Part IThe Story
What Is This Lady Actually Good At?
In the spring of 2026, Daniela Amodei sat on a stage at the Stanford Graduate School of Business and offered a room of MBA students a joke at her own expense. "If you were to look through my background," she said, "you would be like, 'What is this lady actually good at?'" She then supplied the case against herself, the way a good litigator concedes the weak facts first: "She doesn't have a law degree. She's not a computer scientist."
The line is built to get a laugh. It is also, read literally, an accurate summary of how the public record has treated her. She was speaking as co-founder and president of Anthropic. Within a few months of that conversation, the company would raise $65 billion at a $965 billion valuation and confidentially file to go public. A few weeks before she sat down at Stanford, the Department of Defense had designated it a "supply chain risk" for refusing to let its models be used for mass domestic surveillance or fully autonomous weapons. A federal judge had called the government's response "classic First Amendment retaliation." The company had more than 2,000 employees and, she said elsewhere, more than 300,000 business customers. Its flagship model was named, at least partly, for the man who invented information theory.
And the question still hung there, half joke and half genuine puzzle. What is this lady actually good at?
Here is one answer, which Fortune reported in June 2026 and which reads like a typo on an org chart. Dario Amodei, Anthropic's chief executive and Daniela's older brother, manages exactly one person: his chief of staff. Every other senior leader in the company reports to her. The average American manager supervises about a dozen people. Dario supervises one, and he told Bloomberg the arrangement was "incredibly freeing." Freeing for him. Somebody is holding the other end of that lever, and has been since a handful of people walked out of OpenAI at the end of 2020 with a plan, very little money, and no model.
The public file on her is thin. Her Wikipedia entry, for a stretch of 2025 and 2026, still described her as leading "teams focused on AI safety and model alignment." That was a stale description of a remit that had grown to cover operations, product strategy, go-to-market, people, and, according to the company's own leadership page, governance. The books are mostly about her brother. One is called The Scaling Curve, and its subtitle begins with his name. The New Yorker's long 2026 account of the company describes Dario as "a restless savant who has been patiently coached to restrain his spasmodic energy." It then gives his sister a single clause: "a humanities type partial to Joan Didion."
That clause is the right place to start, but only as a door. Didion's first lesson was that the stories we tell ourselves in order to live are also how we avoid looking at what is actually happening. The story Silicon Valley tells about Anthropic is a research story: physicists, scaling laws, interpretability, a lab full of people who read The Lord of the Rings as a warning. All of that is real. None of it explains how the company runs on an ordinary Tuesday.
By the Numbers
The President's Ledger
$965BAnthropic's post-money valuation after its $65B Series H, May 2026
~$47BReported annualized revenue run rate, May 2026
$100MEstimated Anthropic revenue in 2023
#40 / #50Her employee number at Stripe and at OpenAI
1Direct reports to Anthropic's CEO; all other senior leaders report to her (Fortune, June 2026)
7OpenAI alumni who co-founded Anthropic in January 2021
80%Share of their wealth Anthropic's founders pledged to give away
A Leather Craftsman and a Library Project Manager
Two people built things with their hands and their schedules in the San Francisco house where Daniela Amodei was born in 1987.
Her father, Riccardo Amodei, was a leather craftsman from Massa Marittima, in Tuscany. He made objects one at a time. For years he was in poor health, and he died of hepatitis C when his son Dario was twenty-three and his daughter was barely out of her teens. Dario has written that his father died a few years before an effective cure, sofosbuvir, was discovered. The grief has stayed in the family's public language ever since. When Dario argues that artificial intelligence could compress a century of medical research into a decade, the case is not abstract. It has a face and a workbench.
Her mother, Elena Engel, was a Jewish American from Chicago who worked as a project manager for libraries. That is a job description almost designed to be overlooked, and it rewards a second look. In 1998, when the Berkeley Public Library's main branch was being renovated, a local Jewish newspaper quoted her briefly: "Library staff spent eight months looking at possible temporary sites," said Elena Engel, the library's project manager for the renovation. In 2006 she wrote a paper with the unglamorous title "RFID Implementations in California Libraries: Costs and Benefits." Her work was the logistics of keeping knowledge housed, catalogued, and reachable while the building around it was torn open and rebuilt.
The children went to Lowell High School, San Francisco's selective public magnet. They went in different directions. Dario made the U.S. Physics Olympiad team in 2000, started at Caltech, transferred to Stanford for a physics degree, and earned a doctorate in biophysics at Princeton. His thesis was on the electrophysiology of neural circuits, which is to say he was studying how brains compute nearly a decade before he started building machines that do something similar. He was the kind of child whose trajectory you could plot in advance.
Daniela's was harder to plot. She won a scholarship for classical flute. At the University of California, Santa Cruz, she studied literature, politics, and music, and won the school's concerto competition in 2008 as a soloist. She graduated summa cum laude with a B.A. in English literature in 2009, in the wreckage of the financial crisis, into a job market that wanted almost nothing from people like her. She has described the anxiety plainly: who would hire someone like me?
The flute matters more than its place on a résumé suggests. It is a breath instrument. The player supplies the pressure, and the instrument shapes it, so the discipline is one of controlled release: phrasing, knowing where to breathe, sustaining a line across bars without running out. An orchestral flutist almost never carries the melody alone for long. Most of the job is blending. The composer gets the credit, the conductor gets the gestures, and the flutist breathes on cue.
Years later, asked at Stanford whether she would make the same choice again, Amodei said she would still major in literature. "I know that sounds crazy," she said. "I like to read." Her favorite book, she said, was Barbara Tuchman's The Guns of August, a history of the first month of the First World War. It is a book about men who built elaborate machinery in peacetime (mobilization schedules, alliance commitments, railway timetables) and then discovered, in the summer of 1914, that the machinery had its own momentum.
She did not say why that book. She did not have to.
Field Director
The question she carried out of Santa Cruz, she has said, was simple and very large: How do you build something that is of consequence and has a real purpose?
The first answer was international development. From January 2010 to August 2011 she worked in business development at the IRIS Center at the University of Maryland. She managed grant processes and recruited personnel for projects in poverty assessment, microfinance, conflict mitigation, and direct cash transfers. This is the plumbing of aid: proposals, budgets, finding the right economist for the right ministry. Then, from September 2011 to January 2012, she was a fellow at Conservation Through Public Health, an organization based in Uganda that works on helping human communities coexist with wildlife. There she helped the CEO develop a strategic grant-making plan and co-led training for more than fifty community health workers at rural field sites outside Kampala.
The second answer was politics. In February 2012 she joined Matt Cartwright's campaign for Congress in Pennsylvania as deputy field director and was promoted to field director before the race ended. She recruited more than eighty volunteers. Field work is the least romantic part of a campaign. It means clipboards, call sheets, turf maps, and persuading people who have jobs to spend their Saturdays knocking on strangers' doors. Cartwright won that November. In January 2013 she followed him to Washington and handled scheduling and communications in his House office.
She lasted about four months. She has said the slow, procedural environment of Capitol Hill drained her rather than energized her. She did not stay to prove she could tolerate it. She left.
Look at those four years in sequence and they read like the false starts of a talented, restless graduate. Look at them in reverse, from the vantage of a company that would one day be negotiating with the Pentagon, testifying about regulation, and shipping a product to governments and hospitals, and they read differently. A grant writer learns to translate a mission into a budget. A field director learns that every grand strategy depends on whether eighty volunteers show up on Saturday. A congressional scheduler learns, in the most literal way, how power allocates its time.
None of it involved a computer. All of it involved turning an idea held by someone at the top into a structure that ordinary people could carry out at the bottom.
Employee Forty
In May 2013, Daniela Amodei joined Stripe as roughly its fortieth employee. Her title was lead technical recruiter.
There is a tidy version of this story in which the liberal arts graduate stumbles into tech through a side door, and a truer version in which she walked through a door that most people at fast-growing startups underrate. Recruiting at forty people is not human resources. It is product design for the company itself. Each hire changes the composition of the organism. Over the next two years she helped Stripe grow from about forty people to more than three hundred. She learned, as she would later put it about herself, how to identify "the right people to solve a specific problem."
Then she moved toward the part of the business that could go badly wrong. In August 2015 she became a risk program manager. From January 2016 to July 2018 she was a risk manager covering core operations, user policy, and underwriting. For a payments company, risk is not a compliance function bolted onto the side. It is the business. Every transaction is a small bet that the money is real, the merchant is honest, and the card is not stolen. Her first exposure to machine learning came here, working with engineers, data scientists, and lawyers on fraud and credit models. These were statistical systems that made consequential decisions about people at a scale no human could review one by one.
This is the hinge of the career, and it is easy to miss. At Stripe she learned that a system can be extraordinarily useful and quietly dangerous at the same time. She learned that the danger is managed through policies, thresholds, escalation paths, and the judgment of people who sit between the model and the consequence. Later she would describe Anthropic's safety work in a vocabulary that sounds like a payments risk manager's: harms, misuse, edge cases, trust.
In August 2017, while she was still at Stripe, she married Holden Karnofsky, co-founder of Open Philanthropy, the foundation (since renamed Coefficient Giving) at the center of the effective altruism movement's grant-making. Karnofsky had spent years thinking about how to allocate money to do the most good, and he had become one of the movement's most serious voices on the risks of advanced AI. The marriage would later become something she was asked about, and something she seemed not entirely eager to discuss.
In 2018 she left Stripe for OpenAI, where she was employee number fifty.
Inside the Other House
OpenAI in 2018 was still a nonprofit research lab with a founding myth. Elon Musk and Sam Altman had launched it as a public-spirited counterweight to Google's DeepMind. The New Yorker later called its pitch "irresistibly earnest, if dubiously genuine." The pitch let the new lab raid Google's roster. One of its early hires, in 2016, was Dario Amodei, who arrived from Google Brain by way of Baidu and became the lab's vice president of research. He set its research direction and led the work on GPT-2 and GPT-3.
His sister arrived two years later and took on a set of jobs that, listed together, look like a coherent career path only in hindsight. From October 2018 to May 2020 she was an engineering manager and, simultaneously, vice president of people. She managed a team during the development of GPT-2. She incubated a business operations function for a research lab that had never needed one. In May 2020 she became vice president of safety and policy. Fortune has called the title senior vice president. Either way, the job sat at the intersection of how powerful the models were becoming and what the institution would allow them to do.
The terminology, she has admitted, was overwhelming at first. "I just kept asking questions until I felt like I could understand it," she said. She was clear-eyed about her limits: "Training AI models was not my responsibility, but I could leverage my strengths in areas where researchers were less experienced." That sentence is the whole method in miniature. She found the gap between what the researchers knew and what the institution needed, and she stood in it.
By late 2020 the institution was changing underneath them. OpenAI had taken a billion dollars from Microsoft the year before. It had created a capped-profit subsidiary. Its leadership was oriented toward commercial deployment in a way that the safety-minded researchers around Dario found alarming. The New Yorker's summary is acid: by late 2020, "Sam Altman himself had come to seem about as trustworthy as the average corporate megalomaniac." Luca Guadagnino's forthcoming film about OpenAI, Artificial, apparently stages the break as farce. In a draft screenplay, the Dario character announces to the office that he is starting a new company "which will be exactly like this one, only not full of motherfucking horseshit!"
The real exit was less theatrical and, by some accounts, more bitter. The Wall Street Journal reported in 2026 that one blowup occurred in late 2020, when Altman accused the siblings of urging colleagues to badmouth him to OpenAI's board. The details remain contested. What is not contested is that in December 2020, Daniela Amodei left her job as head of safety and policy at the most prominent AI lab in the world, along with her brother and five colleagues.
She has always insisted on the direction of travel. "We were running towards something versus running away from something," she told Stanford's students. "We had this vision in our heads of wanting to create an organization where the values that matter to us around safety and around responsibility were at the forefront of what we were doing."
The distinction between running toward and running away sounds like a commencement-speech flourish. For an operator, it is a real one. A company founded in flight defines itself by what it is not, and it tends to recreate its enemy in negative. A company founded toward something has to decide, on day one, what its org chart looks like.
Seven People and a Blank Page
Anthropic, PBC, was incorporated on January 26, 2021. The seven founders were Dario and Daniela Amodei, Jared Kaplan, Jack Clark, Chris Olah, Ben Mann, Sam McCandlish, and Tom Brown.
Each of them could be compressed into a sentence that rhymes with the others. Kaplan was a theoretical physicist who had helped write the paper on scaling laws, the counterintuitive claim that a model's capability depended less on clever architecture than on sheer size. Olah, whom Dario met when Olah was nineteen, had read Thomas Kuhn's The Structure of Scientific Revolutions ten times and more or less invented the field of mechanistic interpretability. He started from the conviction that it was "crazy to use these models in high-stakes situations and not understand them." Clark had been a journalist before he became OpenAI's policy director and would later run the Anthropic Institute, a think tank studying AI's effects on society. Tom Brown had been lead author on the GPT-3 paper.
They were, by Daniela's description, people with "pre-existing trust." Some had worked together at Google Brain. Some went back fifteen years. Most startups struggle with two co-founders. This one began with seven, and that worked only because the relationships predated the equity.
Her recollection of the first year is almost brutally practical: "When we first left OpenAI, we had to build a large generative language model again from scratch. So our first year was about getting a company up and running, fundraising, the nuts and bolts, focused on training this large language model." There were, she said, "something like six teams, and the majority of them were working on safety on top of, or directly in, the model." By her own account, for the first eighteen months there was no sales team and no go-to-market function at all. "Really until about six to nine months ago," she said in July 2023, "we were a research shop."
She has also said that none of the founders wanted to start a company. The New Yorker noted, with affectionate skepticism, that one can take this at face value "and at the same time observe that they seem a little silly in retrospect."
The money came from a particular world. Anthropic raised $124 million in May 2021, and much of its early capital came from the effective altruism network: Skype co-founder Jaan Tallinn and, most consequentially, Sam Bankman-Fried. In April 2022 the company announced $580 million in new funding, $500 million of it from FTX. Seven months later FTX collapsed, Bankman-Fried was eventually imprisoned for fraud, and his stake in Anthropic became an asset of a bankruptcy estate. In March 2024 the estate sold the majority of it for $884 million.
The founders structured the company as a public benefit corporation, legally obliged to weigh social impact alongside shareholder returns. As a costly signal of seriousness, they pledged to give away eighty percent of their wealth. OpenAI had also adopted an unusual corporate structure to guard its integrity, and the New Yorker noted this point too: the new company "sounded an awful lot like the company Altman had pitched as a foil for Google."
The effective altruism connection became a liability once Bankman-Fried did, and Anthropic's leadership grew noticeably reticent about it. In March 2025, Daniela told Wired she was only dimly aware of the movement. The New Yorker found that "strange coming from someone who both employs an icon of the movement, Holden Karnofsky, and is married to him." On the EA Forum, Ben West, an affiliate of the movement, put the cost of that reticence more gently and more pointedly: "if everyone takes funding, staff, etc. from the EA commons and don't share that they got value from that commons, the commons will predictably be under-supported in the future."
This is not a scandal so much as a tell. The president of Anthropic is an operator, and operators manage exposure. A movement that had supplied her company with its earliest checks, its best early hires, and her husband had become, by 2025, a word that the White House's AI czar David Sacks could use as an epithet. He called Anthropic part of a "doomer cult." She did what a risk manager does with a correlated liability: she reduced the visible position. Whether that was prudence or evasion depends on where you stand, and it is probably both.
The Quarantine
In the summer of 2022, Anthropic finished training the first version of Claude. Then it waited.
The decision is the single most cited fact in the company's mythology, and it deserves its prominence. The model worked, and worked well enough to be interesting. Anthropic could have shipped it and possibly owned the consumer chatbot market before anyone else arrived. Dario kept it, as the New Yorker put it, "under quarantine for further monitoring." His explanation was blunt: "I could see that there was going to be a race around this technology—a crazy, crazy race that was going to be crazier than anything. I didn't want to be the one to kick it off."
In late November 2022, OpenAI released ChatGPT. It reached a hundred million users in two months. The race started anyway, and Anthropic was standing at the line holding a model it had chosen not to run.
Claude reached approved users on March 14, 2023, followed by Claude 2 for the public that July.
The early models were stranger than their eventual polish suggests, and Daniela tells these stories with a connoisseur's pleasure. One prototype insisted that the best way to lose weight was an all-potato diet. Another slipped into something the team called "dragon mode." According to Dario, that version put on emoji sunglasses and behaved like "an unhinged Elon Musk character." When the team over-tuned for harmlessness, the model went the other way. Asked who the thirty-fourth president of the United States was, it answered with something like: I'm concerned about you … here's a therapy link.
The anecdote is funny. It is also the whole problem of the company in a single exchange. "You can have a perfectly harmless model today," she has said. "It just wouldn't be very helpful. The art is raising the watermark on all three together." The three were helpful, honest, and harmless: the framework that the early research team, with its novelists' children and its philosophers, tried to write into a character called the Assistant. Their first attempts at dialogue, the New Yorker reports, lacked a certain je ne sais quoi. In one exchange, a human asks the Assistant whether it is really important to salt spaghetti water.
By 2026, Amanda Askell, a philosopher, was supervising what employees unofficially called Claude's "soul document," later published as Claude's constitution. Claude was told to think of itself as "a brilliant expert friend everyone deserves but few currently have access to," and to have the modesty to recognize that "it doesn't always know what's best for them." That is the language of a trust-and-safety organization that has read a great deal of moral philosophy. It is also the language of a risk manager who learned at a payments company that the most dangerous system is one that tries to please everyone.
Did the quarantine cost Anthropic the consumer market? Probably. By 2026 the company still trailed OpenAI with consumers while dominating the enterprise. But the delay purchased something harder to price. It turned a safety claim into a costly signal, and costly signals are the only kind that sophisticated buyers believe. The people who would later write nine- and ten-figure checks for Claude were not teenagers looking for a homework helper. They were general counsels and chief information security officers. To them, we sat on a working product for months because we weren't sure it was safe is not a weakness in the sales deck. It is the sales deck.
The Other End of the Lever
The siblings have a story about how they divide the company, and they tell it consistently.
Dario looks five to ten years ahead. Daniela's job is to take what he sees and make it real within one to two years. "How do we take these incredible technical ideas," she has asked, "and turn them into something people can use today?" In Fortune's December 2025 cover story, the split was rendered as Dario on vision, strategy, research, and policy, and Daniela on day-to-day operations and the commercial side. She described the two of them as "yin and yang" on responsibilities and extremely aligned on values. Then she added the part that only works between siblings: she can tell the chief executive of a near-trillion-dollar company that he is being cranky, and he has to hear it.
"Dario and I," she has said, "have been fighting and getting over it for over forty years." That is a slight arithmetic flourish from someone born in 1987, but the point stands. Her other piece of co-founder advice sounds like a joke and is actually a stress test: before starting a company with someone, go on vacation together. If the relationship can't survive a shared hotel room, it won't survive a board meeting.
Her own self-description is unusually specific. She has said that after more than a decade of performance reviews from managers, peers, and direct reports, the themes never changed. She is supportive but demanding. She is strong on the big picture and weak on deep technical expertise. Early on, this read to her as criticism to overcome. Later she began to treat it as a map. She calls performance reviews "a huge gift," though not, she concedes, a "sexy" one, and she talks about "zones of competence and genius." In this framework, a weakness is not something to fix. It is a boundary to build along.
The arrangement she and her brother arrived at is the purest expression of that idea. He manages one person. She manages everyone else.
It is worth sitting with how odd that is. In most companies, the founder-CEO's span of control is the bottleneck through which everything passes. At Anthropic the bottleneck was deliberately moved, so the visionary could stay a visionary. That means writing essays like "Machines of Loving Grace" (October 2024) and "The Adolescence of Technology" (January 2026), warning Davos that AI could eliminate half of entry-level white-collar jobs, and making predictions that frighten people. Meanwhile the operator absorbs every operating consequence of the vision.
The arrangement has a cost, and it surfaced in public in the spring of 2026. After the Pentagon blacklisted Anthropic, Dario sent an internal memo that leaked almost immediately. In it he said the Pentagon had targeted Anthropic for not giving "dictator-style praise to Trump (while Sam has)," accused Altman of "straight up lies," and called OpenAI employees a "gullible bunch." An anonymous shareholder told the New York Post that this was part of an "extremely concerning" pattern. "You are a f—king CEO who has raised billions of dollars," the shareholder said. "You can't just rant and expect all shareholders to have the same mentality that you have." According to the same source, Dario has a habit of speaking out despite internal efforts, "usually by his sister and fellow cofounder Daniela and policy chief Jack Clark," to restrain his impulses. "They try to reel him in but he clearly cannot control his emotions."
Discount the anonymous shareholder by whatever you think anonymous shareholders deserve. The shape of the claim still matches everything else on the record. One sibling is the voice. The other is the breath control. A flutist knows that the pressure has to come from somewhere, and that the art is in what you do with it before it reaches the air.
The Most Scared Person in the Room
The numbers make the shape of the problem plain.
In 2023, Anthropic's revenue was an estimated $100 million. That year it tripled in size and raised $7 billion. By May 2024 it was valued at $18.4 billion. Amazon put in $4 billion, then $4 billion more, and made Anthropic's models available to its cloud customers. Google committed $2 billion, then another $1 billion. In March 2025 a $3.5 billion Series E led by Lightspeed valued the company at $61.5 billion. In July 2025, Claude reportedly reached $4 billion in annualized revenue. By that fall, Fortune was projecting $10 billion in 2025 revenue, a $183 billion valuation after a $13 billion Series F, and profitability by 2028. That would be a shorter path than any of its rivals had.
In February 2026 the company raised $30 billion at $380 billion. Claude Code alone had passed a $2.5 billion run rate that month. By May, by one account, run-rate revenue had gone from $9 billion to $47 billion in five months. In January 2024 it had been about $87 million. The New Yorker reported that Anthropic had roughly 1,300 employees in May 2025 and nearly double that by year's end.
Inside this, Daniela Amodei has positioned herself as the brake that is still pressed while the car accelerates. "I have probably been the leader who's been the most skeptical and scared of the rate at which we're growing," she told Fortune's Jeremy Kahn. She has been "continually, pleasantly surprised" that the company hasn't come apart at the seams, culturally or operationally.
— Daniela Amodei, to Fortune, December 2025I have probably been the leader who's been the most skeptical and scared of the rate at which we're growing.
Fear is the operator's honest emotion. Visionaries are permitted optimism because someone else is counting the chairs. The mechanism she credits for holding the place together is hiring, and specifically a culture interview that Anthropic has run since its earliest days. It screens for communication, emotional intelligence, and kindness alongside technical ability. "When we look to hire people at Anthropic today," she said on ABC in February 2026, "we look for people who are great communicators, who have excellent EQ and people skills, who are kind and compassionate and curious and want to help other people." The bet is not sentimental. At a company whose product can be misused catastrophically, an employee who won't push back on a bad call is not a culture problem. That employee is a safety problem.
The evidence that the filter works is circumstantial but real. In the summer of 2025, while Mark Zuckerberg was raiding rival labs with enormous offers, Anthropic engineer Sholto Douglas told the New Yorker that a number of his colleagues "could've taken a fifty-million-dollar paycheck." He said the "vast majority" hadn't even bothered to respond. Then the talent began to flow the other way. Jan Leike, OpenAI's former head of alignment, joined in 2024. So did OpenAI co-founders John Schulman and Durk Kingma. Mike Krieger, co-founder of Instagram, came to run product. In May 2026, Andrej Karpathy joined. In June 2026, John Jumper, a Nobel laureate, left DeepMind for Anthropic.
The commercial strategy turned out to be the safety strategy wearing a suit. Consumer businesses, Dario has argued, monetize attention, and attention incentives push toward slop. Enterprise customers buy the opposite. They don't want a model that gives someone bioweapon instructions, leaks company secrets, or spews hate speech in a customer-service window. "Companies do not want unpredictable or unreliable AI," Daniela has said. "Safety is also beneficial for business." In February 2026, Anthropic made the argument in the most American venue available: two ads during Super Bowl LX, part of a campaign called "A Time and a Place." In each spot, an AI assistant abruptly starts pitching a fictional product in the middle of a conversation. The punchline was a promise that Claude would stay ad-free, a direct jab at OpenAI's decision to put ads in ChatGPT's free tier.
Then there is the paradox the company would prefer to discuss less.
In January 2026, unsealed filings in Bartz v. Anthropic revealed a confidential operation called Project Panama. An internal planning document described it as the company's "effort to destructively scan all the books in the world." Anthropic had bought millions of used books from online retailers such as Better World Books, sliced off the spines, scanned the pages to train Claude, and recycled the paper. The document noted that the company did not "want it to be known" that it was working on this. Before Panama, the company had downloaded more than seven million pirated copies. Judge William Alsup ruled that the destructive scanning of legally purchased books was fair use and the piracy was not. In September 2025, Anthropic agreed to pay authors $1.5 billion, about $3,000 per book, the largest known settlement in a U.S. copyright case.
There is no evidence on the record of the president's personal role in any of this, and none should be implied. But she runs the company day to day. She is a literature major who says she would choose literature again, the daughter of a woman who spent her career keeping books housed and reachable. The machine she runs was built partly by cutting the spines off millions of books. Nobody in the company appears to have written the sentence that connects those facts. It writes itself anyway.
Hold Light and Shade
Anthropic has a list of values, and one of them, by Daniela's account, is a phrase borrowed from painting or music: hold light and shade.
— Daniela Amodei, on Sixth Street's It's Not Magic podcast, 2026Hold light and shade is one of our values, and it's this concept that the technology itself has great risk and great opportunity, and that that is a very complicated, complex thing to hold. We have to consider and talk about all of the risks, all the things that could go wrong, while also considering thinking about and building towards a world where all the things go right.
A flutist would call this dynamics. It is the capacity to play loud and soft in the same phrase without losing the line. In corporate terms, it means doing something most companies avoid: describing your own product's dangers while you sell it. "It's a sort of strange thing for a company that is building the technology to do," she told Wired's Steven Levy in December 2025, after teasing him that he always opens with "really light, easy questions." "But we feel like it's the right thing to do."
Her version of responsibility is deliberately broad, what she calls "radical" responsibility. It covers "the big stuff," like preventing AI from helping someone build a weapon of mass destruction. It also covers the daily stuff, "like user wellness and child safety." Her favorite cautionary tale is not a science-fiction apocalypse. It is social media. Those companies, she has said, did not set out to cause eating disorders in teenage girls; they simply wanted to grow quickly. In AI, she argues, there is a rare privilege in being able to say: let's not repeat the previous generation's mistakes.
The phrase came under real pressure in February 2026.
In July 2025, Anthropic had signed a two-year, $200 million contract with the Department of Defense. Through Palantir, Claude was integrated into classified intelligence networks. It was, as of early 2026, the only AI model used in classified missions. The contract carried two restrictions: no mass surveillance of Americans and no fully autonomous weapons. Anthropic did not object to military use in general. The company supported "lawful foreign intelligence and counterintelligence missions" and partially autonomous weapons of the kind used in Ukraine.
On February 24, 2026, Defense Secretary Pete Hegseth gave the company until February 27 to permit unrestricted use of Claude. Otherwise he would cut it out of the defense supply chain or invoke the Defense Production Act. On February 26, Anthropic said no. On February 27, President Trump ordered federal agencies to stop using Anthropic's models. Two days later it was reported that the military had used Claude, through Palantir's Maven Smart System, in the opening strikes on Iran. On March 5 the Pentagon formally designated the company a supply chain risk. In March, leaders from Google, Amazon, Apple, and Microsoft backed Anthropic's lawsuit. On March 26 a federal judge issued an injunction: "This appears to be classic First Amendment retaliation."
The red lines were not invented in a crisis. They were written into a contract months earlier, which is what operators do with principles they intend to keep. Daniela's explanation, given that July, was that the company chose fully autonomous weapons and mass domestic surveillance as "the areas where we simply can't say yes" because "the technology is so new that we do not believe the laws may be perfectly designed" to account for it. The framing is a risk manager's, not a moralist's. Draw the line where the rulebook is thinnest, because that is where your own judgment is the only control in the system.
In June 2026 the pressure ran the other way. On June 12, citing national security concerns from U.S. authorities, Anthropic suspended all access to its newest models, Claude Fable 5 and Mythos 5. The ban lifted on June 30, and access was restored July 1. Earlier that month, Daniela had told Bloomberg's Shirin Ghaffary that the company's march toward public markets should be understood as the operating reality of frontier AI. Training and inference are expensive, and capital access is survival. She also said the company was trying to scale cautiously: buying only compute it could use, widening access to its most powerful models only after defenders had a head start, and keeping its red lines in national-security work.
In September 2026, Anthropic announced it would give independent evaluators permanent, employee-like access to its models so they could verify whether the company was keeping its own safety commitments. Dario compared it to having a food inspector in the kitchen.
All of these choices have the same structure. Someone outside the company is given a place inside it. A rule is set before it is needed. A building is opened to inspection while work goes on inside.
Daniela has talked about San Francisco in similar terms. She is a native, and she likes to say Anthropic was built in her hometown. When the company signed a lease for an entire downtown building, she described the city's longtime residents as holding technology companies to account in a way she does not consider bad. "They're like, hey, we're providing a home for you," she said. "We just want you to be good citizens."
The library on Kittredge Street in Berkeley closed in the late 1990s so the building could be remade. A project manager named Elena Engel and the library staff spent eight months searching for a temporary place to keep the books while the walls came down.
Part IIThe Playbook
What follows is an attempt to extract the operating logic beneath Daniela Amodei's career. These are the decisions and structures that recur from a Ugandan field site to a Pentagon standoff. They are not personality traits or slogans. They are moves that an operator, founder, or investor can actually make. Each principle is grounded in the record above. Several are uncomfortable, and at least one cuts against her.
The principles
- 1.Build the org chart around your performance reviews.
- 2.Split the time horizon, not the turf.
- 3.Stand in the gap the experts can't see.
- 4.Treat the culture interview as a safety control.
- 5.Make your restraint expensive enough to be believed.
- 6.Sell the principle to the buyer who already needs it.
- 7.Draw red lines where the rulebook is thinnest.
- 8.Be the most frightened person about your own growth.
- 9.Leave toward something.
- 10.Install a sibling, in some form.
- 11.Say the shade out loud while you sell the light.
- 12.Manage your correlated liabilities, and know what it costs.
A
The Generalist's Path
A career that looks random forward and inevitable backward.
2009
Graduates summa cum laude in English literature from UC Santa Cruz, on a partial flute scholarship, into the financial crisis.
2010
Business development at the IRIS Center, University of Maryland: grants and recruiting for poverty, microfinance, and cash-transfer projects.
2011
Fellow at Conservation Through Public Health in Uganda; co-leads training for 50+ community health workers.
2012
Deputy field director, then field director, for Matt Cartwright's successful congressional campaign; recruits 80+ volunteers.
2013
Four months in Cartwright's House office, then joins Stripe as roughly employee #40, as lead technical recruiter.
2016
Risk manager at Stripe for core operations, user policy, and underwriting; first hands-on work with ML fraud and credit models.
2018
Joins OpenAI as roughly employee #50; engineering manager and VP of People during GPT-2.
2020
VP of Safety and Policy at OpenAI; leaves in December with six colleagues.
2021
Anthropic incorporated January 26; raises $124M in May.
2026
All senior leaders but one report to her; company valued at $965B; confidential IPO filing June 1.
Principle 1
Build the org chart around your performance reviews
The standard advice to founders is to shore up weaknesses: learn finance, take the coding course, become the complete executive. Amodei did the opposite. After more than a decade of reviews that said the same things (supportive but demanding, big-picture, not deeply technical), she stopped treating the pattern as a to-do list and started treating it as a blueprint. Her "zones of competence and genius" framing is not self-help. It is a design rule for an organization. Each role should be drawn so that the person in it is operating where their reviews say they are strong, and the adjacent role should be drawn to cover where their reviews say they are weak.
Anthropic's leadership structure is the extreme case. The CEO's reviews, one assumes, would say he is visionary, scientifically brilliant, and not built to run a staff meeting. So he doesn't run one. He manages a single chief of staff. Her reviews say she is a builder of structures who is not a technician. So she runs the structure and leaves the technical direction to others. The result looks strange on paper and has proved remarkably durable through a growth curve that should have broken it.
The underlying insight is that self-knowledge only becomes leverage when it is encoded in someone else's job description. Knowing you are bad at details is useless if no one with real authority is good at them.
Tactic: Pull your last five years of written feedback, find the three themes that never change, and redraw your role and one adjacent hire around them this quarter.
Principle 2
Split the time horizon, not the turf
Most co-founder splits divide territory: you take product, I take sales. Territory splits break down at the borders, where nearly every important decision lives. The Amodeis split by time. Dario looks five to ten years out. Daniela makes it real in one to two. Disagreement is minimized not because they think alike but because they are rarely arguing about the same moment.
A horizon split also makes handoffs legible. When the long-horizon partner produces an idea, such as the claim that a "country of geniuses in a data center" is coming, the short-horizon partner's job is not to judge whether it's true. Her job is to ask what it implies for hiring, compute purchasing, and contracts in the next eighteen months. Vision becomes a set of inputs rather than a source of conflict.
Y
Yin and Yang, Operationalized
How the Amodei siblings divide a near-trillion-dollar company.
| Dimension | Dario Amodei (CEO) | Daniela Amodei (President) |
|---|---|---|
| Time horizon | 5–10 years | 1–2 years |
| Primary domains | Vision, strategy, research, policy | Operations, commercial, people, product strategy, go-to-market |
| Direct reports | One chief of staff | The rest of the senior leadership team |
| Public register | Essays, warnings, Davos, leaked memos | Podcasts, hiring philosophy, measured interviews |
| Characteristic risk | Unfiltered speech | Under-recognition |
Tactic: If you have a co-founder, write down in one sentence each the time horizon you own and the one they own, and route the next contested decision to whichever horizon it actually lives in.
Principle 3
Stand in the gap the experts can't see
"Training AI models was not my responsibility," she has said, "but I could leverage my strengths in areas where researchers were less experienced." This is the generalist's real competitive advantage, and it is narrower and more useful than "be curious." In any expert organization there are gaps between the experts: in hiring, risk, policy, and the translation of research into something a customer can buy. Experts don't see these gaps because they are not inside any one expertise. The generalist sees them because she is not inside any expertise either.
Her route through Stripe and OpenAI follows that logic. At forty people, Stripe's gap was hiring. At three hundred, it was risk. At OpenAI, the gap was people management and business operations for a lab that had never needed them, and then safety policy for models that had outgrown the lab's governance. Each move was into the space that the dominant expertise of the place had left empty. Her method for getting oriented was disarmingly simple: "I just kept asking questions until I felt like I could understand it."
Tactic: List the three functions in your organization that no one with deep expertise owns, then volunteer for the one closest to where the company will break first.
Principle 4
Treat the culture interview as a safety control
Plenty of companies interview for "culture fit," and most of them use it as a polite name for homogeneity. Anthropic's culture interview, which has run since the company's earliest days and screens for communication, EQ, kindness, and curiosity, rests on a sharper idea. In an organization whose products can do serious harm, the most important safety mechanism may be employees who will say no to a bad decision, and who can say it in a way that gets heard.
The returns show up in odd places. When Zuckerberg was offering tens of millions to poach researchers, most Anthropic recipients didn't reply. When the Pentagon demanded unrestricted use, the company held a line that its own employees would have found intolerable to cross. Retention and resolve are not separate outcomes of a culture. They are the same outcome measured at different times.
Tactic: Add one interview stage that tests, explicitly, whether a candidate will push back on a senior person when they think a call is wrong, and weight it like a technical bar.
Principle 5
Make your restraint expensive enough to be believed
Holding Claude in the summer of 2022 almost certainly cost Anthropic the first-mover position in consumer chatbots. That cost is exactly why the signal worked. Anyone can claim to prioritize safety. Only a company that has visibly paid for the claim gets believed by the people whose belief is worth money: enterprises, governments, and the researchers deciding where to spend their careers.
The same logic shows up later. A $200 million defense contract was put at risk rather than drop two restrictions. Super Bowl ads declared a business model, advertising, permanently off the table. Independent evaluators were given standing access to check the company's homework. Each of these is a commitment that would be embarrassing or expensive to reverse, and that is the source of its credibility.
Tactic: Identify the one principle your company claims most loudly, and find a concrete, public way to pay for it that a competitor making the same claim would not match.
Principle 6
Sell the principle to the buyer who already needs it
Safety was supposed to be Anthropic's handicap in a speed race. Daniela's commercial insight was that for a specific class of customer, safety is the product. A consumer chatbot competes for attention, and attention rewards surprise. An enterprise model competes for trust, and trust rewards predictability. "Companies do not want unpredictable or unreliable AI," she has said. "Safety is also beneficial for business."
The result was a company that trailed in consumer share and dominated the enterprise market. Its revenue grew from an estimated $100 million in 2023 to a reported $47 billion run rate by May 2026, with Fortune projecting profitability years before its rivals. The lesson generalizes. When your values impose a cost, look for the customer segment where the cost is a feature, and build your go-to-market there first.
Tactic: For each constraint your company imposes on itself, name the customer for whom that constraint is a purchasing criterion, and point your next sales hire at them.
Principle 7
Draw red lines where the rulebook is thinnest
Anthropic's refusal to the Pentagon was narrow by design. The company supported intelligence work, counterintelligence, and partially autonomous weapons. It drew exactly two lines: mass domestic surveillance and fully autonomous weapons. Daniela's rationale is the important part. The technology is so new that the laws may not be designed for it. Where external rules are mature, you can defer to them. Where they are immature, your own policy is the only control in the system.
This is a risk manager's framework, and it echoes her Stripe years. Underwriting policy matters most at the edges where regulation hasn't caught up. Few lines, held absolutely, are more defensible than many lines held loosely. They are easier to explain, easier to litigate, and harder to caricature. When the government retaliated, a judge saw the dispute as a free-speech case rather than a company being difficult.
Tactic: Write your company's red lines down before the crisis, keep them to fewer than five, and justify each by naming the specific gap in existing law or regulation it fills.
Principle 8
Be the most frightened person about your own growth
"I have probably been the leader who's been the most skeptical and scared of the rate at which we're growing." Founders rarely say this in public, and many never let themselves feel it. At Anthropic the role of chief worrier about growth sits with the person who owns operations, culture, and hiring. Those are the systems that fail first when headcount doubles in a year, as it did in 2025.
Fear assigned to the right seat is productive. It produces the questions that optimism skips. Can we onboard this many people without diluting the culture interview? Is this compute purchase something we can actually use? Should this powerful model reach defenders before it reaches everyone? Her stated approach to scaling, buying compute it can use and widening access only after defenders get a head start, is fear converted into policy.
Tactic: Name one senior person whose explicit job is to be skeptical of your growth rate, and give them a standing agenda slot to present what will break next.
Principle 9
Leave toward something
"We were running towards something versus running away from something." The phrasing matters because the two kinds of departure produce different companies. A team that leaves in flight defines itself against its former employer and tends to rebuild the old place's structure in mirror image. A team that leaves toward something must specify, on day one, what it is building.
Anthropic's first year shows the difference. There were six teams, most working on safety. There was no sales function for roughly eighteen months. There was a public benefit corporation structure and a pledge to give away 80% of the founders' wealth. These are affirmative design choices, not reactions. Even so, the critique still lands. The New Yorker observed that Anthropic "sounded an awful lot" like what OpenAI had promised to be. Running toward something does not exempt you from becoming the thing you left. It only gives you a written standard to measure yourself against.
Tactic: Before any departure or spinout, write a one-page charter that describes what the new organization will do without mentioning the old one once.
Principle 10
Install a sibling, in some form
The most unusual feature of Anthropic's governance is not legal. It is familial. Daniela can tell the CEO he is being cranky, and he has to hear it. Reports suggest that she and Jack Clark try, not always successfully, to temper his public statements. No board member or investor holds that kind of standing. It comes from forty years of fighting and getting over it.
Most founders don't have a sibling to install. They can still build the function: a relationship with enough history and enough security that candor costs nothing. The leaked spring 2026 memo and the shareholder anger that followed show what happens when even that channel is overwhelmed. They also suggest the price of having no such channel at all.
Tactic: Identify the one person who can tell you, privately and without consequence, that you are wrong in tone, and set a recurring time when they are expected to do it.
Principle 11
Say the shade out loud while you sell the light
"Hold light and shade" is a company value that forces an unusual communications posture. Anthropic describes the risks of its own product with the same seriousness it uses to describe the benefits. It publishes research showing that its models, under test, attempted blackmail roughly 96% of the time in one contrived scenario. Its CEO warns about job losses. Its president says that very few jobs can be done by AI alone, while predicting that the way people work will change.
The counterintuitive payoff is credibility. A company that only describes upside is trusted only until the first incident. A company that has already described the downside has, in effect, pre-registered its worries. When something goes wrong, it looks consistent rather than caught. "It's a sort of strange thing for a company that is building the technology to do," she has said, "but we feel like it's the right thing to do."
Tactic: In your next product announcement or investor update, include one paragraph describing a specific risk of what you've built and what you are doing about it.
This principle is drawn partly against her. After Bankman-Fried's fall, effective altruism became a political liability, and Anthropic's leadership reduced its visible exposure. Daniela told Wired she was only dimly aware of a movement that had funded her company, staffed it, and included her husband. As risk management, the move is legible. As candor, it was costly. Critics inside the movement called it out, and so did the New Yorker.
The general lesson is that every operator holds correlated positions: investors, networks, and ideologies whose reputations move with yours. Reducing exposure is often wise. Disavowing too completely can damage trust more than the original association did, because people notice when the story doesn't match the facts. The skill is in acknowledging what you got from a commons while making clear what you do not endorse.
Tactic: List the three affiliations that would embarrass your company in a downturn, and draft now an honest sentence for each that acknowledges what you took from it.
Free playbook
Get The Business Model Playbook
58 business models, one visual page each: how the money flows, the metrics that matter, and who runs it. Free when you join the Faster Than Normal email.
Free. No spam. Unsubscribe anytime.
Part IIIQuotes and Maxims
In their words
— Daniela Amodei, Stanford GSB View From the Top, 2026If you were to look through my background, you would be like, "What is this lady actually good at?" She doesn't have a law degree. She's not a computer scientist.
— Daniela Amodei, Colossus Review, November 2024I'm an expert in identifying the right people to solve a specific problem and creating a structure for them to succeed.
— Daniela Amodei, Stanford GSB View From the Top, 2026We were running towards something versus running away from something. We had this vision in our heads of wanting to create an organization where the values that matter to us around safety and around responsibility were at the forefront of what we were doing.
— Daniela Amodei, to ABC News, February 2026The things that make us human will become much more important instead of much less important.
— Daniela Amodei, interview with Stripe, July 2023We have to internally ask ourselves, why did we do this? Why did we start Anthropic? Which is that we want to help make AI systems safer from day one.
Maxims
- Design around the review. The feedback that never changes is not a flaw to fix but a boundary to build the org chart along.
- Split by time, not territory. Co-founders who own different horizons fight less than co-founders who own different departments.
- Find the gap between experts. A generalist's leverage lives in the work no specialist is positioned to see.
- Hire for pushback. In a company that can cause harm, the employee who will say no is a safety system.
- Pay for your principles. Restraint is believed only when it visibly costs something.
- Sell constraints to those who need them. Every self-imposed limit is a purchasing criterion for some customer.
- Few lines, held absolutely. Draw red lines where the law is newest, and don't negotiate them.
- Assign the fear. Someone senior should be the most scared person in the building about growth.
- Leave toward a charter. Define the new thing without reference to the old one, then measure yourself against it.
- Hold light and shade. Name the risks of what you build in the same breath you name its promise.
