What Anthropic's Prospectus Reveals: Dilution, Off-Balance-Sheet Obligations and a Thin Balance Sheet

Anthropic's widely anticipated prospectus, the S-1, has finally been made public. The presenter says the document has been leaked in parts beforehand and that it contains headline figures such as revenue growth of over 1,000%, along with discussion of improvements and of existential risk to humankind.

Richard says he decided to actually read the bloody thing instead of relying on summaries. He stresses that the video is not an investment recommendation, and that the prospectus is in many ways a marketing piece for the company.

The presenter points out that the document does not say what the company will be valued at. Rumors say the company is aiming for a valuation between $1.5 and $2 trillion and hopes to raise $100 billion with its initial sale. Richard warns viewers that the filing does not contain enough information for a comprehensive analysis.

He notes that Anthropic, the maker of the Claude model, is set to be the first pure-play AI company to go public and could be one of the biggest IPOs to date. It is expected to list on the NASDAQ under the ticker NT within about a month.

The filing opens with a letter from the founders. It states: "We believe AI is one of the most consequential technologies in human history. If managed well, it could rapidly accelerate scientific innovation, greatly increase economic growth rates, bring cures for major diseases within reach, and solve problems that have resisted generations of human effort."

The letter highlights promising areas for AI as well as top risks, including the development of bioweapons and government repression. It ends by asking investors not to think of their purchase in purely transactional terms but to be part of something greater, which Richard jokingly compares to a pitch from an MLM salesman.

The risk section is just shy of 100 pages and, according to the presenter, is not all about AI. Much of it is very generic, though it does include discussion of catastrophic or existential risk to humanity, which the video says it will set aside for now.

Richard explains that Anthropic is set up as a public benefit corporation under Delaware law. Rather than being legally obligated only to prioritize shareholder returns, the company must balance that priority against another public benefit purpose, which the filing describes as the responsible development and maintenance of advanced AI for the long-term benefit of humanity.

To enforce this, the company has set up a long-term benefit trust filled with independent AI experts, which holds special Class T shares. These allow the trust to elect Class T directors, who will constitute a majority of the board. Because the trust is not invested in the company, Richard says it could at times be at odds with shareholders over how the company grows.

One Class F share held by a new entity called Founder LLC makes up 50.1% of the company's voting power, and its vote is controlled collectively by the founders. Richard says that, in shareholder votes, the founders will therefore have majority control in addition to the regular shares they own.

Planned Class E shares for employees offer some limited voting opportunities. Richard notes that the Class T, F and E shares carry no economic value but take a great deal of power away from Class A shareholders, the shares available to the public in this IPO.

Class N shares carry economic value but limited voting power and are largely issued to strategic business partners and to the company's new philanthropy arm, the public benefit fund. That fund has already been given 5% of outstanding shares and is expected to receive an additional 2% each time Class A shares appreciate 50%.

On top of these arrangements, the filing lists $18.6 billion in convertible notes and $157.5 billion in convertible preferred shares, which together represent the company's debt. All of it is set to be converted into shares immediately before the IPO. Richard summarizes it as "Lots of dilution set to come."

Richard says the prospectus includes many operational details, some new and some already widely known. He highlights that about 80% of revenue comes from consumption-based sales versus about 20% from subscriptions. Enterprise customer growth has been rapid, with the filing tracking the number of enterprise customers spending over $100,000.

Anthropic has become increasingly reliant on selling through cloud providers. Amazon, Google and Microsoft sell Anthropic's offerings to their own customers and split the fee. Those cloud providers accounted for about 48% of revenue, compared with 11% in 2023. Richard notes this is an interesting dynamic because these companies are not only suppliers and investors in Anthropic but also competitors.

The filing estimates that 5 to 7% of revenues come from customers violating Anthropic's usage policy. It does not say which terms are being violated, whether that means something like generating images of Mickey Mouse or something more serious. Richard says it does open the company up to liability, since Anthropic could be on the hook for damages caused by users inappropriately using its model.

He also notes that the company borrows a page from SpaceX's playbook by showing case examples of large businesses using its AI to improve operations, which he describes as dozens of them.

Richard says the balance sheet is quite limited. Around 91% of its assets are cash or short-term investments, while liabilities include $108.6 billion in convertible notes and $157.5 billion in redeemable convertible preferred stock. Counting both as debt puts the debt-to-asset ratio at 1.8 times and a shareholder deficit of $127.6 billion. Richard adds that the convertible notes and preferred shares are set to be converted, so this value will largely shift into shareholder equity rather than being repaid in cash.

On adjusted figures, the company reports an operating margin of 10% and a higher net income margin thanks to interest earned on a cash balance of $16.4 billion. Richard points out that these are not GAAP figures. To reach adjusted operating profit the company added back $2.3 billion in stock-based compensation and charitable contributions, and to reach adjusted net income it added back $81 billion in fair value changes. He explains that issuing more shares still hurts investors even when those costs are excluded.

Looking at the GAAP-compliant income statement, Richard says operations are not as bad as most people probably expected. Revenue rose from $386 million in 2024 to $4.6 billion across all of 2025 and $16.3 billion in the first half of 2026, which is 14 times the same period in 2025.

The operating profit margin was negative 4.3% in the first half of 2026 but turned positive in the second quarter. Revenues are outpacing costs, and compute and infrastructure is the single largest operating expense.

Below the operating line, a large item sits at 4.4 times revenue and eats up any operating profit the company might have generated, due to fair value changes. These represent the convertible notes and preferred shares issued to various investors to raise money. Because they convert into Class A or N shares later, and the value of those shares rises with the company's value, the cost of conversion rises each time the company becomes more valuable, which Richard says is why the line has skyrocketed.

As noted, all of these convertible securities are set to convert right before the IPO, which will represent massive share dilution.

Richard says investors might assume profits will improve once the conversion dilution is past. The prospectus, however, says the company expects operating margins to fluctuate. This is partly because it will soon realize more upfront expenses from its infrastructure buildout and partly because it plans to enter more short-term compute agreements, which are generally more expensive than longer-term arrangements. After going public, Richard says the company is very likely to continue raising capital to meet its large future obligations.

Richard says there are sizable future obligations, most of which are not reflected in the balance sheet. The aggregate figures are missing from the table, but the document gives examples of how large the deals are. As of June, purchase commitments stood at $265.6 billion. Upcoming leases starting in the next two years carry payments totaling $371.6 billion over their life, and an upcoming finance lease has payments of $161.2 billion excluding interest. Most of these are non-cancellable and sit alongside hundreds of billions of dollars of other potential spending.

Richard contrasts these amounts with revenues over the past 12 months, which were under $20 billion.

Richard says the company is making so little money at present that it is absolutely dependent on capital raises to survive. The filing highlights $125.2 billion in upcoming lease payments to Broadcom over five years, excluding interest, which he says alone is six times the company's revenue. If revenue growth slows or capital dries up, he says Anthropic would very quickly find itself in a liquidity crisis.

The video also mentions risks described in the filing, including models potentially pursuing deceptive or self-preserving behaviors, and advanced models being manipulated to assist malicious actors. The filing cites a Chinese state-sponsored group using Claude Code to infiltrate 30 global targets, which it describes as the first documented large-scale cyber attack executed without substantial human intervention.

The filing also describes the risk of the AI making an irreparable blunder for customers, such as deleting a critical file, which could open the company up to liability. No dollar amount for this potential liability is provided. Richard says the company faces a wave of copyright infringement and other cases over how it trained its models and how those models have been used.

Richard argues that many of these risks could independently spell financial ruin for Anthropic, in addition to the existential risks the filing mentions. He adds that the company's ease in raising capital by issuing shares is itself a risk for investors, since the large dilution is expected to be followed by sales of shares by holders such as Google, Microsoft and Amazon, which have their own AI ambitions. The unlocking schedule for additional shares has not been disclosed.

No IPO valuation has been set yet. Richard says that if it lands near $2 trillion, the price-to-sales ratio would be around 100 times. That could be justified by rapid growth, but by the company's own admission current revenue growth likely reflects early adoption and will slow substantially at some point. Richard says the stock could still do well and be an attractive investment if exceptional growth lasts a few more years, but high-return opportunities carry a lot of risk.

Richard's high-level takeaway is that the income statement was better than he expected, while the balance sheet was much worse. The company's recent cash raise shows how much money it has burned historically. He says its off-balance-sheet obligations are similar to those other big tech companies have signed up for, even though Anthropic has only a fraction of their current profits or revenue. He concludes that Anthropic might catch up quickly to high-revenue companies such as Google or Meta, but it will be a roller coaster to follow.