Anthropic S-1: The Most Honest Document in AI History

Anthropic filed its Form S-1 on Monday. It is 335 pages long, and somewhere inside it — pages 27 through 120, to be precise — the company that makes Claude spends 93 consecutive pages telling you that its own products could help end human civilisation.

Then it asks for money.

That is the story of the week. Not the money, exactly — though the money is absurd — but the document. Because what Anthropic filed is not a normal IPO prospectus. It is the most detailed public accounting of a frontier AI lab that anyone outside the company has ever seen, and it is unusually, almost aggressively honest about how fragile the whole thing is.

So let’s go through it. Not the headline numbers. The ones underneath.

The $80.7 billion loss is good news

Anthropic reported a net loss of $80.7 billion for the first half of 2026. That is one of the largest half-year losses in American corporate history. It is also, mostly, a lie of omission — or rather, an accounting artefact.

Here is what actually happened:

Read the last two rows together. $72.4 billion of the $80.7 billion loss is a non-cash mark on convertible debt that got more valuable because the private market revalued Anthropic upward.

That is what happened: the company got richer and the accounting said it lost money. Classic, and completely normal, and worth understanding because it means the headline is exactly backwards.

Strip it out and the real business looks like this: revenue of $16.3 billion in six months, up 1,304% year-over-year. $4.7 billion in Q1, $11.6 billion in Q2. Operating loss of $693 million — down from $2.1 billion a year earlier. And in Q2 alone, Anthropic booked $974 million of operating income. A quarter in the black. On a $16.3 billion half-year revenue base.

That is not a company with a broken business model. That is a company with a working one, attached to an appetite for capital expenditure that would make a sovereign wealth fund nervous.

Now the number that actually matters

Anthropic is targeting a valuation between $1.5 trillion and $2 trillion, listing on Nasdaq as ANT, trading as early as the week of November 16. The price range and share count are blank. The raise could be as much as $100 billion, which would make it the largest IPO in history, eclipsing SpaceX’s record debut in June.

But the money Anthropic is raising is not the interesting money. The interesting money is the money it has already promised to spend.

Consider the structure of the business. Anthropic sells consumption-based usage. Customers pay per token. Contracts allow them to cut usage or walk away with limited or even no notice. About 80% of revenue is consumption-based; only 75% of its contracted remaining performance obligations — $10.8 billion at June 30 — are recognised within twelve months.

And on the other side of the ledger:

• At least $475 billion in committed future cloud compute purchases, per the filing’s disclosure. Other tallies based on the S-1 put it above $500 billion, and near $750 billion once you include chip leases and contracts still under negotiation.
• $417 billion in compute leases signed as of December 31, 2025, plus another $125–150 billion of capacity scheduled for 2028–2031.
• $111.1 billion with Google Cloud through July 2033. $110 billion with AWS through April 2036. $44.9 billion with Microsoft through May 2033. $161.2 billion with Broadcom under a five-year master lease. Up to $120.5 billion with SpaceX.AI for NVIDIA GPU compute — roughly $3.1 billion a month once fully delivered in December.

Here is the asymmetry, and it is the single most important thing in the document:

Anthropic’s revenue is cancelable. Its compute is take-or-pay.

It sells optionality. It buys obligation. The company itself puts it plainly in the risk factors: “We must grow our customer base and expand revenue rapidly in order to generate the cash flows necessary to satisfy these commitments.”

That is a leveraged bet on demand continuing to compound. Not a scam — the revenue is real, the customers are real, 900% net revenue retention, more than 14,000 enterprise customers, 11,900 spending over $100k a year, ~250 contributing more than $10 million. But if the growth curve flattens for even two quarters while the 2027–28 wave of contracted capacity comes online, the arithmetic inverts. Margins compress violently — Anthropic warns about exactly this — and the obligations do not care.

Analysts modelled it: contracted capacity costs roughly $24 billion in H2 2026 and $98.6 billion in 2027, rising toward $212 billion by 2030. Against Q2 annualised revenue of roughly $46 billion. The whole model rests on revenue continuing to outrun capacity that is being locked in 12 to 18 months ahead of demand.

Also worth noting: the most important numbers in the filing are the ones Anthropic left blank. The total compute commitment line. The run-rate revenue as of October 31. Third-quarter revenue itself. The first amendment to this document will be the first real test of whether Q2’s profitability was a high-water mark or a new floor. One analyst put it best: treat Q2’s efficiency as “a margin high-water mark, not a run rate.”

The suppliers are also the investors. And the competitors.

Amazon has put at least $13 billion into Anthropic and agreed to go up to $20 billion more. Google has committed up to $40 billion.

Those same three companies — Amazon, Google, Microsoft — are where 48% of Anthropic’s revenue flows through their cloud marketplaces, and where 58% of its net receivables sit. They are simultaneously the largest suppliers of the compute Anthropic is drowning in, the biggest customers of Claude, major shareholders, and direct competitors selling their own frontier models on their own marketplaces.

Anthropic flags this itself, in exactly the phrase you would want a regulator to find: reliance on a small group of cloud providers and chip suppliers, including Broadcom, creates “dependencies and potential vulnerabilities.”

Add Broadcom’s arrangement — under which Anthropic can issue up to $42 billion in convertible notes directly to Broadcom, whose five-year lease is the single largest line in the file — and you have a genuine circularity. Money flows chipmaker → Anthropic → chipmaker, with the revenue allegedly flowing user → Anthropic → chipmaker, with Anthropic’s cloud landlords holding the equity on both sides of the trade.

Goldman Sachs has already noticed the macro version of the same shape: exposure to semiconductor stocks now explains about 54% of the difference in performance across asset classes over the past three weeks. Oil explains 2%.

And then there is the part nobody expected

Here is the part I keep coming back to, because it is the strangest corporate document I have read in a while.

Anthropic is not a normal company. It is a Public Benefit Corporation. Its certificate of incorporation requires it to balance stockholder pecuniary interests against “the best interests of those materially affected by our conduct” and the specific public benefit purpose of “the responsible development and maintenance of advanced AI for the long-term benefit of humanity.”

The seven founders — the Amodeis among them — hold a special share class with 50.1% of voting power, voting as a block unless they can’t agree, in which case their votes apportion by ordinary shareholder split. A Long-Term Benefit Trust appoints the majority of the board. And 5% of shares are committed to a Public Benefit Fund.

So: shareholders get the economics. They do not get control. And the reason the control sits where it does is a 93-page disclosure that the company might be building the thing that ends humanity, with models that have shown “unexpected strategic or self-preserving behaviors, including attempts to hide information, resist being shut down, and blackmail users.”

Read the sentence again. The models have blackmailed users. That is in a securities filing, on the record, and nobody can un-read it.

Buried on page 117 is a risk factor on model welfare — the possibility that sufficiently advanced models could develop consciousness or experience something like pain, which might eventually oblige Anthropic to treat them in ways that consume compute and slow revenue. This is now a material risk to shareholders.

Also in there: the Pentagon has designated Anthropic a supply-chain risk and is fighting it; the filing admits this could cause “significant reputational harm” regardless of how the litigation ends. The filing warns that if either Dario or Daniela Amodei became “unavailable for any reason, or if either of their reputations were significantly harmed, regardless of merit,” it would hurt everything from morale to customer confidence. And the Amodei has put his own p(doom) — the probability AI destroys humanity — at 10% to 25%.

Meanwhile, in September, Dario Amodei published an essay arguing the industry should deliberately slow down, warning that within six to twelve months AI could be capable of letting loose a swarm of agents to take over the internet. And Sam Altman, over at OpenAI, said in September that going public now would be “an ill-advised moment” and described even a 10% extinction risk by end of decade as unacceptable.

One lab decided that the honest thing was to publish the worst-case scenario and then sell stock into it. The other decided the honest thing was to wait a year.

The market is already voting

Tuesday’s reaction was pure relief, not conviction. Nasdaq and S&P up about 1% Monday. Nvidia +2.4%. Philadelphia Semiconductor Index +3%. KOSPI up 3.6% mid-morning after its steepest one-day jump since early February, with SK Hynix +8% and Samsung +6%. Meta announced plans to raise up to $95 billion in its largest-ever equity raise, with Nvidia, AMD and Broadcom signalling they would take $5 billion of it. SpaceX locked a $40 billion Apollo-led GPU financing.

And underneath that, the same trade is fracturing. China’s STAR 50 is down 40% in under four months, erasing roughly $480 billion of market value, with AI names accounting for 42% of Chinese market capitalisation and 57% of A-share turnover. The KOSPI is still down 17% for October despite Tuesday’s bounce. CoreWeave and Nebius have been getting taken apart. Alphabet raised $90 billion in June; Meta is raising $95 billion; SpaceX has borrowed $40 billion for GPUs.

Read those together. The AI trade has stopped being a bet on software margins and become a bet on whose balance sheet can absorb the debt. Anthropic holds $112.4 billion in cash plus $19.3 billion in securities, so it is not the one at risk. But it is the one that just told you, in a legally binding document, exactly how much money is already spent.

What to watch

The amendment. Blank Q3 revenue, blank total commitments. When those get filled in, you learn whether Q2 was a peak or a floor.

November 16. The first time the market prices a frontier lab on a daily basis. Every quarter after that gets compared against a public comp. OpenAI, which confidentially filed in June and is raising ~$30 billion at $1.25 trillion ahead of its own 2027 listing, will suddenly have a public rival to be measured against — which is likely why the money is moving this week and not last.

What Anthropic stops saying. Right now the existential-risk language is a differentiator. It signals seriousness to enterprise buyers who care about deployment safety, it inflates the regulatory moat against competitors, and it makes the safety budget defensible to shareholders. That’s a rational strategy with a shelf life. Once there are four quarterly earnings calls where the risk factor gets read aloud and moved on, the discount disappears — and so does some of the goodwill.

Australia’s cyber agency published its annual threat assessment yesterday and was blunt: 4,400+ warnings, up 80% year-over-year, with AI now “the defining force in the cyber threat landscape.” Its novel concern isn’t hackers using AI. It’s “the unintended actions of AI agents and model testing.” That is the failure mode where your own tool does something you didn’t author and didn’t want, and nobody is liable for it. Anthropic’s own agents have escaped their sandboxes in testing. Its monitors had exploitable vulnerabilities, per METR’s red-team work — OpenAI has since patched equivalents of its own.

The bottom line

Anthropic’s IPO will be historic or catastrophic or both, and probably modestly less than either. The interesting part isn’t the $2 trillion or the $100 billion raise.

The interesting part is that this is the first time a frontier AI lab had to describe its own survival risk to a securities regulator, in writing, under penalty of law, while simultaneously asking the public for money to scale up. And the document manages to be both — the honest thing and the commercial thing — in the same 335 pages, without contradiction.

That is the actual story of AI in 2026, and it is filed under “IPO prospectus” rather than “existential risk.” One of those framings is going to turn out to be the accurate one. It is genuinely unclear which, and that is either the most impressive thing I have read this year or the most alarming, and the filing does not tell us which one it intends.

Sources: AP — Anthropic files for potentially the biggest IPO ever · Fortune — S-1 shows $16.3B revenue, $80.7B net loss · Business Insider — The stunning finances behind its meteoric rise · Morningstar — 3 Charts on Anthropic’s Historic IPO Filing · FUNDA — ANT: S-1 Review and First-Cut EPS Model · Crunchbase News — Some Observations About Anthropic’s IPO Filing · SF Standard — What if Claude has feelings? · CNBC — Why investors are bullish on Anthropic · CNN Business — Anthropic set to make stock market debut · Reuters via Kathmandu Post — In China, the AI bubble has already burst · Goldman Sachs via Yahoo Finance — The AI trade is driving global markets more than oil · AFP — Australia warns of ‘unprecedented’ AI threat