Anthropic

IWhat the business is

Anthropic builds Claude, a family of large language models (LLMs). An LLM is software trained on very large amounts of text and code so that it can write, reason, analyse and, increasingly, carry out multi-step work on a computer. The company offers access to Claude to businesses, developers and individuals. It was founded in January 2021 by seven co-founders, and the filing says all seven are still with the company²⁴¹³.

The product line has four model classes, named Mythos (which includes Fable), Opus, Sonnet and Haiku. Each sits at a different point on a trade-off between capability, cost and speed: larger models do harder work and cost more per task⁴. Customers reach them in three ways. They can call the models directly through the Claude Platform, a developer toolkit. They can use Anthropic’s own applications, chiefly Claude Code, a coding agent that plans, writes, tests and fixes software, and Claude Cowork, which does the same kind of multi-step work on a desktop for non-programmers. Or they can obtain Claude through Amazon Web Services (AWS), Google Cloud (GCP) or Microsoft Azure, the three big cloud platforms².

Anthropic is not a public company yet. It filed a registration statement on Form S-1 with the US Securities and Exchange Commission (SEC) on 26 October 2026, then Amendment No. 1 (S-1/A) on 2 November 2026, and says it has applied to list Class A common stock on the Nasdaq Global Select Market as ANT¹²³²⁶. A Form S-1 is the document a company files to register shares for a first public sale¹²¹³⁴. The amendment adds third-quarter and nine-month results, but the number of shares, the price range, the proceeds, the dilution table and the pro forma balance sheet are still blank, and the SEC has not declared the filing effective¹²¹³⁴. Until it does, shares cannot be traded on an exchange. This page uses the latest filing (the S-1/A) and calls it the S-1; where the original of 26 October differs, this page says so²⁵. Numbers are the filing’s own unless marked as my arithmetic. The company’s newsroom lists the announcement under 26 October 2026²⁵.

Scale, as the company reports it at 30 September 2026: more than 11,900 customers whose trailing-twelve-month revenue is at least $100,000, up from more than 950 a year earlier; more than 1,900 above $1 million; about 250 above $10 million; and more than half of the Fortune 500 above $100,000²⁴. A further group of about 40 million customers each spend less than $100,000²⁴. Revenue per employee over the trailing twelve months is stated as over $6.6 million, against $1.4 million a year before².

Anthropic is a public benefit corporation (PBC). In plain terms, its directors must balance shareholders’ money interests against the interests of people affected by its conduct and a stated purpose: the responsible development of advanced AI for the long-term benefit of humanity¹¹³. The company says it has delayed or limited some releases on safety grounds, for example releasing Mythos first only through a restricted programme (Project Glasswing)⁴. In June 2026 the US Department of Commerce imposed worldwide export restrictions on Fable 5 and Mythos 5, and the company disabled both for all customers until the restrictions were lifted¹¹. These show that regulators and the company’s own safety rules can switch off revenue-generating products.

IIHow it makes money and the unit economics the S-1 reveals

There are two revenue models. The main one the company calls pay-for-intelligence: customers pay for what they use, metered in tokens (small chunks of text) and priced by the model chosen. The second is a subscription for a set level of usage, mostly for the Claude applications, generally for 12 months or less⁴. Consumption-based revenue was $26.3 billion, or 79% of revenue, in the first nine months of 2026; subscription revenue was $6.9 billion, or 21%⁴. In 2024 the split was 75% and 25%, and in 2025 it was 83% and 17%⁴. A growing number of large customers sign annual or multi-year contracts with a minimum spend, but these still mainly produce consumption revenue⁴.

Distribution matters for the margin. Anthropic keeps the direct contract and sets the price, even when a customer purchases through a cloud marketplace, but it pays the cloud a fee. Marketplace sales were about 11%, 32%, 47% and 44% of revenue in 2023, 2024, 2025 and the first nine months of 2026, and 52% of net accounts receivable at 30 September 2026 sat with sales made through those platforms²¹⁰. Sales, marketing and partnerships expense, mostly fees to channel partners, rose from $322 million to $5.5 billion between the first nine months of 2025 and 2026, and fell from 22% of revenue in the first quarter of 2026 to 14% in the third as channel terms changed³⁶. That fee is the price of distribution, and it is paid to companies that also compete with Anthropic.

Growth inside existing customers is the heart of the story. The filing reports Net Revenue Retention of about 900% at 30 September 2026¹. By the filing’s definition, it takes customers that paid in the three months ending a year earlier, excludes the fixed-price Claude Pro and Max subscribers, and divides their latest three-month revenue by their revenue a year before¹. A figure of 900% means that group spends about nine times as much⁴. The company says its Fortune 500 customers spend under 0.1% of their expenses with it on average, which it reads as room to grow⁴.

The cost side is where the filing is most unusual. Management splits operating expenses by function. For the first nine months of 2026, cost of revenue (substantially all compute used to serve customers) was $4.8 billion on $33.2 billion of revenue, which implies a gross margin of about 85%¹⁸. The same measure was about 29% in the first nine months of 2025 and about 40% for 2025 (my arithmetic on the filing’s figures)¹⁸. Research and development, mostly compute for training plus staff, was $19.2 billion, or 58% of revenue; sales and marketing was $6.6 billion (20%) and general and administrative $3.8 billion (12%)¹⁸. Two cautions apply. Compute is bought as one shared pool and split across serving, research and internal use by management judgement, so the line between ’cost of revenue’ and ’research’ is not audited fact, and a different split changes the gross margin sharply⁴. And total compute and infrastructure expense was $18.8 billion, or 57% of revenue, in the nine months, down from 172% a year earlier, but it rose from 48% of revenue in the second quarter to 58% in the third as $5.0 billion of costlier short-term compute arrived⁶.

Other figures that bear on revenue quality: model versions released in the six months to 30 September 2026 produced over 80% of third-quarter revenue, so customers move fast to the newest model and an older model’s revenue can fade quickly⁴. Price reductions on new models and contract discounts cut third-quarter revenue by about $600 million against the second quarter, a sign of price pressure⁶. Customer spend fell about 20% in the two weeks to 2 January 2026 against the prior two weeks, which the company ascribes to holiday code freezes⁴. And the company estimates that 5% to 7% of nine-month revenue came from customers using its models in breach of its usage policy or commercial terms, recognised before it can detect and cut them off⁴.

Selected operating metrics reported in the S-1/A

IIIThe financials in the S-1

The audited statements cover 2023, 2024 and 2025³. The nine-month figures are unaudited³. Revenue rose from $41 million in 2023 to $386 million in 2024 and $4.6 billion in 2025, then reached $33.2 billion in the first nine months of 2026 against $2.5 billion a year earlier³⁵. Quarter by quarter, revenue was $376 million, $787 million, $1.38 billion and $2.04 billion in 2025, then $4.73 billion, $11.56 billion and $16.96 billion in the first three quarters of 2026⁶. The third quarter grew 47% on the second, slower than the 144% of the quarter before⁶. The chart plots those seven quarters.

The two press stories that disagree are both describing true numbers. The ’profit’ is operating profit, adjusted. GAAP operating income was $974 million in the second quarter of 2026, but the third quarter was a GAAP operating loss of $567 million (minus 3% of revenue), which the company ties to $5.0 billion of short-term compute and $2.0 billion of non-cash donation-matching charges⁴⁶⁷. Its own adjusted operating income, which removes stock-based compensation, non-cash charitable contributions and the 2025 Bartz copyright settlement, was $2.2 billion (19%) in the second quarter and $2.5 billion (15%) in the third⁴⁶⁷. For the nine months GAAP operating loss was $1.26 billion (minus 4%) and adjusted operating income $4.1 billion (12%)³. The ’loss’ is the net loss of $104.0 billion, almost all non-cash. $94.3 billion is the change in fair value of convertible notes, derivative liabilities and SAFEs (simple agreements for future equity), which the company marks to market each quarter: when its valuation rises, the investors’ right to convert into shares is worth more, and the company books that as a loss³⁶. Another $10.4 billion of other expense is the value change on the partner financing agreements³⁶. The loss measures how much more the equity became worth, not cash leaving the business.

Adjusted measures flatter the picture. Stock-based compensation was $3.3 billion in the nine months, against $591 million a year earlier, and it is a real cost to owners because it creates shares³. The filing shows $24.1 billion of expense not yet recognised on restricted stock units (RSUs) whose service condition is not yet met; the performance condition is satisfied when the registration takes effect, so a large catch-up charge is due at the IPO³. Non-cash donation-matching charges were $2.0 billion in the third quarter alone and are settled in Class N shares, 28.0 million of them in October 2026⁶¹⁸. The company warns that margins may compress when large blocks of compute arrive, and the third quarter shows it: the adjusted margin fell from 19% to 15%⁴⁶.

Cash is the more reassuring measure. Net cash from operating activities was positive $6.6 billion in the nine months, against minus $2.0 billion a year earlier⁹. Investing outflow was $36.5 billion, mostly short-term investments and servers, and financing inflow $110.7 billion, mainly preferred stock⁹. At 30 September 2026 the company held $95.6 billion of cash and $36.3 billion of short-term investments, plus $2.6 billion of restricted cash, and had $14.8 billion of undrawn credit facility, for liquidity of $146.6 billion⁴⁹. Total assets were $154.3 billion and total liabilities $146.7 billion, up from $117.5 billion at 30 June³. Redeemable convertible preferred stock of $157.5 billion sits outside equity, leaving a stockholders’ deficit of $149.9 billion; the pro forma columns that assume the preferred converts at the IPO are blank³.

The original S-1 left third-quarter revenue blank³⁴⁶. My earlier bracket of $16 billion to $18 billion, from two ratios the filing gave (trailing revenue growth of about 1,200% and revenue per employee over $6.6 million with about 5,300 staff)³⁴, proved right: the amendment reports $16.956 billion, inside that range⁶. Annualised, the third quarter is $67.8 billion (my arithmetic). Trailing twelve-month revenue from customers above $100,000 was $26.1 billion, 74% of the total, which implies trailing revenue near $35 billion⁴.

Summary income statement, US$ millions (audited for 2023 to 2025, unaudited for nine-month periods)

IVCompute, suppliers and cloud commitments

Compute is the central cost and the central risk. Training and running frontier models needs very large numbers of specialised chips (graphics processors and similar accelerators) housed in power-hungry data centres. Anthropic says it manages compute as one fungible pool across four chip families: AWS Trainium, Google TPUs, NVIDIA GPUs and AMD GPUs, and routes work to whichever is best on price and performance with its own orchestration software. It argues this took years to build and is hard to copy, and that it reduces dependence on any one vendor²⁴.

Most compute is bought through multi-year commitments placed 12 to 18 months before the capacity arrives⁴⁶. When demand outruns that capacity the company procures short-term compute, which the filing says costs significantly more per unit; $5.0 billion of it arrived in the third quarter⁴⁶. Because the expense starts when capacity is delivered, not when demand shows up, the cost base is largely fixed in the near term. Compute spending as a share of revenue fell from 251% in the first quarter of 2025 to 48% in the second quarter of 2026, then rose to 58% in the third; a slowdown in revenue would push it up fast⁶.

The size of the commitments is large relative to anything else in the filing. The amendment now states it: aggregate committed spend of about $895.4 billion as of and after 30 September 2026, substantially all for compute, of which about $447.5 billion falls due before the end of 2030 and $69.8 billion in the next 12 months²⁹. The total is made of $320.0 billion in the contractual obligations table, $451.2 billion of leases signed but not yet started, $100.2 billion of commitments entered after 30 September and about $24 billion of short-term contracts (my sum of the filing’s items). It excludes a further $150 billion to $175 billion of estimated payments under arrangements with no fixed dollar amount, which includes the Google TPU capacity estimated at $125 billion to $150 billion⁹. About $213.8 billion of the $895.4 billion is subject to termination rights or conditions before it starts, such as a counterparty securing financing or facilities, which leaves about $682 billion without such outs (my arithmetic)⁹. Named agreements include Microsoft at least $44.9 billion to June 2033⁹³⁴; Google at least about $121.3 billion to July 2033 (up from $114.5 billion in the original filing)⁹³⁴; Amazon at least $110 billion to April 2036⁹³⁴; SpaceXAI up to $120.5 billion into 2029⁹³⁴; Nscale up to about $44.6 billion⁹³⁴; Lambda about $41.3 billion⁹³⁴; AMD above $20 billion; and Broadcom equipment leases with rent of about $161.2 billion⁹³⁴. Against third-quarter annualised revenue of $67.8 billion, total commitments are about 13 times a year of revenue (my arithmetic).

Most of the big contracts are non-cancellable and payable whatever the usage. The Microsoft, Google and Amazon commitments apply whatever actual usage turns out to be, with true-up payments if spending falls short⁹. The SpaceXAI contracts differ: after initial periods that end in January and March 2027, either side can cancel on 90 days’ notice, which gives flexibility but means the capacity is not guaranteed⁹.

Financing for all this comes from partners as well as customers. The company has three note purchase agreements for future financing, together up to $77.0 billion: Google up to $20 billion (draw period to December 2030), Amazon $15 billion in $5 billion steps tied to compute milestones, and a Broadcom-linked Infrastructure Partner agreement for up to $42 billion, issued only if needed to pay lease rent⁸. Broadcom’s notes carry 9% interest paid in kind (added to principal) and can convert at 90% of the recent average share price; they bear 13% and convert at 85% if revenue-growth conditions are not met, and on a change of control Broadcom can demand 150% of principal in cash⁸. AMD has committed to purchase up to $5 billion of Class N stock as GPU deliveries reach milestones, the first expected in 2027⁸. A $15 billion secured revolving credit facility runs to September 2031 and requires at least $5.0 billion of liquidity at each quarter-end from the quarter ending 31 December 2026⁸.

The filing is plain about a conflict. Amazon, Google and Microsoft are suppliers, distribution channels, customers, investors and competitors at once, and the filing warns that this can affect access to compute, its price and the priority given to Anthropic’s workloads⁴. A portion of the build-out has been supported by credit from cloud and chip partners, and reducing that reliance depends on getting financing on the company’s own credit⁴⁸.

Compute and infrastructure commitments (S-1/A, Contractual Obligations). Amounts as stated; named agreements are subsets of the total

VThe moat and how it could be disrupted

I rate the moat ’contestable’. That is a judgement about durability, not about quality: Anthropic’s models are, on the evidence of its revenue, among the best available. But the things that protect a business over ten years are costs of switching, network effects, regulation or a scale that rivals cannot reach. Anthropic has some of each, and none is strong enough on its own.

What it has. First, model quality and research speed. The company says the newest models drive most revenue (over 80% of third-quarter revenue came from versions released in the prior six months), and that Claude now helps write the next version of Claude²⁴. That is a lead, but a lead in a product that is replaced every few months is a moving target. Second, compute scale and the multi-chip software layer. Few firms can sign hundreds of billions of dollars of contracts. That limits entry, but the competitors named in the filing, OpenAI, Google and other large technology companies, can do it too¹⁹. Third, distribution through all three major clouds, where customers can spend from existing cloud budgets, and through consultancies and data platforms². Fourth, developer habit: Claude Code gave the company a base of engineers who recommend tools inside their employers².

What is weak. Switching costs between model providers are low for many uses, and the filing itself says customers often use the full range of Claude models and route each task to the best-fitting one⁴. The same routing habit lets a customer send a task to a cheaper model from another maker. The three clouds that offer Claude also offer competing models and could favour them. The company’s largest rivals own their own chips, clouds and consumer products, so they can cross-subsidise. Open-weight models (whose parameters anyone can download and run) are cheaper, and the filing lists them as a competitor¹⁹. The company’s own risk factors say that if customers build their own solutions or cut reliance, spending can fall at short notice¹⁰.

How it could break. A rival releases a model that matches Claude at much lower price, and the premium the company earns for being at the frontier disappears. A large customer or cloud moves a major share of volume elsewhere. The company cannot get enough power, chips or memory, or is slowed by export controls, as the June 2026 restrictions on Fable 5 and Mythos 5 showed¹¹. The government dispute widens: a Department of War supply-chain-risk designation that a district court found unlawful but the D.C. Circuit upheld on one statute on 25 September 2026¹¹¹².

The closest private rival is OpenAI. In February 2026 OpenAI announced a $110 billion round at a $730 billion pre-money valuation²⁷, and its chief financial officer said in March that the round had grown to more than $120 billion²⁸. TechCrunch later put the total at $122 billion and the last valuation at $852 billion post-money, without saying where the figure comes from³⁵. That matches $730 billion plus $122 billion (my arithmetic). OpenAI said in a June 2026 blog post that it had filed confidentially for an IPO and had not decided on timing³⁵, and its chief executive told Fortune in September that he would say an IPO is not coming in 2026³⁶. So a rival with comparable funding exists, and it is in the same race for compute. The test for a moat is whether profits survive competition. Reported gross margin has widened sharply (about 85% on management’s allocation in the first nine months of 2026), which suggests pricing power today, yet third-quarter price reductions of about $600 million show it is already giving some back⁶¹⁸. The company itself says it cut customers’ cost per task by roughly 40% between Opus 5 and Opus 5.5 while preserving its margin².

VIThe S-1: the pros

1. Very fast growth, with revenue up sequentially in every quarter shown: $376 million in the first quarter of 2025 to $17.0 billion in the third quarter of 2026, and nine-month revenue 13 times the prior year (S-1/A, MD&A, Quarterly Results, pp. 162 to 164)³⁶.
2. Operations are near break-even on GAAP and clearly positive on the company’s adjusted measure: nine-month GAAP operating loss of $1.26 billion, a $974 million operating profit in the second quarter, and adjusted operating income of $4.1 billion (S-1/A, Summary Financial Data, pp. 24 to 26)³⁶.
3. Cash from operations turned positive: $6.6 billion in the first nine months of 2026, against minus $2.0 billion a year earlier (S-1/A, Cash Flows, p. 173)⁹.
4. Breadth of customers⁴. More than 11,900 above $100,000, about 250 above $10 million, no customer above 10% of revenue, and the share of revenue from outside the top five rising from about 64% in the first quarter of 2026 to about 79% in the third (S-1/A, MD&A, p. 145)⁴.
5. Deep expansion inside accounts⁴. Net Revenue Retention of about 900%, and Fortune 500 customers spending under 0.1% of their expenses with Anthropic on average, which the company reads as a long runway (S-1/A, MD&A, Adoption of Claude, p. 137; customers, p. 145)⁴.
6. Compute efficiency⁴. In the third quarter revenue grew 1,127% year on year while compute and infrastructure expense grew 382% (S-1/A, MD&A, p. 138)⁴.
7. A large balance sheet. $131.8 billion of cash and short-term investments at 30 September 2026, $14.8 billion of undrawn credit, and up to $77.0 billion of note purchase capacity from Google, Amazon and Broadcom (S-1/A, MD&A, Liquidity and Capital Resources)³⁴⁸.
8. Diversified compute and distribution²⁴. Four chip families, three clouds and several independent capacity providers lower the chance that one supplier failure stops service (S-1/A, MD&A, p. 138)²⁴.
9. Stable leadership. All seven co-founders remain with the company, and the board includes the former chief executive of Netflix and a former chief financial officer of Microsoft (S-1/A, MD&A, Attracting and Retaining Technical Talent⁴¹⁵; Management)⁴¹⁵.
10. A first-instance legal win on the government dispute¹². On 27 August 2026 a federal district court largely granted summary judgment for the company, finding the Department of War and presidential directive actions retaliatory and unlawful (S-1/A, Legal Proceedings)¹². The appeal outcome differs; see the cons.

VIIThe S-1: the cons and risk factors that matter most

1. A very large GAAP loss³. The nine-month 2026 net loss is $104.0 billion and the accumulated deficit $156.2 billion³. Most of this is non-cash, but it also means the company carries obligations tied to its own valuation, and net loss swings with the share price until the notes convert; total liabilities rose to $146.7 billion (S-1/A, Summary Financial Data, pp. 24 to 26)³.
2. Fixed compute obligations⁹. Aggregate committed spend is about $895.4 billion, of which $447.5 billion falls due before the end of 2030, with a further $150 billion to $175 billion of estimated payments excluded. About $213.8 billion is conditional or terminable; the rest is largely non-cancellable and payable whatever the usage (S-1/A, Contractual Obligations, pp. 174 to 177)⁹. This is the single largest source of fragility.
3. Margins already moved the wrong way once⁴⁶. The adjusted operating margin fell from 19% to 15% and GAAP operating income swung from a $974 million profit to a $567 million loss between the second and third quarters, because of $5.0 billion of costly short-term compute⁴⁶. The company expects more of it in the near term (S-1/A, MD&A, pp. 136 and 139)⁴⁶.
4. Growth and price pressure⁴⁶. The filing says growth may fall significantly and faster than the company expects, and recent growth may reflect early adoption that is not sustainable. Third-quarter growth slowed to 47% from 144%, and price cuts and discounts took about $600 million off the quarter (S-1/A, MD&A, pp. 141 and 162)⁴⁶.
5. Customer and channel concentration¹⁰. The top five customers are about 25% of nine-month revenue, many are not locked in by long contracts, and cloud marketplaces are about 44% of nine-month revenue and 52% of net receivables (S-1/A, Risk Factors, pp. 58 to 59)¹⁰.
6. Suppliers, investors and customers who are also competitors²⁴. Conflicts could affect compute access, cost and priority (S-1/A, Risk Factors Summary²⁴; MD&A, Key Factors)²⁴.
7. Government and export-control risk¹¹¹². The Department of War supply-chain-risk designation, a presidential directive to stop federal use, the split 2 to 1 D.C. Circuit decision of 25 September 2026 upholding the Section 4713 designation (the company has asked the full court to rehear it), and the June 2026 export restrictions on Fable 5 and Mythos 5 (S-1/A, Risk Factors¹¹¹²; Legal Proceedings)¹¹¹².
8. Revenue quality⁴. The company estimates that 5% to 7% of nine-month revenue came from customers in breach of its policies, and expects enforcement to cut it (S-1/A, MD&A, Components of Results of Operations)⁴.
9. Copyright litigation¹². A series of suits over training on copyrighted material; the Bartz class action settled for $1.5 billion, approved by the district court in July 2026, and other cases continue (S-1/A, Legal Proceedings)¹².
10. Dilution already in the pipe²³⁸¹⁷. Amazon’s outstanding notes of $6.2 billion convert into up to 1.84 billion Class N shares at fixed prices of $2.09 and $5.05 per share, subject to a 12% ownership cap²³⁸¹⁷; Series H investors paid $58.90 per common-equivalent share, a figure I derive from the 10-for-1 split²³⁸¹⁷. A further $77.0 billion of note capacity, AMD’s $5.0 billion, Google’s $10 billion IPO option, donation-matching shares, $24.1 billion of unrecognised RSU expense and up to 2% more shares to the Public Benefit Fund for each 50% rise in the share price all add shares (S-1/A, Liquidity²³⁸¹⁷; Prospectus Summary; Summary Financial Data)²³⁸¹⁷.
11. Control sits with others¹¹³. Class A shares get one vote each but collectively only 49.9% of the votes; a single Class F share held by an LLC of the seven co-founders carries 50.1%; and a trust elects a majority of the board¹¹³. The company is a controlled company under Nasdaq rules (S-1/A, cover page¹¹³; Governance)¹¹³.
12. Tail risk and insurance¹¹. The filing says that after a catastrophic AI-related incident, no single insurer, and possibly not the insurance industry as a whole, might have the capacity to cover the exposure, and that its models can behave in unintended or autonomous ways (S-1/A, Risk Factors, AI harm and insurance)¹¹.

VIIIStructure, governance and who owns what

Anthropic’s structure is unlike an ordinary listed company. It is a Delaware public benefit corporation, so directors must weigh shareholders’ financial interests against the interests of people its conduct affects and its stated purpose. There is limited court precedent on how that balance is judged, the company measures its performance against objectives its own board sets, and it must report on this at least every two years¹³. The board therefore has room to take decisions that cost shareholders money.

The Long-Term Benefit Trust (LTBT) is a Delaware purpose trust with five authorised trustees, one seat currently vacant. It holds the only share of Class T common stock, which has no economic rights and does not vote on ordinary matters, but lets the trust elect and remove the directors who form a majority of the board. The trustees named are Neil Shah (chair, chief executive of the Clinton Health Access Initiative), Richard Fontaine (chief executive of the Center for a New American Security), former Federal Reserve chair Ben Bernanke (since June 2026) and Harvard physicist Christopher Stubbs (since October 2026)¹³. Each is paid $400,000 a year in cash, trustees may not hold company securities, and the company paid the trust about $8.4 million in the nine months to 30 September 2026¹³. The trust is consolidated into Anthropic’s accounts as a variable interest entity¹³. Changing the trust’s terms needs 80% of the board and of the trustees, or a combination of founder, employee and Class A votes, and the trust has no say in day-to-day strategy¹³.

Voting power is deliberately split from economic ownership. Class A shares, the ones offered to the public, have one vote each. A single Class F share, issued just before the IPO to a limited liability company whose members are the seven co-founders, carries 50.1% of all votes on general matters, including director elections other than the Class T seats, and has no economic rights¹³. Public holders of Class A therefore share the other 49.9% with every other Class A holder¹³. The Class F vote is cast by a majority of the co-founders; if they cannot agree it is voted in proportion to Class A votes. If fewer than three Continuing Founders remain for 90 days, a 450-day transition starts, after which Class F and Class E shares are redeemed for nothing¹³. Class E shares, for employees, vote only on limited matters. Class N shares, held mainly by strategic partners such as the cloud companies, have no vote and convert to Class A on a sale to an unrelated party. The company says it is a controlled company under Nasdaq rules, a status that can relieve a listed company of some board-independence requirements¹.

Who owns what is only partly visible, because the economic share counts are blank. The filing does state that 846,840,960 Class N shares, 5% of fully diluted shares at 30 June 2026 assuming Amazon’s notes convert, were issued to the Anthropic Public Benefit Fund, a wholly owned subsidiary. From that I infer about 16.9 billion fully diluted shares (my arithmetic). The Series H price of $589.0095 per preferred share converts into ten common shares after the 10-for-1 split of 17 September 2026, which is $58.90 per common-equivalent share and an implied value of about $998 billion, about 3% above the $965 billion the company announced^2171822; I cannot tell whether the gap comes from how donated shares or the option pool are counted^2171822. Each of the seven co-founders reports beneficial ownership of about 232 million to 245 million Class A shares, about 1.69 billion in total, roughly 10% of my estimate, worth about $99 billion at the Series H price (my arithmetic)¹⁴. Amazon’s notes of $6.2 billion could convert into about 1.84 billion Class N shares, roughly 11% of my estimate, subject to a 12% cap; Google’s cap is 14.99%⁸. No investor appears as a holder of more than 5% of the voting Class A, because large strategic and venture holdings are mostly non-voting preferred that converts to Class N¹⁴¹⁸.

The Public Benefit Fund’s managers are the two Amodeis, Jack Clark and Neil Shah, who also chairs the trust⁴. The board has approved issuing up to a further 2% of shares to the fund each time the share price rises 50% above the previous milestone, starting from the IPO price, if the closing price stays above the threshold on 90 of the last 120 trading days⁴²⁴. The co-founders have also pledged 80% of their personal equity to philanthropy⁴²⁴. Existing owners are diluted in a pre-set way as the share price rises, without a shareholder vote.

The company has never paid a dividend, expects to keep all earnings, and is limited by its credit agreement in paying one²¹. Lock-ups limit resale after the IPO: co-founders and holders of more than 0.5% of fully diluted shares are restricted until the earlier of the second trading day after results for the quarter ending 30 September 2027 and 365 days after the prospectus date; other holders until the earlier of the second trading day after first-quarter 2027 results and 180 days²⁰. A staged early-release schedule starting on the 27th day has blank share amounts, so a large supply can reach the market within a year²⁰.

Share classes after the offering, as described in the S-1

IXThe offering as filed

This is an initial public offering of Class A common stock. The company is offering an unstated number of shares, and certain existing stockholders are offering more. The ownership table lists Tom Brown, Jack Clark, Jared Kaplan, Sam McCandlish and Christopher Olah, five of the seven co-founders, under ’Other Selling Stockholders’. The word ’Other’ implies that people listed above them, which could include the Amodeis, may also offer shares; the ’shares being offered’ column is blank for everyone, so the filing does not yet say who offers how much. The company receives no money from shares offered by stockholders¹¹⁴.

The cover leaves blank the number of shares, the price range and the underwriting discount, and the amendment does not fill them in. The summary also leaves blank the net proceeds, the pro forma balance sheet, the dilution table and the share counts after the offering¹³²¹. The company’s announcement of 26 October 2026 says the number of shares and the price range have not been set²³. The SEC’s filing index lists the public S-1, the S-1/A of 2 November 2026 and six earlier confidential draft submissions, the first on 1 June 2026²⁶. The 22 underwriters on the cover are led by Goldman Sachs and Morgan Stanley, followed by J.P. Morgan, Citigroup, Barclays, Wells Fargo Securities, BofA Securities, Deutsche Bank, RBC, UBS and twelve others¹.

Use of proceeds, as stated, is general corporate purposes: working capital, operating expenses and capital expenditure including compute, hiring and retention, and possible acquisitions, though there are no agreements for any now. A portion is also earmarked to pay tax withholding on employees’ RSUs (the dollar amount is blank)²¹.

Dilution at the offering is not yet quantifiable. What the filing does set out is who else is getting shares at or around the IPO. Amazon’s notes convert at the lower of fixed prices ($2.0928 and $5.0478 per share) and the IPO price⁸. If the IPO price is above those fixed prices, Amazon receives shares at them, far below the $58.90 Series H price per common-equivalent share I derive, subject to its 12% ownership cap. Google holds an option to purchase up to $10 billion of Class N stock in a concurrent private placement at the IPO price, and AMD and Amazon have milestone-based commitments after the IPO⁸. The Public Benefit Fund already holds 5%². Lock-up terms are in the governance section²⁰.

Press talk of an IPO value near $1.5 trillion could not be verified in any source opened for this page, so it is not relied on. For scale only: on my estimate of 16.9 billion fully diluted shares, $1.5 trillion would be about $89 per common-equivalent share, about 50% above the Series H price, and about 22 times the third-quarter annualised revenue (my arithmetic).

XWho runs it and how they are paid

The executive officers named in the filing are Dario Amodei, chief executive and co-founder (age 43)¹⁵; Daniela Amodei, president, chair of the board and co-founder (39); and Krishna Rao, chief financial officer since April 2024 (43)¹⁵. Dario and Daniela are siblings¹⁵. Dario Amodei was previously vice president of research at OpenAI; Daniela Amodei held several roles at OpenAI and earlier worked at Stripe; Rao was chief financial officer at Fanatics Commerce and Cedar and held finance roles at Airbnb¹⁵. Other co-founders hold named roles: Tom Brown as Chief Compute Officer, Jared Kaplan as Chief Science Officer, Sam McCandlish as Chief Architect, Chris Olah leading interpretability research and Jack Clark leading the Anthropic Institute¹³¹⁴.

The board has six directors now and seven authorised after the offering: the two Amodeis, Reed Hastings (lead independent director), Christopher Liddell, Vas Narasimhan (chief executive of Novartis) and Yasmin Razavi (a general partner at Spark Capital)¹⁵. Hastings, Liddell and Narasimhan are Class T directors, elected through the trust; Razavi is elected by Class A and Class F. The audit committee has Hastings and Liddell¹⁵. Spark Capital funds bought $440 million of Series H stock, and Narasimhan’s family company bought $10 million, which are related-party transactions the filing discloses¹⁷. The company also gave $450,000 to the Center for a New American Security, whose chief executive is an LTBT trustee¹³.

Pay for the three named executives in 2025 was: Dario Amodei about $18.0 million ($654,167 salary, $9.3 million of RSUs, $4.0 million of options and $4.0 million of other pay, of which $2.7 million was personal security)¹⁶; Daniela Amodei about $16.4 million; and Krishna Rao about $0.72 million with no new equity awards that year¹⁶. RSUs need both time and a performance condition tied to the IPO, so the $24.1 billion of unrecognised RSU expense is due to start flowing through when the registration takes effect³¹⁶. Executives and directors as a group (seven people) report about 496 million Class A shares including options exercisable within 60 days¹⁴.

What cannot be judged from the filing: the quality of the research team, the depth of the bench below the three named officers, or how the co-founders would behave if they disagree. The voting mechanism gives the seven co-founders the final say over the non-trust board seats and over a sale of the company, and they are the same people running it.

XIValuation: is the last price supported?

Equity value in US$ billions. Not a share price: a private company has none. Latest mark: primary round (Series H, $65.2 billion gross per the S-1, closings May to June 2026), post-money $965 billion as announced on 28 May 2026.

What the last valuation assumes

At $965 billion the price is about 14 times the third-quarter 2026 annualised revenue of $68 billion (my arithmetic on $16.956 billion)⁶. Working backwards with my own inputs (12% discount rate over 4.2 years, 15% dilution), to be worth $965 billion today the equity must be worth about $1.8 trillion at the end of 2030. At 20 times earnings that is about $91 billion of annual net income; at a 25% net margin it needs roughly $365 billion of revenue in 2030, or about 62% growth a year from my 2026 estimate; at a 30% margin and a 25 times multiple, roughly $245 billion. A price of $1.5 trillion, which could not be verified, would need about $570 billion of revenue at a 25% margin and 20 times earnings.

Bear, base and bull

• Bear. Revenue grows from my 2026 estimate of about $53 billion (nine months reported at $33.2 billion plus an assumed $20 billion for the fourth quarter) to $100 billion in 2030, about 17% a year, as prices fall and rivals close the gap. Net margin reaches 10% because fixed compute commitments absorb the gain. Terminal multiple 14 times net income, in line with Alphabet’s trailing 14.6³¹. Discount rate 12% a year over 4.2 years, my assumption for a large company with high execution risk. Dilution 20% from partner notes, equity pay, public-benefit issuances and capital raised at lower prices. Result about $70 billion.
• Base. Revenue reaches $200 billion in 2030, about 39% a year from the 2026 estimate, so growth falls steadily from the 47% quarter-on-quarter of the third quarter. Net margin 22%, well below the 85% functional gross margin because of research compute, channel fees and stock pay. Terminal multiple 20 times net income, between Alphabet’s 14.6 and Microsoft’s 26.5 and NVIDIA’s 25.6 trailing multiples²⁹³¹³². Discount rate 12%. Dilution 15%. I do not add the $131.8 billion of cash and short-term investments (about 28% of this value) because the filing’s $895.4 billion of commitments is expected to consume it³⁹; adding it all back would lift the base to roughly $600 billion. Result about $465 billion.
• Bull. Revenue reaches $450 billion in 2030, about 71% a year, which would exceed NVIDIA’s trailing twelve-month revenue of about $303 billion today²⁹. Net margin 30%. Terminal multiple 25 times net income, matching Microsoft and NVIDIA²⁹³². Discount rate 12%. Dilution 12%. Result about $1.85 trillion.

Comparables used

Where I come down

The latest valuation is above my base case. $965 billion is roughly twice my base value of $465 billion, and about 2.8 times my margin-of-safety value of $349 billion (base less 25%). It sits between my base and my bull case ($1.85 trillion). The Series H price was set by large institutional investors on 28 May 2026²² and is five months old in a business whose quarterly revenue has since grown from $11.6 billion to $17.0 billion⁶. To make the price right you have to believe three things together: revenue of $245 billion to $365 billion by 2030, a growth path steeper than any large company has held; net margins of 25% to 30% after paying for part of $895 billion of compute commitments⁹; and a market that still pays 20 to 25 times earnings. The market could be right even if this looks high. Revenue has grown every quarter, third-quarter growth was still 47% on the second, operating cash flow is positive at $6.6 billion for the nine months, and if growth fades more slowly than in my base case the 2030 numbers needed for $965 billion are reachable⁶⁹. Against that, the third quarter showed the squeeze: GAAP operating income went from a profit to a loss, the adjusted margin fell from 19% to 15%, price cuts took $600 million off revenue, and compute spending rose to 58% of revenue⁶. My weighting of 30% bear, 50% base and 20% bull gives about $620 billion, still below the Series H price, but a different weighting or discount rate moves that a lot. The best public comparables (Alphabet at 7.0 times sales, Microsoft at 11.1, NVIDIA at 16.0) are far more profitable and diversified, so the price pays for growth not yet seen in the accounts²⁹³¹³². This is research, not advice, and nothing here says where the shares would be priced.

What I cannot judge

The S-1/A was read on SEC EDGAR itself. It is still a preliminary draft: the share count, price range, proceeds, dilution table, pro forma balance sheet and who offers how much are blank, so no IPO price or per-share figure can be valued¹²¹. Press talk of an IPO value near $1.5 trillion could not be verified. Model quality against rivals, the safety work, the outcome of the government and copyright litigation, and how management splits compute between serving customers and research (which sets the 85% gross margin) cannot be judged from outside. The 10-for-1 conversion is assumed to apply to every preferred class, and the fully diluted share count is inferred. My 2030 inputs are judgement and the result moves a lot with the multiple and margin: a one-point change in terminal multiple moves the base value by about $23 billion.

XIIThe IPO

IPO filed publicly. Anthropic, PBC filed a public Form S-1 on 26 October 2026 (registration number 333-299633) and Amendment No. 1 on 2 November 2026, to list Class A common stock on the Nasdaq Global Select Market under ANT¹²³²⁶. The SEC filing index also shows six earlier confidential draft submissions, the first on 1 June 2026²⁶. Goldman Sachs and Morgan Stanley are listed first among 22 underwriters¹. The number of shares, the price range and the offering size are still blank¹²³. The company is offering shares and certain stockholders, including five co-founders, are offering shares; the company gets no money from the latter, and the amounts are blank¹¹⁴. A concurrent private placement of up to $10 billion to Google at the IPO price is an option, and Amazon, Google, Broadcom and AMD have standby financing tied to post-IPO share prices⁸. A roughly $1.5 trillion IPO valuation could not be verified.

What to watch

• An amended S-1 with the price range, share count and use of proceeds
• Fourth-quarter 2026 revenue, compute cost as a share of revenue and the adjusted margin after the third-quarter drop
• The D.C. Circuit decision on rehearing in the Department of War case
• How much of the Google private placement and the Amazon note conversion is taken up at the IPO price
• The lock-up expiry schedule once share amounts are filled in
• The size of the catch-up stock-compensation charge when the registration takes effect
• Any change to the 5% to 7% of revenue the company attributes to policy-violating customers

XIIIWhat could go wrong

The filing lists the risks it sees. These are the ones that matter most here, in the order worth asking about.

What most likely destroys the business. A mismatch between fixed commitments and revenue. The filing puts total committed spend at about $895.4 billion, of which about $682 billion has no termination right or condition (my arithmetic), running as far as 2045²⁹. Revenue only has to stop growing for a couple of years, or the price of a unit of model output has to fall faster than volume grows, for fixed costs to take the margin from 15% (the adjusted figure in the third quarter) to nothing⁸. The third-quarter margin drop shows how fast that can start. The partners that finance the build-out are also its main suppliers. This is a mechanism, not a prediction. The filing says the revenue-growth conditions attached to the Broadcom note facility are well below the growth achieved so far⁸.

What an owner can actually take out. Public shareholders would own Class A shares with no dividend, a vote capped at 49.9% of the total that cannot change who runs the company, a board majority chosen by a trust, and a claim on future earnings diluted by notes, option pools, public-benefit issuances and partner commitments¹³⁸. The only ways to realise value are a rise in the share price or a sale of the company, and a sale needs approval of a board majority elected by the trust and, under the filing’s vote rules, the Class F holder¹³. Many paths have to be open at once.

What has to go right together. These are my judgement calls, not company statements, but they show the logic of multiplying. (a) The company stays at or near the frontier of model quality against OpenAI, Google and others for the next four years: about 55%. (b) The chips, power and data centres it has signed for arrive on time and on budget, and its partners keep financing them: about 80%. (c) Customers keep paying prices that leave net margins of 20% or more rather than competing them away: about 60%. (d) Government, copyright and export-control disputes do not materially cut off any big revenue source: about 85%. Multiplied, 0.55 times 0.80 times 0.60 times 0.85 is about 22% (my arithmetic). They show how much has to go right; they are not a measurement. Even at 22% the upside case is valuable, which is why a high price is not irrational.

The next best alternative for the same dollar. Large profitable companies that own competing models trade at far lower multiples of revenue today. On data pages opened on 2 and 3 November 2026, Alphabet traded at about 7.0 times sales and 14.6 times trailing earnings, Microsoft at 11.1 times sales and 26.5 times earnings, and NVIDIA at 16.0 times sales and 25.6 times earnings²⁹³¹³². Anthropic’s $965 billion private price was about 14 times its third-quarter annualised revenue of $68 billion (my arithmetic on⁶). The comparison is not like-for-like: those companies are far more profitable and diversified, and Anthropic is growing much faster.

Other risks the filing calls out: dependence on a small group of researchers; harm from model misuse and the liability that follows; regulation that could slow releases; start-up customers whose spending depends on their own fundraising; and a need to raise capital sooner than planned²⁴¹⁰. The company says its cash and facility should cover the next 12 months⁸.

What would change my view. Better: a price range that leaves room; fourth-quarter revenue and margins that show the third-quarter squeeze was temporary; commitments that fall or become cancellable; a favourable full-court D.C. Circuit ruling. Worse: revenue growth well below the 47% quarter-on-quarter of the third quarter without a margin gain; a price-cut cycle from a rival; compute delivery delays; another government action; or new financing on harsher terms than the Broadcom notes.