Why Anthropic's Founders Left OpenAI

Twenty twenty, inside OpenAI. The GPT-three beta is running, and Dario Amodei, the vice president of research, asks Sam Altman a simple question. Who is actually using this thing? Altman gives him a list.

And I'm guessing the list is wrong.

The list is short. Amodei says he found a second one later. Longer. And he began to suspect that what he'd been shown had been... filtered.

That feels like a small thing to quit a company over.

Trust is made of small things. December twenty-ninth, twenty twenty. Amodei walks out of OpenAI after nearly five years. Co-builder of GPT-two and GPT-three. And he does not walk out alone.

How many went with him?

Six colleagues immediately, including his sister Daniela. Eventually around a dozen ex-OpenAI researchers founded Anthropic in twenty twenty-one. People have argued ever since about what they were running from. Start with the version you've probably heard.

The Microsoft money. The billion dollar deal. The safety people got squeamish about the cash.

Dario says no. Flatly. He's said there's a lot of misinformation out there, and quote, people say we left because we didn't like the deal with Microsoft. False. He'd been directly involved in commercialising GPT-three. He was not anti-business.

So the founding myth is just... wrong?

The clean version is wrong. What he describes instead is trust. His line is, it is incredibly unproductive to try and argue with someone else's vision. That's not a fight he lost. It's a fight he decided wasn't winnable from the inside.

That is a very polite way to say I'm leaving.

The less polite version turned up this year. Kevin Roose has a book out, The AGI Chronicles — the two lists story is his reporting — and it describes a power struggle before the exit. A so-called countries plan, selling advanced AI to governments. Amodei emailing senior leaders in blunt language.

Okay, but that's one book, one side. Has OpenAI answered any of it?

Not that I can point you to. So hold it the way I'm holding it. Reported, not proven. Roose's reporting, Amodei's recollection.

Fine. They leave. And now they need money.

They need a LOT of money. April twenty twenty-two, it arrives from the strangest possible place. Sam Bankman-Fried puts in roughly five hundred million dollars, leading a five hundred and eighty million dollar round. Anthropic valued at about four billion.

The safety company. Funded by FTX. That aged like milk.

Seven months later FTX collapses, because — as prosecutors established — he'd been taking customer money. So that stake becomes this strange asset sitting inside a bankruptcy. March twenty twenty-four, the estate agrees to sell most of it. Eight hundred and eighty-four million dollars.

To who?

Biggest buyer, a group aligned with Mubadala, the Abu Dhabi sovereign wealth fund. Nearly five hundred million dollars' worth. Second biggest? Jane Street. The trading firm where Bankman-Fried started his career.

It goes back to his old shop? Come on.

Plus Fidelity funds. And the Ford Foundation.

The Ford Foundation owns a piece of Claude. That's genuinely strange.

And by then Anthropic had already found the real money. Amazon committing up to four billion dollars, Google adding billions more. Amazon's own filings this year disclose roughly ninety-eight billion dollars of Anthropic convertible notes, plus about ninety-two billion in non-voting preferred stock.

Hang on. A hundred and ninety billion? On one balance sheet?

On one balance sheet. The crypto bet became the seed of one of the largest equity positions in corporate America.

So five years out from that walkout — who's winning?

February twenty twenty-six, Series G. Thirty billion raised at three hundred and eighty billion. May, Series H. Sixty-five billion at nine hundred and sixty-five billion. June first, a confidential S-one filed with the SEC. And Bloomberg, on September fifteenth, put annualized revenue at sixty-five billion dollars.

From the company whose whole pitch is that the product might be dangerous.

Which is the part that hasn't changed since twenty twenty. Reuters first reported the prospectus on Monday, September twenty-eighth. It warns the technology could pose a catastrophic or existential risk to humanity. That sentence is in the document they're selling shares with.

Buy our stock. Read our disclaimer.

And weeks before that, Amodei published an essay arguing AI companies have to pace the frontier. Slow down. Together.

Which only works if everybody actually does it.

Right. And on September twenty-fourth, Reuters reported the co-founders plan to keep voting control after listing. Same instinct as twenty twenty. Never again be outvoted on somebody else's vision.

So it all comes back to the second list. The whole company is a trust issue, incorporated.

That's it. And none of the rest is confirmed yet. The reported target is November, on the Nasdaq, up to a hundred billion raised, at around two trillion dollars. If it holds, the biggest IPO in history.

Run by the man who left because he thought the others were moving too fast.

He walked out over the pace of the race. In November, they hand him the fastest car on the track.