Part IThe Story
The letter went up in the second week of April 2018, and it read like nothing its author had ever built. For six years Zhang Yiming had been designing machines on one premise: people should never have to say what they want. A patient enough system could watch a thumb pause over a headline, register the half-second someone lingered on a joke, and work out the rest without asking. Now Zhang was the one who had to declare himself. China's National Radio and Television Administration had ordered his flagship news app, Toutiao, and his first app, a jokes-and-memes platform called Neihan Duanzi, temporarily removed from the country's app stores. The regulator accused Neihan Duanzi of hosting "vulgar" and "improper" content and of "triggering strong sentiments of resentment among internet users." The next day Neihan Duanzi announced it was shutting down permanently. It had more than 200 million users.
Zhang's response was a public self-criticism. He wrote that the app had been "incommensurate with socialist core values," that his company's implementation of Xi Jinping Thought had been "weak," and that ByteDance would "further deepen cooperation" with the Party to better promote its policies. The China Media Project, which tracks the language of the Chinese state the way seismologists track tremors, filed the episode under a phrase that stuck: "Tech Shame in the 'New Era.'" On April 29 the company's own Party committee published a statement through the state-affiliated outlet The Paper, under the heading "Prioritize orientation and responsibility." ByteDance then announced that it would give preference to Communist Party members in hiring and expand its content-review staff from 6,000 to 10,000.
That same month, almost unnoticed, the company's two-year-old short-video app Douyin added a feature that prompted users to take a break every ninety minutes.
Hold these facts side by side. A thirty-five-year-old engineer who had spent his career removing the friction between a person and their desires publicly confessed that he had removed too much. A recommendation engine built to be indifferent to everything except attention was told it must hold an ideology. And a product so compulsive that it had to remind people to stop was being readied to go to the rest of the world, which was where Zhang had always believed the real prize lay. Six months later, with more than a billion monthly users across its apps, ByteDance was valued at $75 billion and passed Uber as the most valuable privately held startup on earth. In October 2018 the Sydney Morning Herald ran a profile under a headline that was accurate and, in hindsight, slightly absurd: "The unknown 35-year-old behind the world's most valuable startup."
He would not stay unknown. He would try, with remarkable discipline, to stay unseen. Those are different projects, and the distance between them is roughly the distance between Beijing and Washington.
By the Numbers
The Attention Factory, Measured
$155BByteDance revenue, 2024
$300BEstimated valuation, November 2024
>50%Zhang's share of ByteDance voting rights
~120,000ByteDance employees, 2025
7 monthsTime to develop Douyin, the app that became TikTok
19.9%ByteDance's retained stake in TikTok's U.S. joint venture, 2026
$105.5BZhang's net worth when Fortune named him Asia's richest person, September 2026
The News Reads You
The thesis came first, and it was a complaint. In 2012 Zhang believed that Chinese smartphone users were struggling to find information in the apps available to them, and that Baidu, the country's dominant search engine, was mixing its results with undisclosed advertising. Search, in his view, was the wrong grammar for a phone. Search made the user do the work: think of a question, type it with a thumb, wade through results that might be paid placements. His alternative reversed the direction of the transaction. Content would be pushed to the user, chosen by recommendations generated by artificial intelligence. The user would only have to look.
He made this argument to venture capitalists, and most of them did not buy it. The story told about the founding of ByteDance is a story of rejection, and like most such stories it is told by the people who were right. Sequoia Capital passed. Funding came only when Susquehanna International Group, the Philadelphia-area firm built by Jeff Yass, agreed to invest. Remember that name. It returns twelve years later in a very different role.
In early 2012 Zhang and his friend Liang Rubo, along with several employees from their previous company, rented a four-bedroom apartment in Zhongguancun, the Beijing district that had become the country's shorthand for its technology ambitions. ByteDance was incorporated on March 13, 2012. Its Chinese name, 字节跳动, pairs the word for "byte" with a verb of leaping or pulsing. The first product shipped that same month: Neihan Duanzi, roughly "profound gags," a place to pass around jokes, memes, and funny videos. It was a modest beginning, and it was the product that six years later would require the apology.
Toutiao followed in August 2012. The name means "headlines," and it looked like a news aggregator, which was misleading in the useful way. Toutiao had no editors in any meaningful sense. It used algorithms to classify articles by user preference and served each user a different front page. Within two years it had more than 13 million daily users. In January 2014 the company launched Toutiaohao, a platform that let content creators publish directly into the feed. Later that year it added video. Its distribution was interest-based and decentralized, which meant obscure writers with small subjects could find audiences that no human editor would have assigned them.
Sequoia came back. In 2014 the firm that had turned Zhang down led a $100 million investment in the company.
What the market had missed was neither the size of the opportunity nor the quality of the engineering. It had missed the inversion. Every major Chinese internet company of that era was built on a destination: a search box, a messaging app, a storefront. Toutiao had no destination. It was a mirror that slowly learned your face. The Wall Street Journal, writing about the app in August 2017, found the headline the product deserved: "The News Reads You in China—and People Can't Get Enough of It." A month earlier The Information had run a story whose title captured what this meant for incumbents: "How a News Startup Caught China's Tencent by Surprise." The surprise would become litigation. Starting in 2018 ByteDance and Tencent sued each other repeatedly. ByteDance alleged that Tencent was blocking its content. Tencent accused ByteDance of using WeChat and QQ profiles without authorization and illegally crawling data from public WeChat accounts, and it won an injunction. As of early 2024 ByteDance's suits had still not been resolved, stalled by disputes over jurisdiction.
The reading went one way, though. Toutiao read its users. Its founder was harder to read, and he seemed to want it that way.
The Whampoa Class of Kuxun
The biography itself is short and deliberately plain.
Zhang Yiming was born on April 1, 1983, in Longyan, in the southeastern province of Fujian. His parents were civil servants and he was their only son. Note the parents' profession. The man who would later write a self-criticism in the language of the state grew up in a household that worked for it. In 2001 he enrolled at Nankai University in Tianjin, studied microelectronics and then software engineering, and graduated in 2005 with a degree in computer engineering. He met his wife there. Nothing in the record suggests a dramatic adolescence, a dropout's legend, or a garage. The early life of the most consequential Chinese founder of his generation is as orderly as a recommendation engine's training data. Its inputs are visible. Its outputs could not have been predicted from them.
In February 2006 he joined a travel search site called Kuxun as its fifth employee and first engineer, and was promoted to technical director a year later. Kuxun is worth a paragraph because of what it produced. A Chinese business paper, The Economic Observer, later described the company as the "Whampoa Military Academy" for a cohort of post-1980s bosses, Zhang among them. The reference is to the officer school that trained both Nationalist and Communist commanders before they turned on one another. The analogy is half flattery and half prophecy: a small institution that gives its graduates a common language and then sends them off to become rivals.
In 2008 Zhang left Kuxun for Microsoft and found the corporate rules stifling. He did not stay long. He moved to a startup called Fanfou, an early Chinese microblogging service, which failed. In 2009, as Expedia was about to acquire Kuxun, Zhang took over its real-estate search business and turned it into his first company, 99fang.com. His co-founder was Liang Rubo.
Liang had been Zhang's college roommate. When the trade press eventually profiled him, KrASIA summed him up in one word, the "executor," and the label fits the role he played for a decade. Zhang was the founder who argued with venture capitalists about the future of information. Liang was the founder who made the apartment run. They left 99fang together, about three years after starting it, to build the thing in Zhongguancun. They would trade places once, a decade later, in a handoff that looked effortless from the outside.
Microsoft keeps reappearing in this story, the way a lesson does when it has not quite been learned. Zhang left the company because it felt like a bureaucracy. In March 2016 ByteDance founded its research arm, the ByteDance AI Lab, and chose Wei-Ying Ma to run it. Ma was the former assistant managing director of Microsoft Research Asia. In January 2020, when TikTok needed its first global general counsel, it hired Erich Andersen, a Microsoft in-house attorney. The large American company Zhang found too rule-bound as a twenty-five-year-old became, a decade later, a source of the expertise he needed most: research depth, and the legal fluency to survive inside the American system. He took the people and left the structure behind.
One-Fifth of the World
Most Chinese internet founders of Zhang's era built for China because China was enough. It was enormous, it was walled off from American competitors, and it rewarded companies that understood its regulators better than its users. Zhang drew a different conclusion from the same numbers.
— Zhang YimingChina is home to only one-fifth of Internet users globally. If we don't expand on a global scale, we are bound to lose to peers eyeing the four-fifths. So, going global is a must.
The plan appeared remarkably early. In January 2013, less than a year after the apartment, a four-part plan for the company's future was presented to executives. Part four was an English-language version of Toutiao, built to win users in English-speaking countries. At that point the company had one hit and one jokes app.
The international expansion that followed was less a strategy than a series of probes, some of which worked. TopBuzz, an overseas content platform for videos, articles, and GIFs, launched in 2015 and was abandoned in 2020 as its business dwindled. Former employees later alleged that it had been used to push soft messaging sympathetic to China to foreign audiences, which ByteDance denied. In December 2016 the company invested in BABE, an Indonesian news recommendation platform. In February 2017 it bought Flipagram, which it renamed Vigo Video in July. Vigo was shut down on October 31, 2020. In November 2017 it bought News Republic from Cheetah Mobile. In India it built Helo, and in December 2018 it sued the Chinese tech site Huxiu for defamation after Huxiu reported that Helo was spreading fake news. When the company built Lark, its workplace collaboration suite, Zhang insisted that it target American, European, and Japanese markets rather than China alone, as had originally been proposed.
The probe that mattered had been running quietly at home. Douyin, whose name translates roughly as "shaking sound" or "vibrating sound," launched on September 20, 2016, under the name A.me, and took its permanent name that December. It was built in nearly seven months. Within a year it had 100 million users and more than a billion video views a day. A year after its domestic launch, in September 2017, ByteDance released an international version under a different name. TikTok and Douyin would share an interface and almost nothing else. Their content was walled off from each other, and their servers sat in the markets they served. One machine, two populations.
Then came the purchase. On November 9, 2017, ByteDance agreed to buy Musical.ly, a Shanghai-headquartered startup with an office in Santa Monica, whose lip-sync videos had already captured American teenagers. The price is reported variously, as "up to $1 billion" and elsewhere as $800 million. I have not been able to reconcile the two figures, and given the company's habits, the ambiguity may be the most authentic detail in the deal. What is not in dispute is what ByteDance did with it. On August 2, 2018, it shut Musical.ly down as a separate app, folded its accounts and data into TikTok, and kept the TikTok name. Musical.ly brought the audience and features like the "duet." TikTok brought the engine. In October 2018 TikTok became the most downloaded app in the United States. On Apple's App Store it was the most downloaded app of both 2018 and 2019, ahead of Facebook, YouTube, and Instagram.
The Musical.ly acquisition followed a logic that would come back later in a much harsher form. Zhang bought a beachhead in a foreign country and replaced its insides. The shell stayed and the brain changed. Years later Washington would demand the same operation in reverse.
India became the largest proof of concept. In 2019, 43 percent of TikTok's new users came from India. On June 29, 2020, after a border clash with China in the Himalayas, the Indian government banned TikTok, CapCut, and 57 other Chinese apps. The ban was made permanent in January 2021, and two months later Indian authorities froze ByteDance's bank accounts over alleged tax evasion, which the company disputed. The four-fifths of the world, it turned out, were not a market waiting to be won. They were made up of countries, and countries have borders and memories.
The Shaking Sound
It helps to describe the product precisely, because most accounts of TikTok describe its effects and treat its mechanics as magic.
Early TikTok videos ran from a few seconds to a minute. The app opened directly onto the "For You" feed, which did not wait for you to follow anyone. Conventional social networks built recommendations on the relationships between users: who knew whom, who liked whose post. TikTok built them on behavior. As Wired reported in 2020, it monitored a far wider range of signals during viewing than the likes, clicks, and follows other platforms relied on. In 2021 the Wall Street Journal concluded that it understood users' preferences and emotions better than its rivals did. Each video was an experiment and each swipe was a result. Toutiao had read the news to you. TikTok read you.
The numbers that followed belong in a geology textbook more than a business plan. TikTok passed two billion downloads in April 2020. In July 2020, excluding Douyin, it reported nearly 800 million monthly active users after less than four years of existence. It reached a billion users in September 2021, faster than Facebook, Instagram, or YouTube had reached the same mark. In the first quarter of 2022 the average user spent an hour and twenty-five minutes a day on the app and opened it seventeen times. In 2021 TikTok was the most popular website in the world, ahead of Google.
Culture bent around it. In 2019 Lil Nas X's "Old Town Road" spread on the platform and became the longest-running number-one song in the history of the Billboard Hot 100.
— Lil Nas X, on TikTok's role in the success of "Old Town Road"[TikTok] really boosted the song ... I credit them a lot.
Fleetwood Mac's "Dreams" returned to the Hot 100 in 2020, forty-three years after its first run, carried by a skateboarding video. Kate Bush's "Running Up That Hill" topped the UK singles chart thirty-seven years after its release. Songs were sped up so more of them would fit inside a clip, and some artists, including SZA and Steve Lacy, released official sped-up versions to match. The market inverted, as markets do. In 2022 Halsey said her label would not release a new song unless she agreed to "fake a viral moment on TikTok."
The commercial aftershocks reached as far as feta cheese, whose sales rose 200 percent in 2021 after a baked-pasta recipe went viral. Stanley tumbler revenue went from $73 million in 2019 to $750 million in 2023. BookTok put six Colleen Hoover novels in the top ten of the New York Times paperback fiction list in October 2022. A BookScan analyst called the community "the industry's most important platform for discovering new writers," and in 2023 ByteDance decided to capture that value directly by founding a publisher, 8th Note Press, which closed abruptly in June 2025 and returned its rights to authors. The darkest example came from Australia's drug regulator, the Therapeutic Goods Administration. It wrote that when videos about rapid weight loss with the diabetes drug Ozempic "went viral on TikTok," the trend "triggered a huge demand for the product that the manufacturer was not prepared for, and it quickly developed into a worldwide shortage."
Every American platform rebuilt itself in TikTok's image. Instagram launched Reels in 2020. YouTube launched Shorts in 2021. By 2022 Instagram had spent $120 million to lure creators to Reels, and Snapchat had paid out $250 million to creators the year before. Facebook added a personalized video tab. Spotify redesigned its home screen as a scrolling feed. Reddit added a video feed. The shaking sound became the house style of the internet.
The machine was not as neutral as it looked. In January 2023 Forbes reported on a tool called "heating," which let TikTok staff manually promote particular videos. It accounted for one to two percent of daily views, by the magazine's account, and had been used for the FIFA World Cup and for Taylor Swift. TikTok said heated videos made up less than 0.002 percent of feeds. In 2020 The Intercept reported that moderators had been told to suppress posts from the For You feed when the people in them were deemed "too ugly, poor, or disabled." The algorithm had a hand on it. Once that was known, people started asking whose hand it was.
The economics had their own irony. All that attention was cheap. American users were monetized at about $0.31 an hour, a third of Facebook's rate and a fifth of Instagram's, or $67 a year against Instagram's more than $200. In March 2024 The Information reported that it was an open secret among investors that TikTok was losing billions of dollars a year. The most captivating product of its time was, outside China, not yet a very good business. Douyin, with its in-video shopping, hotel booking, and a search feature that could find more videos of a person by their face, showed what the machine looked like once fully monetized. TikTok showed what it looked like abroad: loved, distrusted, and underpriced.
Horse Racing
Inside the company, Zhang built an organization designed, like his products, to remove the friction of deference.
His management style was modeled on American technology companies, Google in particular. He held town hall meetings every two months. He discouraged employees from calling him "boss" or "CEO," the usual Chinese convention. Talent was bought aggressively. In September 2017 Bloomberg reported that the company was luring engineers with pay packages worth $3 million. Lark, the collaboration suite, began as an internal tool and became the company's main channel for communication before it was offered to outside customers in 2019. In 2021 the company launched BytePlus to sell the software framework underneath TikTok, so that others could build similar apps. Its internal infrastructure kept turning into products.
The most distinctive practice was competition among teams building the same thing. In March 2023 the Wall Street Journal reported that, according to former employees, ByteDance assigned several teams to build the same product, a practice known internally as "horse racing." The team whose version performed better got more support. This is the For You feed applied to the organization. Ideas, like videos, would be tested against each other and the data would decide.
The costs fell on people. Former employees described a workweek that started on Sunday to line up with Chinese time zones, after-hours work that was encouraged, back-to-back meetings that left no time even for basic bodily needs, and personal tolls that included marriage therapy. The company said it was committed to offering employees "support and flexibility." In June 2022 the Financial Times reported that a culture clash at the London office had caused a staff exodus. In August 2025 the South China Morning Post reported that ByteDance's own data put employees' median length of service at about three years.
The flat hierarchy and the absence of the word "boss" did not mean the absence of a boss, and the clearest proof came from outside. On May 19, 2020, ByteDance and Disney announced that Kevin Mayer, the head of Disney's streaming business, would join ByteDance. A short sketch of Mayer is enough. He was the consummate American corporate operator, a dealmaker from the most brand-conscious entertainment company on earth, hired to be the face that would reassure Washington. From June 2020 he was CEO of TikTok and chief operating officer of ByteDance, reporting directly to Zhang. He resigned at the end of August, after roughly three months. Business Insider's headline described him as "boxed out by ByteDance during TikTok's biggest moment of crisis." Vanessa Pappas, an Australian executive, took over in the interim.
Mayer did not fail at anything he had been hired to manage. He was hired to look like he was in charge of something that was not his to run. Six years later, in August 2026, Disney and TikTok announced a content partnership that would let TikTok creators use approved material from Disney films and shows, with selected videos appearing inside Disney+. The institution came back to TikTok. The man it had sent did not.
Deepen Cooperation
There were always two companies inside ByteDance. One was the flat, Google-style meritocracy of town halls and horse races. The other had been there since 2014, when, according to state media, the Communist Party set up a branch inside the company.
The committee is led by Zhang Fuping, a ByteDance vice president and the company's Party Committee Secretary. A report submitted to the Australian Parliament quoted his statement of purpose. Another account, from Quartz, described committee members gathering regularly to study Xi Jinping's speeches and pledging to follow the Party in technological innovation. Zhang Fuping's line sums up the parallel institution better than any outside analyst could: ByteDance should "transmit the correct political direction, public opinion guidance and value orientation into every business and product line."
Every business and product line. Including, presumably, the ones built to be indifferent to everything but engagement.
After the 2018 letter the second company became much easier to see. By 2019 ByteDance's Beijing headquarters had an office where cybersecurity police were stationed so that illegal content could be reported immediately. In April 2019 the Ministry of Public Security held a ceremony launching its "new media matrix" of accounts on Toutiao and Douyin, and in November the ministry's publicity bureau signed a strategic cooperation agreement with ByteDance that included unspecified "offline activities." In 2018 ByteDance had helped found the Beijing Academy of Artificial Intelligence, backed by the Ministry of Science and Technology and the city government. In 2019 it formed joint ventures with Beijing Time, a publisher controlled by the Beijing municipal Party committee, and with Shanghai Dongfang, a state media firm. In 2021 ByteDance said the Shanghai venture had never operated and dissolved it.
The regulators did not let up. In November 2019 the Cyberspace Administration of China ordered ByteDance to remove "slanderous" material about Fang Zhimin, a Communist martyr, from Toutiao. In April 2020, in the middle of a pandemic that had made remote-work software essential, the agency ordered ByteDance to take down Lark because it could be used to get around internet censorship. The tool Zhang had insisted be built for America, Europe, and Japan was too open for China. In January 2021 regulators fined the company over "vulgar information." In March 2021 the market regulator fined a subsidiary for antitrust violations. In April 2021 the People's Bank of China named ByteDance among thirteen platforms that had to "self-examine" and "rectify." In May 2021 the Cyberspace Administration said the company had illegally collected and misused personal data. In August 2021 its children's English-tutoring service Gogokid was shut down after the crackdown on after-school tutoring.
That same month the state became a shareholder. The China Internet Investment Fund, a state-owned entity, bought a one-percent "golden share" in ByteDance's main domestic subsidiary and placed a board member there: Wu Shugang, a government official who, according to the South China Morning Post, had spent most of his public-sector career in propaganda since joining the Ministry of Education in 2007. The Economist noted that his résumé showed no clear business experience but "a background in communist propaganda." One percent of the domestic company, a seat on its board, and a vocation.
The requests grew more ideological. In April 2022 ByteDance said it would report users whose posts on Toutiao and Douyin amounted to "historical nihilism," meaning departures from the official Party account of history. In November 2022, during the COVID-19 protests, the Cyberspace Administration directed the company to step up censorship of the demonstrations. In November 2023 Forbes reported that Chinese authorities had inspected Feishu, the domestic version of Lark, which held material on network security, data security, personal information, and daily operations, in the run-up to the Party's 20th National Congress. In September 2025 the Cyberspace Administration launched a crackdown on Toutiao itself, citing content that promoted "negative outlooks on life such as world-weariness."
World-weariness. A government had found an attitude worth correcting in a news feed that was, by design, only reflecting back what its users already felt.
Some allegations reached Zhang personally and remain unresolved in the public record. In May 2023 the New York Times reported on a lawsuit by a former ByteDance executive accusing the company of "lawlessness" and accusing Zhang of facilitating bribes to Lu Wei. The same former employee alleged in a court filing that Chinese Communist Party members had accessed data on Hong Kong users, including protesters, in 2018. ByteDance denied the claims, saying he had worked on a defunct project and that TikTok had pulled out of Hong Kong in 2020. He later claimed in a sworn statement that his father in mainland China had been detained in retaliation.
The Reuters headline from September 2020 described the trend line in one clause: "TikTok troubles narrow gap between Beijing and ByteDance founder Zhang Yiming." The founder who wanted the four-fifths kept being drawn back toward the one-fifth.
The Mouthpiece
Washington got to TikTok later than Beijing did and with less precision, but with more noise.
It started with children. On February 27, 2019, the Federal Trade Commission fined the company $5.7 million for collecting data from users under thirteen through Musical.ly, in violation of the Children's Online Privacy Protection Act. The penalty is a footnote now. The settlement was reopened by the FTC and the Justice Department in August 2024 and resolved in August 2026 for $400 million. The Committee on Foreign Investment in the United States began examining the 2017 Musical.ly merger. In January 2020 the Army and Navy banned the app from government devices.
On August 6, 2020, President Trump signed an executive order that would ban transactions with TikTok in forty-five days unless ByteDance sold it. On August 14 he gave ByteDance ninety days to sell or spin off the American business, citing "credible evidence" that it "might take action that threatens to impair the national security of the United States." The New York Times later reported that CIA analysts had concluded the Chinese government could obtain user data from the app but had found no evidence that it had. Beijing answered on August 28 with updated export-control rules that experts said would give Chinese authorities a say in any sale of ByteDance's technology. The algorithm was now formally a national asset, and the founder could not sell it on his own. In September Oracle was confirmed as TikTok's cloud partner. That same month, in a legal filing, the U.S. Justice Department called Zhang Yiming a "mouthpiece" of the Chinese Communist Party.
The courts blocked Trump's orders. In June 2021 President Biden revoked them and ordered a Commerce Department review in their place. The concern stayed and grew sharper.
— FBI Director Christopher Wray, testifying to U.S. lawmakers, November 2022the possibility that the Chinese government could use [TikTok] to control data collection on millions of users or control the recommendation algorithm, which could be used for influence operations
The case against TikTok was built partly from evidence and partly from rivals' self-interest, and the two are hard to separate. In June 2022 BuzzFeed News published reporting by Emily Baker-White based on leaked audio from eighty internal TikTok meetings. It showed that employees in China had accessed American user data, including through a "master admin" who could see "everything." TikTok then confirmed that China-based employees could access U.S. data and announced it would route American traffic through Oracle's cloud. In December 2022 ByteDance acknowledged that employees in China and the United States had accessed the data of two journalists, from Forbes and the Financial Times, and their contacts while trying to identify sources of leaks. Four employees were fired. The Justice Department and the FBI opened an investigation. The BBC later reported that one British journalist had been tracked through her cat's account. It is the kind of detail that makes the abstract case concrete: a pet's profile used as a surveillance tool.
The rivals were also busy. In March 2022 the Washington Post reported that Meta had paid Targeted Victory, a Republican consulting firm, to run a campaign portraying TikTok as "a danger to American children and society." The campaign placed letters to the editor and promoted stories about dangerous teen trends such as "devious licks," some of which had actually started on Facebook. Reddit's CEO, Steve Huffman, had already called TikTok "fundamentally parasitic." One analysis estimated that a TikTok ban could be worth between $2.46 billion and $3.38 billion in advertising revenue to Meta. American incumbents were not neutral observers.
ByteDance answered with the tools of the American political system. It spent more than $17.7 million on lobbying from its first filing in 2019 through July 2023, and $8.7 million in 2023 alone. Its lobbying operation, led by Michael Beckerman, included former senators Trent Lott, a Mississippi Republican, and John Breaux, a Louisiana Democrat. The company proposed Project Texas, a plan to isolate American data inside a subsidiary, TikTok US Data Security, that would report to CFIUS rather than to ByteDance, with Oracle reviewing code and data flows. CFIUS took the proposal and said little. In March 2024, as Congress moved, ByteDance spent $2.1 million on an advertising campaign in Senate battleground states. The New York Times summarized the cast in its headline: "TikTok Turns to Nuns, Veterans and Ranchers in Marketing Blitz."
It did not work. The House passed a divest-or-ban bill on March 13, 2024. A revised version rode through the Senate inside a foreign-aid package on April 23 and was signed the next day. The deadline was January 19, 2025. The Supreme Court upheld the law in TikTok, Inc. v. Garland.
On the night of January 18, 2025, at about 10:29 p.m. Eastern, TikTok went dark across the United States. CapCut, Lemon8, Gauth, Hypic, and the card game Marvel Snap went with it. American users, looking for somewhere to go, signed up in large numbers for a different Chinese app, Xiaohongshu, known in English as REDnote. The next day TikTok came back after President-elect Trump assured the company he would not enforce the law. Two months earlier, in November 2024, Trump had come out against a ban after a donation to his campaign from Jeff Yass, the investor whose Susquehanna had written the check no one else would write in 2012. The first money into ByteDance had become, more than a decade later, part of its rescue.
The reprieves became policy: seventy-five days in January 2025, seventy-five more in April, ninety in June. On September 14, 2025, the Wall Street Journal reported that Washington and Beijing had agreed on the framework of a deal involving Larry Ellison's Oracle. It closed on January 22, 2026. TikTok USDS Joint Venture LLC would run TikTok in the United States. Oracle, MGX, and Silver Lake each took 15 percent. Michael Dell and Vastmere Strategic Investments shared 35.1 percent. ByteDance kept 19.9 percent, and the global ByteDance-controlled TikTok entity kept control of e-commerce, advertising, and marketing on the U.S. platform. Adam Presser, formerly TikTok's head of operations, became chief executive. Under the deal the recommendation algorithm would be "retrained" on American users' data.
This was the Musical.ly operation run backward. In 2018 Zhang had kept an American shell and replaced its brain. In 2026 Americans kept the shell and set out to reeducate the brain.
Within days, users reported that videos critical of Trump and of Immigration and Customs Enforcement were failing to upload, and that the word "Epstein" seemed to be suppressed. TikTok denied censoring anything and blamed power outages at a U.S. data center. California's governor, Gavin Newsom, announced an investigation. UpScrolled, a small competitor, rose to second place on Apple's App Store. The fear had always been that a foreign government would put its hand on the algorithm. In the first month after the sale, the hands in question were American.
In May 2021, almost exactly a year after the Mayer announcement, Zhang said he would step down as chief executive. He would be succeeded by Liang Rubo, the roommate, the co-founder of 99fang, the executor. Reuters' headline used Zhang's own explanation: "'I'm not very social.'" Leaked internal posts, Protocol reported, showed employees giving Liang a warm welcome. The handover was completed on November 4, 2021. The same month Liang's appointment was announced, Shou Zi Chew, who had previously been chief financial officer of Xiaomi and then of ByteDance, was named chief executive of TikTok. Chew, a polished, numerate finance executive, became the company's public face in the West, the man who would sit in front of Congress in March 2023. Zhang let other people do the talking.
The record of the withdrawal is mostly a record of absences. Business Insider called him "the secretive Chinese billionaire behind TikTok." When ByteDance listed its board of directors in November 2024, it named five people: Liang Rubo; Arthur Dantchik of Susquehanna; William Ford of General Atlantic; Xavier Niel, the French telecom entrepreneur; and Neil Shen of HongShan. Zhang was not on the list. He kept, according to Reuters, more than half of ByteDance's voting rights. Ownership was split roughly 60 percent to investors outside China, 20 percent to the founders and Chinese investors, and 20 percent to employees. In practice he controlled the company while appearing in almost none of its public documents.
The money kept arriving anyway. In 2024 TikTok's global growth made him the richest man in China. ByteDance's revenue that year was reported at $155 billion, and Bloomberg reported in December 2025 that the company was on track for $50 billion in profit. He became the first Chinese centibillionaire. In September 2026 Fortune reported that he had overtaken Gautam Adani as the richest person in Asia, with a net worth of $105.5 billion. The man who said he was not very social had built the most socially consequential machine of his decade and then kept as far from it as a controlling shareholder can.
The company did not slow down, and neither, apparently, did his conviction that machines should do the work. ByteDance formed an AI team called Seed in 2023 and launched a chatbot, Doubao, in August of that year. In December 2023 The Verge reported that ByteDance had been using OpenAI's API to help build its own models. OpenAI suspended the account pending investigation. ByteDance said its use had been licensed for markets outside China and that ChatGPT-generated data had been removed from its training sets. Volcano Engine, the cloud unit founded in 2021, released AI video generators in 2024. In February 2025 the company demonstrated OmniHuman-1, which could produce realistic video from a single photograph and an audio clip. On December 1, 2025, it announced a Doubao-powered voice assistant for Chinese smartphones, starting with ZTE's Nubia M153. In February 2026 it released Seedance 2.0, a video model whose clip of an AI-generated "Tom Cruise" fighting an AI-generated "Brad Pitt" went viral. The Motion Picture Association denounced what it called "massive" infringement. Reuters framed the release as China looking for its "second DeepSeek moment."
By March 2026 ByteDance was reportedly working with Aolani Cloud to deploy about 500 Nvidia Blackwell systems in Malaysia, roughly 36,000 B200 chips, computing power placed in a third country, outside both of the jurisdictions that had made the company's life difficult. Bloomberg reported that it was building out an AI team in the United States. In April 2026 a British investor, Fred Blackford, was reported to have built a stake in ByteDance worth more than $500 million.
The pruning went on alongside the expansion. LetsChat, a messaging app marketed mainly in Nigeria for its low data use, reached 7 million downloads before it was discontinued in March 2024. TikTok Music, formerly Resso, shut down in November 2024. TikTok Notes, a photo app, lasted from April 2024 to May 8, 2025. ByteDance announced a retreat from gaming in November 2023, and Moonton, the Mobile Legends studio it had won in a contest with Tencent in 2021, was sold to Savvy Games Group in 2026. In August 2026 TikTok said it would close its Nashville office, where about 250 people, including members of its content moderation operation, worked. That same month it introduced a setting that lets users decide what happens to their accounts after they die.
The machine had started planning for the end of its users. Its founder had already arranged to be almost absent from his own.
Toutiao Helped Find Him
Go back to 2018, the year of the letter.
For anyone trying to understand Zhang Yiming, the Attention Factory account of ByteDance's rise and the archive of Western investigations into what came after point in the same direction: to a founder who preferred building systems to explaining them, and who explained himself publicly, at length, only once, when the state required it. The letter of April 2018 is still the most revealing document he has produced. It is also the least characteristic. It is a man who built machines to infer desire being made to state his allegiance in writing.
Over the years ByteDance added features to Toutiao that had little to do with headlines or engagement. One of them was a missing-person alert system. It used the same location-aware, interest-aware distribution that sent jokes and news to hundreds of millions of people to send a stranger's face to the phones of the people most likely to have seen them. By June 2020 those alerts had helped find 13,116 missing people.
In February 2018, two months before the letter, during the Lunar New Year travel rush, when hundreds of millions of Chinese workers crowd onto trains and buses to go home, a migrant worker lost track of his fifty-six-year-old brother, who was mentally ill. The South China Morning Post told the story in a single headline: the man had lost his sibling in the travel crush, and Toutiao helped find him.
Part IIThe Playbook
What follows separates Zhang Yiming's operating principles from the circumstances in which he applied them. Some of these principles are portable. Others only worked because ByteDance was the right company at the right moment in history, and some came with costs that most operators should not want to pay. Each principle is grounded in the record above and ends with a tactic.
The principles
- 1.Reverse the direction of the transaction.
- 2.Treat rejection as a pricing error, not a verdict.
- 3.Size your ambition to the four-fifths.
- 4.Buy the audience, keep the engine.
- 5.Let the data choose between your own teams.
- 6.Flatten the title, concentrate the vote.
- 7.Turn internal plumbing into product.
- 8.Build twins, not one global product.
- 9.Price the sovereign in as a co-founder.
- 10.Kill fast, and keep killing.
- 11.Hand the operating job to the executor.
- 12.When forced to sell, keep the brain's dividend.
Principle 1
Reverse the direction of the transaction
ByteDance's founding insight was not "use AI" or "build for mobile." Both were commonplace in 2012. It was that the dominant interaction model, search, asked users to do work they increasingly did not want to do, and that a phone made that work more expensive. Zhang's critique of Baidu had two parts: users struggled to find information in mobile apps, and search results were contaminated by undisclosed advertising. His answer did not improve search. It removed the search box. Toutiao pushed content. Douyin and TikTok pushed video. The user's job shrank to swiping.
The general lesson is to look for markets where the customer is doing labor the product could absorb. The most valuable reversals tend to look, at first, like a loss of control for the user. That is exactly why incumbents dismiss them, and why venture capitalists in 2012 dismissed this one.
The risk is built into the benefit. A product that removes the burden of choosing also removes the user's sense of having chosen, and that opens the door to arguments about addiction, manipulation, and influence that ByteDance spent the next decade fighting on three continents.
Tactic: List the three most labor-intensive steps your customer performs before getting value, and prototype a version of your product in which the system performs the most tedious one by default.
Principle 2
Treat rejection as a pricing error, not a verdict
Most of the venture capitalists Zhang approached did not share his view. Sequoia turned him down, then led a $100 million round in 2014 once Toutiao had more than 13 million daily users. Susquehanna was the outlier who bet early, and Susquehanna's managing director Arthur Dantchik sits on ByteDance's board today.
Rejection here was not information about the idea. It was information about how far the idea was from the investor's mental model, and that distance was itself the opportunity. Zhang did not soften the thesis to raise money. He found the one investor who understood it, shipped, and let usage make the case. The investors who said no were not wrong about the facts they had. They were looking at a recommendation engine through a search-era lens.
The long tail of that first check is the part founders rarely think about. Twelve years later the Susquehanna relationship mattered in Washington. The earliest believers in a company are not just capital. They become the company's longest-running constituency.
Tactic: When an investor passes, write down the specific assumption behind the "no," then design a metric you can hit in the next two quarters that directly disproves it.
Principle 3
Size your ambition to the four-fifths
Zhang's statement that China held only one-fifth of global internet users, and that a company that did not expand would "lose to peers eyeing the four-fifths," is a sizing argument disguised as a strategic one. It set the denominator. A domestic champion measures itself against domestic rivals. Zhang measured ByteDance against the world. That is why an English-language Toutiao was part four of a plan presented in January 2013, and why he insisted Lark target America, Europe, and Japan.
The global frame changed concrete decisions: which acquisitions to make (Musical.ly, Flipagram, News Republic), where to invest (BABE in Indonesia), and how to set up the corporate structure (a Cayman Islands holding company, with TikTok based in Singapore and Los Angeles). A founder who sizes ambition to the home market builds an org chart for the home market, and that org chart becomes a ceiling.
The four-fifths, though, are not one market. India was 43 percent of TikTok's new users in 2019 and a permanent ban by 2021. Global ambition multiplies the number of sovereigns whose patience you depend on.
Tactic: Re-run your TAM analysis with the global user base as the denominator, and identify the one decision you are making today that would change if you believed it.
Principle 4
Buy the audience, keep the engine
The Musical.ly acquisition is a lesson in being clear about what you are buying. ByteDance did not need Musical.ly's technology. It already had the recommendation engine that powered Douyin. It needed Musical.ly's users, its creators, its American presence, and a few features (the duet among them). So it bought the company, merged accounts and data into TikTok on August 2, 2018, retired the Musical.ly brand, and kept its own engine. Two months later TikTok was the most downloaded app in the United States.
Many acquirers make the reverse mistake. They buy for the audience, then let the acquired product's inferior core survive out of politeness or integration fear. ByteDance's discipline was to treat the acquired company as distribution and its own system as the irreplaceable part.
The same logic was later used against ByteDance. In the 2026 joint venture, American investors kept the TikTok shell and set out to retrain the algorithm on U.S. data. Anything you can do to an acquisition, a regulator can do to you.
Tactic: Before any acquisition, write a single sentence naming the one asset you are buying, and commit in advance to retiring everything that competes with your own core system.
⏱
The Decisions That Built ByteDance
A founder's timeline, read as a sequence of bets rather than a biography.
2009
Takes over Kuxun's real-estate search unit as 99fang.com with Liang Rubo.
2012
Rents a four-bedroom apartment in Zhongguancun; ByteDance founded March 13; Toutiao launches in August.
2013
Four-part plan presented to executives includes an English-language Toutiao.
2014
Sequoia, which had passed, leads a $100 million round.
2016
AI Lab founded under Wei-Ying Ma; Douyin launches September 20 after nearly seven months of development.
2017
TikTok launches internationally; Musical.ly acquired November 9.
2018
Neihan Duanzi shut down; public self-criticism letter; Musical.ly merged into TikTok August 2.
2020
Kevin Mayer hired and gone within roughly three months; Trump orders; India ban.
2021
Zhang hands CEO role to Liang Rubo; state takes 1% golden share in domestic unit.
2023
Seed AI team formed; Doubao launches.
2025
TikTok goes dark in the U.S. on January 18, returns the next day.
2026
U.S. joint venture closes January 22; ByteDance keeps 19.9%; Seedance 2.0 released.
Principle 5
Let the data choose between your own teams
"Horse racing," as former employees described it to the Wall Street Journal, assigns several teams to build the same product and gives more support to whichever version performs better. It is the For You feed applied to the organization: generate variants, measure, allocate attention to the winner. It turns internal politics into an experiment, at least in principle.
The practice fits the founder. A man who distrusts editors will distrust managers who act as editors of ideas. Competing teams answer the question "who decides?" with "the users, measured." It also lets a company explore several product hypotheses at once without betting on one leader's taste.
The costs are real and documented: duplicated work, burnout, a culture former employees described in punishing terms, and a median tenure of about three years. Horse racing works when the races are short, the metrics are unambiguous, and the losers are redeployed instead of discarded. Without those conditions it becomes an attrition machine.
Tactic: For your next high-uncertainty product bet, fund two small teams with the same goal and a fixed four-week deadline, agree on the deciding metric in writing before either starts, and redeploy the losing team rather than releasing it.
Principle 6
Flatten the title, concentrate the vote
Zhang discouraged employees from calling him "boss" or "CEO," held bimonthly town halls, and modeled the company's culture on Google. He also kept more than 50 percent of ByteDance's voting rights, and still did after leaving both the CEO title and, by November 2024, the published board list. These are not contradictions. They are a design.
Informal culture speeds up the flow of information. Concentrated control protects the long-term strategy from that same flow. The Kevin Mayer episode shows the split clearly. Mayer had the titles, CEO of TikTok and COO of ByteDance, and lasted roughly three months. Formal authority and actual authority lived in different places, and everyone inside knew which mattered.
For founders the lesson is not to hoard control for its own sake. It is to be honest, at least with yourself, about where decisions are really made, and to make sure the people hired to carry titles understand that before they arrive.
Tactic: Write down, for each senior role you are hiring, the three decisions that person will make alone and the three that will remain with you, and share that document before the offer is signed.
⚖
Conventional Wisdom vs. the ByteDance Approach
Where Zhang's choices diverged from the standard founder playbook.
| Conventional wisdom | Zhang's approach |
|---|---|
| Improve search; help users find what they ask for. | Remove the search box; push what the system infers they want. |
| Win the home market, then expand. | Plan for the "four-fifths" in year one; English Toutiao in the 2013 plan. |
| Integrate acquisitions carefully, preserving their product. | Absorb the audience, retire the brand, keep your own engine. |
| Pick one team and one leader per product. | Race several teams; let measured performance decide. |
| Visible founder as chief storyteller. | Founder steps back; roommate runs the company; spokesman faces Congress. |
| One global product, one brand. | Twin apps with separate content and servers: Douyin and TikTok. |
Principle 7
Turn internal plumbing into product
Lark began as ByteDance's internal communication and collaboration tool and was released to outside users in 2019. BytePlus, announced in 2021, sold the software framework underneath TikTok to other companies. Volcano Engine turned the company's compute into a cloud business. Each time, a capability built to run ByteDance became something to sell.
This works because internal tools at a hypergrowth company are stress-tested at a scale few vendors can match. The plumbing has already survived a flood. Selling it spreads the cost of infrastructure you had to build anyway, and gives the company a second, steadier revenue base under a volatile consumer business.
The record also shows the limits. In April 2020 China's internet regulator ordered Lark taken down because it could be used to get around censorship. In 2023 authorities reportedly inspected Feishu, the domestic version, ahead of a Party Congress. Infrastructure that holds a company's secrets is also infrastructure a state may want to read. Productizing internal tools exposes them to more customers and to more scrutiny.
Tactic: Identify the internal system your team would be most upset to lose, and spend one sprint scoping what it would take to offer it, with security and data-residency controls, to a single design partner.
Principle 8
Build twins, not one global product
Douyin and TikTok share an interface and nothing else: separate content, servers in their own markets, different features. Douyin offers face-based search, shopping, hotel booking, and geotagged reviews. In China it carries a government-mandated "teenage mode" with whitelisted content and screen-time limits for users under fourteen. TikTok has a different feature set, different moderation guidelines (eventually rewritten by local teams), and a different relationship with regulators.
Compartmentalization let ByteDance follow incompatible regulatory regimes at once and shield each product's users from the other's rules. It also let Douyin go deeper into commerce than TikTok could. Running one engine under two personas was the only way to be both a Chinese company and a global one at the same moment.
The twins never fully reassured anyone. American critics read the gap between the two apps' treatment of minors as evidence of malice. Former employees described overlaps in personnel and product development that blurred the separation. Leaked audio showed data moving across the line. Separation is a structure, and a structure is believed only as far as it is enforced.
Tactic: If you operate across jurisdictions with conflicting rules, define in writing what data, staff, and code may never cross between your regional products, and commission an outside audit of that boundary before a regulator asks for one.
Principle 9
Price the sovereign in as a co-founder
The April 2018 letter is the moment ByteDance's real cap table became visible. Before it, the company had a Party committee (since 2014, according to state media). After it, it had police officers stationed at headquarters, a 1 percent golden share in its main domestic unit, a board member with a background in propaganda, and a standing obligation to report "historical nihilism." On the other side of the Pacific it had CFIUS, an FTC consent decree, a law forcing a sale, and in the end a 19.9 percent stake in its own American business.
For a company at this scale, governments are not background risk. They are participants with veto power, and they hold it unevenly. Zhang's mistake, if it was one, was not underestimating Beijing. It was building a product whose central asset, a model of what hundreds of millions of people want, was exactly what every government would come to see as strategic.
Founders working in sensitive categories (attention, data, finance, health) should model regulators the way they model competitors: their incentives, their likely moves, their timelines. ByteDance's lobbying spend of more than $17.7 million by mid-2023 came after the political fight had already started. Earlier and quieter would have been cheaper.
Tactic: Add a "sovereign" column to your risk register for every country you operate in, naming the agency most able to shut you down, the trigger most likely to make it act, and the concession you would offer first.
✂
The Pruning Record
ByteDance kills and sells products as readily as it launches them.
| Product | Status | What happened |
|---|---|---|
| Neihan Duanzi | Killed | Shut down in 2018 after regulator order; had 200M+ users. |
| TopBuzz | Killed | Launched 2015, abandoned 2020 as business dwindled. |
| Vigo Video (Flipagram) | Killed | Acquired 2017, shut October 31, 2020. |
| Gogokid | Killed | Closed 2021 after China's tutoring crackdown. |
| LetsChat | Killed | 7M downloads, mostly Nigeria; discontinued March 2024. |
| TikTok Music | Killed | Shut November 2024. |
| 8th Note Press | Killed | Founded 2023; closed June 2025, rights returned to authors. |
| Moonton | Divested | Acquired 2021; sold to Savvy Games Group in 2026. |
| CapCut | Scaled | 200M+ monthly users by 2023; #2 U.S. download in March 2023. |
| Doubao / Seed | Scaling | AI assistant launched 2023; voice assistant on phones by late 2025. |
Principle 10
Kill fast, and keep killing
The pruning record above is not a list of failures. It is evidence of a system. ByteDance launches into many categories (news, music, publishing, gaming, education, messaging, photo-sharing) and withdraws quickly when an experiment stalls or the environment turns hostile. TikTok Notes lasted about a year. Nuverse's gaming push was reversed within roughly four years. 8th Note Press lasted two.
The habit follows from the company's operating logic. A business that runs on experimentation has to be equally fluent at ending experiments, or it drowns in its own portfolio. Quick closure also limits the reputational and regulatory exposure of products that drift into trouble. The 8th Note Press closure returned rights to authors, which at least ended the venture cleanly.
The risk is to trust. Partners, creators, and authors who build on a product that vanishes will be more cautious next time. A company that kills easily should say so early and design exits that protect the people downstream.
Tactic: For every new product line, set a dated review checkpoint and a pre-written shutdown plan, including how partners and users will be made whole, before launch.
Principle 11
Hand the operating job to the executor
Zhang's succession was unusually clean. He chose Liang Rubo, his college roommate and co-founder of two companies, whom the press called "the executor," and explained the decision with candor: "I'm not very social." The Mayer experiment the year before had shown the opposite approach failing. An outside star installed as a public face, without real authority or a shared history, lasted about three months.
The lesson concerns what founders are for. Zhang's comparative advantage was thesis and system design: recommendation over search, global over domestic, data over hierarchy. Running a company of more than 100,000 people under pressure from two superpowers is a different job, and the person best suited to it may be the co-founder who has always done it quietly. Stepping aside while keeping majority voting control let Zhang move his attention, by every outward sign, toward the next system, artificial intelligence, without giving up direction of the company.
Succession by a trusted insider has a cost. It keeps the company's culture and blind spots intact. A clean handoff is not the same as a fresh perspective.
Tactic: Write down the two or three activities where your own leverage is highest, then name the person who could run everything else, and test them with full authority over one division for six months before any formal transition.
Principle 12
When forced to sell, keep the brain's dividend
The U.S. joint venture that closed on January 22, 2026, looks like a defeat, and in Washington's framing it was one. Read the terms. ByteDance kept 19.9 percent. The ByteDance-controlled global TikTok entity kept e-commerce, advertising, and marketing on the U.S. platform. The American investors got governance of data and the job of retraining the algorithm. ByteDance kept a large share of the revenue and kept selling into the market it had been forced to give up.
The most valuable negotiating position in a forced sale is to separate control from economics. Regulators usually care about control: who sees the data, who signs the code, who sits on the board. Sellers can often keep a large share of the value by giving up control willingly while defending revenue functions and equity hard.
Control has its own consequences. Within weeks, the new owners faced accusations of censorship, which TikTok blamed on a data-center power outage, and a competitor surged on the backlash. Even a well-structured exit leaves the seller's brand tied to decisions it no longer makes.
Tactic: In any regulator-forced divestiture, separate the asks into "control" and "economics," concede the former early and visibly, and hold the line on equity and revenue-generating functions.
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
— Zhang Yiming, public letter on the shutdown of Neihan Duanzi, April 2018…incommensurate with socialist core values…
— Zhang Yiming, explaining his decision to step down as ByteDance CEO, as quoted by Reuters, May 2021I'm not very social.
— Zhang Fuping, ByteDance vice president and Party Committee Secretary, as cited in a report to the Australian Parliamenttransmit the correct political direction, public opinion guidance and value orientation into every business and product line.
— Shou Zi Chew, TikTok CEO, internal memo on the U.S. joint venture, January 2026our U.S. users to continue to discover, create and thrive as part of TikTok's vibrant global community and experience
Maxims
- Remove the box. The biggest products often come from eliminating the step customers hate most, even when the step is the one they think they need.
- A "no" is a coordinate. An investor's rejection tells you how far your idea is from their model, which is often the size of your opportunity.
- Size to the four-fifths. The denominator you choose in year one decides the org chart you build for the next decade.
- Know what you bought. In an acquisition, name the single asset you need and retire everything that competes with your own core.
- Let the race decide. Competing teams with a shared metric replace politics with evidence, provided the races are short and the losers are redeployed.
- Titles are cheap; votes are not. Informal culture and concentrated control can coexist, but say plainly where decisions actually live.
- Your plumbing is a product. Infrastructure that survived your own hypergrowth is worth selling, and worth securing before you sell it.
- Every state is a shareholder. At scale, governments hold vetoes, so model them as carefully as you model competitors.
- Kill on schedule. Experimentation without disciplined closure becomes clutter, so write the shutdown plan before the launch plan.
- Sell the control, keep the cash. In a forced exit, give regulators the governance they want and fight for the economics they don't.
