Part IThe Story
Twenty Billion, Countered
In the late spring of 2025, Mark Zuckerberg was unhappy. That is the fact everything else rests on. His company's newest large language model, Llama 4, released in April, had landed with a thud, and then with something worse than a thud: an accusation that Meta had tuned a version of the model to flatter a popular benchmark. A larger model, codenamed Behemoth, was meant to restore the company's standing against OpenAI, Anthropic, and Google. It was delayed amid internal doubts about whether it was any good. So the founder of the largest social network in human history did what founders do when the institution beneath them stops working. He went around it. He started a WhatsApp group chat with senior leadership to recruit researchers. He set a target of roughly fifty people. He began inviting candidates to his homes in Lake Tahoe and Palo Alto. One researcher, receiving an unsolicited message from Zuckerberg, assumed it was a hoax and didn't reply for several days.
Among the visitors to those houses was a 28-year-old from Los Alamos, New Mexico, who had never trained a frontier model, never published a research paper of consequence, and never finished college. What was said in Tahoe and Palo Alto has not been reported in full. The numbers have. According to The Information, Zuckerberg was prepared to put about $5 billion into Scale AI, the data-labeling company the young man had co-founded at 19. Alexandr Wang countered with $20 billion.
They settled at $14.3 billion, with some reports putting it at $14.8 billion, for a 49 percent non-voting stake. The stake was non-voting for a reason: a federal judge, James Boasberg, was weighing the government's antitrust case against Meta, and a full acquisition would have invited exactly the kind of scrutiny the company could least afford. So the deal was structured as an investment, Scale was declared independent, and its founder walked out the door and into Menlo Park as Meta's first-ever Chief AI Officer. His personal cash and equity in the transaction were worth more than $5 billion, by Bloomberg's count. On June 12, he stepped down as Scale's CEO. On June 30, Zuckerberg sent a memo to the company.
"Alex and I have worked together for several years," it read, "and I consider him to be the most impressive founder of his generation."
Note the order of operations. Not we bought a company and acquired its CEO. Something closer to the reverse. A Bloomberg headline put it with unusual bluntness: Scale AI's Wang Brings to Meta Knowledge of What Everyone Else Is Doing. Time was blunter still: "the real prize isn't his company—it's Wang himself." Google, Scale's largest customer, reportedly moved to sever ties within days. An analyst described Zuckerberg as a "wartime CEO." The phrase stuck, as phrases about war tend to stick to Wang, who had been using that vocabulary for years.
There is an old question buried in this transaction, and the rest of this piece will circle it without quite answering it, because Wang's career has been a sustained refusal to let it be answered. The question is: what, exactly, is a person worth when the thing being built is supposed to replace people? Scale AI made its fortune on the premise that machines could not learn without humans telling them what they were looking at — hundreds of thousands of humans, many of them in the Philippines and Kenya and Venezuela, drawing boxes around pedestrians for less than a cent per task. Meta paid $14.3 billion, and nine-figure packages to the researchers who followed, on the premise that a very small number of humans were irreplaceable. Wang has been, in one form or another, the broker of both markets. He priced the cheapest human labor in the AI economy and then, at a lakeside house in the Sierra Nevada, priced the most expensive.
Twenty billion. Countered.
By the Numbers
The Data Broker
19Wang's age when he co-founded Scale AI in June 2016
$14.3BMeta's investment for a 49% non-voting stake in Scale, June 2025
$29BScale's valuation implied by the Meta deal
$870MScale revenue in 2024, up from roughly $250M two years earlier
240,000+Contract workers in Scale's labeling army, per Time
15BHuman decisions Scale says it has captured to train AI models
~100People in Wang's Meta lab after 60 days — by his account a tenth the size of rivals
A Name With Eight Letters
The first thing anyone notices is the missing letter. He is Alexandr, not Alexander, and the reason is arithmetic: without the second e, the name has eight characters, and eight is a number associated with good fortune in Chinese culture. Somebody, at the very beginning, edited a boy's name for luck. His Chinese name, 汪滔, Wāng Tāo, is rendered in two characters, and the second one means a great surging wave. One should not read too much into a name. One should probably read a little.
The parents were Chinese immigrants who came to work at Los Alamos National Laboratory, the institution on a New Mexico mesa where the first atomic bombs were designed. Wang, describing them on a TED stage in April 2023, was precise in the way children of scientists are precise about their parents' fields: "My dad's a physicist and my mom's an astrophysicist. Their scientific work in plasma fluid dynamics will have deep implications for how we understand our universe." The Observer would later describe them, less tenderly, as "two weapons physicists." Both descriptions are probably true. Los Alamos is a town where those categories have never been cleanly separable.
He was born there in January 1997, and he describes it, depending on the audience, two ways. To policymakers, it is the birthplace of the bomb — "I grew up in the birthplace of a technology that defined the last era of warfare, the atomic bomb," he told TED, and he was off, within seconds, into autonomous fighter jets and drone swarms. To an auditorium of teenage founders at Y Combinator's Startup School in July 2026, it was something humbler: "Los Alamos, New Mexico, which is now Oppenheimer-famous, but it really was the middle of nowhere."
Both again. The middle of nowhere, built around the most consequential object of the twentieth century.
The clearest record of his childhood is not in any magazine. It is a post on the website of the University of New Mexico's Department of Mathematics & Statistics, written in the summer of 2022 by a staff member named Cristina, who had been surprised to open the Spanish newspaper El País one morning — June 9 — and find a former contestant of hers being hailed as the world's youngest self-made billionaire. She went back through the records of the UNM-PNM Math Contest. Alexandr Wang had first competed in the 2007–08 contest, as a fifth grader. He finished third overall in 2009–10. From then on, through tenth grade, he was first overall. In 2013–14, as an eleventh grader, he slipped to fifth — "By then he must have been already thinking about bigger things," Cristina wrote. His older siblings had competed and won, too; the family had been part of the contest for at least a decade. "Sometimes the love for math does run in families," she noted, and signed off with a single word that captures the strange vertigo of watching a local kid go supernova: "wow!!"
The national credentials stacked up in parallel. USA Computing Olympiad finalist in 2012 and 2013. The Math Olympiad Program in 2013. The U.S. Physics Team in 2014, an achievement the Los Alamos Daily Post announced with an exclamation point of its own. This is the biography of a particular kind of American child — the child of immigrant scientists in a government town, measured and ranked from the age of ten, very good at problems that have answers. It is worth holding onto the rankings. Third, then first, then first, then first, then fifth. Wang would spend the next decade building a company whose entire business was converting the world's ambiguity into labeled, ranked, sortable data, and he would eventually build the benchmark, Humanity's Last Exam, by which the world's most powerful AI systems are now graded. Some people escape the contest hall. Some people buy it.
In ninth grade, he and a friend started a Google Doc of startup ideas for when they were old enough. He quoted from it, with some embarrassment, in a 2016 post: "superpopular app. any ideas?" Then: "iTunes replacement (I swear it always freezes)." Then, presumably some days later: "actually nvm spotify is pretty dope." When a venture capitalist later asked him when he'd first known he wanted to start a company, he said, he was caught off guard and "rattled off some silly answer about being inspired by The Social Network."
The film about the founding of Facebook. He was a child when it came out. He would one day report to its protagonist.
An API for Human Intelligence
What the Los Alamos years produced, more than any single skill, was velocity. A friend from back home, who had gotten an internship at Palantir after high school, showed him the path out. At 17 he was in Silicon Valley as a software engineer at Addepar, a wealth-management software company, building and maintaining financial models. Then Quora, where he became a tech lead "directly responsible for all speed projects" — a job description that reads now like prophecy — and where he met a product designer named Lucy Guo. He took what he calls a gap year. Then MIT, in 2015, where by the account Index Ventures later published he received perfect grades in his first year.
"I remember this period from 17 to 19," he told Garry Tan at Startup School. "I felt like I was constantly changing." He was drinking, he said, "from the fire hose pretty constantly."
The fire hose had a direction. In December 2015, a freshman, he was interviewing "everywhere, perhaps indiscriminately" — DeepMind, Snapchat Research, Opendoor, Stripe. Offers came. None felt irrefusable. "It really felt like I was going to go work at yet another tech company," he wrote a year later in a LinkedIn post titled "What I learned in 2016," which remains the most candid document of his early life, partly because it was written before anyone was watching.
The conversation that broke the indecision was with Eric Wu, the co-founder and chief executive of Opendoor, then a startup of about thirty people trying to persuade a 19-year-old to choose it over Stripe or Dropbox. Wu was making a recruiting pitch. Wang heard something else.
— Eric Wu, Opendoor CEO, as recalled by Alexandr Wang in 2016As you get older, you're only going to become more and more risk-averse. You'll get married, buy a house, have kids, etc. That means today, you're the most risk-seeking you'll ever be. So, if there's a risk today that you aren't willing to take, then you might never end up taking that risk.
"Overall," Wang wrote, "it persuaded me against accepting any offers at all." The anecdote has the comic shape of so many founding myths: a CEO delivers a speech so persuasive it persuades the listener not to join his company. Wang framed what followed in the language of his discipline. He had been "minimizing my regret, an important concept in decision theory and machine learning." And then the line that might serve as his epitaph if he were the type to get one: "I believe that if you don't have opportunities you regret not taking, you aren't chasing after enough opportunities."
In January 2016, working as an algorithm developer at Hudson River Trading, a high-frequency trading firm, he started hacking on side projects. One of them was a camera inside his refrigerator.
The accounts vary charmingly. In one version, told to Index Ventures, he wanted to know when to restock the milk. In another, told to his investors at Accel, the camera was meant to catch roommates eating his food. Either way, the system needed to look at the inside of a refrigerator and understand what it saw, and after a few weeks he understood that it never would — not because the algorithms were wrong, but because he could not possibly gather enough labeled images to teach it. "We're going to build incredible things with AI," he recalled thinking. "What are the hurdles? It became really clear that data was going to be one of those."
He was 19. The compute was available for rent from Amazon. The algorithms were published in open papers. What nobody was selling, at least not well, was the thing the fridge camera lacked: someone to tell the machine what it was looking at.
The first version of Scale was stranger and more ambitious than the company it became. Wang and Guo entered Y Combinator's Summer 2016 batch with something he later described as "an API for human labor," aimed initially at chatbots for doctors. The phrase he used to describe the core idea, years later on YC's Lightcone podcast, is the most succinct statement of the business he would ever make: "What if you could call a human with like an API?" Software would send a request; somewhere, a person would do the task; the answer would come back as if from a machine. Delegation inverted. The human as subroutine.
The medical idea didn't survive the summer — YC partner Jared Friedman steered him off it — and within months a fellow YC company called Cruise, then building self-driving cars, became Scale's first major customer. Autonomous vehicles needed exactly what the fridge camera needed, at a scale of millions of frames: someone to draw a box around every pedestrian and every stop sign and, as Forbes later put it, to teach the machine "the difference between a paper bag and a pedestrian." By year's end, Wang was writing that Scale was "helping amazing companies like Uber, Alphabet, Houzz, and P&G with our API for human intelligence."
He never went back to MIT. The company, as one account put it, had started "as a summer gig."
Then a small accident. Scale turned up, unplanned, on Product Hunt, and a partner at Accel named Dan Levine sent Wang a cold email. Levine was the kind of investor who reads launch boards for signal — early, patient, willing to bet on teams before decks. The founders were so nervous about venture capitalists that they seriously considered ignoring him. They didn't. The courtship took about a month. Levine never saw a pitch deck before investing. Accel led Scale's Series A in July 2017, and soon Wang and Guo were working out of Levine's basement in San Francisco — two founders, 19 and 21, building infrastructure for the most hyped technology of the decade underneath an investor's house.
The go-to-market strategy was as unglamorous as the product. "We would be super scrappy," Wang said of the early days at CVPR, the major computer vision conference. "We'd go from booth to booth with a laptop with a demo on it." Zhichun Li, one of Scale's first employees, who would later lead its Rapid labeling product, remembered walking into a big client meeting in the South Bay with two colleagues, all young, all wearing backpacks. Someone remarked that it looked like a field trip. "It was just funny," she said, "how you bring in a small group of very young, talented people and hand them this problem. And we're just going to solve it."
Youth was a liability, and Wang said so. "It definitely made it harder to hire," he told Forbes. One candidate asked him directly: "Explain to me why I should entrust my career to you." Years later, as a billionaire, he was still getting carded at bars.
The People Inside the Machine
Here is the part of the story that does not fit on the booth demo.
An API for human intelligence requires humans. In 2017, Scale established a crowdworking platform called Remotasks to supply them, focused on computer vision and autonomous vehicles, with facilities in Southeast Asia and Africa. In 2019 it set up an entity called Smart Ecosystem Philippines to run Remotasks in that country, where many hires were freelance contractors not covered by labor laws. Remotasks did not always advertise that it belonged to Scale. As the platform expanded into India and Venezuela, "vicious competition" pushed the pay for some annotation tasks below one cent. Late payments, according to The Washington Post's 2023 investigation into what it called "digital sweatshops," were "commonplace"; some workers received only a few percent of what they had been promised. A 2022 study by the Oxford Internet Institute found Remotasks met the "minimum standards of fair work" in one out of ten criteria. In early 2024 the platform abruptly stopped operating in Kenya, Nigeria, and Pakistan, citing administrative and operational considerations. Workers learned by losing access.
Scale has said late payments are "exceedingly rare" and that it is always improving. Its website now reports $1 billion paid to contributors globally, and fifteen billion "human decisions to train AI models." Both numbers are meant to impress. Both are, read a certain way, a confession of how much of artificial intelligence is not artificial at all.
Wang has never pretended otherwise, which is his peculiar honesty. "The only model that we have in the world for the level of intelligence that we seek to create is humans, or is humanity," he told Andreessen Horowitz's David George in 2024. Producing the data for the next generation of models, he said, "will be one of the great human projects of our time." It would look like the internet, he said, "on steroids." The sentence is both true and a little chilling, because the internet, too, was assembled by billions of people who were mostly not paid for it.
The person who seems to have argued hardest about this inside the company was Lucy Guo. Guo's micro-biography rhymes almost too neatly with Wang's: a Quora alumna, a Thiel Fellow — Peter Thiel's program pays young people to skip college — two years older than her co-founder, named alongside him to the Forbes 30 Under 30 list in 2018. Within months she was gone. Some accounts say she was fired; Forbes reported she left over "differences in product vision and road map." Asked by Time about it in 2025, she named the workers.
— Lucy Guo, Scale AI co-founder, to Time (2025)I was like, 'we need to focus on making sure they get paid out on time.'
Wang, she said, was more concerned with growth. It was one of their "clashing points."
The irony of Guo's departure is that she kept her equity, and the equity did what Scale's equity did to everyone who held it. She later became the world's youngest self-made woman billionaire, unseating, per a Forbes headline that could only have been written in the 2020s, Taylor Swift. The two co-founders who could not agree on how quickly to pay people a cent per bounding box both became billionaires off the bounding boxes. The contractors did not.
In December 2024 a former worker sued Scale alleging wage theft and misclassification. A second suit followed in January 2025, and then a third, from contractors alleging psychological harm from exposure to disturbing content. By then Scale's labelers were increasingly a different population — its Outlier platform recruited people with advanced degrees, journalists, native speakers of many languages, to tutor chatbots — and the company could plausibly say the old Remotasks model was being superseded. But the pattern was set early, and it is the pattern of nearly every infrastructure fortune: the leverage comes from making labor invisible to the customer, and the invisibility does not stay contained.
The same year The Washington Post published its sweatshop investigation, a reporter visited Wang at Scale's sixth-floor headquarters in downtown San Francisco and watched "the occasional (partly) self-driving car" zip by below. The parenthetical is the whole story. Partly self-driving. Partly human. Wang later estimated that tele-operators still supervise robotaxi fleets at ratios of three to five cars per person. Somebody is always in the loop. The question has only ever been who, and at what rate.
Credential or Cult
By November 2020 Scale had hired more than two hundred people, and Wang had personally interviewed every one who received an offer. He wrote a memo to the company before a hiring push that would take it from roughly 75 people to 150, and later published it under a title that could not have been run through a communications department: "Hire people who give a shit."
The memo is short, unpolished, and more revealing than anything he has said on a stage since. He screened for two things, he wrote: caring about Scale, and caring about work in general. The second test had a specific interrogation attached. What's the hardest you've ever worked on something? How many hours were you working a week? "If someone is applying to Scale and has never been deeply obsessed about something before," he wrote, "then it's a bad bet to think Scale will be the first."
Then the passage that, read now, sounds like a premonition he would spend the next five years either fulfilling or fleeing.
— Alexandr Wang, internal memo (November 2020)One very scary thing to me is Scale becoming a credential rather than a cult. An institution like Harvard or MIT is a credential—people go there because it signals to others that the person is smart, capable, or otherwise prestigious.
A recruiting team that functions like a college admissions office, he wrote, "is certain death for a startup." Recruiting done right "looks more like courtship. You need there to be a spark, and sometimes that is preceded by months of convincing and pursuit."
The MIT dropout — the boy who had won the contest seven years running and then walked out of the most prestigious credential available to a young American engineer — was writing against the thing he had declined to finish. And the language of courtship, of months of pursuit, of finding "people you'd be willing to spend every waking moment with," describes exactly what Mark Zuckerberg would do to him, in Tahoe, five years later.
The ethos hardened over time into doctrine. "Ambition shapes reality" became one of Scale's core values, a phrase Wang wrote himself. He took a one-on-one call with every new employee even as headcount climbed into the hundreds, and later told Time his approach was "one of very high attention to detail, being very in-the-weeds." He thought of leadership in tiers: "There's level one—can you accomplish the things that are right in front of you? Level two is: are the things that you're doing even the right things?" Level three, he said, was culture, "probably the most important."
In June 2024 he published a blog post, "Meritocracy at Scale," that introduced a hiring framework he called MEI — merit, excellence, and intelligence — framed explicitly against corporate diversity, equity, and inclusion programs. Elon Musk and Coinbase's Brian Armstrong praised it. DEI researchers said it mischaracterized how diversity programs actually worked. The episode marked Wang's arrival as a political figure in a way the defense contracts had only implied; the acronym was a meme by design, and he knew how memes travel.
In October 2024 came the fullest statement of the creed, a blog post in capital letters: "DO TOO MUCH: How to be a leader." "As a leader, you are the upper bound for how much anyone in your company will care," he wrote. "You need to do more, care more, attempt more than would seem reasonable." He offered a list of reframings, presented with the flat confidence of axioms:
Overoptimism is just optimism.
Overcommunicating is just communicating.
Overdelivering is just delivering.
Micromanagement is just management.
Ruthless prioritization is just prioritization.
"I have never seen ordinary effort lead to extraordinary results," he wrote. "Actually living this way will seem crazy, and that's ok." And then, unexpectedly, a softer note, almost Romantic: "Creating something meaningful is a beautiful, and yes, scary and painful thing. And if you're not overdoing it, you're underdoing it."
Keep the fourth line of that list in mind. Micromanagement is just management. It will come back.
There is a novel that Wang finished in March 2023 and wrote about with an enthusiasm he rarely extends to anything that isn't a scaling curve. Gabrielle Zevin's Tomorrow, and Tomorrow, and Tomorrow follows two young friends who build video games together, fall out, and keep circling one another. Wang said it had given him "a renewed love of fiction," and that "nothing else has captured the naivéte & immaturity of being young and attempting foolishly bold things." He quoted a passage from the end: "No one was watching us, and we weren't even watching ourselves. What we had was our impossibly high standards, and your completely theoretical conviction that we could make a great game."
Two young co-founders. An early, unwatched period. A partnership that does not hold. He did not say whether he was thinking of anyone in particular.
Picks, Shovels, and the Steepest Curve
Scale's strategic history can be told as a series of bets on which curve would go vertical next, and the striking thing is how early Wang placed them.
The first was self-driving cars, which investors told him was too small a market. Then the market was flooded with capital, and it wasn't. The second, quieter, bet came in 2019, during what he calls the GPT-2 era, when Scale began working with OpenAI on language models — years before the public had heard the phrase "scaling laws." In August 2019, Peter Thiel's Founders Fund put in $100 million, valuing Scale above $1 billion. Bloomberg's Ashlee Vance ran the headline: "Silicon Valley's Latest Unicorn Is Run by a 22-Year-Old."
On June 21, 2021, Wang opened Scale's first Transform conference with a short speech that functions as a manifesto. Model performance, he said, increases logarithmically with compute and dataset size. Chips were scaling; algorithms were scaling. "This leaves the creation of massive scale high-quality datasets as the largest remaining bottleneck." Scale's goal, he said, was "to actualize Moore's Law for annotation." And then a line borrowed from an older vocabulary of military logistics: "In machine learning, the evidence is clear: quantity has a quality of its own."
A month later, a Series E led by Greenoaks, Dragoneer, and Tiger Global valued Scale at $7.3 billion. Wang owned about 15 percent. Forbes named him the youngest self-made billionaire in the world, at 24. He had, by then, assembled a senior bench that tells you whom he wanted to become. In April 2021 Jeff Wilke, the former head of Amazon's consumer business — the operator who had run the largest retail machine on earth — joined as an advisor to the CEO; Forbes later called him one of Wang's most trusted. In May, Michael Kratsios, who had served as Chief Technology Officer of the United States in the first Trump administration, joined as managing director and head of strategy. A retail logistician and a former White House technologist. Bret Taylor, then Salesforce's co-CEO, compared Scale's rise to Snowflake and Datadog. The comparisons were all to infrastructure. Nobody was comparing Scale to anything consumers had heard of.
Then the curve he had been riding flattened. In January 2023, Scale laid off 20 percent of its workforce. One trade publication's headline said Scale "was scaling till clienteles refused to buy it anymore." The autonomous vehicle industry was contracting; the money that had made the first bet look prescient was draining away.
Except the third bet was already placed. When OpenAI's image and chat models broke into the public consciousness in 2022 — Wang has called it the moment that prompted Scale to "aggressively pivot" — he redirected most of the company's workforce toward large language models. Some called the move extreme. It meant the company abandoning the business it was known for in favor of one that was still, for most enterprises, a demo. In August 2023 OpenAI named Scale its "preferred partner" to fine-tune GPT-3.5, and it emerged that Scale's services had helped create ChatGPT itself. That same month, Scale's evaluation platform ran the first generative AI red-teaming event at DEF CON, the hacker convention, where thousands of attendees tried to break models from multiple companies.
"We're the picks and shovels in the generative AI gold rush," he told Forbes. Scale had pulled in roughly $250 million in revenue the year before, at a time when many AI startups weren't making anything. In 2024, it brought in $870 million. In March 2024, an Accel-led round valued it near $13 billion; in May, a $1 billion Series F with new investors including Amazon and Meta pushed it to about $14 billion. The infrastructure metaphor, he said, was Amazon Web Services — the "infinite market," the business behind the business.
The fourth bet was evaluation. If Scale knew how every lab trained its models, it also knew how to test them. In January 2025, Scale and the Center for AI Safety released Humanity's Last Exam, a benchmark built from the hardest problems top researchers could devise — problems, crucially, that could not be found on the internet. When Wang discussed it in mid-2025, frontier models were scoring around 20 percent, up from seven or eight. The math-contest kid had become the contest. Every major lab in the world was now, in a sense, sitting his exam.
And here a pattern becomes visible, the kind you notice only from far enough away. Scale's moat was never a product. It was a position. To label a lab's data is to see what the lab is building; to grade its models is to know where they fail. "He's just really good at knowing what's going on in the industry," Bloomberg wrote, almost admiringly. The company that began as an API for calling a human became, without ever quite saying so, an API for calling the entire AI industry and asking what it was up to.
The Last Eighty Years
In 2018, Wang took a trip to what Forbes called his "ancestral homeland." He sat in a room and listened as China's brightest engineers gave impressive presentations on artificial intelligence. What unsettled him was what they didn't say.
"They were really dodgy on what the use cases were," he said. "You could tell it was for no good."
It is impossible to know from the outside how much of the subsequent political persona was born in that room and how much was later attributed to it. What is knowable is the shape of what followed. Scale took its first Department of Defense contract in 2020. In January 2022 it won a $250 million contract to give federal agencies access to its tools. In February 2022, after Russia invaded Ukraine, it built an Automated Damage Identification Service that analyzed satellite imagery to geotag destroyed buildings for humanitarian groups — Time put it on its Best Inventions list that November. Later, Scale quietly scrapped a deal with the Chinese-owned TikTok over security concerns. In May 2023, Scale signed with the U.S. Army's XVIII Airborne Corps and became the first AI company to deploy a large language model, called Donovan, on a classified network. In February 2024 the Pentagon picked it to set the testing and evaluation path for military LLMs. In March 2025 came Thunderforge, a Defense Innovation Unit project — Anduril and Microsoft also involved — to use AI to "plan and help execute movements of ships, planes, and other assets," starting with Indo-Pacific Command and European Command.
Wang explained the philosophy with a historical argument that sounds, coming from a 26-year-old in gorpcore and a sixth-floor office, either grandiose or unusually clear-eyed.
"If you think about the history of humanity, it's mostly been punctuated by war except the last 80 or so years, which have been unusually peaceful," he told Forbes. "A lot of that has been because of American leadership in the world."
Eighty years. Count back from 2023 and you land, give or take, at 1945 — the summer the work done on his hometown's mesa ended the Second World War. He did not say this, but the arithmetic says it for him. The peace he wanted to defend began with the weapon designed in the town where he was born, by the institution where his parents worked.
At TED in April 2023, in a talk that has been played nearly 1.4 million times, he made it explicit. "As we saw in World War 2 with the atomic bomb, the country that is able to most rapidly and effectively integrate new technology into war fighting wins." He claimed China was outspending the United States on military AI by a factor of ten, adjusted for total budget. The American military threw most of its sensor data away, he said: "In an AI war, everything boils down to data." And he named the cultural enemy: "Somewhere along the line, tech leaders decided that working with the government was taboo." Three months later, on July 18, 2023, he testified before a House Armed Services subcommittee hearing titled "Man and Machine: Artificial Intelligence on the Battlefield."
Notice what happened to the vocabulary. In April 2022, at TEDxBerkeley, he had given a talk titled "Why AI Will Never Replace Humans." A year later, the talk was about autonomous fighter jets. The Washington Post summarized the shift in a headline: "Some tech leaders fear AI. ScaleAI is selling it to the military."
He was, by January 2025, among the tech founders and chief executives at Donald Trump's second inauguration. Scale bought a full-page advertisement in The Washington Post addressed to the president: "America must win the AI war." Days later at Davos, he warned that China's DeepSeek had caught up with American models. In February he met Keir Starmer, Narendra Modi, Emmanuel Macron, and House Speaker Mike Johnson. That same month Scale signed a five-year deal with the government of Qatar to modernize its public services with AI. He was building, in effect, a foreign policy.
In March 2025 he co-authored a policy paper with Eric Schmidt, Google's former chief executive, and Dan Hendrycks, director of the Center for AI Safety. Superintelligent AI, it warned, "would amount to the most precarious technological development since the nuclear bomb."
The bomb, again. Always the bomb. The son of Los Alamos physicists had found a way to make every argument about the future an argument about his hometown. Whether this is insight or narrative strategy is difficult to separate, and the critics who have tried have tended to choose one. The National, writing in 2026, called the "moral imperative" framing "hollow." A Substack writer called him a fraud and a product of "PR-built nepotism," which ignores the seven years of math contest results on a University of New Mexico web page. Neither critique quite lands, because neither accounts for the possibility that a person can believe something sincerely and also understand precisely how useful it is to be seen believing it.
The Network Is the Product
During the worst months of the pandemic, Sam Altman lived in Wang's apartment. The arrangement lasted months, by Time's account. Altman had run Y Combinator when Scale went through it in 2016; by the time he moved in, he was running OpenAI, which would become one of Scale's most important customers. The detail is the kind that sounds like gossip and is actually a balance sheet.
Altman, one source told Bloomberg, once jokingly told Wang to tone down his networking a little. "Fortunately for Wang," Bloomberg observed, "he ignored the advice." Since founding Scale at 19 he had befriended the founders of Plaid, Figma, and Brex. He joined the board of Expedia Group in 2023 — the seat had been held by Altman, who left it that June. He was photographed at the 2024 Met Gala. He posted Instagram photos with the actor Kiernan Shipka, listened to what Forbes called "sad girl" musicians like Gracie Abrams and Billie Eilish, and told his Accel podcast host that "algo-raving" was the best way to experience music. He cultivated, according to Bloomberg, right-wing influencers and Republican lawmakers in Washington, partly by hammering on China. In January 2025 he sat for a long conversation with the comedian Theo Von.
It is tempting to read all of this as surface. It is more accurate to read it as structure. Scale's business made Wang the one person every lab had reason to talk to; Wang's temperament made him the one person every lab wanted to talk to. By early 2025 he had become close with Zuckerberg — the CNBC-published memo says they had "worked together for several years," which tracks with Scale's role in Meta's 2023 Purple Llama security initiative and the Defense Llama product the two companies built together for military use. When Zuckerberg went looking for a way out of the Llama 4 debacle, he did not need to search for Wang. Wang was already at the house.
What Meta paid for, then, was not a lab or a model or a team of researchers. It was a vantage point. The Guardian summarized Bloomberg's argument: Wang "brings to Meta knowledge of what everyone else is doing." The structure of the deal let Meta pull this off with a 49 percent non-voting stake, an "intentionally muted role," as one account put it, "in order to avoid scrutiny from the Federal Trade Commission." Forbes reported the deal included a five-year commercial agreement worth hundreds of millions annually. Scale's competitors understood immediately what had happened. If Google kept using Scale, it would be strengthening a company half-owned by a rival; if it left, it would have to rebuild its pipeline. Reuters reported Google would cut ties. Time reported OpenAI was severing its deals as well, though Bloomberg reported OpenAI intended to stick around. Either way, the interim CEO, Jason Droege — a former Uber executive and Scale's chief strategy officer — spent his first week publicly insisting the company was not "winding down."
Then the rest of the shopping list. Meta tried to buy Safe Superintelligence, the startup of OpenAI co-founder Ilya Sutskever, and was rebuffed. It held talks with Thinking Machines Lab and Perplexity that fell through over price and strategy. It hired Daniel Gross, SSI's chief executive, and Nat Friedman — the former chief executive of GitHub at Microsoft, a programmer-turned-investor who had sat on Meta's advisory group for a year — and arranged to partially buy out their venture fund. Offers to OpenAI and Google researchers reportedly ran from $1 million to $100 million. Andrew Bosworth, Meta's chief technology officer, said it plainly on CNBC: "The market is setting a rate here for a level of talent which is really incredible and kind of unprecedented in my 20-year career as a technology executive."
To pay for it, Meta began putting advertisements into WhatsApp.
The labelers in the Philippines were paid less than a cent per task when competition was vicious. The researchers in Menlo Park were paid nine figures when competition was vicious. Same market logic. Same broker in the middle of both ends.
Suffocating
The lab Wang built first was small on purpose. In his first sixty days, he told the hosts of the TBPN podcast, John Coogan and Jordi Hays, he had assembled about a hundred people — "smaller and more talent dense than any of the other labs." "The other labs are, like, 10 times bigger," he said, but Meta's had "cracked" coders. It was called TBD Lab, a name so provisional it sounded like a placeholder someone forgot to change. In August 2025, Meta restructured the whole superintelligence organization into four groups: TBD Lab under Wang, building the large models; FAIR, the old fundamental research unit; Products and Applied Research under Friedman; and infrastructure under Aparna Ramani.
He was effusive about the assets. "We have the business model to support building literally hundreds of billions of dollars of compute," he said. Meta's smart glasses, he argued, would be "the natural delivery mechanism for superintelligence" — "It will literally feel like cognitive enhancement. You will gain 100 IQ points by having your superintelligence right next to you." He told an audience of adults that thirteen-year-olds should spend all their time vibe-coding. "It's impossible to understate the degree to which I've been radicalized by AI coding," he said — meaning, presumably, overstate, a slip the transcribers let stand. "Literally all the code I've written in my life will be replaced by what will be produced by an AI model."
Every code-golfing hour in Los Alamos, every speed project at Quora. Replaced, and glad of it.
The institution he had entered was older and heavier than any he had run. Its research culture had been shaped for more than a decade by Yann LeCun, the French computer scientist who pioneered convolutional neural networks, won the Turing Award, and came to Facebook from NYU in 2013 to found FAIR. LeCun is the closest thing the field has to a living founding father, and for years he had been saying in public that the approach Meta was now betting its future on — scaling large language models — was a road to nowhere. Wang was 28. LeCun was 65.
In October 2025, Meta cut 600 jobs from the superintelligence organization. On November 20, LeCun left to found Advanced Machine Intelligence Labs, a startup devoted to "world models" that learn from video and physical data rather than text. In January 2026, he gave an interview to the Financial Times and said what he thought of the young man who now ran the place.
— Yann LeCun, former Meta chief AI scientist, to the Financial Times (January 2026)He learns fast, he knows what he doesn't know. There's no experience with research or how you practice research, how you do it. Or what would be attractive or repulsive to a researcher.
It is a careful insult, the kind a professor gives — two compliments as a cushion and then the knife. LeCun said Zuckerberg had "basically lost confidence in everyone who was involved" after the Llama 4 benchmark scandal and had "sidelined the entire Gen AI organization." "A lot of people have left," he said. "A lot of people who haven't yet left will leave." And of the strategy itself: "LLMs basically are a dead end when it comes to superintelligence. I'm sure there's a lot of people at Meta, including perhaps Alex, who would like me to not tell the world that."
The more interesting leak had come a month earlier. In December 2025, the Financial Times reported, citing multiple people familiar with the matter, that Wang had complained to associates about Zuckerberg's management style. The word he reportedly used was "suffocating." The complaint, in essence, was micromanagement — that the chief executive's involvement was keeping teams from moving fast and experimenting freely.
Fourteen months earlier, in capital letters, Wang had written: Micromanagement is just management.
It would be cheap to call this hypocrisy, and also inaccurate. It is something more structurally interesting. The DO TOO MUCH doctrine is a founder's doctrine; it assumes the leader is the upper bound on caring because the leader is at the top. Wang had never before worked for a founder. He had worked as one, from the age of nineteen, with a board that had every reason to let him run. Inside Meta he was, for the first time since Quora, someone else's direct report — and the someone else was the most famous practitioner of founder mode alive, a man who had personally DMed researchers until one thought it was a prank. Wang had imported the creed into a building where the creed's true believer already occupied the corner office. Two upper bounds. One company.
The work, meanwhile, shipped. On April 8, 2026, Meta Superintelligence Labs released its first model, Muse Spark — internally codenamed Avocado — a multimodal reasoning model that now powers the Meta AI assistant. It was closed-weight, API-only, a sharp turn for a company whose AI identity had been built on giving Llama away. Wired's headline said it gave Zuckerberg "a Seat at the Big Kid's Table." Muse Spark 1.1 and Muse Image followed about two months later; Muse Video arrived in July; and on August 10, 2026, Meta released Muse Glimmer, a 30-billion-parameter open-weights model under the Apache license, small enough to run on a personal computer. Wang would later describe Muse Spark 1.1 as "Opus"-level capability at one-eighth the cost.
Muse Image brought the other kind of headline. Meta shipped it with a feature that let users generate images by @-mentioning public Instagram accounts — users were automatically opted in, meaning their photos could be used as reference material by strangers. The backlash was immediate, and within days Meta pulled the feature: "We've heard the feedback that this feature missed the mark, so it's no longer available." For a man whose company had spent nine years invisibly harvesting human judgment to teach machines, it was a strange echo — the people inside the machine, suddenly, were the users themselves, and they noticed.
By June 2026, the Financial Times, The New York Times, and Business Insider were reporting that several engineering groups under Wang's oversight — infrastructure, training data systems, and strategic projects including Avocado and another called Mango — were being reassigned to a new applied engineering division under Maher Saba, a senior executive from Reality Labs, the glasses-and-headsets unit. The reorganization was presented as a move to turn breakthroughs into products. It was read, widely, as a narrowing. Some employees questioned whether running a fast-growing startup had prepared him to run research at a global technology company. LeCun's question, asked again by people who did not have to say it on the record.
Scale, the company he left, kept going. In March 2026 it launched Scale Labs, an expanded research division. On July 30, 2026, it named Francis deSouza, Google Cloud's chief operating officer, as CEO — a veteran operator, the type Wang once borrowed from Amazon as an advisor, now running the company he founded in a basement.
Theoretical Conviction
On July 29, 2026, Wang walked onto the stage at Y Combinator's Startup School to what Garry Tan called "full rockstar treatment." The audience, Tan noted, was full of people who were "just like us, but when we were 18 or 20." There were sixteen-year-olds in the room.
He was 29 now, a year into Meta, the leaked word suffocating trailing him like a tin can. He talked about the startup landscape shifting from "David vs. Goliath" to "Mecha-Goliath vs. Goliath" — small companies armed with agents taking on incumbents. He predicted that intelligence and agency would soon become abundant and that the scarce resource would become "vision and ambition." He spoke of 10x, 100x, even 1,000x expansions in GDP growth. He described a future not of one machine god but of billions of personal agents talking to billions of business agents, and said that the people doing technical work would increasingly be organizing "armies" of agents — millions, trillions. He said the debate over when superintelligence arrives was a waste of time. The title of the session became: "This is a once-in-a-civilization opportunity."
Tan asked him what he would tell his eighteen-year-old self. The answer was the most consistent thing he has ever said, in the sense that he has been saying it since he was nineteen. Develop your own internal compass for how the future will unfold. Hold conviction in it against the noise. Successful founders, he said, hold beliefs that contradict early consensus. Investors had called data unsexy. Now data was the most valuable business in AI.
Your completely theoretical conviction that we could make a great game.
By then a small shelf of instant biographies had appeared — Alexandr Wang: Codebreaker – The Quiet Genius Behind AI's Next Frontier is one; Biography of Alexandr Wang: How One Visionary is Shaping the Future of Artificial Intelligence and Global Innovation is another — the kind of books that come into existence when a person becomes a story faster than anyone can report one. They tend to tell it as an ascent. Los Alamos, MIT, billionaire at 24, chief AI officer at 28. Up and to the right, the way he once described model performance scaling with data.
The record is lumpier. It includes a co-founder who left fighting over contractor pay, workers in Kenya and Nigeria who lost access without explanation, a TEDx talk titled "Why AI Will Never Replace Humans" followed a year later by a talk about drone swarms, a memo warning against becoming a credential followed by the most expensive credential in the history of the industry, and a creed that micromanagement is just management followed by a report that someone else's micromanagement was suffocating. It also includes a nineteen-year-old who saw, before almost anyone, that the bottleneck in artificial intelligence was not cleverness but labeled examples — that every machine that seems to understand the world has been shown the world, one judgment at a time, by a person.
He came to that idea through a refrigerator. A freshman in Cambridge, hacking on side projects between trading-firm shifts, mounted a camera inside his fridge and asked it a simple question — is the milk running low? — and found it could not answer, not because it could not see the carton, but because no one had yet told it, thousands of times over, what a carton looked like when it was nearly empty.
Part IIThe Playbook
What follows is an attempt to extract the operating logic beneath the biography — not the motivational poster version of Wang's career, but the structural decisions that compounded, and the ones whose costs arrived on a delay. Several of these principles cut both ways. That is deliberate. Wang's playbook is unusually effective and unusually costly, and an honest reading has to hold both.
The principles
- 1.Find the bottleneck nobody wants to own.
- 2.Sell the human before you sell the automation.
- 3.Spend risk at the age you have the most of it.
- 4.Let the first real customer choose your market.
- 5.Pivot the whole company, not a skunkworks.
- 6.Defend against becoming a credential.
- 7.Be the upper bound on care — and know whose ceiling it is.
- 8.Take the customer everyone else finds taboo.
- 9.Convert market position into information position.
- 10.Price yourself, not just the company.
- 11.Treat the supply side as product, or it becomes your scandal.
- 12.Keep the lab small enough to know everyone in it.
Principle 1
Find the bottleneck nobody wants to own
Wang's founding insight was a triage. Of AI's three inputs — compute, algorithms, data — two had obvious owners. Nvidia and the cloud providers sold compute; the labs and academia produced algorithms, mostly in public. Data had no owner because it was tedious, labor-intensive, and unsexy, which is exactly why investors initially dismissed it. "This leaves the creation of massive scale high-quality datasets as the largest remaining bottleneck," he told his own conference in 2021, and the sentence is a business plan disguised as an observation.
The pattern generalizes. In any fast-growing technical field, glamour pools around the visible layer and the bottleneck migrates to whatever nobody wants to do. A founder who can stomach that work — who can go booth to booth at CVPR with a laptop demo for a labeling service while everyone else pitches models — inherits a position the glamorous players will eventually depend on.
The fridge camera is instructive here. Wang didn't arrive at the insight through market analysis. He tried to build the glamorous thing, failed, and asked why he failed. The bottleneck was the answer to that question.
Tactic: Build something ambitious in your target field, fail at it, and then sell the thing that made you fail to everyone else who is about to hit the same wall.
Principle 2
Sell the human before you sell the automation
"What if you could call a human with like an API?" Scale's original architecture put people behind a software interface, so customers experienced human judgment as if it were a machine call. This let Scale ship a product immediately — no waiting for the technology to work — while capturing the workflow, the data formats, and the customer relationships that automation would later absorb.
It is the inverse of how most AI startups think. They wait until the model is good enough to sell. Scale sold the outcome first and treated automation as a margin improvement to be layered in over time. The interface was the moat; what sat behind it could change.
The cost, as Principle 11 explores, is that the humans behind the interface become invisible by design, and invisibility has consequences.
Tactic: If your product can't be fully automated yet, wrap human labor in a clean programmatic interface and sell the outcome now — then automate behind the API without the customer ever needing to change anything.
Principle 3
Spend risk at the age you have the most of it
Eric Wu's argument — you will only become more risk-averse, so take the risk now — is a simple observation about the shape of a life. Wang's elaboration was more technical. He framed his choices as regret minimization, a concept he borrowed from decision theory and machine learning, and added a corollary: "if you don't have opportunities you regret not taking, you aren't chasing after enough opportunities."
The corollary is the non-obvious part. Most career advice optimizes for avoiding bad outcomes. Wang treated regret as a signal of sufficient exploration — a sign the search space was being covered. He interviewed indiscriminately in December 2015 (DeepMind, Snapchat, Opendoor, Stripe), collected offers, and turned all of them down. Snapchat and Opendoor later did well. He noted it without flinching.
Note too what came before the leap: a year at Quora, stints at Addepar and Hudson River Trading. "From the outside, you have no idea how companies work," he told Startup School. The risk was large but not blind.
Tactic: Treat your early twenties as the exploration phase of a search algorithm — work inside real companies to calibrate, then take the highest-variance bet you can find while your obligations are lowest.
Principle 4
Let the first real customer choose your market
Scale entered Y Combinator to build medical chatbots. It left as a data supplier for self-driving cars because Cruise — a batchmate — showed up needing labeled images at volume. Wang didn't argue with the demand. He followed it, even though investors told him autonomous vehicles were too small a market.
They were wrong for a reason worth remembering: market size is a lagging indicator in a capital-flooded field. When the money arrived in AVs, the market expanded to fit it. Wang later told SaaStr that TAM "in the traditional sense barely matters," and his career is the case study.
S
The Curve-Jumping Record
Each bet was placed before consensus; each was abandoned before it fully peaked.
2016
Scale launches at YC as an API for human labor aimed at medical chatbots; pivots within months after Cruise becomes first major customer.
2017
Accel leads Series A; Remotasks founded to supply labeling labor for computer vision.
2019
Work with OpenAI begins in the GPT-2 era; Founders Fund invests $100M, valuation passes $1B.
2020
First Department of Defense contract.
2021
$7.3B valuation; Wang named youngest self-made billionaire at 24.
2022
Most of the workforce redirected to large language models; $250M federal contract.
2023
20% layoff in January; Donovan deployed on a classified network; OpenAI preferred partner for GPT-3.5 fine-tuning.
2025
Humanity's Last Exam released; Thunderforge contract; Meta invests $14.3B and Wang becomes Chief AI Officer.
2026
Muse Spark ships April 8; reports of friction and reassigned teams follow by June.
Tactic: Ship a generic capability early and watch which customer pulls hardest — then reorganize the company around that customer even if the market looks small, because in a capital-flooded field size follows money.
Principle 5
Pivot the whole company, not a skunkworks
When generative AI broke through in 2022, many incumbents spun up small innovation teams. Wang redirected most of the company. The move was costly — a 20 percent layoff landed in January 2023, as the autonomous-vehicle business softened — but it put Scale in position to become OpenAI's preferred partner within a year and grow revenue from roughly $250 million to $870 million in two.
The logic is about organizational gravity. A skunkworks inside a company built for the old market gets starved by the old market's priorities. Only a whole-company pivot reallocates the attention, the hiring, and the founder's time. Wang's "DO TOO MUCH" philosophy has a strategic analogue here: half-pivots are just underdoing it.
Tactic: When you're convinced the next curve has arrived, move the majority of the company onto it at once and accept the short-term pain, rather than building a small team that the existing business will eventually smother.
Principle 6
Defend against becoming a credential
Wang's 2020 hiring memo identifies a failure mode most founders never name: success makes a company attractive for the wrong reasons. Early on, nobody joins for prestige, so everyone who joins genuinely cares. Later, applicants arrive "for the brand rather than the substance," the top of the funnel swells, and "the ROI of the hires drops dramatically." The company starts to resemble a university — "a constant churn of smart but uninvolved people."
His countermeasures were personal and labor-intensive: interviewing every offer recipient for years, a one-on-one with every new hire well into the hundreds, interview questions designed to surface past obsession ("What's the hardest you've ever worked on something?"). The MEI framework of 2024 was the politicized descendant of the same impulse, and drew both praise and serious criticism for how it characterized the alternative.
The deeper irony, of course, is that Meta is the ultimate credential, and Wang's own value there was partly as a symbol. The principle applies to institutions and to people.
Tactic: Write down, before you're famous, what your company's draw will be once it is — and design your interview process specifically to filter out candidates who are there for that draw rather than the work.
Principle 7
Be the upper bound on care — and know whose ceiling it is
"As a leader, you are the upper bound for how much anyone in your company will care." This is Wang's most quoted idea and his most effective one. Founders who personally review data quality, as he says he once did for all final checks, set a standard that propagates downward.
But the doctrine has a structural blind spot, revealed at Meta. "Upper bound" implies a single ceiling. In an organization with two founder-style leaders — one of them the CEO — the bounds collide. Wang's reported complaint that Zuckerberg's micromanagement was "suffocating" is the creed, micromanagement is just management, experienced from below.
W
The Over-Doing Doctrine, Tested
Wang's 2024 reframings, and what each looks like from the other side of the org chart.
| Wang's reframing (2024) | From the top | From below | Risk |
|---|---|---|---|
| Overcommunicating is just communicating. | Alignment at speed | Noise, meetings | Manageable |
| Overdelivering is just delivering. | Standards rise | Burnout | Manageable |
| Micromanagement is just management. | Quality control | "Suffocating" | Structural |
| Ruthless prioritization is just prioritization. | Focus | Abandoned teams, abrupt exits | Structural |
| Overoptimism is just optimism. | Morale, ambition | Missed forecasts | Low |
Tactic: Before you adopt a founder-intensity operating style, map where it will bump into another person's intensity — a co-founder, a board, an acquirer's CEO — and negotiate those boundaries explicitly before the first collision.
Principle 8
Take the customer everyone else finds taboo
"Somewhere along the line, tech leaders decided that working with the government was taboo," Wang told TED in 2023. Scale took its first Defense Department contract in 2020, when much of Silicon Valley was still recoiling from the employee revolts over military work. The result was a nearly uncontested position: a $250 million federal contract in 2022, the first LLM on a classified network in 2023, the Pentagon's LLM evaluation mandate in 2024, Thunderforge in 2025.
The principle is not "work with the military." It is that a customer category which peers avoid on cultural grounds can be underpriced in competitive terms. Wang paired the commercial move with a coherent public argument — the eighty years of peace, the atomic precedent, the China threat — that converted a reputational risk into a political asset. He also hired for it: Michael Kratsios, a former U.S. CTO, as head of strategy.
The cost is that the argument becomes part of your identity, and critics will test it for sincerity.
Tactic: Identify the customer your competitors avoid for reasons of culture rather than economics, build a principled public case for serving them, and hire someone who already speaks their language.
Principle 9
Convert market position into information position
Scale's moat was never really labeling technology. It was that labeling a lab's data and grading its models gives you a cross-sectional view of the entire industry. Bloomberg's summary of what Meta bought — "knowledge of what everyone else is doing" — names the asset. Humanity's Last Exam turned the position into a public institution: every frontier lab now reports scores on Scale's test.
This is the deepest pattern in Wang's career. Infrastructure players see everything that flows through them. A founder who understands this builds products that maximize what they see, and relationships — Altman in the apartment, Zuckerberg in Tahoe — that compound it.
The fragility is equally clear. When Scale's neutrality ended with the Meta deal, Google reportedly walked, and the information position started to decay immediately.
Tactic: If you sell infrastructure to competitors, treat your neutrality as your most valuable asset — and understand that the day you trade it away, you are cashing out the view, not just the company.
Principle 10
Price yourself, not just the company
The Meta negotiation — $5 billion offered, $20 billion countered, $14.3 billion settled — only makes sense if you understand that both sides were pricing Wang as much as Scale. Meta took a non-voting stake, left the company nominally independent, and got its founder. Wang walked away with more than $5 billion in cash and equity and the most powerful AI job at one of the world's largest companies.
In an era of "not quite acquisitions" — Microsoft, Amazon, and Alphabet had all paid large sums to license startups and hire their talent — Wang read the market correctly: the scarce asset was a person, and persons can be priced aggressively. His counter at four times the opening offer was not reckless. It anchored the conversation in what Zuckerberg actually wanted.
Tactic: When an acquirer is courting you, determine whether they are buying the company or the founder — and if it is the founder, negotiate as the scarce asset rather than as the seller of a business.
Principle 11
Treat the supply side as product, or it becomes your scandal
Lucy Guo has said that paying workers on time was one of her clashing points with Wang. The record that followed — sub-cent tasks, late and partial payments, the Oxford Internet Institute finding Remotasks met one fair-work criterion in ten, abrupt country exits, three lawsuits in roughly a month — became the counter-narrative to every one of Scale's growth milestones. Time's profile at the moment of Wang's greatest triumph led its critique with the contractors.
For any marketplace or labor-backed business, the supply side is a product with its own users, and neglecting their experience creates a liability that compounds as the company becomes famous. The very invisibility that made the "API for human intelligence" elegant made the people behind it easy to underserve.
Tactic: Assign an owner, metrics, and a quality bar to your supply-side experience — payment latency, communication, access changes — with the same seriousness you apply to customer experience, because journalists will eventually measure it for you.
Principle 12
Keep the lab small enough to know everyone in it
At Meta, Wang built TBD Lab to roughly a hundred people in sixty days — "smaller and more talent dense than any of the other labs," which he said were "like, 10 times bigger." It is a continuation of his Scale instinct: personally interviewing everyone, knowing every hire. Muse Spark shipped within about nine months of his arrival.
The principle is that density beats headcount in frontier work, and that a leader can only uphold standards across a group small enough to see. The limits showed up around it: LeCun's critique that Wang lacked feel for research culture, 600 jobs cut from the broader organization, and the 2026 reassignment of several teams to a new applied division. A small, dense core can still be undone by the large, diffuse institution around it.
Tactic: Cap your core team at a size you can personally know, protect it fiercely — and invest separately in your relationship with the larger organization surrounding it, because that is where a dense team's mandate is won or lost.
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
— Alexandr Wang, What I Learned in 2016I believe that if you don't have opportunities you regret not taking, you aren't chasing after enough opportunities.
— Alexandr Wang, Scale Transform opening remarks (June 21, 2021)In machine learning, the evidence is clear: quantity has a quality of its own.
— Alexandr Wang, to Forbes (2023)If you think about the history of humanity, it's mostly been punctuated by war except the last 80 or so years, which have been unusually peaceful. A lot of that has been because of American leadership in the world.
— Alexandr Wang, DO TOO MUCH: How to be a leader (October 2024)As a leader, you are the upper bound for how much anyone in your company will care. You need to do more, care more, attempt more than would seem reasonable.
— Alexandr Wang, to David George of Andreessen Horowitz (September 2024)The only model that we have in the world for the level of intelligence that we seek to create is humans, or is humanity.
Maxims
- Fail at the glamorous thing first. The bottleneck that stopped you is the business everyone else will need.
- Put a person behind an API. Sell the outcome before the automation exists, then automate behind the interface.
- Regret is a coverage metric. If you regret nothing you passed on, you didn't search widely enough.
- Let demand outrank market size. In a field flooded with capital, the market grows to meet the money.
- Pivot with the whole body. Half-moves get smothered by the business you're leaving.
- Fight to stay a cult. Once people join for the brand, the quality of hires quietly collapses.
- Your intensity has a ceiling — and so does your boss's. Founder mode works until it meets another founder.
- The taboo customer is the underpriced one. Cultural avoidance by competitors is a moat you can walk into.
- Infrastructure sees everything. Neutrality is the asset; sell it only knowing you are selling the view.
- The people inside the machine are users too. Neglect the supply side and it will become the story told about you.
