Anthropic's S-1 IPO Filing: What It Reveals

Anthropic filed its S-1 with the SEC on Monday, October twenty-sixth. It's going public on the Nasdaq under the ticker ANT, and it could be the biggest IPO in history.

Bigger than SpaceX? I thought that one was untouchable.

SpaceX raised a record eighty-six billion dollars in June. The AP says Anthropic could eclipse that, with a debut as soon as next month.

Okay. So what's actually in the document? Because that's the part nobody reads.

Three hundred and eighteen pages of prospectus, three hundred and thirty-five in the full filing. Revenue for the first half of twenty twenty-six: sixteen point three billion dollars. That's up about thirteen hundred percent year over year.

Thirteen hundred percent. For a company founded in twenty twenty-one.

Five years old. And in that same six months, it lost eighty point seven billion dollars.

Wait, it lost five times what it made? How does that even work?

Most of it isn't cash. Seventy-two point four billion is an accounting charge, reflecting the rising value of convertible notes Anthropic had issued to investors.

So on paper they owe their own investors more because the company got more valuable. That's a weird kind of pain.

Strip that out and the picture flips. Operating losses actually narrowed, to six hundred and ninety-three million for the half, from two point one billion a year earlier. On an adjusted basis Anthropic says it made one point six billion in operating income.

And real cash? Not adjusted-this, excluding-that.

Positive operating cash flow for the first time: three point nine billion dollars in, versus one and a half billion out the year before.

Okay, that's a genuinely different story than "lost eighty billion."

The number that made finance people lose it, though, is net revenue retention. Around nine hundred percent as of September thirtieth.

Meaning what, in plain English?

Existing customers spent about nine times more this year than last. The median public cloud software company sits at a hundred and eight percent.

So everyone else grows a customer by eight percent and Anthropic grows them by eight hundred.

Anthropic's explanation is that companies start with one use case, then Claude spreads across departments and workflows. And as models get cheaper per task, customers just buy more of them.

Who are these customers? Because I assumed this was a chatbot business.

Mostly businesses, and the filing says so flatly. The line is, frontier AI is most valuable in real-world settings, because deployment across enterprises is where economic impact is greatest. That's the split with OpenAI, which is far more consumer.

So where's the money going? Because that revenue can't just be profit.

Compute. It's the single biggest operating expense. Five point six billion in the second quarter alone, against eleven point six billion of revenue that quarter.

Two dollars back for every dollar spent. That's thin but it's not insane.

Two dollars and eight cents, and they track it as a metric. But before I give you the commitment number... want to guess what they've promised to spend on compute and infrastructure going forward?

Given the quarterly numbers? I'll say... eighty billion.

Between five hundred and forty-two and five hundred and sixty-seven billion dollars.

Half a trillion? That's more than thirty times their annual revenue.

On December thirty-first of last year they signed a four hundred and seventeen billion dollar compute deal with various providers, running one to twenty years. Then another hundred and twenty-five to a hundred and fifty billion of capacity starting between twenty twenty-eight and twenty thirty-one, with pricing not even finalized.

And if demand doesn't show up, they're still on the hook.

That's the bear case. The filing also admits they've underestimated their own compute needs in the past, and that forecasts are inherently uncertain. Too little and they fall behind the frontier, too much and it sits idle.

What about the safety stuff? Dario Amodei talks about existential risk constantly. Does that survive contact with a prospectus?

It's in there. They wrote that moving quickly brings AI's benefits sooner, but moving too quickly could outrun the world's ability to make AI systems safe. They warn new models could be DELAYED over cybersecurity and safety threats.

Telling investors "we might hold back the product" is a bold opening move.

And they say their safeguards against things like bio and chemical weapons misuse are incomplete, and have been circumvented before. More than ninety pages of risk factors in total.

So who controls this thing once it's public? Because public markets usually want the brakes off.

This is the strangest part structurally. It's a Delaware public benefit corporation. The co-founders get a single share of Class F stock through a Founder LLC, carrying fifty point one percent of voting power... with no economic rights at all.

They own the votes but not the money?

And a Long-Term Benefit Trust elects a majority of the board through yet another class of stock. Public shareholders buy Class A, one vote each, and they're structurally outnumbered.

So what's the number everyone's actually waiting for?

The price. The S-1 leaves share count and price range blank. Anthropic was valued at nine hundred and sixty-five billion in May, and CNBC reports it's seeking one point five to two trillion.

Which would make it a top-ten US tech company on day one, having never traded before.

Goldman Sachs and Morgan Stanley are running it. The registration isn't effective yet, so nothing can be sold until the SEC clears it, and then we get the pricing range.

And if rates stay high and the AI mood sours between now and then?

Then that two trillion number gets a lot of scrutiny. Watch for the amended filing with the price range, probably in the next few weeks.

I'll be refreshing EDGAR like it's a sports score.