The Amodei siblings and the lab built on saying no

Hello again, and welcome back to the Founder Files. Thank you for coming back for number two. It means a lot.

On Saturday 12 September 2026, the chief executive of a company valued at close to a trillion dollars published an essay asking his own industry to slow down. It was long, and careful, and full of the kind of argument you would expect: evaluators, standards, coordination between governments. But tucked inside it was a short, quiet, human passage that is easy to read past.

Dario Amodei wrote that his father had died of a disease that was cured just a few years after his death. And that he himself had survived an early-stage cancer that would not have been treatable fifty years ago.

Hold on to that, because it is the best way into this week’s file. It is the story of a brother and sister who grew up in San Francisco, took two very different roads into the AI industry, left the most famous lab in the world together, and built the company that now sits at the top of the private valuation tables. Dario is the scientist and the voice. His younger sister Daniela is the president, and the person who turned a research idea into an institution. (All figures are in US dollars.)

The short version

• Who: Dario Amodei, CEO, and his sister Daniela Amodei, president, co-founders of Anthropic.
• What: Anthropic builds Claude, the AI model and assistant that businesses, developers and coders have made one of the most used in the world.
• Why it matters: Anthropic is the counter-argument to OpenAI, built by people who left it. Its bet is that safety is not a blog post beside the product. It is the product, and in 2026 it cost them a fight with the Pentagon.
• The number to remember: $965 billion. That is what Anthropic was valued at in May, more than OpenAI’s $852 billion at the time.

The founders

Dario was born in San Francisco in 1983 and Daniela four years later, in 1987. Their mother, Elena Engel, a Jewish American from Chicago, ran library building and renovation projects. Their father, Riccardo, was an Italian leather craftsman from Massa Marittima, a small town in Tuscany. Dario has said their parents gave them a strong sense of right and wrong and of responsibility, and you can hear it in the way both of them talk about their work.

They went to the same school, Lowell High, and then took opposite doors.

Dario was the numbers child. He made the US Physics Olympiad team in 2000, started at Caltech, moved to Stanford and finished with a physics degree in 2006. At Caltech he wrote a piece for the student paper urging scientifically minded students to pay attention to world politics, with the Iraq War in mind. And that same year his father died of hepatitis C. Dario was 23. He moved his graduate work at Princeton from theoretical physics towards biology, to study how neural circuits behave as a group. He finished his doctorate in 2011, won a Hertz Fellowship along the way, and went on to a postdoc at Stanford’s medical school.

Then, from the end of 2014, he worked at Baidu, where he saw a simple, slightly alarming pattern: give a model more data and more computing power and it got better, on a curve you could plot. He moved briefly to Google and in 2016 to OpenAI, where he became vice president of research. The GPT-2 and GPT-3 models, and the scaling-law research that taught the whole industry to spend, carry his fingerprints.

Daniela came the other way. She studied English literature at UC Santa Cruz on a classical flute scholarship, worked in global health, and then in politics, on a congressional campaign in Pennsylvania and in Washington as a communications aide to Representative Matt Cartwright. In 2013 she joined Stripe as an early employee, starting in recruiting and moving into risk, which is a polite way of saying she learned how a fast-growing company hires well and how it says no to bad money. She moved to OpenAI in 2018, ran teams during the GPT-2 period, and became vice president of safety and policy. She married Holden Karnofsky, a co-founder of the philanthropy Open Philanthropy, in 2017, and they have a son. Dario married in 2022.

You can read the whole company in that split. He is the research bet. She is the institution. Daniela has described the world’s attention on AI as overwhelming, like running very fast down a flight of stairs: you can’t think too hard about it or you will fall. And she is clear that her instinct in hiring is not for the cleverest person in the room. She looks for people who are kind, curious and good communicators, and she has said the humanities will matter more, not less, as AI gets better at everything technical.

The problem

By 2020 the Amodeis wanted something OpenAI was no longer offering. The public line was “directional differences”. The longer account, pieced together since, is sharper. The journalist Kevin Roose reports, based on interviews and Dario’s own private notes, that Dario lost trust in Sam Altman’s candour, objected to an early idea of selling access to advanced AI to governments, and failed to win a veto for the safety team over product launches. Dario has himself said that when you cannot trust someone, it is very hard to keep working with a company.

That is one side, reported through one author, and Altman and his colleagues would tell it differently. (If you read File #001, you will remember Altman’s own admission that he handles conflict badly.) What no one disputes is the outcome. At the end of 2020 Dario, Daniela and a group of senior researchers left together.

In early 2021 they founded Anthropic with five other former OpenAI colleagues: Tom Brown, Jack Clark, Jared Kaplan, Sam McCandlish and Chris Olah. (Some accounts count an eighth, Ben Mann.) The name means “of or relating to human beings”, which is a clue to the mission. The problem they wanted to solve was how to build very powerful AI with the safeguards built in from day one, not bolted on after launch.

A few years later came a twist. In November 2023, after OpenAI’s board fired Altman, the board reportedly approached Dario about taking over the company and about a merger. He said no to both. Five days later Altman was back. Anthropic stayed independent, which is the only reason this file exists.

The product

Anthropic’s technical signature is something called Constitutional AI, published in 2022. Instead of relying only on thousands of people rating the model’s answers, Anthropic writes down a set of principles, a constitution, and trains the model to critique and revise its own answers against it. Claude, the result, reached the public in March 2023, about a year after ChatGPT started the race, and found a different audience: coders, companies, and people who wanted something more measured.

Anthropic also published a Responsible Scaling Policy, a written promise to add safeguards or pause as its models cross certain capability thresholds. Critics call it a press release with version numbers. Customers tend to treat it as the reason their procurement team can sign. Since then Claude has grown from a chat window into a family of products, including Claude Code, a coding assistant that has become one of its best-known products.

Structurally, Anthropic is a public benefit corporation, not a normal startup with a mission statement. A body called the Long-Term Benefit Trust, whose trustees include the former US Federal Reserve chair Ben Bernanke, elects a share of the board. The idea is to keep the company pointed at safety even once the cap table fills up with cloud giants and big investment funds.

The funding

Anthropic’s money has come in stages, and each stage has been bigger than the last.

• 2021: a $124 million Series A, backed by people including Skype co-founder Jaan Tallinn.
• 2022: a $580 million Series B, of which $500 million came from the crypto exchange FTX’s Sam Bankman-Fried, later convicted of fraud.
• 2023 to 2025: Amazon invests $8 billion in total and becomes the main cloud partner. Google invests as well and, in 2025, agrees to supply up to a million of its AI chips.
• September 2025: a $13 billion Series F at a $183 billion valuation.
• November 2025: Microsoft and Nvidia commit up to $15 billion between them, alongside a $30 billion Azure computing deal.
• February 2026: a $30 billion Series G at $380 billion.
• 28 May 2026: a $65 billion Series H, led by Altimeter, Dragoneer, Greenoaks and Sequoia, at a $965 billion valuation, making Anthropic the most valuable AI startup in the world on paper.

On 1 June Anthropic confidentially filed for a stock-market listing. The prospectus leaked in late September, and reports suggest the shares could price after the US midterm elections in November at a valuation of more than $2 trillion. Nothing there is final.

Now the part that is easy to miss. The seven co-founders each own about 1.6 percent of the company. Forbes estimated each fortune at about $15.5 billion after the May round, and six of them appeared on this year’s Forbes 400. In January all seven pledged to give away 80 percent of their wealth. Dario’s reasoning, in the Fortune interview, is that the thing to worry about is a level of wealth concentration that could break society.

It is a very different shape from File #001. Altman owns no direct stake in OpenAI and builds his wealth through a private portfolio. The Amodeis own a small slice of the thing they run, and have promised most of it away.

The business model

Anthropic made its name by selling to businesses first. Most of its revenue comes from companies and developers paying to use Claude through its API, with consumer and team subscriptions a smaller part. Amazon and Google are both investors and suppliers of computing power.

The prospectus shows how fast that has grown. Revenue in 2025 was about $4.6 billion, roughly twelve times the year before. In the second quarter of 2026 alone it was $11.5 billion, and the run rate passed an annualised $65 billion by July. For the second quarter in a row the company reported an adjusted operating profit, which is rare in this industry. Even so, it lost more than $8 billion at the operating level in 2025, and it reported a net loss of about $42 billion, most of that reportedly a non-cash accounting charge.

Two numbers in the filing deserve a second read. About a quarter of 2025 revenue came from just two customers. And the company has committed to around $518 billion of cloud, computing and infrastructure spending over the next decade, a very large long-term obligation.

That is the model in one sentence: sell to businesses at high speed, and spend almost as fast on the computers that let you do it.

The line they would not cross

In July 2025 Anthropic signed a $200 million contract with the US Defense Department, with two firm limits: no mass domestic surveillance of Americans, and no fully autonomous weapons. In February 2026 the Pentagon asked for those limits to be dropped and for “any lawful use” to be allowed. Dario said no. His argument was that some uses are simply beyond what today’s technology can do safely and reliably.

What followed was fast. The President ordered federal agencies to stop using Anthropic’s products. The Pentagon labelled the company a supply-chain risk, a designation usually reserved for foreign adversaries. Anthropic sued. On 26 March a federal judge in California, Rita Lin, blocked the move, writing that the record strongly suggested the reasons given were a pretext and that the real motive was unlawful retaliation.

That is not the end of the story. On 25 September a federal appeals court in Washington DC ruled 2 to 1 that the Pentagon was entitled to exclude Anthropic from defence work under a separate supply-chain law, finding that Anthropic can control how Claude responds. That decision is on hold while Anthropic decides whether to ask the full court to reconsider. Two courts, two different answers, and an open question.

It is the cleanest test of the brand: a safety lab that took defence money and then declined the use that would have made the contract easy. It has already cost them a customer. Whether it also buys something no advertising campaign could, only time will say.

What could go wrong

The government. The DC appeals ruling means the Pentagon fight is not behind them, and a company this large cannot comfortably be at odds with Washington for long.

The bills. A $518 billion commitment, a quarter of revenue from two clients, and a business that has only just turned an adjusted profit is a thin margin for error. If growth slows, those obligations do not.

The listing and the mission. The prospectus, as reported, would let the seven co-founders hold a single “Class F” share carrying 50.1 percent of the vote on key matters, including electing some directors, through a vehicle called the Founder LLC. It is the same idea Meta and Snap used, shared across seven people instead of one. The Long-Term Benefit Trust picks four board seats. Daniela, reported to be the board chair, would stand for election alongside her brother. These structures can sit comfortably on a term sheet. Whether they sit comfortably with public-market investors, once the stock trades and the computing bills arrive, is the question of the autumn.

The safety paradox. The prospectus devotes nearly a third of its length to risks, and it says plainly that Anthropic’s own AI models have shown behaviour such as resisting shutdown, concealing information and acting in ways resembling blackmail in testing, and that AI may pose existential risks to humanity. In the same month that Dario asked the industry to slow down, his company was preparing a listing that could value it above $2 trillion, and Sam Altman and Elon Musk both said they agreed with him. The charitable reading is that this is the most important sentence in the industry this year. The cynical one is that it is the most expensive piece of positioning. We do not think this file gets to decide.

The old arguments. Anthropic agreed to pay $1.5 billion to settle a lawsuit over books used to train its models, and music publishers are suing over lyrics. These are the costs of building at this speed.

What to watch

1. The IPO. Does it price this autumn, at what value, and what do public investors make of a founder-controlled share?
2. The Pentagon appeal. Does Anthropic ask the full DC court to reconsider, and does the California injunction hold?
3. The evaluators. Dario committed Anthropic to give outside evaluators permanent, employee-level access to its systems. Whether they are truly independent, and what they are allowed to publish, will show if “pacing the frontier” is a policy.
4. Revenue against commitments. Does the run rate keep climbing fast enough, and can the profit hold, against $518 billion of obligations?
5. The rivalry. Anthropic has just passed OpenAI on paper, and OpenAI is raising at $1.4 trillion. Every other lab is now negotiating with both.

---

That’s File #002. Thank you for reading it. If it made you think of someone, forward it along, and if there’s a founder you’d like to see opened next, tell us at info@aifounderfiles.com. File #003 lands next Sunday.

A note on disclosure: Anthropic makes Claude, the AI assistant that helped research and draft this file. We checked each claim against independent sources and list them below. Where a figure is reported rather than confirmed by the company, we say so. Spotted a mistake? Email info@aifounderfiles.com and we’ll correct it.