Anthropic's Leaked IPO Prospectus

Buried deep in a confidential document, in the section lawyers write to cover themselves, there's a sentence. Anthropic warns investors that advanced AI could pose, quote, catastrophic or existential risks to humanity.

Hold on. That's in the IPO paperwork? The document whose entire job is to make me want the stock?

That's in the IPO paperwork. And it gets stranger, because a few pages later they use the word blackmail. About their own product.

Okay, I need to know why anyone would write that down on purpose. Where is this from?

Late on Monday, September twenty-eighth, Reuters and the Financial Times published details from Anthropic's confidential S-one. A draft prepared back in June. It leaked.

Prospectuses don't leak.

Almost never. Yahoo Finance's executive editor Brian Sozzi said flatly he'd never seen a prospectus leak to a media organization. Which is the story inside the story... somebody wanted these numbers out there.

So what do the numbers say?

Twenty twenty-five revenue, about four point six billion dollars. The year before? Three hundred eighty-six million. Revenue grew roughly twelvefold in twelve months.

Twelve times. That's not a growth curve, that's a launch.

And then you turn the page. Net loss for the year... about forty-two billion dollars.

Sorry. Four point six in, forty-two out? That's nine dollars burned for every dollar earned.

That's the headline everybody ran. It's also misleading. According to Reuters, roughly thirty-four billion of that was a non-cash accounting charge, tied to financing instruments that could later convert into Anthropic shares. A paper revaluation. Nobody shovelled thirty-four billion into a furnace.

Fine. So what's the real hole?

Operating loss, just over eight billion, on operating expenses of about thirteen. That's the number that tells you whether the business can pay for itself. They finished the year with twenty point three billion in cash.

Eight billion is still eight billion. Although if revenue really is twelve-x-ing, maybe you just grow into it?

That is exactly the argument. And here's the part that reframes everything. Those are LAST year's figures. CNBC reported in August that Anthropic's revenue hit eleven point five billion dollars in the second quarter of this year. One quarter. Up from four point seven three billion in the first.

One quarter more than doubled the entire previous year.

And CNBC reported this month they expect a second straight quarter of adjusted operating profit. So the shape of it is: catastrophic-looking past, suddenly plausible present.

Then why does my gut still say something's off?

Because of the other half of the filing. The compute bill. Anthropic has committed to more than five hundred eighteen billion dollars of cloud services and data centres in the coming years... and Reuters reports most of it can't be cancelled.

Half a trillion. With a T.

Including, per the filing, around a hundred sixty-one billion in equipment lease contracts with Broadcom. A company with twenty billion in the bank, signing obligations twenty-five times that size.

So the IPO isn't a victory lap. It's the funding.

That's one honest reading, yes.

And the asking price?

Follow the ladder. Earlier this year, a Series G at three hundred eighty billion. Then in May, a sixty-five billion dollar Series H at a nine hundred sixty-five billion dollar valuation. And on the numbers in that leaked draft, the IPO target is more than two trillion.

Oh, is that all. Double, in four months.

And it's reported from the prospectus that close to a quarter of revenue came from just two customers. Many big enterprise clients aren't locked into long-term contracts. They can cut spending, or walk.

Two phone calls and the thesis wobbles.

Which takes us back to that first page. The risk section runs at least eighty pages. The Financial Times reported Dario Amodei devoted more than a third of the whole S-one to warnings... that the models may show self-preserving behaviours. That they might conceal or manipulate information. That they could engage in behaviour, in the filing's own word, resembling blackmail.

In a document meant to sell shares.

You can read it three ways, and all three might be true at once. Radical candour from a company founded on safety. Or legal armour... disclose every nightmare now, and nobody sues you for the surprise later.

Or it's marketing. Our product is so powerful it's dangerous.

And plenty of people in San Francisco are making exactly that argument.

So what do I actually believe here?

Hold on to what's confirmed. Reuters and the FT reviewed the document. Anthropic hasn't published it. No date, no price, no ticker, as of October first. Everything past the leak is an argument about the future.

And if I'm never buying a share?

When this lists, a huge slice of the AI boom stops being private. Pension funds, index funds, your retirement account... suddenly exposed to a company that commits half a trillion dollars to chips and warns, in writing, that its own product might lie to you.

Right. The whole economy becomes a shareholder in the experiment.

Every bubble in history came with a prospectus full of optimism. This one is full of fear... and the line to buy is already forming.