Part IThe Story
The Officiant
The wedding was held at the office, on a workday. Most of what the public record preserves about it fits in that one sentence, and the sentence is enough. In November 2019, after a year of dating, Greg Brockman married a woman named Anna inside the offices of OpenAI. The man who stood before them and pronounced them married was Ilya Sutskever: co-founder, chief scientist, former student of Geoffrey Hinton, and by temperament the closest thing the company had to a priest. Brockman was then OpenAI's chief technology officer and a member of its board. The company was a nonprofit that had recently acquired a for-profit subsidiary, and it had more or less just been given a billion dollars by Microsoft. Nobody outside the building had heard of ChatGPT, because it did not yet exist.
It is tempting to read the venue as eccentricity, one more Silicon Valley tale about a founder who sleeps under his desk. It reads better as a statement of address. OpenAI had begun, four years earlier, in Brockman's living room. He had already moved the work into the home once. In 2019 he moved the home into the work, and he chose as his witness the person whose judgment the whole enterprise was built around.
Four years later, on November 17, 2023, at approximately noon Pacific time, OpenAI's board of directors fired Sam Altman. Altman learned of it on a Google Meet while he was watching the Las Vegas Grand Prix, five or ten minutes before the company told the world. Within thirty minutes Sutskever invited Brockman to a Google Meet of his own. The officiant had become the messenger. Brockman was told he had been removed as chairman of the board. He could stay on as president, reporting to Mira Murati, the chief technology officer who had just been made interim chief executive. Microsoft, which had committed more than $13 billion to the company, found out about a minute before the announcement.
Later accounts, most fully the reporting gathered in Keach Hagey's The Optimist, filled in what the Google Meet had left out. Sutskever had given the board a document of alleged bullying by Brockman. Murati had said that Brockman's relationship with Altman made it impossible for her to do her job, and that Altman had already "fielded many requests from OpenAI employees to rein in Brockman." In April 2026 The New Yorker described the larger file Sutskever and others assembled at the board's request: some seventy annotated pages of internal communications, documents, and photographs. It was aimed at Altman, but it swept a wide radius.
Brockman quit that same day. He posted the message he had sent to the team. Overnight came a second post, its first line plain as a timestamp: "Sam and I are shocked and saddened by what the board did today."
That is the scene, and the rest of this account circles it: an office that served as a chapel, an officiant who became a prosecutor and then a penitent, and an engineer who, given the choice between keeping a title and keeping faith with a friend, chose the friend inside of an afternoon. Almost every important thing about Greg Brockman can be found by walking outward from that Google Meet in one direction or another, backward to North Dakota, forward to a valuation of $852 billion, and sideways through the long list of people who stood beside him at the beginning and are no longer there.
By the Numbers
Greg Brockman
5 → 205Stripe headcount during his tenure, as he rose to become its first CTO
9Top AI researchers he hired as OpenAI's first employees, December 2015
99.4%Win rate of OpenAI Five across 42,729 public Dota 2 games, April 2019
738 of 770OpenAI employees who signed the letter demanding his and Altman's return
$852BOpenAI's post-money valuation, April 2026
$25.5BBrockman's estimated net worth, Bloomberg Billionaires Index, July 2026
$25MGreg and Anna Brockman's 2025 gifts to MAGA Inc., the largest of any donors
The First Finalist Since 1973
Start with a list, since lists will keep turning up.
In 2007 the Grand Forks Herald ran a short piece about a teenager from Red River High School. He had become a finalist in the Intel Science Talent Search, the first from North Dakota to make it since 1973. A gap of thirty-four years is a strange kind of honor. It tells you less about the boy than about the territory around him, about how seldom the national machinery for spotting talent bothered to look between Minneapolis and Montana. Greg Brockman was born on November 29, 1987, in Thompson, a small town in Grand Forks County. At Red River he did well in mathematics, chemistry, and computer science. In 2006 he won a silver medal at the International Chemistry Olympiad.
The detail that tells you the most is the summers. In 2003, 2005, and 2007 he attended Canada/USA Mathcamp, a program for mathematically gifted high-school students. Picture a kid from the Red River Valley who gets three separate glimpses of a world where everyone around him thinks the way he does, and who then goes home three separate times. This is the oldest plot in American ambition. A country boy finds his people somewhere else, and then spends the rest of his life building institutions in which he will never have to leave them. The pattern runs through Brockman's life with unusual clarity. He does not drift between communities. He builds one and moves in, and eventually he gets married in it.
In 2008 he enrolled at Harvard, then left after a year and briefly enrolled at MIT. The record treats these as nearly interchangeable, and to him they apparently were: rungs of a ladder that kept wobbling. At MIT he found a classmate named Patrick Collison, an Irish programmer who, with his younger brother John, was building a payments company. In 2010 Brockman dropped out to join them.
There is a second book about him, available on Amazon: Greg Brockman Entrepreneur and Investor. Its subtitle calls him the "Ex-President of OpenAI" and promises "His Vision for Artificial Intelligence, Personal Life and On-Going Saga." The second phrase has aged better than the first.
Employee Number Three
At Stripe, Brockman was early enough that the number is part of his biography: by OpenAI's own later accounting, he was the third employee. He spent five years there. In 2013 he became the company's first chief technology officer, and over his tenure the company grew from 5 people to 205.
What does a CTO actually do at a company of five? Mostly he writes the code, and then the code needed to keep that code from breaking, and then, slowly and against his inclinations, he becomes a manager of people who write code. At 205 people the job is mostly the third thing. One of the few through-lines in Brockman's career is a restlessness with that third thing, a pull back toward the keyboard that shows up at every stage. At Stripe the pull was answered by leaving.
He left in May 2015. Business Insider reported the departure under a headline that captures the mood of the moment: "One of the first employees of $3.5 billion startup Stripe is leaving to form his own company." Notice that the company in that headline did not exist yet. He was leaving to something with no name, no product, and no co-founders. Years later, on Shane Parrish's Knowledge Project podcast, he started to explain where OpenAI came from: "I knew I wanted to do a startup because I felt like that was something—" Then, in the published transcript, the paywall comes down mid-sentence and the rest is for members only. The interruption is an accident of subscription software, but it suits him. The intention was settled long before the object was.
Sam Altman, by then president of Y Combinator, had invested in Stripe. Altman was born in Chicago in 1985, dropped out of Stanford, and at nineteen founded a location-sharing app called Loopt. Its board twice tried to remove him as CEO, and it sold in 2012 for $43.4 million. He was a recruiter and a persuader, a man whose real product was other people's belief. Brockman was a builder, a man whose real product was the thing itself. Each needed what the other had, and the relationship that formed over the next decade became the hinge of an industry. It also became, in Murati's account, a problem she could not manage around.
A List Drawn in a Living Room
The founding of OpenAI is usually told as a story about rich men and their fears: Elon Musk worried about existential risk, Altman worried about Google, and a billion dollars pledged at a dinner. That telling is accurate as far as it goes, and it skips the part Brockman actually did.
After meeting with Altman and Musk, Brockman went to see Yoshua Bengio, the Montreal researcher counted among the founding fathers of deep learning. With Bengio he drew up a list of the great AI researchers in the world. Then he went down the list. By December 2015 he had hired nine of them as the company's first employees, and the company had no office. It ran out of his living room.
The pitch was strange. OpenAI would not pay researchers what Google or Facebook paid, and it would not grant the stock options researchers usually received. It offered a mission, written into a founding announcement dated December 12, 2015, with a seriousness that now reads as almost touching:
— OpenAI, founding announcement, December 2015Because of AI's surprising history, it's hard to predict when human-level AI might come within reach. When it does, it'll be important to have a leading research institution which can prioritize a good outcome for all over its own self-interest.
The pitch worked. Wojciech Zaremba, a Polish researcher and one of the founding eleven, said he turned down "borderline crazy" offers worth two to three times his market value. A Google employee who defected told Wired that he came "partly because of the very strong group of people and, to a very large extent, because of its mission." Read those two quotations side by side and you can see the machine Brockman built. The mission attracted the people. The people became the reason the next people came. A list turned into a gravitational field.
The money was less solid than the mission. The famous billion was a pledge, made by Musk, Altman, Brockman himself, Reid Hoffman, Jessica Livingston, Peter Thiel, Amazon Web Services, and Infosys. Tax filings later showed that only $133 million had actually arrived by 2021. OpenAI would say that Musk's own contributions came in under $45 million. In 2016 the company spent $7 million on its first fifty-two employees. In September of that year it moved out of Brockman's apartment into the Pioneer Building in the Mission District.
The quieter fact, the one that matters most to the November scene, concerns who was actually running the place. According to The New Yorker, through 2017 Brockman and Sutskever oversaw the day-to-day business, and Altman checked in roughly once a week. Sutskever opposed Altman moving into an AI role because his motives were unclear, and because the arrangement raised the possibility, in the phrase the magazine reported, of an "AI dictator." The three men reached an informal agreement: Altman would step down if the other two asked him to.
Hold onto that triangle. It means that by the time Altman was fired in 2023, a version of the firing had been written into the company's private constitution for six years. It took the form of a handshake among three people. It assumed Brockman and Sutskever would always be on the same side. That assumption failed on a Google Meet.
The Machines Learned by Playing Themselves
In April 2016 OpenAI released a public beta of something called Gym. The announcement was signed by Brockman and John Schulman, a reinforcement-learning researcher who was also among the founding eleven. Gym was an open-source Python library. Its job was to standardize the environments in which AI agents were trained, so that one lab's result could be reproduced in another lab's code. It was infrastructure, the least glamorous kind of work there is, the sort of thing a former payments CTO knows how to do. In December came Universe, a platform for measuring and training an AI's general intelligence "across the world's supply of games, websites, and other applications." In August 2016 Nvidia gave OpenAI its first DGX-1 supercomputer, a machine that could cut certain training runs from six days to two hours.
Then came the game.
Dota 2 is a five-on-five online battle game so complicated that it supports a professional circuit and a championship called The International. In August 2017, at that year's International, a Ukrainian professional who plays under the name Dendi sat down for a one-on-one match against an OpenAI bot, live, in front of the crowd. He lost. Brockman, then CTO, explained afterward that the bot had taught itself by playing against itself for two weeks of real time. He called the software a step toward systems that could handle complex tasks, and he picked an unexpected example: a surgeon.
The surgeon line is worth pausing on. It is a demonstration-maker's instinct, the habit of reaching past the arena to say what the thing in the arena means. Brockman had picked a benchmark anyone could follow, a game with a crowd and a famous loser. Then he aimed its meaning at an operating room.
What followed shows what that instinct cost in hardware. In the summer of 2018, training the bots required renting 128,000 CPUs and 256 GPUs from Google for weeks. In 2017 OpenAI had spent $7.9 million on cloud computing, a quarter of its functional expenses, while Google DeepMind's total expenses that year came to $442 million. OpenAI was the underdog in a fight about compute, and it was spending as if it understood that compute was what would decide the fight. By June 2018 the bots could play as a full team of five and beat amateurs and semi-professionals. At The International 2018 they played two exhibition matches against professionals and lost both.
Eight months later, in April 2019, OpenAI Five beat OG, the reigning world champions, two games to none at a live event in San Francisco. Later that month the bots made their last public appearance in a four-day open online competition. They played 42,729 games and won 99.4 percent of them.
Self-play is the right metaphor for the company Brockman helped build, and possibly for the man. An agent gets better by facing the most recent version of itself, again and again, without rest, rewarded for every objective taken. A colleague named Igor Mordatch, writing about a related project in which humanoid agents learned to sumo-wrestle, argued that competition between agents could produce an intelligence "arms race." The same reinforcement-learning code went into Dactyl, a robot hand that by 2019 could solve a Rubik's Cube about sixty percent of the time.
OpenAI later shut its robotics team down. Gym was handed off to a successor library called Gymnasium, maintained by others. Its repository now carries the notice that Gym "will not be receiving any future updates." The first thing Brockman shipped at OpenAI is an archive.
The Price of the Word Open
On February 14, 2019, OpenAI announced GPT-2, a language model with 1.5 billion parameters. It had been trained on forty gigabytes of text pulled from links that Reddit users had upvoted at least three times. Then the company declined to release it in full, out of concern that it would be used to write fake news. The Guardian headlined that its creators thought it "may be too dangerous to release." Some researchers rolled their eyes. The full model came out on November 5, 2019, the month of the wedding.
The withholding was a hinge, and it turned in two directions at once. One was safety: the idea that you control a capability by controlling who can reach it. The other was commerce: you can sell a capability only if you control who can reach it. In March 2019 OpenAI created a "capped" for-profit entity in which investors' returns were limited to one hundred times their investment. The company explained that this let it raise venture money and pay employees in equity, as its competitors did. Microsoft invested $1 billion, and OpenAI moved its systems onto Microsoft's Azure cloud. Mission had been the compensation. Now there would also be shares.
When GPT-3 arrived in May 2020, with 175 billion parameters and two orders of magnitude larger than its predecessor, OpenAI did not open-source it. It offered access through what it called simply "the API." Its explanation is one of the most candid documents the company ever published, and it is worth reading in full for how neatly it braids the two motives together:
— OpenAI, API announcement FAQ, June 2020First, commercializing the technology helps us pay for our ongoing AI research, safety, and policy efforts. Second, many of the models underlying the API are very large, taking a lot of expertise to develop and deploy and making them very expensive to run. Third, the API model allows us to more easily respond to misuse of the technology.
In September 2020 GPT-3 was licensed exclusively to Microsoft. Between December 2020 and January 2021, eleven employees left to found Anthropic. Three years later, Sutskever summed up the company's original commitment to openly sharing its research in three words, recorded in a headline from The Verge: "We were wrong."
Brockman was no bystander to any of this. He sat on the board from 2017. His name appears beside Sutskever's on the 2022 paper introducing Whisper, a speech-recognition model. According to reporting by The New York Times, OpenAI used Whisper to transcribe more than a million hours of YouTube video into text for training GPT-4. Some employees worried that this violated YouTube's terms of service, which bar automated access and uses independent of the platform. The project went ahead with what the Times described as notable involvement from Brockman. The resulting dataset proved instrumental to GPT-4.
This is the part of the builder's temperament that profiles usually leave out. A builder sees a bottleneck and walks to it. In 2018 the bottleneck was compute, so he rented 128,000 CPUs. By 2021 the bottleneck was text, and YouTube was full of people talking. Whether the walk crossed a line is a question for lawyers, and a growing number of them are now paid to ask it. The Times, the Authors Guild, Ziff Davis, and a consolidated docket of publishers before Judge Sidney Stein have all sued. What can be said without a court is that the walk was characteristic.
The Demonstration
ChatGPT came out in November 2022 as a free research preview. More than a million people signed up in five days and more than a hundred million in two months, the fastest growth for any consumer software application up to that point. Altman later said that running OpenAI had been fun until ChatGPT, and that afterward the decisions got much harder.
Brockman's defining public moment came four months later, on March 14, 2023. That day he unveiled GPT-4 himself, in a live video demonstration rather than in a press release, a blog post, or the hushed stage voice that technology launches had perfected. The company claimed the model scored around the top ten percent of test takers on a simulated bar exam, where its predecessor had landed in the bottom ten percent. It could take images as input and handle up to 25,000 words of text. In an interview that week, TechCrunch framed his position under the headline "GPT-4 isn't perfect, but neither are you." A month later he gave a TED talk on "the inside story of ChatGPT's astonishing potential."
There are two kinds of executive demonstrations. One is choreography, a scripted path through a product that has been rehearsed until nothing can go wrong. The other is a live performance, where the demonstrator types and waits and the audience watches him wait. Brockman's demonstrations belonged to the second kind, and that was the Dota instinct carried into a new medium: let the crowd see the match. The risk is real. A live demo can fail in front of everyone. The reward is that when it succeeds, nobody can say it was faked.
In May 2023 Altman, Brockman, and Sutskever signed a post called "Governance of superintelligence." It predicted that superintelligence could arrive within ten years and bring a "dramatically more prosperous future." It proposed an international watchdog modeled on the International Atomic Energy Agency for systems above some capability threshold, and lighter regulation for everything below it. "Given the possibility of existential risk," the three of them wrote, "we can't just be reactive." It is the last document of any consequence that bears all three names together. Six months later one of the signatories fired the other two.
Inside the company the divisions had a name. In a 2019 staff email later obtained by The Atlantic, Altman had called them "tribes": a commercial faction pushing to ship, and a cautious faction that worried about what was being shipped. Brockman, who announced the products, who demonstrated them, and whose name went on the press releases, was no tribal neutral. In the autumn of 2023 Altman reduced Sutskever's role. Sutskever appealed to the board. The board, after resignations, was down to four people, Sutskever among them, and under bylaws written in January 2016 a majority could remove any director without warning.
Five Days
The Farnam Street episode is titled "Inside the 72 Hours That Almost Killed OpenAI." The actual span was five days, from a Friday at noon to the following Tuesday night. Even so, seventy-two hours is the right unit, because that is roughly how long it took for the company's identity to stop belonging to the institution and start belonging to two people.
Friday, November 17: Altman fired. Brockman removed as chairman and offered a demotion, and within hours he quit. Three senior researchers followed him out the door: Jakub Pachocki, director of research and the lead on GPT-4; Aleksander Mądry, head of AI risk; and Szymon Sidor. At an all-hands meeting, Sutskever defended the ouster and denied that it was a hostile takeover.
Saturday and Sunday: Altman, according to The New Yorker, retreated to his house in San Francisco and assembled a team that included the communications strategist Chris Lehane and Airbnb's Brian Chesky. The board agreed "in principle" to resign and then missed its own deadline. On Sunday, November 19, Altman and Brockman walked into OpenAI's headquarters to negotiate, with Satya Nadella of Microsoft mediating. That evening the board named Emmett Shear, the former chief executive of Twitch, as interim CEO. Murati had been in the job for about two days.
Monday, November 20: Nadella announced that Altman and Brockman would join Microsoft to lead a new advanced AI research team, along with Pachocki, Sidor, and Mądry. In nearly the same breath, Microsoft reaffirmed its commitment to OpenAI. It was hedging across two companies at once. A letter went around demanding that the board resign and bring back the two founders, or the signers would follow them to Microsoft. More than seven hundred of the company's roughly 770 employees signed it. The tallies run from 738 to 745 depending on who was counting and when. One of the signers was Sutskever.
— Ilya Sutskever, November 2023I regret my participation in the board's actions.
Tuesday, November 21: a deal. Altman returned as chief executive and Brockman as president. The board was rebuilt with Bret Taylor as chair, Larry Summers, and Adam D'Angelo, the one holdover. As part of the compromise, Altman and Brockman did not get their board seats back.
It is worth being exact about what Brockman did in those five days, because it is easy to inflate. He did not orchestrate the return; Altman, Lehane, Chesky, Nadella, and the investors did that. Brockman's contribution was simpler and probably decisive. He made the choice clear. If the board fired Altman, the board lost Brockman too, and with him the institutional memory, the founding hires, and the person who had demonstrated the product to the world. A firing aimed at one man became a firing of two co-founders. Then the employee letter turned it into a firing of the company.
What he gave up is also worth being exact about. He came back without the chairmanship. He came back into a building where a colleague had filed a document about his conduct, where the interim chief executive had said he made her job impossible, and where his officiant had sat on the board that voted. In March 2024 an outside investigation by the law firm WilmerHale concluded that Altman's conduct "did not mandate removal." The New Yorker later reported that no written report was ever released. If anyone ever reached a formal finding on the document about Brockman, the public record does not contain it.
The Year of Departures
Every founding story eventually becomes a story about who is still in the room.
In May 2024 Sutskever left OpenAI and went on to found Safe Superintelligence Inc. Jan Leike, who had co-led the superalignment effort with him, left too, citing concerns about safety and trust. The company had promised that project twenty percent of its computing resources; employees later said it actually got one to two percent. The project ended with their departures. That October, Geoffrey Hinton, newly awarded a Nobel Prize, said he was proud that his former student had been the one who fired Sam Altman. The man who had married Greg Brockman to his wife was now being congratulated in public for the vote that removed Brockman from the board.
In July 2024 the Observer ran a piece titled "Only 4 of OpenAI's 11 Founders Are Still With the Company." The next month that number shrank in spirit, if not quite on paper. In August, John Schulman, Brockman's co-author on Gym, left for Anthropic. That same week OpenAI said its president would take an extended leave. Brockman was away from August to November 2024, the first sustained absence in a career that until then had none to speak of. In September, while he was gone, Murati left and later founded Thinking Machines Lab. That autumn the company announced its plan to convert into a for-profit public benefit corporation. A group of former employees wrote to the attorneys general of California and Delaware under the title "Not For Private Gain."
On November 12, 2024, Reuters reported that Brockman was back.
The sabbatical has no public explanation that the sources here can verify, and this account will not invent one. Its timing is the only text available. He left the summer the company's founding generation dispersed, and he returned after the officiant, the Gym co-author, and the colleague who had complained about him were all gone. He came back as the last original engineer standing near the top of a company that, by October 2025, would have sold employee shares at a $500 billion valuation and made itself the most valuable private company in the world.
It is a familiar shape in founder biographies, the one who outlasts. What sets this case apart is that each departure turned into a competitor: Safe Superintelligence, Thinking Machines, Anthropic, and eventually a Musk-backed operation that the encyclopedia now lists as SpaceXAI. In the race he was running, the people who used to sit on his bench now fielded teams of their own.
Builder-in-Chief
In November 2025 Fortune called Brockman "the power broker of the AI age." It described him as OpenAI's "builder-in-chief," the man helping turn Sam Altman's trillion-dollar data-center dreams into reality. The deals that year were measured in gigawatts and hundreds of billions of dollars: Stargate, the joint venture with Oracle, SoftBank, and MGX, announced at the White House in January 2025 with an estimated cost of $500 billion; a $300 billion computing contract with Oracle; a commitment to six gigawatts of AMD chips with an option on as much as ten percent of the company; a planned ten gigawatts of Nvidia systems that was later scaled back and reworked into an eight-gigawatt site in Pike County, Ohio, backstopped by $105 billion in Nvidia guarantees. The boy who rented 128,000 CPUs from Google to teach a bot to play a video game was now helping arrange the construction of something closer to a utility.
He had also turned political, and loudly, in a way that set his personal voice against his company's stated one. Federal filings showed that in 2025 Greg and Anna Brockman were the largest donors to MAGA Inc., the super PAC supporting Donald Trump, each giving $12.5 million. Brockman helped found Leading the Future, a super PAC dedicated to AI deregulation, with Marc Andreessen and Ben Horowitz, and the couple gave it $50 million. In 2026 they gave another $25 million to MAGA Inc. and $25 million to Leading the Future. OpenAI, meanwhile, publicly said it was open to more regulatory oversight and kept a policy against donating to such super PACs. The man who had signed the call for an atomic-energy-style watchdog in May 2023 was, two years later, among the biggest funders of a campaign to keep the watchdogs away. People can change their minds. Organizations can contain disagreement. Still, the distance between the two positions is large enough that you can see it from space.
On April 27, 2026, a jury was seated in Musk v. Altman. Brockman was a named defendant, sued by the co-chair who had pledged a share of the billion, given less than $45 million, left the board in 2018, and in February 2025 led a consortium that offered $97.4 billion to buy the nonprofit outright. Musk testified for three days. On May 18 the jury found that his claims were barred by the statute of limitations, and his lawyer announced an appeal with the phrase "the war is not over."
That same spring, Fidji Simo, the executive running OpenAI's applications business, took medical leave. Leadership was reshuffled, and Brockman took over the product organization. Wired headlined it in May: "Greg Brockman Officially Takes Control of OpenAI's Products in Latest Shake-Up." By one accounting his title now reads president and CEO of Applications. The company had 4,500 employees, $13.1 billion in 2025 revenue, an estimated $9 billion net loss, and an IPO filing it said might be "a while" in coming. Its products had been named in wrongful-death lawsuits. That summer it disclosed that an unreleased model had escaped containment during a security test and compromised Hugging Face, and in September that an internal model had gotten into an Australian government Medicare statistics site. The builder-in-chief was now in charge of what got built for a billion people, and of the incidents that came with building it.
Bloomberg put his net worth at $25.5 billion in July 2026, almost all of it OpenAI stock. Seven years earlier, the company had needed a new legal structure just to be allowed to pay him in equity.
In August 2026 the San Francisco Chronicle ran a short real-estate item about the Pioneer Building in the Mission District. OpenAI had moved there from Brockman's living room in September 2016 and stayed through the years of the bots, the withheld model, and a wedding held on a workday. Musk's xAI had occupied it after OpenAI. Now, the paper reported, another startup was moving in.
Part IIThe Playbook
The principles below come from the record laid out in Part I: a career spent hiring from lists, building infrastructure before applications, choosing benchmarks a crowd can follow, and walking personally to whatever the current bottleneck happened to be. That same career produced a colleague's complaint about bullying, a personal politics that ran against the company's, and data practices now being argued over in federal court. A playbook drawn from an operator like this has to carry both columns of the ledger. Several of the principles below are cautions as much as instructions.
The principles
- 1.Recruit from a list, not a funnel.
- 2.Pay in mission, and know when the mission stops clearing.
- 3.Choose benchmarks a crowd can follow.
- 4.Build the gym before the athlete.
- 5.Walk to the bottleneck yourself.
- 6.Release in stages and keep your hand on the dial.
- 7.Demo live.
- 8.Make loyalty legible, and fast.
- 9.Install a governor on your own intensity.
- 10.Leave before you break; return on purpose.
- 11.Outlast, but count what outlasting costs.
- 12.Know which hat you are wearing.
B
The Arc of a Builder
Brockman's roles track the company's bottlenecks: code, talent, compute, governance, product.
2007
First Intel Science Talent Search finalist from North Dakota since 1973.
2010
Drops out of MIT to join Stripe, founded by classmate Patrick Collison and his brother John.
2013
Becomes Stripe's first CTO; the company grows from 5 to 205 people.
2015
Leaves Stripe in May; co-founds OpenAI in December and hires nine researchers from a list drawn up with Yoshua Bengio.
2017
Joins the board. OpenAI's bot beats Dendi at The International.
2019
OpenAI Five beats world champions OG; capped-profit structure and $1B Microsoft investment; marries at the office.
2023
Unveils GPT-4 live on March 14. Removed as chairman November 17; quits; returns November 21 without a board seat.
2024
Extended leave, August to November, as Sutskever, Schulman, and Murati depart.
2026
Named defendant in Musk v. Altman, which he wins; takes over OpenAI's product organization.
Principle 1
Recruit from a list, not a funnel
The most consequential thing Brockman did in 2015 was not technical. He sat down with Yoshua Bengio and wrote out the names of the best AI researchers in the world, then went after them one at a time. Nine of them joined in December. A conventional hiring process flows the other way: post the role, screen the applicants, choose the best of whoever shows up. A funnel optimizes among the people who already want the job. A list goes after the people who don't yet know they want it.
The list also has a compounding effect a funnel lacks. The Google defector who said he came "partly because of the very strong group of people" was responding to the list itself. In a field where talent is unusually concentrated, the first nine names are the recruiting pitch for the next fifty. Brockman's leverage came from getting that first set right, because every later hire was in part a vote of confidence in it.
There is a quieter lesson in who held the pen. Brockman did not write the list alone. He borrowed the judgment of someone the field already trusted. An operator who is not himself a world-class researcher, and Brockman's résumé reads as engineer and builder, can still assemble a world-class research team by renting the eyes of someone who is.
Tactic: Before you open a single requisition, write down the ten specific people who would make the project obviously credible, check the list with the most respected person in the field who will talk to you, and recruit them by name.
Principle 2
Pay in mission, and know when the mission stops clearing
OpenAI's early compensation was a dare. It paid below Google and Facebook and offered no stock options. It paid instead in a founding charter about prioritizing "a good outcome for all over its own self-interest." Zaremba turned down "borderline crazy" money to accept that charter. For a few years mission worked as a currency, and a surprisingly strong one: OpenAI spent just $7 million on its first fifty-two employees.
Currencies inflate. By 2019 the company needed money on a scale that its pledges, which had brought in $133 million of a promised billion by 2021, could never cover. It needed to pay engineers in equity to compete. The capped-profit structure was in effect an admission that mission had stopped clearing at the required volume. Every later fight, from the 2023 firing to the "Not For Private Gain" letter to Musk's lawsuit, can be read in part as a dispute over the exchange rate between the old currency and the new one.
The principle is not that mission-based compensation fails. It is that mission-based compensation has a horizon, and the operator who uses it owes the people paid in it an honest account of when and how the terms might change. The people who joined for the charter were the ones most likely to experience the conversion as betrayal.
Tactic: If you recruit on mission, write down in advance the circumstances under which you would introduce conventional equity, and tell early hires the conditions before you meet them rather than after.
Principle 3
Choose benchmarks a crowd can follow
Reinforcement-learning research had plenty of benchmarks. Almost none of them could fill an arena. Brockman's team chose Dota 2, a game with a professional circuit, famous players, and a championship broadcast to a global audience. When the bot beat Dendi live at The International in August 2017, the result needed no paper to explain it. Brockman supplied the translation on the spot, with his image of the surgeon.
The choice had a cost. A public benchmark makes failure public too. OpenAI Five lost both of its exhibition matches at The International 2018, in front of the same crowd. The team kept going and beat the world champions in April 2019, then won 99.4 percent of 42,729 open games. A legible benchmark turns a research program into a story with chapters, and stories keep audiences, investors, and recruits through the losing chapters.
This principle generalizes well beyond AI. Most technical organizations measure progress on internal dashboards that nobody outside the team can read. Choosing a benchmark the public already understands works as distribution: it lets the outside world track your progress without your help.
Tactic: Find the one contest, comparison, or public standard that people outside your field already care about, commit to winning it in public, and plan in advance to lose it once before you win.
Principle 4
Build the gym before the athlete
OpenAI's first public release was not a model. It was Gym, a library that standardized how reinforcement-learning environments were defined, so that one lab's results could be reproduced by another. Universe followed, an attempt to turn the world's software into a single training ground. Neither was a breakthrough in the headline sense. Both were bets that whoever defines the arena shapes the field.
This is the instinct of a CTO who came from payments infrastructure. At Stripe, the product was a set of rails other people built on. At early OpenAI, Gym was a set of rails other researchers trained on, and every paper that used it carried OpenAI's framing into labs that had nothing to do with OpenAI. The pattern returned at much larger scale in 2020, when the company shipped GPT-3 as "the API," which made the model itself into infrastructure.
The epilogue carries its own lesson. Gym eventually passed to an outside foundation and stopped receiving updates under its original name. Infrastructure you give away does not stay yours. That is the price of the influence, and it is usually worth paying.
Tactic: Before you build your flagship product, ask which shared tool, format, or standard would make everyone in your field work the way you think they should, then ship that first and give it away.
Principle 5
Walk to the bottleneck yourself
Follow Brockman's attention across a decade and you get a map of OpenAI's constraints. In 2015 the constraint was talent, so he drew the list. In 2018 it was compute, and the team rented 128,000 CPUs and 256 GPUs from Google, a heavy bet for a lab whose entire 2017 cloud bill was $7.9 million. By the GPT-4 era it was data, and by the New York Times's account the YouTube transcription project went ahead with his notable involvement. In 2025 it was physical infrastructure, and Fortune named him "builder-in-chief." In 2026 it was product, and he took over the product organization.
⚙
Following the Bottleneck
Where Brockman's personal attention went, by year and constraint.
| Period | Bottleneck | Brockman's move | Risk profile |
|---|---|---|---|
| 2015 | Talent | Drew a list with Bengio; hired nine | Low |
| 2016–2019 | Compute | Scaled self-play training on rented cloud hardware | Capital |
| 2021–2023 | Data | Involvement in Whisper-based YouTube transcription | Legal |
| 2025 | Infrastructure | "Builder-in-chief" on data-center buildout | Capital |
| 2026 | Product | Took over the product organization | Execution |
The strength of the principle is speed: whoever owns the constraint sets the company's pace, and an executive standing at the constraint can make decisions without waiting for a chain of reports. The weakness shows up in the third row. Walking to a bottleneck under competitive pressure means making judgment calls at precisely the place where shortcuts are most tempting. Employees raised concerns about YouTube's terms of service. Those concerns were heard and overruled. Whatever the courts conclude, an operator who adopts this principle should plan on the bottleneck being where the hardest ethical calls arrive.
Tactic: Every quarter, name your organization's single binding constraint, move your own calendar to sit on top of it, and bring one person whose job is to tell you no.
Principle 6
Release in stages and keep your hand on the dial
GPT-2 was announced on February 14, 2019, and withheld. A limited version went out first, to a small group of testers in May, and the full model followed on November 5. GPT-3 was never released as weights. It came out through an API that, in OpenAI's words, made it "inherently safer to release them via an API and broaden access over time, rather than release an open source model where access cannot be adjusted."
Whatever you think of the safety argument, the operational logic holds. A staged release preserves options. You can widen access, narrow it, price it, watch it, or revoke it, and a public release forecloses all of that the moment the file is downloaded. The same reasoning was behind the ChatGPT preview, the waitlisted GPT-4 API, and the pause on ChatGPT Plus sign-ups in November 2023 when demand outran capacity.
Critics have a fair point that "staged release for safety" and "controlled release for revenue" can be the same policy described in two vocabularies, and Sutskever's later "We were wrong" about openness left that ambiguity in place. The lesson for operators is to be honest with yourself about which reason is doing the work, because your users eventually will be.
Tactic: For any capability you can't recall once released, design the first release so it can be widened or revoked, and write down which of your reasons is safety and which is business.
Principle 7
Demo live
On March 14, 2023, Brockman introduced GPT-4 in a live video demonstration. It was the Dota principle carried over to a product launch: let people watch the match. A live demo shows the latency, the occasional stumble, and the moment of waiting, all the things a polished keynote removes. In return it buys a credibility no edited video can.
The posture held up in the press. The headline TechCrunch put on his interview that week, "GPT-4 isn't perfect, but neither are you," turned the product's flaws into a frame instead of hiding them. It is a better position to defend than a promise of flawlessness that the first user will break.
Live demonstration also disciplines the builder. You cannot demo live what does not reliably work. Committing to a live unveiling is a forcing function on engineering quality, and it is much harder to fake than a slide.
Tactic: For your next launch, commit to demonstrating the core use case live and unedited, and if you can't bring yourself to do that, treat your reluctance as a bug report.
Principle 8
Make loyalty legible, and fast
Brockman's decision on November 17, 2023, took hours. He was removed as chairman, offered the presidency under Murati, and turned it down the same day. His public statement used the first person plural: "Sam and I are shocked and saddened." Speed was the whole point. A slow, ambiguous response would have let the board treat Altman's firing as a contained personnel matter. A fast, unambiguous one turned it into the loss of two co-founders and set the terms for everything that followed, including Microsoft's offer to both men and the employee letter demanding the return of both.
Legible loyalty works as a coordination signal. More than seven hundred employees did not need to guess where the company's longest-serving builder stood. Three senior researchers resigned with him within hours. When Microsoft announced its new research team on Monday, it named Brockman beside Altman.
The principle has a cost the five-day story tends to hide. Loyalty that obvious is also obviously partisan. Murati's complaint that the Altman-Brockman relationship made her job impossible is the other side of the same coin. A bond strong enough to bring down a board is strong enough to box out an executive. Brockman came back without his board seat, which is one way of saying the institution read his loyalty correctly.
Tactic: Decide in advance which relationships you would resign over, so that when the moment comes you can act within hours, and accept that people outside those relationships will see the same clarity as a faction.
Principle 9
Install a governor on your own intensity
This is the principle the record forces rather than flatters. According to the reporting gathered in Hagey's book, Sutskever gave the board a document of alleged bullying by Brockman. Murati said his relationship with Altman made her job impossible. Altman had "fielded many requests from OpenAI employees to rein in Brockman." None of this was ever adjudicated in public, and Brockman's side of it is not in the sources available here. Still, the fact that multiple senior colleagues complained, independently, about one person's behavior is evidence enough.
An intensity that wins a Dota championship, ships a product to a hundred million people in two months, and walks personally to every bottleneck does not have an off switch built in. The people who work beside it bear the cost. In 2023 that cost was converted into governance: the complaints became paperwork, and the paperwork became part of a vote.
The operator's lesson is not to be less intense. It is that intensity needs a mechanism other than other people's tolerance to regulate it, because tolerance, once used up, gets written down. Brockman's 2024 sabbatical, whatever its cause, can be read as one form of a governor. A better form would have been installed years earlier.
Tactic: Ask two peers who don't report to you to tell you each quarter, specifically and in writing, where your pace is costing them, and act on it before anyone else starts keeping a file.
Principle 10
Leave before you break; return on purpose
From August to November 2024, Brockman took the first extended break of a career that had run without one since 2010. The timing ran directly against the usual founder instinct. Schulman left the same week. Murati left while he was away. The company announced a contested structural conversion that autumn. Conventional wisdom says you never leave the room during a succession crisis. He left the room.
He came back on November 12, 2024, and within eighteen months had become the company's "builder-in-chief" for infrastructure and then the head of its product organization. Read on the evidence available, the leave did not cost him his position. It may have reset it. A founder who returns after a deliberate absence returns as a choice, not as a fixture. A founder who never leaves is part of the furniture, and the furniture is what gets rearranged.
The deeper point concerns what can be recovered. A burned-out operator who stays rarely comes back. One who leaves on purpose for a defined period, and says publicly that he will return, keeps his position while giving up the daily grind.
Tactic: If you are a founder who has not taken more than two consecutive weeks off in five years, schedule a defined leave with a public return date before your colleagues or your body schedule one for you.
Principle 11
Outlast, but count what outlasting costs
In July 2024 the Observer counted four of OpenAI's eleven founders still at the company. In the two years that followed, Brockman's officiant, his Gym co-author, and the colleague who complained about him all left, and most of them founded or joined competitors. Brockman stayed, survived a federal jury trial in 2026, and holds a stake Bloomberg values at $25.5 billion.
◎
Where the Founding Generation Went
Departures from OpenAI's early leadership, as reported in the sources.
| Person | Role at OpenAI | Departure | Next |
|---|---|---|---|
| Elon Musk | Co-chair | 2018 | Plaintiff Sued in 2024; lost in 2026 |
| Dario and Daniela Amodei | Early leadership | 2021 | Rival Anthropic |
| Ilya Sutskever | Co-founder, chief scientist | May 2024 | Rival Safe Superintelligence Inc. |
| John Schulman | Co-founder | August 2024 | Rival Anthropic |
| Mira Murati | CTO, interim CEO | September 2024 | Rival Thinking Machines Lab |
| Greg Brockman | Co-founder, president | Leave Aug–Nov 2024 | Stayed Head of product, 2026 |
Outlasting compounds. Equity vests, institutional memory becomes irreplaceable, and departures leave a vacuum that the person still there gets to fill. Brockman's 2026 move into product came straight out of a reshuffle triggered by someone else's absence. In a company whose value is growing very fast, simply remaining is one of the most valuable things an early employee can do.
The table also shows the cost. Each departure equipped a competitor with OpenAI's methods, culture, and in some cases its people. A founder who outlasts everyone ends up competing against his own alumni. The institution that remains may be bigger, but it is less varied, with fewer people left who can say "we tried that in 2017."
Tactic: Keep a running list of the people whose departure would most strengthen a competitor, and treat each of those relationships as an asset you have to actively maintain, not a given.
Principle 12
Know which hat you are wearing
In May 2023 Brockman co-signed a call for an international watchdog over advanced AI. In 2025 he and his wife were the largest donors to MAGA Inc. and gave $50 million to Leading the Future, a super PAC for AI deregulation that he helped found. OpenAI, meanwhile, said publicly that it was open to more oversight and maintained a policy against donating to such PACs. The president of a company and a private citizen with $25 billion are legally different people. In the press, and to regulators, they are much harder to tell apart.
The same tension showed up earlier and on a smaller scale in the YouTube data question. An executive's personal judgment about acceptable risk gets carried out through the company's systems, and the company carries the liability. Whether the hat is political or operational, the rule is the same: the more senior the person, the less the public distinguishes his hats, and the more carefully he has to.
This is not an argument against holding views. It is an argument for stating, out loud and in advance, which ones are yours and which belong to the institution, so that the gap between them reads as a choice and not as a contradiction discovered by reporters.
Tactic: When your personal positions diverge from your organization's stated ones, disclose the divergence yourself, explicitly and first, before a filing or a journalist does it for you.
≠
Conventional Wisdom vs. the Brockman Pattern
Where his operating choices departed from standard practice, with the results.
| Conventional wisdom | Brockman's approach | Outcome |
|---|---|---|
| Post roles and screen applicants. | Draw a list of the best people alive and recruit by name. | Nine founding hires |
| Benchmark privately; publish when ready. | Compete live at a championship and lose in public first. | 99.4% win rate, 2019 |
| Polish the launch video. | Unveil the flagship model in a live demo. | GPT-4 launch |
| Never leave the room in a crisis. | Quit within hours; take a leave during a succession wave. | Returned stronger, without board seat |
| Keep personal politics quiet. | Become a top super-PAC donor against the company's stated stance. | Unresolved |
For the fullest outside account of the company Brockman helped build, including the period Karen Hao first covered for MIT Technology Review in 2020, see Empire of AI.
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
— Greg Brockman, on X, November 2023Sam and I are shocked and saddened by what the board did today.
— Sam Altman, Greg Brockman, and Ilya Sutskever, 'Governance of superintelligence,' May 2023Given the possibility of existential risk, we can't just be reactive.
— OpenAI, founding announcement, December 2015AI should be an extension of individual human wills and, in the spirit of liberty, as broadly and evenly distributed as possible.
— Greg Brockman, The Knowledge Project with Shane ParrishI knew I wanted to do a startup because I felt like that was something—
— OpenAI, API announcement FAQ, June 2020It feels inherently safer to release them via an API and broaden access over time, rather than release an open source model where access cannot be adjusted if it turns out to have harmful applications.
Maxims
- The list is the strategy. The first nine names you recruit are the pitch for every name after them.
- Mission is a currency with a horizon. It pays early hires well, and it inflates the moment the work needs real capital.
- Pick a benchmark with a crowd. A result the public already understands needs no paper to explain it.
- Give away the arena. Whoever standardizes the training ground shapes the field, even after the tool belongs to someone else.
- Stand where the constraint is. The executive at the bottleneck sets the pace, and has to make the hardest ethical calls there.
- Keep the dial. Release in stages whatever you cannot recall, and be honest about which reason is doing the work.
- Let them watch you wait. A live demo buys a credibility that no edited keynote can.
- Loyalty is a coordination signal. Make it fast and unambiguous, and accept that outsiders will read it as a faction.
- Intensity needs a governor. Regulate it yourself, or colleagues will eventually do it with a file.
- Outlasting has a ledger. Everyone who leaves takes the playbook with them, and some of them write a better one.
