Stealth AI Startups and Why They Don't Go Public

Nine frontier AI labs closed two hundred forty point six billion dollars of equity in the first nine months of this year. That's from BlockWest Research, data through September thirtieth. Two hundred forty billion. Not one of those companies was listed on a stock exchange.

None of them. So where does that money even come from?

Softbank. Nvidia. Amazon. Sovereign wealth funds. Which raises the question I want to land by the end of this: if private money is that easy, why on earth is Anthropic about to do the one thing everybody else is avoiding?

Right, because an IPO is supposed to be the prize. You ring the bell, you made it.

That was the old deal. Born small, raise privately, graduate to the public markets. Andreessen Horowitz's David George put numbers on how broken that is. Highly valued private tech is now around five trillion dollars of market cap. Almost a quarter of the S and P five hundred. And the ten biggest private companies are forty percent of it.

Five trillion sitting in a market I can't buy into.

Worse than that. George says for recent IPOs, fifty-five percent of the value creation happened before listing. A decade ago it was twelve percent.

So by the time retail gets in, the growth has already been eaten.

And in George's universe of public companies, only three are growing over thirty percent a year. Three. If you want hypergrowth, it's not on the exchange. It's in somebody's data room. Now here's the part founders actually talk about privately. Aron Bohlig, who runs the investment bank ComCap, told Observer back in twenty twenty-four that AI companies simply don't have good revenue forecasts.

Hang on. These are the most valuable companies on earth and they can't forecast revenue?

His exact line was, "I wouldn't want to be the CFO of OpenAI and on a quarterly basis need to have disclosures to investors and change my strategy every six or nine months because the industry is evolving so quickly."

That's a hell of a thing to admit.

And look at the data. Cost of frontier model access has fallen more than ninety-nine percent in two years while capabilities roughly double every seven months. Try building a four-quarter guidance model on top of that.

So secrecy isn't some James Bond thing. It's that you literally can't promise a number.

Partly. Partly it's that strategic investors don't ask. Microsoft put thirteen billion into OpenAI. Amazon four billion into Anthropic, Google two. Those backers want collaboration and compute deals, not quarterly earnings. Joe Endoso at Linqto Capital put it plainly: it lets them access capital without the pressure of public markets. And then the scale broke everyone's brain. March thirty-first, twenty twenty-five, OpenAI closes forty billion dollars at a three hundred billion valuation. Largest private tech round ever. Thirty of that from SoftBank.

What was the previous record?

Ant Group, fourteen billion, twenty eighteen.

So they nearly tripled it in one round.

Then twenty twenty-six happened. OpenAI closed one hundred twenty-two billion dollars on April first at eight hundred fifty-two billion post-money. Anthropic did thirty billion, then sixty-five billion, landing at nine hundred sixty-five billion in May. Those two took ninety percent of all the money the nine labs raised.

That's not a market. That's two companies and a rounding error.

And there's a tell buried in the Amazon commitment. Fifty billion to OpenAI. Fifteen up front... thirty-five tied to triggers, and one of the triggers is an IPO.

Wait, so the private money is quietly pulling them toward going public anyway?

Eventually. OpenAI isn't expected to list before twenty twenty-seven at the earliest. Which brings us to the one that blinked first. Anthropic's prospectus, seen by Reuters, reported by CNBC on September twenty-eighth. Revenue up twelve-fold in twenty twenty-five, to nearly four point six billion. Net loss for the year: forty-two billion dollars. Future cloud and compute obligations: five hundred eighteen billion.

Five hundred eighteen billion in obligations. That's the answer, isn't it. You can't private-round your way out of that.

That's the whole loop. Staying private works beautifully until your compute bill is bigger than most countries. Reuters reported on October twenty-second that Anthropic is pressing ahead for mid-November, marketing to investors the week of November ninth, at a valuation some think could reach one and a half trillion.

And everyone else?

Gone. Switch, Nscale, SB Energy all pushed back. AirTrunk and Nvidia-backed Firmus pulled or paused. Reuters says the fall tech IPO calendar is down to Anthropic alone — partly because nobody wants to compete with it for investor attention.

So it's not courage, it's necessity. And a clear runway.

And the scrutiny arrives instantly. On October fifteenth, New York City Comptroller Mark Levine — who oversees more than three hundred twenty-five billion in pension assets — wrote to Dario Amodei asking him to delay, citing safety. The city holds about seven million dollars of Anthropic exposure.

Seven million out of three hundred twenty-five billion and he still writes the letter.

That's the price of the bell. The moment you list, a pension fund gets a vote on how fast you build superintelligence. Everyone else stayed private partly to avoid exactly that conversation.

So the question isn't why they hide. It's how long the hiding can last.

Until the invoice comes due. Anthropic's is five hundred eighteen billion dollars, and in November, it's being passed to you.