s Largest AI Company Squeezes Its Own Bubble

Anthropic has formally filed for an IPO, and the prospectus is now public.

01

Ledger

First look at income.

Anthropic launched Claude in Q1 2023, beginning commercialization.

The subsequent annual revenue is:

By quarter, the four quarters of 2025 are 376 million, 787 million, 1.382 billion, and 2.040 billion USD; Q1 2026 is 4.731 billion, Q2 2026 is 11.555 billion. The sequential growth rate jumps from 48% to 132%, then to 144%.

Q3 revenue is not stated directly, but the prospectus mentions that Q3 2026 revenue is 86 times that of Q4 2024 ($197 million). Based on this, it is estimated at about $16.9 billion, with the quarter‑over‑quarter growth slowing to roughly 47%.

The prospectus itself is also cooling down the growth rate.

So how did the net loss of 80.7 billion, which has been widely discussed on social media lately, arise?

The main reason is two accounting issues.

The first item, $72.388 billion, represents the fair‑value changes of convertible bonds, derivatives, and SAFEs, with the bulk coming from Amazon.

In 2023 and 2024, Amazon lent money to Anthropic through two note purchase agreements. As of June 30, $6.203 billion of principal remained unconverted, all in non‑interest‑bearing notes. Under the agreement, they will convert at the IPO into non‑voting Class N shares at $2.0928 and $5.0478 per share (or the IPO price, whichever is lower), totaling about 1.84 billion shares.

These notes, with a principal of 6.2 billion, had a fair value of $108.566 billion at the end of June. Each time the company’s valuation rises, this liability is re‑measured, and the difference is recorded as a loss. By the end of June, out of the cumulative $132.9 billion loss, $113.2 billion stemmed from such fair‑value changes.

Two years, from losing almost ¥1 for every ¥1 sold to keeping ¥0.85 for every ¥1 sold.

However, most of the $8.945 billion in compute and infrastructure costs incurred in the first half were not recorded as operating expenses but were capitalized into R&D—such as training, research, internal use—and a portion was allocated to sales and marketing to support free users. R&D expenses in the first half totaled $9.741 billion, about 60% of revenue.

02

The concentration is as follows:

In the first half of the year, 48% of revenue came from cloud marketplaces such as AWS, Google Cloud, and Azure, and Anthropic must pay channel fees for this. Sales, marketing, and partnership expenses totaled $3.168 billion in the first half, with channel partner costs up $2.7 billion from a year ago.

Growth is mainly driven by new models. Opus 4.5, released in November 2025, increased the month‑over‑month growth rate from 48% to 132%. Models released between April and September 2026 contributed more than 80% of Q3 revenue. In September, Opus 5.5 cut single‑task costs by about 40%, Fable 5.1 reduced typical task costs by about 25%, and complex tasks saw reductions of up to 45%, yet the prospectus says the profit margin remained intact.

The prospectus also notes a few less flattering points:

• In the first nine months of 2026, about 5% to 7% of revenue came from customers later found to be violating usage policies or commercial terms. Those who were blocked may return with a new account. After stricter enforcement in August and September, this portion of revenue will decline, potentially dragging down subsequent growth. The company believes this situation is common across the industry.

• Usage-based billing cannot avoid holiday effects. In the two weeks around New Year’s Day, total customer spending dropped by about 20% compared with the previous two weeks.

Computing Power

There is a sentence in the risk factors:

• Microsoft: at least $44.9 billion, from November 2026 to May 2033;

• Google: at least about $114.5 billion, from April 2026 to July 2033; also pledged to purchase TPU compute delivered from 2028 to 2031, estimated $125 billion to $150 billion;

• Amazon: at least $110 billion, from May 2026 to April 2036, irrevocable;

SpaceXAI (formerly X.AI, a wholly owned subsidiary of SpaceX): uses NVIDIA GPUs, term extended to 2029, up to a total of $120.5 billion. Capacity is expected to be fully in place by December, then about $3.1 billion per month thereafter. The initial irrevocable periods end in January and March 2027 respectively; after that either party may cancel with 90 days written notice;

• Nscale: up to approximately $44.6 billion;

• Lambda: about $41.3 billion;

• Broadcom: five‑year equipment lease, total rent about $161.2 billion;

• AMD: starting in 2027, purchase rack systems based on AMD Helios; if fully purchased, spending will exceed $20 billion.

Anthropic is also starting to build its own data centers, with the first one coming online this year, and some facilities powered by on‑site natural‑gas generation. It has also made voluntary commitments, such as covering grid‑upgrade costs and procuring new generation capacity, to reduce the impact of data‑center electricity use on local communities and electricity customers.

Such a large investment is not avoided by Lianchuangxin. They wrote that these investments “may seem irresponsibly large at first glance.”

They use pharmaceuticals as an analogy: the R&D stage requires huge investment, and returns are only visible once the drug is completed, but with Anthropic this process is much faster and on a much larger scale. They also say they are willing to sacrifice some upstream to gain downstream protection, meaning they will resell excess compute when demand is overestimated, improve efficiency, and raise funds if necessary.

Money

• Series A: $2.57

• Series B: $11.23

• Series C: $11.23, 2023, $445 million

• Series D: $30, 2024, $1.2 billion

• Series E-1: $56.09, Q1 2025, $3.57 billion

• Series F-1: $140.97, Aug–Sep 2025, $13 billion

• Round G-1: $259.14, Jan–Feb 2026, $23.5 billion; concurrently G-2 has an additional $21.7 billion

• Round H-1: $589.0095, May–Jun 2026, $45.7 billion; concurrently H-2 has an additional $19.5 billion

• Google: signed in March. In addition to buying about $10 billion of G‑2 shares, it also pledged up to $30 billion in debt or equity financing by the end of 2030, with $20 billion remaining as of the end of June; whether it can be drawn is tied to the conditions of the third‑party compute agreement and revenue targets. Google also has an option: at IPO, privately subscribe for up to $10 billion at the offering price.

• Amazon: signed in April. About $5 billion to purchase G‑2 shares, and it also pledged up to $20 billion in financing, with $15 billion remaining as of the end of June, unlocking a tranche for every $5 billion delivered according to Amazon’s compute capacity milestones. Amazon’s shareholding is capped at 12%; any excess will be issued as notes.

• Broadcom: $42 billion, which can only be used to pay the quarterly lease payments for Broadcom equipment. 9% annual interest, paid in kind; if the revenue target is not met, the rate rises to 13%, and the conversion price drops from 90% of the 30‑day average price to 85%. This revenue target is set far below Anthropic’s actual quarterly growth rates in past quarters.

• AMD: additionally commits to subscribe for up to $5 billion of Class N shares in batches, based on the progress of GPU delivery and deployment.

Looking at them together, Amazon and Google are simultaneously shareholders, creditors, landlords, channels, and customers.

They are also competitors. The prospectus states that Amazon, Google, and Microsoft sell competing models on their own clouds, and the cloud channels also allow them to see Anthropic’s pricing and commercial terms; some compute power and infrastructure rely on capital and credit support provided or facilitated by cloud providers and chip partners. Reducing this dependence depends on whether it can independently raise funds based on its own credit.

Voting Rights

Anthropic is a public benefit corporation (PBC) in Delaware. The law requires the board to balance the economic interests of shareholders, the interests of those affected by the company’s actions, and the public benefit purpose.

After the IPO, it has five classes of common stock:

The 50.1% is locked in. The prospectus gives an example: for every additional 1 million Class A shares issued, Class F automatically gains about 1,004,008 votes. All Class A shares combined will forever total only 49.9%.

The board has 7 seats, LTBT selects 4 seats, the co‑founders and Class A together select 3 seats. According to Nasdaq rules, this is a “controlled company”.

When will the co-founders hand over their voting rights?

Founder LLC agreement includes a sunset clause: if fewer than three co‑founders remain in the LLC for a continuous 90 days, a 450‑day transition period begins; after the transition period, Class F and employees’ Class E are redeemed by the company at zero consideration. The co‑founder letter states: “Our increased voting power is not permanent.”

When does a co‑founder count as “no longer present”? It requires that they resign, die, become incapacitated, have their shareholding fall below 2% of the IPO level, or be removed for cause, including moving to a competing business. “Incapacity” has two scenarios: one is permanent total incapacity medically; the other is wrongful imprisonment, illegal detention, or capture, expected to last more than 12 months.

Benjamin Mann has been with the company since its founding in January 2021, often referred to as a co‑founder, but he is not part of Founder LLC and does not participate in the 50.1% voting block. The “co‑founders” mentioned in the prospectus refer to the other seven individuals.

LTBT does not have an automatic sunset here. To amend the trust or cancel Class T, it requires 80% approval from both the board and the trustees; if the trustees oppose, the co‑founders, employees, and public shareholders each must meet their respective thresholds; if the co‑founders are also absent, then 85% of Class A approval is required.

The shareholder agreement also contains a clause that as long as Class T remains, the company must give LTBT prior written notice before liquidation, selling more than half of the voting rights or economic interests, or filing for bankruptcy. The same list also includes an item to transfer control of an AI model that exceeds the capability threshold set by the board.

04

22%

Before the IPO, Anthropic placed 5% of its fully diluted equity—846,840,960 Class N shares—into its wholly owned subsidiary, the Anthropic Public Benefit Fund. After the listing, for every 50% increase in the stock price, the board can allocate up to an additional 2%. This money will be used for science and medicine, bio‑security, global health, economic research, and to more broadly distribute the economic value generated by Anthropic’s equity.

The other two items:

Gate

In April 2026, Anthropic trained the then‑strongest model Claude Mythos Preview, which was not publicly released.

Anthropic explained that they judged this model’s cybersecurity capabilities to be too strong, and releasing it widely without new safeguards would pose excessive risk. Therefore, they only opened it to a few organizations for network defense through Project Glasswing, and after adding protections, they released it publicly under the name Claude Fable 5. The letter admits that this delay means forfeiting a large amount of revenue and ceding ground to competitors; they also granted network defenders about $100 million worth of Claude credits.

Just how powerful is this model?

The prospectus included a Mozilla case study. In a two‑week evaluation, Opus 4.6 discovered 22 previously unknown bugs in Firefox, 14 of them high‑severity; later Mythos Preview found another 271, which were fixed in Firefox 150. Firefox’s CTO said that to date they have not found any category or level of complexity of bugs that humans can find but models cannot.

On the ExploitBench conducted by Carnegie Mellon researchers, over about four months, the score rose from 40% on Opus 4.8 to 78% on Mythos 5, and then to 91% on Opus 5.5.

Confession

Risk factors are about 95 pages.

The prospectus states that researchers, regulators, and policymakers both inside and outside the company have raised concerns. Sufficiently advanced AI systems, if lacking proper safeguards, oversight, and safety measures, could pose catastrophic, even existential, risks to humanity.

Specific cases that have occurred:

There are also some points about themselves:

• In September 2026, over 90% of contributions in the codebase were written by Claude; in Q3, the average daily code contribution per engineer was more than 45 times that of 2024;

• Since April 2026, the monthly internal value of using Claude (at list price) has already exceeded the monthly personnel costs (excluding equity incentives);

• The workforce grew from about 940 at the end of 2024 to roughly 5,300, and the average revenue per employee over the past 12 months exceeded $6.6 million.

The corresponding risk factor quote is: “our models write a large and growing share of our code and increasingly perform our research, including our safety research.”

It then notes that the company may lose visibility into certain processes, and monitoring does not cover all internal and external deployments; moreover, evidence of “AI accelerating AI R&D” is mostly lagging indicators, which may only be recognized after a significant acceleration has already occurred.

A company preparing to go public, why does it need to write this?

Part of the reason is legal requirements and insufficient disclosure, which become grounds for future lawsuits. Another reason is the very product Anthropic has been selling—its safety narrative is why enterprise customers choose it.

The prospectus also outlines the costs; these voluntary disclosures go beyond industry norms and regulatory requirements and could be used against the company in lawsuits, regulatory proceedings, and public disputes. Regulators have already, and may continue to, intensify scrutiny based on them.

Selling stocks is also tightly regulated by various rules. For major shareholders holding more than 0.5%, the lock‑up period can be up to 365 days, but it isn’t released all at once at expiry; on day 27 and day 48, 1.8% each is released, on day 62 and day 76, 2.1% each, and thereafter it follows the earnings reports, unlocking tranche by tranche. Starting from day 180, Lianchuang can transfer up to 29% initially.

In the prospectus’s financial footnotes, there is a table of preferred stock. The Series A row shows 48.574 million shares at $2.57 per share, totaling $125 million.

That money was not easy to obtain back then.

Early Anthropic investor Anjney Midha has recounted this experience many times. Tom Brown called him, saying he wanted to start a new lab with Dario and others, needing funding and compute. Starting from early 2021, they met weekly to turn research ideas into a business plan. Midha introduced the team to 22 firms on Dune Road, and 21 said no.

The money ultimately came from a group of individual investors. The Series A round was led by Skype co‑founder Jaan Tallinn, with participation from Facebook co‑founder Dustin Moskovitz and former Google CEO Eric Schmidt, many of whom are closely connected to the effective altruism (EA) community.

Before investing in Tallinn, let Dario get offers from a few VCs first. He wants to confirm that this company isn’t something he could also fund, so his money would only replace a particular VC’s, rather than magically create an AI lab out of thin air.

Midha’s own stake was almost all his savings at the time, most of which were Discord shares.

In 2022, the lead investor in the Series B round was Sam Bankman-Fried, and FTX invested $500 million at $11.23 per share. In November of the same year, FTX collapsed.

The VC’s first lead investment had to wait until the C round in spring 2023. Yasmin Razavi of Spark Capital gave the nod, investing $75 million. In the prospectus, the C round price was $11.2261 per share, the same as the B round.

That was the first time a Silicon Valley firm took a heavy stake in Anthropic.

Razavi is 37 this year and is still on the board, occupying one of the three seats selected by Lianchuang and Class A shareholders.

Those Series B shares later went their separate ways.

In 2024, the FTX bankruptcy trustee sold about two‑thirds of them for $884 million, and the biggest buyer was linked to Mubadala of Abu Dhabi.

Another small portion is the 4,453,900 Series B shares that Anthropic bought back from the U.S. government in October 2025, at $140.9676 per share, and on the same day resold them to nine investors at the same price, totaling about $627 million. The prospectus only states that these shares were confiscated from the original holder in a lawsuit unrelated to the company.

4,453,900 shares multiplied by the Series B price amount to about $50 million. Media reports that former FTX executives Caroline Ellison and Nishad Singh each invested $10 million and $40 million respectively in the Series B round; in the plea agreement, they handed over these shares, and the numbers match.

At the Series H price, this batch of shares is now worth about $2.6 billion.

Now calculate two items using the Series H valuation. The $1 invested in Series A is now worth about $229; Spark’s $75 million is now worth about $3.9 billion.

Midha said that later he invested several hundred million dollars in successive rounds and plans to donate most of it to charitable causes.

Is AI killing “traditional companies”?

What is it that 64% of people never manage to overcome in their lifetime?

A 6 billion‑yuan acquisition that helped a father‑son duo in Suzhou rake in 94 billion yuan

AI content enters the experience era: How does Loopit turn every idea into a playable world?