For years, if you joined Anthropic, Dario Amodei would recommend The Making of the Atomic Bomb. It was the CEO’s favorite book.

Aware of this, Bloomberg asked Amodei whether he saw parallels with Robert Oppenheimer, who led the Manhattan Project.

“The figure I most identified with was Leo Szilard,” said Dario.

Szilard was the first to conceive the idea of a nuclear chain reaction. He urged physicists to hold back their results from Germany, then helped Albert Einstein write to President Roosevelt, warning of a German bomb. Once the bomb was built, he drafted a petition urging President Truman not to use it unless Japan had been told the surrender terms and had refused them. So Szilard conceived the technology, pushed for the side he trusted to build it first, then argued for limits on its use.

In 2016, Dario co-wrote Concrete Problems in AI Safety at Google Brain, on how machine-learning systems could cause accidents. In 2020, he and nine colleagues published Scaling Laws for Neural Language Models. As computing power, data, and model size grew, the models’ error fell at a steady rate the authors could forecast, across more than seven orders of magnitude.

This was Dario’s chain reaction. Similar to how a nuclear chain reaction only needed more atoms to scale, a better AI model only needed more computing power and data to scale.

He had already seen it start. In 2019, he helped build GPT-2 - a model that answered only about 4% of simple trivia questions correctly. He'd later say that most researchers shrugged at the model, but a few realized how far-reaching the consequences were. Dario himself recalls the moment as:

“My God, you just scale this thing up another factor of a million and it’ll be better than humans at everything... This is going to determine the fate of nations and economies and the future of humanity.”

Scaling a model doesn’t make it safer, only more capable, and as Dario explained on the Future of Life Podcast in 2022, new behaviour can appear suddenly as models grow. A quirk that’s harmless in a weak model may not be harmless in one a million times more capable. In 2021, Dario and six colleagues left OpenAI to start Anthropic, sure they could be trusted with this technology more than anyone else. That belief cuts both ways. It is why they held back a working Claude in 2022. It is also why they dropped their pledge to pause in February 2026: stopping, they argued, would hand the lead to labs with weaker safeguards.

Regardless of what one thinks about Anthropic as a firm, its financial track record has been undeniable. By Dario's account, when he forecast $100M of sales in the first year to investors, he was told: “This has never happened in the history of capitalism. You’ve lost all credibility.”

Anthropic hit it, then went from $100 million to a billion the next year.

For a while, they looked unstoppable. But recently, their growth curve bent. Quarterly sales growth fell from 144% to about 47%, while compute spending rose 77%. That wouldn't be so dramatic if it were not for Anthropic's astronomical spending commitments of $900.9 billion invested through 2045. They're open about this:

“If I’m just off by a year in that rate of growth, or if the growth rate is 5x a year instead of 10x a year, then you go bankrupt.” - Dario, February 2026

In its first two days of trading, Anthropic was worth about $2 trillion, then $1.76 trillion, then less again by Friday morning, when the stock was down more than 12%. The world is still trying to put a price on this company. The stock price is an output - a reflection of future value, so a $200+ billion swing in a day says less about Anthropic than about how little anyone can see of its future. What's the input? Anthropic's own roadshow video lists five advantages and puts the mission first: "All of this begins with our mission."

The founders have gone a long way to keep control of where this company goes. Seven people own about 13% of Anthropic and cast 61% of the votes, through a special share that carries 50.1% of the vote on its own. Each holds a near-equal stake, and each agreed in advance to be outvoted by the others. Public buyers, who put up $81 billion this week, get about 5.5% of the votes.

Why have they gone through so much inconvenience to structure their company this way? We see it as a sign of the founders being sure of themselves.

Our deep dive looks at where that conviction comes from, for each of the seven founders, and how much of it holds up against the record, all to answer a simple question:

What is Anthropic's edge, and how durable is it?

Disclaimer: The views and opinions expressed above are current as of the date of this document and are subject to change without notice. Materials referenced above will be provided for educational purposes only. None of the above will include investment advice, a recommendation or an offer to sell, or a solicitation of an offer to buy, any securities or investment products.

Deep Dive PDF below ↓

logo

Become a Member To Read The Full Deep Dive

Members of Social Capital Research get access to this deep dive and other members-only content.

Join 3,000+ Investors and Founders

Already a subscriber? Sign in

A membership gets you:

  • Weekly members-only deep dives
  • Group chat with Chamath
  • Access to 50+ past deep dives
  • Quarterly AMA with Chamath
  • Occasional In-Person Event Invitations