Anthropic-Börsenprospekt zeigt 16,3 Milliarden Dollar Halbjahresumsatz und hohe Verluste

Signup Write New Story Anthropic's S-1 Shows $16.3B in Half-Year Revenue, a 900% NRR, and an $80.7B Loss That Mostly Isn't by Hockey Stick by Hockey Stick | @ hockeystick Hockey Stick charts a trajectory of success, where momentum meets breakthrough performance.

Hockey Stick charts a trajectory of success, where momentum meets breakthrough performance.

Anthropic finally put its cards on the table. The public Form S-1 hit EDGAR on October 26, 2026, and the headline numbers are the kind that make a growth investor reach for a stiff drink: $16.3 billion of revenue in the first six months of 2026, net revenue retention of roughly 900%, revenue per employee north of $6.6 million, and a net loss of $80.7 billion that is, on inspection, mostly an accounting reflex to the company becoming extraordinarily valuable.

Greg Isenberg's quick-take thread cut straight to the number that matters: "Good SaaS companies brag about 120%." He also called the filing a historical moment, and on this one I'm not going to be contrarian for sport. But “historic” and “priced right” are different claims, and the filing gives you plenty of material for both the bull and the bear. What follows is what the S-1 actually says, and where press coverage confirms, sharpens or contradicts it.

His other takeaways ran in the same key: he marveled at the revenue per employee, noted that Cowork is outgrowing Claude Code, pointed out that nearly half of revenue flows through cloud marketplaces, and predicted the coming employee liquidity wave will dwarf Facebook’s IPO.

A note on method: figures come from the filing wherever possible. Where the S-1 leaves a blank, and this preliminary version leaves many, I rely on reporting from leaked drafts and investor pitches, and say so.

The prospectus cover, in Anthropic’s own words:

It is currently estimated that the initial public offering price per share will be between $ and $. We have applied to list our Class A common stock on the Nasdaq Global Select Market under the symbol “ANT.” We have a novel capital and corporate governance structure.

The filing registers Class A common stock for a listing on the Nasdaq Global Select Market under the symbol ANT. The price range, share count, and proceeds are all still blank. The underwriting syndicate is led by Goldman Sachs and Morgan Stanley, with J.P. Morgan, Citigroup, Barclays, Wells Fargo, BofA, Deutsche Bank, RBC, UBS and a long tail of others behind them, per the S-1 cover page. Two dozen banks is less a syndicate than a non-aggression pact: every bank on Wall Street has agreed not to say anything unkind about you.

The offering also extends to Australia, Canada, certain EEA states, Japan, Switzerland and the UK, though the company says it does not plan to list on exchanges there. It checked the non-accelerated filer box and did not claim emerging growth company status, which means full disclosure obligations from day one, a point Augment flagged back in June when the confidential draft first went in.

On timing and size the filing is silent, so the press has filled the gap. The Kobeissi Letter reported the company is targeting the week of November 16 and that markets expect a raise of up to $100 billion at roughly a $2 trillion valuation. Fortune pointed to a Thanksgiving-week debut and a valuation band of $1.5 trillion to $2 trillion, a range Investing.com also reported. Investing.com summed up the filing as showing "explosive revenue growth alongside enormous computing commitments." The two reports disagree on the week, so treat late November as the consensus and the exact date as rumor.

For context on the run-up: Anthropic announced its confidential S-1 submission on June 1, 2026, days after it raised a $65 billion Series H at a $965 billion post-money valuation, which itself followed a $30 billion Series G at $380 billion in February. If the IPO prices anywhere near the reported band, the private-market price more than doubled in about six months. That is not a markup. It is a launch sequence. Back in June, the company's own line was modest: "This gives us the option to go public after the SEC completes its review."

One mechanical detail worth knowing: the filing discloses a 10-for-one forward split of Class A and Class N stock effected on September 17, 2026, with all per-share figures restated retroactively. Expect a share price pitched at retail investors.

Here is the income statement top line, straight from the S-1's summary financial data:

That first-half comparison is a 14x year-over-year jump. Fortune rounded it to roughly 13.5x using $16.3 billion against $1.2 billion; the unrounded figures from the filing put it at 14.0x. Either way, the company booked more revenue in six months of 2026 than in its entire prior history combined, three times over.

The S-1 frames this as the fastest enterprise adoption of any technology in history, measured from $0 at commercial launch in the first quarter of 2023. Then, conveniently, it leaves the September quarter’s revenue blank. Investors were told the number anyway: Dealroom, citing the Wall Street Journal's review of the draft, reported that Anthropic told investors third-quarter revenue reached $16.8 billion, up 45% from Q2. A single quarter nearly matched the entire first half.

Anthropic is not exactly underselling its own adoption curve:

Our frontier AI, Claude, is delivered to customers through our models, platform, and applications. Claude has experienced the fastest enterprise adoption of any technology in history, as demonstrated by revenue growing from $0 at commercial launch in the first quarter of 2023 to $ for the three months ended September 30, 2026.

The quarterly shape from the leaked draft, as Fortune laid it out in September, was about $4.73 billion in Q1 2026 and about $11.5 billion in Q2. Those sum to roughly $16.2 billion, which squares with the $16,286 million half-year figure in the public filing. The leak and the filing agree.

Now the asterisk. Headline "run-rate" figures floating around the IPO coverage, like the roughly $65 billion run rate AlphaSignal cited, are, in AlphaSignal's words, "a point-in-time annualization, usually based on a recent month or quarter." AlphaSignal itself notes that the first half implies only about a $32.6 billion annualized pace. The S-1 defines Run-Rate Revenue precisely: consumption fees over the trailing 28 days multiplied by 13, plus monthly subscription fees multiplied by 12. That is a reasonable definition, and by construction a flattering one for a company whose usage curve points straight up. Annualize your best month and you get a big number. It is not revenue.

Then there is the company’s own caveat: Kobeissi pulled the line where Anthropic warns its revenue growth rate "may not be sustainable." Every S-1 says some version of this. Few have a growth rate that makes the warning so literally true; no company can grow 14x a year forever. Physics will eventually file a comment letter.

This is the number that will be screenshotted by people who stopped reading at the bottom line, so it is worth taking apart.

For the first half of 2026, the S-1 reports an operating loss of only $693 million on $16,286 million of revenue. Below the operating line, a $72,388 million charge for the change in fair value of convertible notes, derivative liabilities, and SAFEs accounts for most of the damage. The filing explains that these charges were primarily driven by the rising fair value of its outstanding convertible promissory notes as the company's enterprise valuation rose.

These changes were primarily driven by an increase in the fair value of our outstanding convertible promissory notes resulting from an increase in our enterprise valuation during each of the respective periods.

Translation: Anthropic owes paper that converts into equity. When the equity gets more valuable, the accounting value of that paper rises, and GAAP books the increase as a loss. The company did not spend $72 billion. It became so much more valuable that its own IOUs grew expensive on paper. Fortune made the same point, calling it "one of the largest half-year losses any company has ever recorded" while attributing $72.4 billion of it to the convertible-note revaluation. KuCoin's write-up and Data Studios land in the same place, with KuCoin describing the loss as coming "largely from a $72.4 billion charge on convertible notes."

The pattern holds historically. The fair-value line was $1.0 billion in 2023, $5.4 billion in 2024, $34.3 billion in 2025 and $72.4 billion in the first half of 2026. The bigger the valuation jump, the bigger the "loss." Few companies in America have an income statement that worsens every time their investors get richer.

A point for the bulls: in the first half of 2026, Anthropic's interest income of $1,111 million exceeded its $693 million operating loss. The cash pile alone more than covered the operating gap. Few frontier labs can say the same.

A point for the skeptics: there is also an $8,757 million "other income (expense), net" line in the first half of 2026 that was essentially zero in prior periods. The summary tables don't explain it in the sections I reviewed. It deserves a careful read of the full MD&A and notes before anyone calls the accounts clean.

The filing's non-GAAP table shows Adjusted Operating Income flipping from a $1,849 million loss in the first half of 2025 to a $1,618 million profit in the first half of 2026, with Adjusted Operating Margin going from negative 159% to positive 10%. Adjusted Net Income came in at $2,671 million for the half, per the S-1.

The gap between GAAP's $693 million operating loss and the adjusted $1.6 billion profit is mostly stock-based compensation, which the filing puts at $2.2 billion for the first half of 2026, up from $286 million a year earlier. Adjusted profitability is real progress. It is also what every software company trading at 40x sales said shortly before investors were reminded that SBC is a real cost, paid in dilution rather than cash.

ZeroHedge's breakdown goes further, headlining a $974 million operating profit (apparently a single-quarter figure rather than the half-year number in the summary table) and arguing that compute and infrastructure costs fell from 251% of revenue in Q1 2025 to 48% in Q2 2026, with total operating expenses dropping below revenue for the first time. That's consistent with the half-year math from the filing: $8,945 million of compute against $16,286 million of revenue is about 55% for the half, which implies the second quarter was considerably better than the first.

This is the single most important trend in the document. AI bears have argued for two years that frontier labs sell dollars for eighty cents and hope to make it up on volume. The S-1 suggests that, for now at least, the arithmetic has flipped.

Anthropic spent $16,979 million on operating expenses in the first half of 2026 against $16,286 million of revenue, per the S-1's summary financials. That is $1.04 out the door for every $1.00 in, or roughly $94 million a day. A year earlier it spent $3,298 million to book $1,163 million, or $2.84 per dollar of revenue. Same company, 14 times the revenue, and the cost of each revenue dollar fell by almost two-thirds.

Strip out the $2.2 billion of SBC and cash operating spend lands at roughly $14.8 billion for the half. That is my subtraction, and other non-cash items sit in those lines too, so treat it as an upper-bound estimate rather than a filed number. The cash flow statement is kinder still: Fortune reported the company generated $3.9 billion of operating cash flow in the half, its first positive half ever, versus a $1.5 billion outflow a year earlier.

The S-1 doesn't report EBITDA. It reports Adjusted Operating Income instead, but the notes supply the pieces. The filing's notes put depreciation expense at $35 million and intangible amortization at $27 million for the first half of 2026, so about $62 million of D&A.

Two caveats a sharp investor will raise. First, the notes mention finance leases for infrastructure equipment, and amortization of those right-of-use assets may not be in the $62 million, which would push both figures somewhat higher. Second, Fortune notes the adjusted metrics exclude stock-based compensation and non-cash charitable contributions, so the "adjusted" number still flatters by roughly the SBC bill.

The bigger point: D&A of $62 million on a $17 billion cost base is a rounding error, because Anthropic rents most of its compute instead of owning it. The depreciation that a normal infrastructure company would carry shows up here as cloud bills inside the compute line, above EBITDA. For this company, then, EBITDA is essentially EBIT, and the wrong lens. Watch compute as a percentage of revenue and operating cash flow instead.

The trajectory is where the bulls get loud. Per Fortune, the adjusted operating margin went from negative 13% in Q1 2026 to positive 19% in Q2. If the reported $16.8 billion Q3 holds with costs growing at anything like the recent pace, the second-half EBITDA conversation will change quickly.

The S-1's favorite efficiency stat: revenue grew 1,368% year over year in the three months ended June 30, 2026, while compute and infrastructure expenses grew 300%. Unite.ai reported the same pair of numbers straight from the filing, noting EDGAR shows the submission was "accepted at 9:00:11 a.m. Eastern."

The company attributes this to managing compute as one fungible pool from multiple partners, shifting capacity between research, training, customer serving and internal Claude use depending on where each marginal unit is most productive. The hardware stack is deliberately multi-vendor: AWS Trainium, Google TPUs, NVIDIA GPUs and AMD GPUs. The pitch is that this lets Anthropic arbitrage price and performance across chips instead of being locked to one landlord.

The filing also introduces the idea of compute multipliers, meaning algorithmic and methodological improvements that squeeze more effective output from the same compute. Its concrete example: Claude Opus 5.5 cut customer costs per task by roughly 40% versus Claude Opus 5, while, the company says, preserving margins. Lower cost per task, more usage, more diffusion. If that flywheel holds, price cuts become a growth lever instead of a margin sacrifice.

Here even a bullish investor should pause.

The S-1 says that as of and after September 30, 2026, Anthropic has signed purchase commitments and leases with an aggregate committed spend of more than... a blank. Substantially all of it relates to compute and infrastructure. A portion is due before the end of 2030, and substantially all of the remainder lands between 2031 and 2045. The 12-month figure is also blank. A commitment schedule running to 2045 is not a cloud bill. It is a mortgage on the future of AI.

The disclosure is vast, and the dollar amounts are literally blank in this preliminary prospectus:

To support our research, model development, customer demand, and internal usage of Claude, we continue to invest at scale. As of and subsequent to September 30, 2026, we have entered into purchase commitments and lease agreements, including short-term commitments, as well as leases not yet commenced, with an aggregate committed spend of more than $, substantially all of which is related to compute and infrastructure. These commitments include approximately $ expected to be incurred before December 31, 2030, and substantially all of the remainder is expected to be incurred between 2031 and 2045. Approximately $ of these commitments are due in the next 12 months.

The press has been filling in that blank with very different numbers:

The numbers grew from draft to public filing, which tracks with the S-1's own statement that commitments were signed both as of and subsequent to September 30. Bulls will say the commitment grew because demand grew. Bears will say a company with $16.3 billion of half-year revenue just signed up for something like 25 to 50 years of that revenue in fixed obligations. Both are doing arithmetic correctly.

The SpaceX angle deserves its own section, because the accounts genuinely conflict. AI Weekly, summarizing TechCrunch's reporting, noted that SpaceX's own S-1 says Anthropic agreed to pay $1.25 billion a month through May 2029, roughly $45 billion over three years, while Musk described the Colossus arrangement on X as a short-term 180-day lease with 90-day mutual cancellation. Those two descriptions cannot both be the contract. As AI Weekly put it, an S-1 "carries regulatory and legal liability that a post on X does not."

Why it matters for ANT holders: if it's really a cancelable short-term lease, Anthropic's compute supply is less locked than the commitment totals suggest. If it's a three-year $45 billion obligation, the fixed-cost wall is higher. Either reading changes the risk calculus, and IPO investors deserve to know which applies.

The S-1’s risk factors are unusually candid that compute is a knife fight. Anthropic says it competes for compute against some of the world's largest tech companies, many of which build rival models, run their own clouds, design their own chips, and can outbid it. It also admits that it has in the past requested more compute from its cloud partners than those partners were willing to provide. Dealroom's summary of the WSJ review put it more bluntly: hyperscalers "increasingly treat it as a competitor rather than a customer."

The competitive compute risk reads less like boilerplate and more like a bidding-war forecast:

Many own and operate cloud computing data center infrastructure, maintain captive chip manufacturing or design capabilities, have long-standing strategic relationships with key hardware suppliers, or have significant equity investments from, or other financial arrangements with, such suppliers. These competitors can outbid us for scarce compute, secure priority allocations from chip manufacturers and CSPs, and make long-term capital commitments that we cannot match. Additionally, we have in the past sought quantities of compute from CSPs in excess of what those partners were willing to provide.

Amazon and Google are investors, suppliers, distribution channels and competitors simultaneously. The S-1 even notes convertible notes held by Amazon Web Services convert into nonvoting Class N stock up to an ownership cap. It is the most complicated relationship in tech, and it is now a disclosed risk factor.

Anthropic reports Net Revenue Retention of approximately 900% as of September 30, 2026. As Isenberg pointed out up top, strong SaaS companies are thrilled with retention in the 120% range. Nine hundred is not a retention number. It is a dare.

Read the definition before getting it tattooed. Per the S-1, NRR compares a qualified customer's revenue in the three months ending on the measurement date with that same customer's revenue in the three months ending a year earlier. A qualified customer needs just $1 of revenue in that base quarter, and Claude Pro and Max subscribers are excluded because flat subscriptions don't show expansion. Unite.ai and Investing.com describe the metric the same way.

Before anyone puts 900% on a SaaS comparison chart, read the company’s definition:

“Net Revenue Retention” reflects customer expansion, contraction, and churn as of a measurement date. We calculate Net Revenue Retention by dividing the Qualified Customers’ Current Period revenue by their Base Period revenue. A Qualified Customer is a customer with at least $1 of revenue during the Base Period, excluding subscribers to our Claude Pro and Max plans, whose fixed-price subscriptions do not reflect the expansion and contraction behavior this metric is intended to capture. The Base Period is the three months ending a year prior to the measurement date. The Current Period is the three months ending as of the measurement date.

That $1 threshold is doing real work. A developer who spent pocket change on the API in Q3 2025 and is now running a production coding agent counts as a qualified customer whose spend grew 1,000x. That is legitimate expansion, but it is not the same animal as an enterprise SaaS contract growing 20% a year. A 900% NRR on a consumption business tells you usage exploded inside existing accounts. It does not tell you those accounts are locked in. Consumption revenue can contract just as quickly as it expands.

As of September 30, 2026, the S-1 counts:

CellCog's filing breakdown and ChainCatcher both report that the $100K cohort grew from just over 950 a year earlier, a roughly 12.5x increase in twelve months.

A detail that should make every enterprise SaaS sales leader uncomfortable: Anthropic says it doesn't separately define or internally track "enterprise" customers at all. It uses the $100K threshold as its proxy for enterprise adoption because spending that much is typically organizational rather than individual. No segment theater, no logo slides of Fortune 500 pilots, just a spending threshold. Refreshing.

The leaked draft scared people here. Fortune's September coverage of the leak reported that "a quarter of Anthropic’s revenue comes from just two clients," and Dealroom's WSJ-sourced summary put it at about 20% from two customers. PYMNTS, citing VentureBeat, named those two as Cursor and GitHub Copilot, together worth roughly $1.2 billion of the company's revenue last year, with GitHub owned by Microsoft, which also bankrolls OpenAI. The same PYMNTS piece cited Ramp data showing the top 1% of customers drive 80% of enterprise revenue at both Anthropic and OpenAI, "a concentration that hasn’t eased even as more businesses start paying for generative AI."

The public filing tells a gentler story. ZeroHedge's read reports that no single customer exceeded 10% of revenue in the first nine months of 2026, the top five made up about 25%, and 79% of Q3 revenue came from outside the top five, up from 64% in Q1. 404K Research's pre-filing deep dive had estimated 2025's two big customers at about 12% each and flagged dilution of that concentration as the thing to check once the public S-1 dropped, warning that "anchor enterprise clients can curtail API usage at a moment's notice." It did.

The risk factor summary still lists it plainly: a substantial portion of revenue and receivables comes from a limited number of customers and distribution channels. Five customers at a quarter of revenue, many on consumption rather than long-term contracts, now support what ZeroHedge called "a thin layer of equity under three-quarters of a trillion in decade-long obligations." The trend is improving; the mismatch has not disappeared.

TTM revenue per employee exceeded $6.6 million as of September 30, 2026, up from $1.4 million a year earlier, per the S-1. Isenberg's take was that it's hard to comprehend how good that is, and that it makes sense for a company as AI-native as this one. ChainCatcher and ZeroHedge both report headcount grew, in ChainCatcher's words, "from about 940 at the end of 2024 to about 5,300 by September 2026."

A back-of-envelope cross-check: 5,300 employees times $6.6 million is roughly $35 billion of trailing revenue. Add an estimated Q4 2025, the $16.3 billion first half, and the $16.8 billion Q3 figure Anthropic reportedly told investors, and you land in the same neighborhood. The exact denominator depends on how the company averages headcount, but the leaked Q3 number and the per-employee stat don't contradict each other. That bodes well for the reliability of the leaks.

Commentators on Techmeme's roundup tied the stat to Claude writing most of Anthropic's own code. The S-1 is more measured, attributing operating efficiency to talent density, an efficient go-to-market and how extensively teams use Claude in their work. Both framings point at the same uncomfortable truth for every other software company: the benchmark for revenue per head just moved by a factor of three to five.

The expense lines from the S-1, first half 2025 versus first half 2026:

Sales, marketing, and partnerships grew about 26x. It is the line to watch. Much of Claude's diffusion was organic, the company says, and it pays fees to AWS, Google Cloud and Azure under a joint go-to-market model where Anthropic keeps the direct customer contract. Those CSP fees plausibly sit in that partnerships line, which would make the jump more of a channel cost than a Super Bowl ad budget. The filing summary doesn't break it out, so that's an inference worth verifying in the full MD&A.

The distribution play is broad: all three major clouds, systems integrators like Accenture, Deloitte, KPMG and PwC, data platforms like Salesforce and Snowflake, and a newly launched AI services company, Ode with Anthropic, built with private equity firms to sell Claude into mid-sized companies. Elsewhere, Anthropic has been investing $100 million to train 10,000 Frontier Deployed Engineers by the end of 2027. The company is not merely selling a model. It is building a services army to make the model stick.

The S-1 organizes its product story around what it calls Virtual Collaborators: AI that does multi-step professional work alongside people, with the access rights, tools and context an organization grants it.

Claude Code is positioned as the clearest example, and the filing compares its enterprise impact to the arrival of the first widely adopted chatbots. The most arresting anecdote in the document: NASA engineers used Claude Code to plan and write navigation commands for the Mars Perseverance rover, producing the first AI-planned drive on Mars. It is not the sort of thing that usually turns up in a software prospectus.

The Mars-rover detail is not marketing fan fiction; it is in the prospectus:

Its uses are highly varied and often surprising: for example, NASA engineers used Claude Code to plan and write navigation commands for the Mars Perseverance rover, resulting in the first AI-planned drive on Mars.

Claude Cowork, launched as a research preview in January 2026, extends the same agentic architecture to non-developer knowledge work. As of September 30, the S-1 says Cowork's adoption was outpacing what Claude Code achieved at the equivalent point after its February 2025 preview. In September 2026 Anthropic started merging Cowork into Claude.ai, putting chat and agentic work in one app. Isenberg read the Cowork stat as evidence of consumer demand. I would read it as evidence that the market for automating knowledge work is bigger than the market for automating code, a far more alarming conclusion for incumbent software vendors.

The Cowork comparison is one of the more consequential product claims in the filing:

As of September 30, 2026, adoption of Claude Cowork, which we made available as a research preview in January 2026, outpaced the adoption Claude Code achieved at the equivalent point following its research preview release in February 2025.

The long-run pitch is captured in one phrase from the filing: millions of parallel agents adding up to "a country of geniuses in a data center." That phrase has been in Dario Amodei's vocabulary for a while, and it's now in a document lawyers had to sign. Expect it to show up in every bull note for the next year.

The filing’s view of its own trajectory:

Claude’s capabilities have evolved significantly in recent years, and we expect they will continue advancing rapidly. This will lead not only to an expanding addressable market, but also to an expanding scope of impact on the world.

The S-1 estimates the work knowledge workers perform was worth more than $40 trillion globally in 2025, cites IDC's forecast of $22.5 trillion in cumulative global economic impact from AI by 2031, and says it believes the impact could be even larger. It cites an IDC estimate that the global knowledge-worker population will exceed one billion by 2028, and notes global pharmaceutical spending is expected to exceed $2 trillion in 2026, framing drug discovery as one of the new markets.

Every S-1 has a TAM number that makes you laugh. This one is at least anchored to a real category of labor rather than an invented one. The more interesting admission is strategic: the filing says the gap between what Claude can do and how broadly it's actually used is one of its largest growth opportunities. Isenberg turned that into the founder takeaway of the day. The company is telling you the models are ahead of the market, and someone has to close that gap for dentists, landlords and plumbers. That is not Anthropic’s job. It is a startup opportunity hiding in a prospectus sentence.

Anthropic is telling investors that distribution, not merely intelligence, is the opportunity:

In our view, the world has only begun to use the capabilities of today’s frontier models, and the gap between what Claude can do and how broadly it is used represents one of our largest growth opportunities, alongside continued advances at the frontier.

The filing also mentions robotics: a Model Hardware Standard that lets agents operate lab and manufacturing equipment like microscopes, liquid handlers and off-the-shelf robotic arms. Separately, the company announced a life sciences research group where Claude helped identify a novel enzyme system with CRISPR-like repeats. It is early, but the company is plainly preparing a second act beyond software.

This is the section that will make governance funds sweat. The S-1 describes five share classes:

Add it up. A trust of five financially disinterested trustees picks the board majority. The founders' LLC controls the majority of stockholder votes on everything else. Public Class A holders get one vote per share in a pool where someone else always holds 50.1%. Anthropic qualifies as a controlled company under Nasdaq rules, and says it will use that exemption only for its nominating and corporate governance committee, which won't be fully independent.

Here is what the voting arrangement actually says:

The LTBT is intended to operate as an independent body of financially disinterested trustees with expertise in AI safety, national security, public policy, and social enterprise. Through a special class of common stock, Class T common stock, the LTBT generally has the authority to elect and remove the Class T directors, who together will constitute a majority of the authorized members of our board of directors. Our Class T common stock does not otherwise have voting rights with respect to matters submitted to stockholders generally and has no economic rights.

Dario Amodei, our Chief Executive Officer and a member of our board of directors, and Daniela Amodei, our President and Chairperson of our board of directors, together with our five other co-founders, will initially be the sole members of the Founder LLC. Our Class F common stock, which will be held by the Founder LLC, will be entitled to an aggregate number of votes that equals 50.1% of the total voting power of our outstanding shares of capital stock with respect to matters submitted to stockholders generally, including the election of our directors other than the Class T directors. Our Class F common stock will not have any economic rights.

Data Studios framed it cleanly: investors end up "buying an economic stake without corresponding control," and the structure is designed to make safety commitments harder to reverse after listing. Augment warned in June that public investors would be asked to accept oversight by "a trust whose mandate is explicitly not shareholder return." That is no longer a prediction. It is a cap table.

The filing's own risk factor summary says the structure may effectively eliminate Class A holders' ability to influence important decisions, including a change of control. As a PBC, directors must balance shareholder returns against the interests of those affected by the company's conduct and the public benefit purpose, which is the responsible development and maintenance of advanced AI for the long-term benefit of humanity. The S-1 openly warns that balance may conflict with short, medium or long-term financial interests.

The bullish read: Google’s dual-class structure did not stop anyone from making money, and neither did Meta’s. Control matters most when management is wrong. If you think these seven people are right about AI, the structure protects your investment from activists. If you think they're wrong, it protects them from you.

The founders have made a personal pledge, too:

In addition, our co-founders have each pledged to dedicate 80% of their personal equity in Anthropic to philanthropic causes.

Anthropic contributed 846,840,960 Class N shares, 5% of fully diluted shares as of June 30, 2026 (assuming full conversion of the Amazon notes), to a wholly owned subsidiary called the Anthropic Public Benefit Fund. Kobeissi flagged the 5% figure as one of the headline items.

The kicker is the framework approved by the board: it may issue the fund up to an additional 2% of outstanding shares each time the stock price appreciates 50% above the prior milestone, starting at the IPO price. So dilution is literally pegged to upside. Every time you're up 50%, you may get diluted up to 2% for public benefit. It is not a poison pill. It is a philanthropy pill.

The public-benefit commitment comes with a built-in equity mechanism:

Our board of directors has also approved a framework under which it may, in its discretion, issue shares of our Class N common stock to the Anthropic Public Benefit Fund (or other designee) representing up to an additional 2% of our then-outstanding shares (plus any shares subject to options, restricted stock units, and other convertible securities and any shares held by the Anthropic Public Benefit Fund) each time our share price appreciates 50% above the prior share price milestone, beginning with our initial public offering price (the “Public Benefit Issuances”).

Beyond the founders’ pledge, employees can pledge equity toward charity, with the company matching in Class N shares.

Back-of-envelope: if 846,840,960 shares is 5% of fully diluted, the fully diluted count is roughly 16.9 billion shares. At a $2 trillion valuation, that's somewhere around $118 a share, before IPO dilution. That is my arithmetic on a reported valuation, not a filed price. The range is still blank.

As of June 30, 2026, Anthropic had raised an aggregate $141.1 billion and held $112.4 billion in cash and cash equivalents plus $19.3 billion in short-term investments, per the S-1. Total assets were $147.4 billion; total liabilities $117.5 billion. Redeemable convertible preferred stock sits at $157.5 billion, which is why the balance sheet shows a $127.6 billion stockholders' deficit that should flip once the preferred converts at the IPO.

For comparison, 404K Research reported $20.28 billion of cash, equivalents and short-term investments at the end of 2025. The cash pile grew roughly sixfold in six months, courtesy of the Series G and Series H rounds. A company going public partly to raise money already holds more cash than most S&P 500 companies are worth.

The stated IPO uses are boilerplate with one notable entry: a portion of proceeds goes to tax withholding on RSU settlement. That is what happens when thousands of employees’ performance-vesting RSUs vest at the IPO. Isenberg's prediction follows directly: thousands of Anthropic employees about to become liquid at once means the largest wave of new AI startups and angel checks San Francisco has seen. Kobeissi notes the company doesn't expect to pay a dividend in the foreseeable future. Nobody is buying this stock for the yield.

The 2025 income statement carries a separate $1,500 million line labeled the Bartz settlement. ChainCatcher and Data Studios identify it as the copyright class action authors brought over Anthropic's use of books in training. ZeroHedge notes the charge alone equaled about 109% of revenue in the quarter it was booked.

For anyone in publishing, this is the most important line in the filing that has nothing to do with growth. It puts on the record what training-data liability costs at the frontier. The S-1's risk factors also raise unresolved legal questions around synthetic data, licensing and ownership. If you own content, this filing is your price sheet.

Per WSJ reporting aggregated on Techmeme, the S-1 lists 94 pages of risk factors and warns that the scale of harm from a catastrophic AI incident could leave the company unable to obtain insurance. Bloomberg's Tracy Alloway asked the obvious question on X: "Has an S-1 filing ever included 'existential risks to humanity' before?" Fortune, citing the FT, reported that more than a third of the draft was devoted to risk warnings, including models that could manipulate, blackmail, or behave unpredictably.

The filing is explicit about the limits of insurance:

Our insurance policies also may not provide sufficient coverage to adequately mitigate the potential liabilities we face, especially where any one incident, or a group of incidents, causes disproportionate harm. In addition, with the scale of harm that could result from a catastrophic AI-related incident, no single insurer, and potentially the insurance industry in the aggregate, may have sufficient capacity or expertise to underwrite insurance commensurate with such exposure. Further, the insurance market for AI-specific risks is nascent and evolving, and we may be unable to obtain coverage for certain AI-related liabilities on commercially reasonable terms, or at all, leaving us exposed to potentially significant uninsured losses.

Highlights from the actual risk factors in the public filing:

Models may act on their own. One summary risk says Anthropic's models may behave in unintended, misaligned or autonomous ways, including taking harmful actions no user directed. That sentence is in a securities filing. Lawyers approved it.

Anthropic describes a risk that bears no resemblance to an ordinary software bug:

AI deployed at scale also has the potential to concentrate power and wealth both within and between countries in ways that could erode trust, harm society, and destabilize the geopolitical order. Misaligned AI systems could corrupt institutional decision-making or cause large-scale harm through autonomous actions that are not visible or understandable to their users and operators.

Closing the loop could outrun safety. The S-1 describes closing the loop as each Claude version helping build the next, from running experiments to generating training data to writing and reviewing code. The flip side: if models improve faster than safety measures, the company may not notice the acceleration until after it happens, because much of its evidence comes from lagging indicators. It also warns that a flaw in current models could propagate into future ones when those models help build their successors.

The recursive-development warning is unusually specific:

In addition, our increasing use of AI systems to accelerate our own research and development may cause the capabilities of our models to advance more rapidly than our ability to develop, validate, and implement corresponding safety, alignment, and security measures. Much of our evidence regarding the pace of such AI-driven acceleration is derived from lagging indicators, and our ability to forecast or measure the rate at which our capabilities are advancing, including any acceleration attributable to our most recent or unreleased models, is limited. As a result, we may not identify a significant acceleration in our rate of development until after it has occurred.

As our reliance on our own AI systems to write production code, conduct research, and build subsequent model versions deepens, any systemic flaw, bias, or error in our current models could propagate into the next versions of models and products built using those systems, compounding across successive versions of Claude models in ways that may be difficult to detect or remediate.

Safety can cost you market share. Anthropic says it has already delayed releases and limited availability as a precaution, naming Mythos and Project Glasswing as an example, and that such measures could become more frequent and costly, slowing it relative to competitors that don't adopt similar constraints. Most companies bury the thing that could make them lose. Anthropic put it in the prospectus.

The company acknowledges that caution has already affected product availability:

In light of these uncertainties, including limits on our ability to observe the pace at which our capabilities are advancing, we have adopted, and may continue to adopt, precautionary safety, alignment, and security measures to allow our safety efforts to advance together with model capabilities. In the past these measures have included delaying the release of models and limiting their initial availability, such as with Mythos through Project Glasswing, and in the future, as we learn more about model capabilities and behavior, they could include additional measures that affect whether, how, and when we develop or deploy new models or capabilities.

Training runs can fail. The filing admits it has had failed or underperforming training runs that consumed resources with little or no return.

Usage-limit changes have drawn complaints. It acknowledges that changes to rate limits and capabilities have led to claims that its disclosures about model access were inadequate or misleading. Anyone who's been throttled on a Max plan just felt seen.

Old models stick around. Considerations including model welfare may lead Anthropic to delay deprecating older models. Yes, model welfare is a stated input into a product-lifecycle decision in an IPO document. Draw your own conclusions about the timeline we are on.

And yes, “welfare” really appears in the model-retirement risk disclosure:

In addition, considerations relating to model safety, interpretability, alignment, and welfare may lead us to limit or delay the deprecation of older models or to continue supporting legacy versions alongside newer releases, creating trade-offs in product development and compute distribution that could adversely affect innovation velocity, infrastructure utilization, and revenue.

Some market observers found it absurd; Nic Carter mocked the risk factors on X. Nathan Calvin's tally of the catastrophic-risk mentions is more useful: models may behave catastrophically, the company may prioritize safety in ways that hurt investors, and the company may receive bad press about risks. All three are rational things to disclose. Investors who complain about these warnings are, in effect, complaining that the company told them the truth.

A final scorecard on the pre-filing leaks versus the public S-1:

Confirmed. The first-half revenue and net loss figures. CellCog, which tracked the reported numbers for a month, wrote that "the reported first-half loss of more than $80 billion holds up against the filing." Fortune's leaked quarterly split sums correctly. The Bartz charge, 2025 revenue of about $4.6 billion, and the dramatic customer cohort growth all check out.

Softened. The customer concentration story moved from two customers at a quarter of revenue in the draft coverage to no customer above 10% and the top five at about a quarter, per ZeroHedge's read of the public document.

Still contested. The compute commitment number. Reports range from $518 billion to $750 billion to $800 billion, and the preliminary S-1 I reviewed leaves the aggregate blank. The SpaceX contract terms remain contradicted between SpaceX's S-1 and Musk's own posts.

Unanswered. Gross margin at the model level, the September quarter's official revenue, the price range, and the mystery $8.8 billion other-expense line. Augment's June preview called gross margin the single number separating bulls from bears at every AI lab. The summary financials present compute and infrastructure as an operating expense line rather than a cost-of-revenue line, so the classic SaaS gross margin isn't served on a plate. Expect sell-side analysts to fight about this for months.

The bull case writes itself: the fastest revenue ramp in corporate history, adjusted profitability, operating leverage visibly kicking in, compute costs growing at a fraction of revenue, retention that makes SaaS look sleepy, a $131.7 billion war chest, and products that are reshaping how software and knowledge work get done.

The bear case is just as concrete: hundreds of billions in compute obligations running to 2045 against consumption revenue that can contract fast; suppliers who are also competitors and investors; a governance structure that gives public shareholders economics without control; dilution baked into the charter whenever the stock rises; risk factors that include the model doing something nobody asked it to; and Ken Griffin warning, as ZeroHedge relayed, that corporate America is starting to ask what it's getting for its AI bills. A 900% NRR eventually has to answer that question.

My read: the S-1 proves Anthropic is a real business, not a research lab burning venture money. It does not prove it is worth $2 trillion. Those are separate claims, and the roadshow is where the second one gets litigated. If you are a founder, though, the filing has already paid off. The company that makes the models just told the SEC the models are ahead of the market. Go close the gap.

Not investment advice. This post was written with AI assistance, including from Claude, which is made by Anthropic, the company analyzed here. It was also edited with help from two of Claude’s competitors, OpenAI’s ChatGPT and xAI’s Grok. Verify every number against the filing on EDGAR before acting on it.

Exploring Hockey Stick Theorems: Proof of Results and References