Part IThe Story
Wait, Let Me Check the Number
About a quarter of the way into a closed-door meeting with investors on May 20, 2026, a meeting that ran close to four hours, the founder of the most consequential artificial intelligence lab outside the United States stopped in the middle of a sentence to do arithmetic.
He had been explaining how DeepSeek priced access to its models. The rule, he said, was simple. Buy a batch of servers on the open market and charge enough to earn back their cost in ten months. The accountants might depreciate the machines over three or five years, but commercially ten months was enough. "OK, we think that's enough." Then he made a remark that would have cleared the room at most American AI companies. At DeepSeek's prices, he said, demand was inelastic. Raise the price by half, or double it, and token consumption would barely change. Double the price and revenue would nearly double too. The transcript, assembled afterward by machine and corrected by hand, records what came next in the flat way transcripts record everything: 稍等一下，我对一下。哎呀，太好了。 Wait a moment, let me check. Oh, great.
He did not say what he checked. A number on a second screen, maybe a message, maybe a calculation about money he had chosen not to collect. Then Liang Wenfeng went back to explaining why he would leave it on the table.
Neither DeepSeek nor Liang has confirmed that the transcript is authentic. It surfaced in July on X under a handle that cannot be printed here. Within days it was on GitHub, on Reddit, on mirror sites in several languages, and on chapter-by-chapter annotation pages with English translations and timestamps. Chinese platforms began taking copies down. One Reddit poster who had translated the full text advised readers to fork the repository quickly: "you know how it is." In the transcript, an investor addresses Liang at one point as "杨哥," Brother Yang, which is almost certainly a speech-recognition error. The most guarded founder in the industry had become, for four hours, a text of uncertain provenance, transcribed by software and partly misheard.
According to Bloomberg, he was unhappy about the leak, and DeepSeek suspended a second funding round it had begun discussing in mid-July at a valuation of roughly $70 billion. The first round had closed weeks earlier. It raised about 50 billion yuan, around $7.4 billion, at a post-money valuation in the low-to-mid $50 billions, and it was structured so that most of the money flowed into a limited partnership Liang controlled. The investors received no voting rights and agreed to a five-year lock-up. The largest contributor, at about 20 billion yuan, was Liang himself.
The rest of the transcript reads like that pause for arithmetic. A man who could charge more explains, patiently and in plain words, why he won't. Over the course of the afternoon he describes himself as one of "a group of very ordinary people." He calls his strategy 克制, restraint, and says plainly that restraint is not modesty but a way of winning. "You take less," he says in one form or another, "and you are more likely to get it done."
Taken literally, it is an odd thing to hear from the founder of a company valued above $50 billion, and odder still from a man Bloomberg's Billionaires Index had just named the world's wealthiest founder of an AI-model company, worth $36 billion and ahead of Dario Amodei and Greg Brockman. Taken seriously, it is an operating philosophy, and possibly the most complete one any frontier-lab founder has put on the record. He never meant to put it on the record. It got there anyway.
By the Numbers
The Restraint Ledger
10,000Nvidia A100 GPUs High-Flyer acquired before U.S. export controls
10 monthsDeepSeek's target payback period on server hardware, by Liang's account
$592.7BNvidia market value erased in a single day after R1, January 27, 2025
95%+Price cuts by Alibaba and Baidu after DeepSeek-V2 set off a price war
~20,000H100-equivalent GPUs DeepSeek held as of the May 2026 meeting
$52BPost-money valuation after DeepSeek's first outside round, 2026
160Employees at DeepSeek in 2025
The Village That Sells Zhuangyuan Juice
Mililing is in Qinba town, on the outskirts of Wuchuan, a small city administered by Zhanjiang in the far southwest of Guangdong. Zhanjiang is a port city of shipyards and engineering works, and it has never been rich. Until early 2025 almost nobody in China had heard of Mililing. Then visitors began arriving by the hundreds each day. Villagers, many of them also named Liang, set up stalls near his childhood home and sold souvenirs, drinks, and zhuangyuan sugar cane juice. Zhuangyuan was the title given to the top scorer in the imperial examinations, the poor scholar who rose by intellect alone and came home famous. The local government paid to plant trees, widen roads, and demolish whatever made the place look shabby. A hamlet was getting a makeover because a man who barely appears in photographs had become a national symbol.
The region already had the mythology waiting. Local boosters describe Wuchuan as the "only hometown of a top scholar in western Guangdong." Over the centuries it produced one imperial zhuangyuan, 20 jinshi, and 165 juren. The sugar cane juice simply put a new face into an old slot.
The face was harder to fill in than the slot. Both his parents taught primary school in the town, and by one account both taught Chinese language. People who knew them use the same words for them: honest, simple. People who knew him as a boy remember two things side by side. He read comic books, and he was very good at mathematics. A middle-school teacher identified in local reports as Mr. Rong recalled that by middle school "he had already mastered high school mathematics and was delving into university-level concepts." Rong also added something that the hero narrative tends to drop: Liang was not a stereotypical bookworm, and he learned in "unique and innovative" ways. A graduation photograph of Wuchuan No. 1 Middle School's Class of 1999, Grade 9, Class 5, has since circulated online. He stands in the back row, sixth from the right. You have to be told where he is to find him.
In high school he took up mathematical modeling and small research projects. He placed in the top three on every math exam. On the gaokao, the national university entrance exam, he scored 806, the highest in the Zhanjiang region. That score cleared the bar for Tsinghua, which for a rural Guangdong student in 2002 was the obvious end of the story.
He chose Zhejiang University instead, because it had the program he wanted. He enrolled at seventeen.
The choice matters later. Twenty-four years afterward, explaining to a room of investors why DeepSeek is a story about ordinary people, he says: "I myself am just a university graduate, and not from the very top school either." It is literally true. It leaves out that he declined the very top school. The sentence is restraint carried into autobiography: a fact about himself, presented so that it shrinks him.
The Man Who Declined the Drone
A classmate from Zhejiang University's Electronic & Information Engineering program, class of 2002, later wrote on Zhihu that he barely knew Liang. They were in different sections and different dorms and overlapped mainly in an electronics design contest. One impression stayed with him. In sophomore year, while everyone else attended lectures and handed in problem sets, Liang was already teaching himself digital and analog circuits and building his own engineering projects. It is the same pattern as the middle-school math: the school's sequence was there, and he went past it on his own.
He earned his bachelor's in electronic information engineering in 2007 and stayed for a master's in information and communication engineering under Xiang Zhiyu. He finished in 2010 with a thesis titled "Study on Object Tracking Algorithm Based on Low-Cost PTZ Camera." The title is easy to skim past, and it shouldn't be. A pan-tilt-zoom camera is a cheap surveillance device. The research problem was how to get sophisticated tracking out of commodity hardware. The word "low-cost" is right there in the title, years before it would become his signature.
During his student years Wang Tao, the founder of DJI, asked Liang to join him as a co-founder. DJI became the world's dominant drone maker. Liang said no. The reason he gave was that he wanted to apply artificial intelligence to financial markets. He has since said that the people around him had entrepreneurial instincts, while he himself valued academics. That is a curious thing for a future billionaire founder to say, unless you take it at face value.
The financial interest had a date attached. In 2008, as the global financial system came apart and margin calls went out across the Pacific, Liang, then twenty-three, formed a team with classmates to collect market data and macroeconomic indicators and to test whether machine learning could trade. He and two classmates began trading domestic stocks that year. None of them had overseas or institutional experience, which was unusual among the founders of Chinese quant funds. They tried discretionary trading. They tried arbitrage. They kept going.
In 2009 he interned at Shanghai Aiqi Info Tech, a company founded by a fellow Zhejiang alumnus, Zhou Chaoen. Zhou later said Liang was quickly promoted to manager of the new technologies department, at a monthly salary of 16,000 yuan, and that he spent whole days in the office on AI video and imaging problems. Zhou's description has the plainness of a reference letter written by someone who meant it: "He is calm and straightforward, with a rigorous approach to products and technology, exemplifying the typical characteristics of a technical expert." Reserved, Zhou added, but clear when the discussion was technical.
After graduating he did not take a programming job at a large software company, as most of his classmates did. He moved into a cheap flat in Chengdu, in Sichuan, and tried applying AI to one field after another. They all failed. A high school friend identified as Mr. Chen remembers him building automated embroidery machines at one point, starting from scratch and refining as he went. The image fits him well: a machine-vision engineer from a schoolteacher's household, in a rented room in a provincial city, teaching a machine to stitch.
The idea that eventually worked was the one he had started with. In 2013 he and Xu Jin, another Zhejiang alumnus, founded Hangzhou Jacobi Investment Management. The name, Yakebi in Chinese, honors the nineteenth-century mathematician Carl Jacobi. Two years later the same pair founded Hangzhou Huanfang Technology. In February 2016, Liang and two engineering classmates registered Ningbo High-Flyer Quantitative Investment Management Partnership.
There Must Be a Way to Model Prices
The Chinese name of the fund is Huànfāng, 幻方, "magic square": a grid of numbers whose rows, columns, and diagonals all add up to the same total. Tradition says the object was first discovered in China. It is the kind of name a mathematician gives a hedge fund, half joke and half statement of belief, the belief being that if the numbers are arranged correctly, the order underneath them shows itself.
The founding myth is precise to the day. High-Flyer's own website says the firm's first trade generated by a deep learning algorithm was executed on October 21, 2016, and that by 2017 deep learning was fully deployed in its trading. The fund had started with models built on price and volume factors. Machine learning let it find new factors and what the firm called "non-linear" relationships among them. Assets grew from about 1 billion yuan in 2016 to more than 10 billion by 2019. One data provider ranked it China's top stock hedge fund that year. By the end of 2021 it managed more than 100 billion yuan.
The patron saint was American. In February 2021 a Chinese edition of Gregory Zuckerman's The Man Who Solved the Market: How Jim Simons Launched the Quant Revolution was published, with a preface by Liang. Simons was a mathematician who founded Renaissance Technologies, one of the most successful funds in history, and he became the model for a whole generation of systematic traders. Liang wrote that whenever he hit difficulty at work, he would think of a line of Simons': There must be a way to model prices. Four years later, after DeepSeek had wiped hundreds of billions off American market capitalization, Zuckerman wrote in the Wall Street Journal that he had discovered his Chinese preface writer only belatedly and had been trying to reach him without success. Like Simons, Zuckerman observed, Liang was secretive and very hard to contact. The biographer of one recluse had been blurbed by another and couldn't get him on the phone.
Liang's single real public appearance in that era came on August 30, 2019, at the China Private Equity Golden Bull Awards in Shanghai. He gave a keynote titled "The Future of Quantitative Investment in China from a Programmer's Perspective." It went viral among Chinese institutional investors, which a speech about asset management almost never does. He opened by calling American asset managers China's "teachers." He showed that hedge-fund leaderboards had flipped from mostly discretionary in 2004 to mostly quantitative by 2018, and argued that a top Chinese team could someday manage 200 to 300 billion yuan. Then he offered a definition with no hedging in it. What makes a fund quantitative, he said, is not its tools but who decides. In a real quant fund, nobody does. There is no portfolio manager. There are servers. He added that on the technical side, programs "already far surpass most human capabilities," and that the investing capability of quant funds was doubling roughly every eighteen months, a Moore's Law for alpha. The stated mission of High-Flyer, he said, was to make China's secondary market more efficient.
The servers were real, and he built them himself. In 2019 High-Flyer brought up a computing cluster called Fire-Flyer and retired it after a year and a half. Its successor, Fire-Flyer 2, had by 2022 grown to 5,000 PCIe A100 GPUs across 625 nodes, eight to a node, linked by 200-gigabit interconnects in a dual fat-tree topology. The firm wrote much of its own software stack: a distributed file system called 3FS, built for random reads that would never be repeated; a communication library called hfreduce, written to replace Nvidia's own collective communications library; and a training framework called HaiScale. A hedge fund was quietly rewriting parts of Nvidia's infrastructure because Nvidia's versions weren't efficient enough for its workloads. Few people outside the building knew. Those who did mostly assumed it was about trading.
Then the magic square stopped adding up. In the turbulent Chinese market of 2021 and 2022, High-Flyer lost money and apologized to its investors at length. It was an unusual stumble for a firm built on the idea that the machine knows better, and its assets fell by more than a third from a peak above $12 billion. The quant who had told a ballroom in Shanghai that there are no portfolio managers, only servers, found that servers can be wrong too, and that a fund's investors want a human to say sorry.
Ten Thousand Cards and a Terrible Haircut
During those same difficult years, Liang began buying GPUs on a scale that made no sense for a trading firm.
Through 2021 he bought them in the thousands for what looked like a side project, while still running a fund under pressure. By the time Washington moved to restrict sales of advanced Nvidia chips to China, High-Flyer had reportedly accumulated 10,000 A100s. According to a report in Finance Eleven, fewer than five companies in the country owned more than 10,000 GPUs. The others were the tech giants. Ten thousand A100s was widely treated as the threshold for training a serious large language model. A quant fund in Hangzhou had crossed it before most people in the industry had stopped thinking of chatbots as a curiosity.
His partners did not get it. One of them described the early impression to the Financial Times, and it is the most vivid description of Liang anyone has given on record:
— A business partner of Liang Wenfeng, to the Financial TimesWhen we first met him, he was this very nerdy guy with a terrible hairstyle talking about building a 10,000-chip cluster to train his own models. We didn't take him seriously. He couldn't articulate his vision other than saying: "I want to build this, and it will be a game change." We thought this was only possible from giants like ByteDance and Alibaba.
When asked about it in 2023, Liang refused to dress it up. "Many people would think that there is an unknown business logic behind this," he said of the chip purchases, "but in fact, it is mainly driven by curiosity." This is the second recurring pattern of his career. The first is low cost, and the second is that he declines to supply a strategic story for a move whose most strategic feature was timing. "Curiosity" is a strange explanation for the decision that, in hindsight, kept the most important AI lab in China alive through the export controls.
On April 14, 2023, High-Flyer announced on WeChat that it would concentrate resources to "explore the essence of AGI." Days later a managing director told Yicai that High-Flyer would not use AGI to trade stocks: AGI, he said, "has nothing to do with finance." On July 17 the lab was spun out as Hangzhou DeepSeek Artificial Intelligence Basic Technology Research Co., Ltd. Its Chinese name, Shēndù Qiúsuǒ, means something like "deep seeking," and the second half of the phrase carries a literary echo of searching without end.
Venture capitalists were not interested. An AGI lab with no product and no exit horizon looked to them like a philanthropy project. So High-Flyer funded it. By May 2024, corporate filings showed Liang holding an 84 percent stake in DeepSeek through two shell partnerships registered in Ningbo. The lack of outside money turned out to be structural. With no investors, there was no board demanding a product roadmap, no quarterly targets, and none of the KPIs that Liang would later say were the first thing he eliminated. "With DeepSeek's funding model," as one account put it, "it lacks commercial pressure."
In the May 2023 interview with Waves, a sub-brand of 36Kr, which was translated a year and a half later by a former DeepSeek intern named Zihan Wang, then a first-year computer science PhD student at Northwestern, Liang laid out a plan that read at the time like the manifesto of a well-funded eccentric. The lab would not just replicate ChatGPT. It would do research. It would hire for curiosity. It would treat the collective curiosity of a group of people as its main weapon. That a former intern did the translation fits the story. DeepSeek's case to the English-speaking world was first made, in effect, by one of its own junior people.
The Accidental Catfish
The first models were unremarkable. DeepSeek-Coder came out on November 2, 2023, with an architecture very close to Meta's Llama. DeepSeek-LLM followed on November 29, then a mixture-of-experts model on January 9, 2024. In April a math model introduced Group Relative Policy Optimization, a leaner variant of a standard reinforcement-learning method. Outside a small circle, almost nobody noticed it. A year later almost everybody did.
In May 2024 came DeepSeek-V2, built on a new mechanism called multi-head latent attention. MLA compresses the memory a model needs while generating text, and by doing so it cuts the cost of serving it. DeepSeek priced V2 to match. China's largest companies then did something they had not planned to do. Alibaba and Baidu cut their own model prices by more than 95 percent. A Chinese financial outlet ran a headline asking, with some disbelief, whether the "butcher" in the price war was really a quant fund.
Liang's explanation, given to Waves in July 2024, was close to embarrassed. The team, he said, had not expected pricing to be so sensitive. In a phrase that circulated widely afterward, DeepSeek had not meant to become a catfish (the fish dropped into a tank of sardines to keep them moving) but had become one by accident. He also described MLA's origin in a way that says a lot about how he runs things. The architecture, he said, came from a young researcher's personal interest. Nobody assigned it. Someone was curious, the company had compute to spare, and the cost structure of a whole industry shifted.
His hiring philosophy came straight out of that. "If you are pursuing short-term goals, it is right to find people with ready experience," he had told 36Kr in 2023. "But if you look at the long-term, experience is not that important. Basic skills, creativity, and passion are much more important." He went further than most founders who say similar things: "When doing something, experienced people will tell you without hesitation that you should do it one way. But inexperienced people will repeatedly explore and think seriously about how to do it, and then find a solution that suits the current actual situation." DeepSeek hired graduates straight out of university and hired people from outside computer science, including literature and advanced mathematics, so the models would absorb more than code. "Everyone has their own unique journey and brings their own ideas with them," he said, "so there's no need to push them."
To understand how unusual this was in China, it helps to know a contemporary who took the opposite view. In March 2023 Robin Li, the CEO of Baidu and for twenty years the face of Chinese search, announced China's first ChatGPT rival. In an interview that year he said China would never recreate OpenAI's success, and that Chinese firms should focus on building applications on top of existing models. It was a reasonable view. It matched how Chinese technology had won for three decades: take a foreign invention, from smartphone apps to electric vehicles, and scale it faster than the inventors could. Robin Li spoke at the premier's symposium in early 2024. Liang would sit at the same table a year later and make the opposite case.
— Liang Wenfeng, interview with Waves (36Kr), July 2024China's AI can't be in the position of following forever. We often say that there is a gap of one or two years between China's AI and the United States, but the real gap is the difference between originality and imitation.
"In the past thirty years," he said in the same interview, the Chinese tech industry "has only emphasized making money, and ignored innovation. Innovation is not solely driven by business, it also needs curiosity and a desire to create." He framed the case in moral and national terms, though never in nationalist ones: "China must gradually transition from being a beneficiary to a contributor, rather than continuing to ride on the coattails of others." On open source, the issue on which Silicon Valley had long assumed it held the cultural advantage, he took that argument over: "Even if OpenAI is closed-source, it cannot stop others from catching up… Open-source is like a cultural practice, rather than a business practice." And with a bluntness that would later be quoted everywhere: "OpenAI is not a god and cannot always be at the forefront."
The real point was underneath the quotes. What China lacked, he argued, was not capital. It was confidence, and knowledge of how to organize dense talent into actual innovation. DeepSeek, he said, had not hired anyone particularly special. Its people were mostly educated in China. The moat was not the weights. The moat was the organization.
In December 2024 the organization released DeepSeek-V3. A month later came R1.
The Week of January 20
The date lined up almost too neatly. On January 20, 2025, DeepSeek released R1, a reasoning model it said performed on par with OpenAI's o1, under an MIT license, with a free app on iOS and Android. On the same day, in Beijing, Liang was one of nine people invited to speak at a closed-door symposium chaired by Premier Li Qiang on the draft of the annual government work report. In the CCTV footage he looks youthful and slight among grey-haired academics, officials, and heads of state conglomerates. For many people abroad it was the first time they saw his face. Bloomberg's Saritha Rai, who covers AI in Asia, later said her team had searched the company's website and the open internet and found no usable photographs of him until the state broadcaster supplied one. Asked what he looked like, she offered: slim, glasses, doesn't talk much. Baby-faced, the host suggested. Yes, she said.
The Guardian reported DeepSeek's claim that it had trained a model comparable to ChatGPT for less than $6 million. For comparison, Microsoft alone planned to spend about $80 billion on AI infrastructure that year. The number was disputed and qualified almost immediately, but the market barely paused over the qualifications. By January 27 DeepSeek had passed ChatGPT as the most downloaded free app on the U.S. App Store. That Monday Nvidia fell roughly 17 percent, a one-day loss in market value of $592.7 billion. The Nasdaq fell 3 percent. Constellation Energy, which had been trading as a bet on AI data-center power demand, fell 20 percent. American commentators reached for "Sputnik."
On the afternoon of January 27, according to his old teacher Mr. Rong, Liang was back in Wuchuan playing football with former classmates.
Chinese New Year's Eve fell on January 28, and he went home for it. He had messaged friends beforehand that he would be back for the holiday but intended to stay out of the spotlight. The village did not go along. Banners went up in the streets: A warm welcome to Wenfeng, the pride and hope of our hometown! Welcome home, Wenfeng! Your hometown is proud of you! Someone put up the red double-happiness arch normally used for weddings, hung with a couplet: Wenfeng returns home with great achievements, bringing new energy to rural revitalization. Visitors came to be photographed at the village entrance marker under the arch. He left on the morning of January 29, the first day of the lunar year. The Year of the Snake had started with a hometown boy treated as a returning examination champion, and the boy had left town by lunchtime.
Three weeks later, on February 17, he sat in the Great Hall of the People at a symposium on the private economy chaired by Xi Jinping, alongside Ren Zhengfei of Huawei, Pony Ma of Tencent, and Jack Ma. Fortune later described the seating as lifting him to "the status of Chinese business royalty." The South China Morning Post compared his place in national life to that of the nuclear and rocket scientists of the 1950s and 1960s. Self-published English-language biographies appeared within weeks. One, Liang Wenfeng: Leading DeepSeek to Shape the Future of AI, came out in hardcover on February 13, about three weeks after the app topped the charts. Another promised, in a subtitle longer than most of his interviews, The Hidden Genius of Liang Wenfeng and "the untold story of a visionary." The story stayed untold. He gave no interviews.
The honors kept coming through 2025: Time's 100 Most Influential People in AI, Fortune's 100 Most Powerful People in Business, Nature's 10, Fortune China's 40 Under 40, Entrepreneur of the Year. In November, with a reported net worth of $11 billion, he joined China's list of the 100 richest people. None of it came with a statement from him.
Abroad, the response hardened. On February 4, 2025, Australia's Department of Home Affairs ordered DeepSeek products removed from all government systems. The U.S. National Defense Authorization Act for fiscal 2026 told the Secretary of Defense and the Director of National Intelligence to remove AI developed by DeepSeek or owned by High-Flyer from their devices and their contractors' devices. The New York Times reported that dozens of DeepSeek researchers had current or past affiliations with People's Liberation Army laboratories and the "Seven Sons of National Defence" universities. From March 2025 the chatbot was adopted in non-combat roles by military hospitals, the People's Armed Police, and national mobilization bodies. In February 2026, Anthropic accused DeepSeek of using thousands of fraudulent accounts to generate millions of conversations with Claude for training. Microsoft Azure and Perplexity hosted DeepSeek models anyway, because the weights were open and anyone could host them. In Nairobi, startups built on them. The models spread through every channel the license allowed, and governments responded through every channel that law and suspicion allowed.
Sesame Seeds and Watermelons
The 2026 transcript shows him thinking through these years from the inside, and the plainest example he gives concerns the user surge that followed R1.
When DeepSeek went viral over Spring Festival in 2025, he told the investors, it had not been in the plan: "that wasn't in our script; we never thought about it." Millions of users arrived, and the standard playbook was obvious. Spend heavily, keep them, monetize them, build the next super app, take on ByteDance and Tencent. He said he could have done it and didn't. He described this as part of restraint and explained it with a farm image any Chinese listener would recognize:
— Liang Wenfeng, investor meeting, May 20, 2026 (translated)There are watermelons ahead; what's in front might just be sesame seeds. I shouldn't grab every sesame seed.
DeepSeek did keep the users, at low cost, "because it might be useful later." He says they kept coming whether DeepSeek wanted them or not: "for a while we didn't even want to maintain those users, but you couldn't drive them away." The phrase 赶都赶不走, you can't even chase them off, shows up more than once and carries an obvious satisfaction. At the time, he said, everyone was fighting for consumer traffic, "and it was taken by someone who wasn't even competing."
Pricing works on the same logic. Early on, worried about demand they couldn't serve, DeepSeek priced a model high, and he says the team was unhappy. He cut the price to a quarter. "Many people in the company group chat cheered." He presents this as evidence of what the company is for: "For our competitors, a price cut certainly isn't good news. They definitely wouldn't cheer. Because your revenue, your ARR, if you cut it in half, ARR drops by half." At DeepSeek it got cheers.
The ten-month rule is the mechanism that keeps this from being merely sentimental. Ten months to earn back a server, Liang says, works out to about six times profit. "Six times profit looks high," he concedes, "but it isn't." Someone in the chat window wrote that ten months was too greedy, and he agreed there was room to go lower. The rule does two jobs at once. It funds the company, and it shuts out rivals: at those prices, he said, independent third parties hosting DeepSeek's open weights couldn't make money, because they couldn't match its serving costs. That is why open-sourcing everything cost him nothing. "If you want to make a hundred times profit, open source will hurt you," he said. At six times, it doesn't. Open source is not charity in this framework. It is safe only for a company already charging close to cost.
The broader theory, delivered near the meeting's two-and-a-half-hour mark, is the closest he comes to a philosophy of history:
— Liang Wenfeng, investor meeting, May 20, 2026 (translated)Those who take more are beaten by those who take less. You don't even have to actually take more; if your vision is to take more, you will be beaten by someone whose vision is to take less.
He works through it like a proof. Suppose AI eventually accounts for 20 percent of world GDP. A company that aims to capture 5 percent will lose to one that asks for 1 percent, which will lose to one content with 0.1 percent. The process stops only when someone takes so little that the business can't survive. OpenAI, he said, had believed from the start that it could monopolize the world and would face challenger after challenger, including from China, "because the Chinese are willing to take less and still provide the service." He does not present this as a principle he picked. He presents it as an objective law, using a Chinese phrase with Marxist-textbook overtones, 客观规律: "If we wanted to monopolize this benefit, we would certainly be abandoned by history."
In the same breath he insists it is not a rationalization. "I believe in this. I'm not trying to find reasons for it, because there's no need. I already do it this way, and there must be a reason. The reason may not be conventional, but I think the company itself is unconventional." It is a circular piece of reasoning, and he seems aware of that and untroubled by it, much like a quant who trusts a signal he can't fully explain.
Seen from Hangzhou, the scale of the ambition is easy to miss. He talks as if DeepSeek wants very little: one slice of the market, no verticals, no chips of its own if they can be bought at a fair price. "If you run a power plant, you don't necessarily need to build generators." But the slice he wants is general intelligence, and the restraint is about how he gets there, not about the destination. "If the AI era will produce many trillion-level companies," he said, "I think we're one of them." The person cheering for price cuts and the person who expects to build a trillion-dollar company are the same person.
The Only Core Interest
Ask a founder what he cannot give up and the answer is usually control, or margin, or speed. Liang's answer, given unprompted, was people.
"Our biggest core interest is to maintain team stability. That is our biggest core interest, arguably the only core interest. As long as I can keep the team stable, I will definitely succeed, definitely achieve AGI. It's that simple." Money was not a problem, he said, and neither were resources. Setbacks would cost six months, maybe a year. Departures could cost everything. The first outside financing, he explained, had largely addressed that risk, because "everyone has received quite a lot of options, and the amounts are relatively large." Keep the most senior people, and everyone else stays even with fewer options, "because they didn't come entirely for the money."
The organization those people work in is barely an organization. "Actually, we have no organization," he told the investors. "It's vision-driven." The vision has never been written down, and he says people inside may understand it differently. He credits the idea to Jack Welch, the GE chief executive he admired most twenty years earlier. Most of what Welch said, Liang allows, has not held up, but one point did: a large company is governed by vision rather than rules, and vision "is what you do, not what you say." It is a nice irony that a Chinese AI founder learned to run a company with no hierarchy from the American executive most associated with ranking employees and firing the bottom tenth.
There is one structural rule. Work splits into two flows, top-down and bottom-up. Top-down is "formal": shipping V4, where everyone gets a piece. Bottom-up is whatever an individual decides to pursue, with no KPIs and nobody supervising. Formal work, Liang said, should not take more than half of anyone's time. Researchers own the other half. He added that there is little overtime, for two reasons. Research needs slack, and DeepSeek is very focused: "We are very focused, which means we have very few things to do." He mentioned, as if it were a feature, that many of DeepSeek's products are unfinished and nobody has gone back to finish them.
Even his roadmap to AGI is chosen partly with ease in mind. As he describes it, the field climbs a staircase. Chain-of-thought reasoning was last year's step, agents this year's. Next comes continual learning: the ability to absorb context over time the way a new hire does in their first two months, so that when you say "call Xiao Wang over," the system already knows who Xiao Wang is. After that comes a gradual, nonlinear "singularity" in which models help build their successors. Only then, embodied intelligence and robots that do housework and care for the elderly. He says the order matters because it is the easiest one: "This roadmap means we don't have to work overtime." Build robots first and the work is brutal. Build the self-improving researcher first and the robots largely follow. Asked what DeepSeek's models are for, he gave an answer few founders would say to investors: "the first goal of the model we build is not for everyone to use well, but for us to use well ourselves."
The organization is starting to strain. Asked about structure as headcount grows, he admitted that "our previous organizational structure was very fragmented, because there was no structure," and that some departments now needed "a relatively strict hierarchy." He said he was already making the change. He also offered a figure that cuts against the image of a lab running on curiosity: about half of DeepSeek's core researchers, "the most important people," were labeling data. China, he explained, has no cost advantage in high-quality annotation, so the people with the most taste were doing the drudge work. And half his company, he said in passing, thinks OpenAI's models are better than Anthropic's.
On the gap with America he was blunt, and unusually specific. It was not talent: the smartest Chinese researchers, he said, did not leave the country disproportionately. Roughly half stayed home, perhaps a little less. The gap was compute. DeepSeek had about 20,000 H100-equivalent GPUs, most of them arrived within the previous month or two, mostly Nvidia. The frontier American models, he estimated, activated around 800 billion parameters. China's largest activated a few tens of billions, an order of magnitude less. Training at the American scale would take something like 50,000 GB300s or 200,000 Huawei 950s, before any research budget. "Even if we spent all 50 billion, we couldn't afford to train it." He summarized DeepSeek's position as a narrative he wanted to change: one to two years behind America using one-twentieth of the compute. The goal was to cut the lag to six months, then three, on a fraction of the compute.
His spending plan for the new money was to turn it into chips as fast as possible. "If I could spend all the money within six months, that would be ideal," he said. A bank paid about 2 percent, and a GPU paid for itself in ten months. If the procurement team managed to spend 20 billion yuan in a year, that would count as outstanding performance. On domestic chips he was more optimistic than the American consensus. Huawei's 950 supernode, he said, could fully substitute for Nvidia's GB200 and GB300. If it cost 100 percent more, or even 200 percent more, that didn't matter. Nvidia's CUDA moat was eroding because AI can now write the software ecosystem that used to take a decade to build, and because of a DeepSeek-built high-level language called TileLang, in which, he said, V3 had already been trained on Nvidia hardware with almost no dependence on Nvidia's software. "Nvidia is digging its own grave," he said. The hardware gap, he estimated, was "four times plus two years."
He was firm that DeepSeek was a company. Asked whether it was a Bell Labs, he said no. Bell Labs did not need to commercialize, and DeepSeek did. "We ultimately have to survive; we are a company after all. The government won't give me a cent." That sentence sits uneasily next to the cap table of the round he had just closed, which reportedly included the Chinese National AI Fund alongside Tencent, CATL, JD.com, and NetEase. He chose those investors, he said, for being "the least hostile to us," the ones who most wanted DeepSeek to succeed. "Not everyone wants us to succeed, because we have still damaged the interests of many others."
What Leaks, and What Doesn't
The money that supported all this restraint was not inexhaustible. In July 2026 a global sell-off in AI chips hit China's momentum-driven quant funds. All nine of High-Flyer's publicly displayed products fell more than 20 percent that month. The worst, High-Flyer CSI 500 Quantitative Aggressive No. 1, lost 22.15 percent, and eight of the nine turned negative for the year. A Rhodium Group analyst, Ciel Qi, described High-Flyer's revenue as "unstable" and noted that maximizing returns in China increasingly meant investing in line with Beijing's strategic priorities. High-Flyer had already moved that way. It took pre-IPO allocations in the memory-chip maker CXMT, about $26 million, or 175 million yuan, and CXMT rose more than fivefold on debut. It took $5.8 million in Unitree Robotics. DeepSeek itself took 2.31 percent of Unitree's offering as one of nine strategic investors and accepted a 36-month lock-up when most backers accepted twelve. The fund that once described itself as servers making all the decisions was now buying into the national industrial strategy, and the lab that would not build robots was holding a robot company's shares for three years.
The pressure was clear. DeepSeek's capital needs were outgrowing what the hedge fund could provide. Bloomberg and the Financial Times reported in July that the company had begun preparing for an IPO, possibly listing as soon as 2027. A man who had said he did not start the company to go public was getting ready to. Asked in the meeting how he would balance pure research against public markets, he said he could have both. A few hundred million dollars of enterprise revenue, plus consumer users, made a commercial base. At a billion dollars in annual revenue, cash flow would cover all research spending. "In the worst case, just selling API could sustain a listed company." In late September a headline in one aggregator claimed that DeepSeek's annualized revenue had doubled to $1 billion within months. If it's true, the sesame seeds had reached the size he said they could.
The products continued arriving roughly every two to three months, the cadence he told investors he found comfortable. V3.1 came on August 21, 2025, V3.2 on December 1. A V4 preview arrived on April 24, 2026: a 284-billion-parameter Flash model and a 1.6-trillion-parameter Pro, each with a million-token context window, under the MIT license, and adopted by Huawei and Cambricon. The official V4-Flash shipped July 31, V4-Pro August 13, and V4.1-Flash on September 10. A research firm called V4-Flash by far the cheapest of the well-known models to run. On August 13 the company also released DeepSeek Harness, an open-source agent framework in which, according to its documentation, everything is a plugin.
Then there was the transcript. Its spread was the only part of the past two years Liang clearly did not want. For a few weeks the internet treated it as the most revealing document about Chinese AI ever published: Coco Feng of the South China Morning Post wrote that it "resonated deeply in a culture that traditionally values humility, discipline and long-term ambition." One blogger worked out the implied returns on a GPU paid off in ten months. A Chinese essayist titled his reaction "Take the Money, Leave the KPIs." An annotator added notes imagined from the year 2030, line by line. It struck many readers as unusually sincere, which is what made the leak sting. A man who had spent his career keeping his reasoning to himself had his reasoning read by everyone, in a form that might have been corrupted at any line, and whose authenticity he has never confirmed or denied.
The obvious way to read him is as the zhuangyuan: the top scorer from the poor village who came home in triumph, as the banners and the sugar cane juice suggest. He has told the story differently every time he has told it. In his version, a random group of ordinary people with little money, few chips, and no fame did something unusual because they wanted less than everyone else and were therefore under less strain. "You think we chose a hard mode," he told the investors. "But we've given up a lot elsewhere, so we're still very relaxed." The two versions do not reconcile, and he doesn't seem to need them to. The nation needed a hero, and he needed people to stay.
He is still mostly out of sight. A few photographs circulate. The best known shows him at the premier's table on the day R1 shipped. The others are older and blurrier: the class of 1999, back row, and one from January 27, 2025, posted by an official account in Zhanjiang, showing a group of men in a provincial town after a pickup football game. Nvidia was about to lose more than half a trillion dollars, and the people of Mililing were putting up a wedding arch. The account posted the picture with a single line underneath:
Can you guess which one is Liang Wenfeng?
Part IIThe Playbook
What follows are the operating principles that emerge from Liang Wenfeng's career, from a rented flat in Chengdu to a $50-billion lab that still says it has no organization. They do not come from a manual, since he has never published one. They are inferred from what he has done and, now, from four hours of what he said in private. Several of them contradict standard startup practice. That is partly the point, and partly the risk.
The principles
- 1.Treat restraint as a weapon, not a virtue.
- 2.Price to a payback period, not to what the market will bear.
- 3.Name the main line, and treat everything else as a by-product.
- 4.Buy the scarce input before it becomes a story.
- 5.Fund the first experiment yourself, and keep control when others arrive.
- 6.Govern by an unwritten vision that shows up in behavior.
- 7.Cap the formal at half.
- 8.Hire the people who don't yet know the "right" way.
- 9.Make team stability the one thing you won't trade.
- 10.Open-source what others can't run as cheaply as you.
- 11.Build your own tools before you need independence.
- 12.Choose the roadmap that is easiest to walk.
Principle 1
Treat restraint as a weapon, not a virtue
Liang's central claim is not ethical. It is a claim about how competition plays out: those who take more are beaten by those who take less, and even a stated intention to take more is a handicap. He walks through the logic explicitly. A company aiming for 5 percent of AI's share of world GDP loses to one asking for 1 percent, which loses to one content with 0.1 percent, until someone takes so little the business collapses. The stable position is just above that floor.
What makes this more than a slogan is that DeepSeek put it into practice under circumstances where most founders would not. It cut prices when demand was inelastic and declined to monetize a flood of users it hadn't expected. It released its strongest models under the MIT license. Each choice gave up short-term revenue for something harder to measure: the team's cohesion ("employees feel a sense of achievement"), goodwill across the industry, and resistance to the coalition that any would-be monopolist provokes. He describes restraint as something that raises "the probability of achieving AGI," which is a quant's way of putting it. He is not giving up money. He is buying a better probability of the outcome he wants.
The lesson generalizes beyond AI. In any market large enough that no single player can capture all of it, the visible appetite for capture is itself a cost. It attracts competitors, regulators, and talent-poaching. A company that visibly takes less tends to get underestimated, and often gets help.
Tactic: Before your next pricing or expansion decision, estimate the value of being seen as the player who takes the least, and count it as an asset on your side of the ledger.
Principle 2
Price to a payback period, not to what the market will bear
The ten-month rule is the most transferable mechanism in the transcript. Rather than pricing to demand, DeepSeek prices so that newly purchased hardware pays for itself in ten months, which Liang says works out to roughly six times profit. It is generous enough to fund the company and aggressive enough to leave third-party hosts of its own open weights with no margin.
The rule works because it is anchored to cost, not to competitors. If costs fall through better architecture, sparse attention, or cheaper chips, prices fall automatically, and the company's efficiency gains show up as cheaper service for customers instead of a cushion that hides waste. That has an organizational effect. A team that sees its optimizations turned directly into lower prices, and gets cheered for it in the group chat, keeps optimizing. A team whose optimizations simply become margin eventually loses interest in them.
Liang is open that this is not profit-maximizing. He told investors that doubling the price would nearly double revenue. He doesn't do it because it would raise the ceiling for every competitor, and because his model of the market says the high-margin player eventually loses.
Tactic: Replace "what will customers pay?" with "what payback period on our assets would satisfy us?", then publish prices that follow from that answer.
Principle 3
Name the main line, and treat everything else as a by-product
Liang's discipline about scope is extreme. DeepSeek does not do 3D, does not do video generation, and does not do world models. These are not bad businesses. Video is "a good business," he concedes, but they do not raise the ceiling of intelligence. Multimodality is "a component," like search. Hallucination is treated as "a product issue." The consumer app and the enterprise API are "byproducts on the road to AGI," worth picking up "casually" and nothing more.
The strategic insight is what he calls 降维打击, a dimensionality-reduction strike: a company working at the technical frontier can produce lower-tier products with very little effort, while a company organized around those products cannot reach the frontier. "In the past three years, any time you came to me to talk about commercialization paths and product lines, it was a waste of time," he said, because the ground kept shifting. Products built around last year's models had short lives. Capability improvements carried forward.
This works only when the main line is genuinely the source of value. Liang's bet is that, for now, "model technology is the most important thing." If that stops being true, the strategy turns into a liability. He admits as much, noting that a product-focused company holds advantages "in products, user service, and traffic."
Tactic: Write down the one capability whose improvement makes every other product easier, and require any project outside it to justify itself as a by-product rather than a goal.
Principle 4
Buy the scarce input before it becomes a story
In 2021 Liang's partners watched him buy GPUs by the thousand and decided he was indulging a hobby. By the time the U.S. export controls took effect, High-Flyer had about 10,000 A100s, enough to put it among fewer than five Chinese firms at that scale. His own explanation, "mainly driven by curiosity," understates how important the timing was. No strategic memo could have bought those chips after the controls arrived.
The pattern repeats in 2026, under different conditions. He wanted to turn the entire financing into compute within six months and was willing to pay a premium, because a GPU returns its cost in ten months while cash in a bank earns 2 percent. The general lesson is that the binding constraint on a future capability is usually visible years ahead to the people actually doing the work, and that once the constraint has a public narrative, it is already priced.
A quant habit is underneath this: treat physical capacity as a position, and accumulate it while it is cheap.
Tactic: Identify the input your field will be rationed on in three years and start acquiring it now, while doing so still looks eccentric.
Principle 5
Fund the first experiment yourself, and keep control when others arrive
Venture capitalists passed on DeepSeek in 2023 because an AGI lab offered no exit. Liang funded it from High-Flyer and owned about 84 percent through two Ningbo partnerships. The absence of outside money was what made the absence of KPIs possible. With no investors there was no board, and with no board there was no quarterly product roadmap.
When outside money finally came in 2026, he structured it to keep that freedom. Most of the capital went into a limited partnership he controlled. Investors got no voting rights and a five-year lock-up. Liang himself was the largest contributor at around 20 billion yuan. He also chose investors for alignment rather than price: those "least hostile to us." He accepted capital without accepting the oversight that normally comes with it.
The cost of this approach is concentration. When High-Flyer's products all fell more than 20 percent in July 2026, the dependence became obvious, and a funding source that had looked limitless turned out to have limits.
Tactic: Get your first proof of concept done without investors, then raise only on terms that leave the decision rights with the people who did the work.
¥
The Capital Stack, Then and Now
How Liang preserved research freedom as DeepSeek's capital needs outgrew its parent.
2016
High-Flyer's first deep-learning-generated trade executes on October 21.
2019
Fire-Flyer cluster built; High-Flyer passes 10 billion yuan in AUM.
2021
AUM exceeds 100 billion yuan; Liang begins buying GPUs by the thousand.
2023
VCs decline; High-Flyer funds DeepSeek, spun out July 17.
2024
Liang holds ~84% of DeepSeek through two shell partnerships.
2026
First outside round: ~$7.4B, investors without votes, five-year lock-up; Liang the largest contributor.
2026
July: High-Flyer drawdowns; leaked transcript; second round suspended; IPO preparations reported.
Principle 6
Govern by an unwritten vision that shows up in behavior
"Actually, we have no organization." Liang's governance model, borrowed in spirit from Jack Welch, is that a company is held together by vision, and that vision is "what you do, not what you say." DeepSeek's vision has never been written down. Employees may each understand it a little differently. What keeps it consistent is that the company's visible choices, such as cutting prices, open-sourcing, and turning down the super-app opportunity, all point the same way, so new hires learn the vision from what they see.
This is harder than it sounds, because an unwritten vision cannot survive a single high-profile contradiction. If DeepSeek had raised prices once revenue mattered, the vision would have been exposed as a slogan. Liang's decision-making reinforces it. He describes his authority as built on consensus: "I first look at what the consensus is… my guidance has a very limited effect."
The limits are already showing. Liang himself acknowledged that some departments need "a relatively strict hierarchy" as headcount grows. A vision-only organization works at 160 people. Whether it works at 1,600 is unproven.
Tactic: Pick the most expensive decision you are willing to make in service of your stated mission, make it visibly, and let that decision explain your values better than any document could.
Principle 7
Cap the formal at half
DeepSeek separates work into two streams. "Formal" work is top-down, coordinated, and deadline-driven, such as shipping V4. Everything else is bottom-up, self-directed, and unsupervised. Liang's one structural rule is that formal work should not exceed half of anyone's time. The other half is for exploration, with no prerequisite beyond whether the company's compute can support it.
The result most often cited is multi-head latent attention, the architecture that cut DeepSeek's serving costs and set off China's price war, which Liang attributes to a young researcher's personal interest. Early research, he argues, is cheap to fund: it "doesn't consume cards, only needs very few," and needs ideas more than headcount. He calls it "摸奖," a lottery draw: "the threshold is low, anyone can draw, but who draws what, I don't know whether it depends on talent or something else."
Pairing this with a no-overtime culture is deliberate. Research, he says, needs a relaxed environment, and DeepSeek can afford one because its scope is so narrow.
Tactic: Set an explicit ceiling on assigned work for your strongest people and protect the remaining time from meetings and deadlines.
Principle 8
Hire the people who don't yet know the 'right' way
"Experienced people will tell you without hesitation that you should do it one way." Liang's preference for new graduates and non-computer-science backgrounds, including poetry and advanced mathematics, rests on the view that experience narrows the search. For short-term goals he agrees experience is right. For long-term goals he prizes "basic skills, creativity, and passion."
He goes further and argues that talent is not the scarce resource at all. Talent shortages, he says, have always been temporary, whether for website builders, server engineers, or pilots: "There has never been a long-term shortage of any particular type of person in history." What is scarce is the ability to organize smart people so that they actually cooperate: "gathering smart people doesn't automatically mean they can cooperate." In his account, the US-China gap is not in people. Of the brightest Chinese researchers, roughly half stayed home.
The approach works because it fits Principles 6 and 7. Inexperienced hires with no existing playbook are exactly the people an unwritten vision and an exploration budget can shape.
Tactic: For your next long-horizon project, staff at least one seat with someone who has never done it before and give them protected time to find their own method.
Principle 9
Make team stability the one thing you won't trade
Restraint everywhere else makes the single exception stand out. Liang told investors that DeepSeek has "only one core interest… we must maintain team stability." Money, compute, and setbacks were all described as matters of time. Departures were the only thing he described as an existential risk.
This explains the financing structure and much else. The 2026 round's large option grants were, by his account, aimed primarily at retention. His reasoning is that if the most senior people stay, everyone else will, because "they didn't come entirely for the money." It also explains why the culture of no KPIs and no overtime is not merely generous. It is a retention mechanism in a market where ByteDance and Alibaba pay as well or better.
The principle carries an implicit warning. A company whose only core interest is its people becomes vulnerable to anything that makes those people visible and recruitable: fame, a leaked transcript, a $70-billion valuation that suddenly gives every employee a number.
Tactic: Identify the few people whose departure would set you back a year, and design compensation, workload, and culture around keeping them before you design anything else.
Principle 10
Open-source what others can't run as cheaply as you
Liang's open-source argument has two layers. The first is ideological: open source is "like a cultural practice," and closed-source moats are temporary. The second is commercial, and it is subtler. Even with the weights public, deploying a model at DeepSeek's cost is "very, very hard," requiring organizational and engineering effort that large companies find difficult to coordinate and small companies can't afford. DeepSeek sits, he says, in a "sweet spot" between them.
So open-sourcing gives away the recipe but not the cost structure. At ten-month-payback prices, third-party hosts make no money, so the open model expands the market without creating a competitor. He goes as far as saying he worries third parties will deploy his models badly, not that they will deploy them at all. He also draws a contrast with a domestic rival, Zhipu, whose open-sourcing "feels forced."
The principle is easy to misapply. Open-sourcing a product whose value lies in the artifact itself does give it away. It works when the value lies in operating the artifact efficiently.
Tactic: Before open-sourcing anything, confirm that your operating cost advantage, rather than the code or weights themselves, is what customers actually pay for.
⚖
Conventional Wisdom vs. the Liang Playbook
Where DeepSeek's choices diverge from frontier-lab norms, and where they carry risk.
| Conventional approach | Liang's approach | Exposure |
|---|---|---|
| Price to demand; capture inelastic surplus. | Price to a ten-month hardware payback. | Margin left on table |
| Convert a viral user surge into a super app. | Keep users cheaply; don't chase them. | Product lag |
| Closed weights as the moat. | Open weights; moat in serving cost and organization. | Distillation, state bans |
| KPIs, hierarchy, 996. | No KPIs, formal work capped at half, no overtime. | Scaling strain |
| Hire proven senior researchers. | Hire curious graduates and non-specialists. | Low |
| Raise from VCs early; cede board seats. | Self-fund; later raise without voting rights. | Parent-fund volatility |
Principle 11
Build your own tools before you need independence
Years before export controls made independence urgent, High-Flyer was already building its own infrastructure: the Fire-Flyer clusters, the 3FS file system, and hfreduce, written to replace Nvidia's own communication library. The motive then was efficiency, not geopolitics. The result was a firm that understood its hardware well enough to optimize around it.
That habit is now DeepSeek's route away from dependence on Nvidia. Liang told investors that V3 was trained on Nvidia cards but almost without Nvidia's software ecosystem, using TileLang, a DeepSeek-built high-level language for writing compute kernels. He accepts a 1 to 2 percent performance loss for the portability, and is porting the stack to Huawei. A company that learned to write its own plumbing in order to trade stocks more efficiently found it had also prepared for a world where the plumbing could be cut off.
The broader lesson is about timing. The cheapest moment to build independence from a supplier is before you need it, when the work can be justified as performance tuning.
Tactic: Rewrite one critical vendor dependency in-house this year, justified by performance, so that leaving that vendor becomes possible rather than theoretical.
Principle 12
Choose the roadmap that is easiest to walk
Liang's AGI sequence is chain-of-thought, then agents, then continual learning, then a gradual self-improving "singularity," and only then embodied robotics. He defends it partly on technical grounds and partly on what it asks of his team. "This roadmap means we don't have to work overtime." Each step builds on the one before, and later steps can be delegated to the systems produced by earlier ones. Doing robotics first would be "data-intensive, labor-intensive," and grueling.
This is an unusual criterion for strategy: pick the path where each step uses the tools the previous step produced. It fits his stated goal that DeepSeek's models should first be useful to DeepSeek, speeding up the building of the next model. Building your own leverage first is not laziness. It is a way of arranging the work so that it compounds.
The risk is that the easiest sequence is not always the one the world rewards. A rival who grinds through the hard path first may get there sooner. Liang's answer is that the relaxed team, better at research and less likely to leave, gets there more reliably.
Tactic: Order your roadmap so that each milestone produces a tool that makes the next one cheaper, even if that means delaying the most visible goal.
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
— Liang Wenfeng, investor meeting, May 20, 2026 (translated)Actually, we have no organization; it's vision-driven. We organize around a vision.
— Liang Wenfeng, investor meeting, May 20, 2026 (translated)As long as we can maintain the stability of the team, I will definitely achieve AGI. It's that simple.
— Liang Wenfeng, interview with Waves (36Kr), 2023If you are pursuing short-term goals, it is right to find people with ready experience. But if you look at the long-term, experience is not that important. Basic skills, creativity, and passion are much more important.
— Liang Wenfeng, interview with Waves (36Kr), July 2024In the past thirty years, (China's tech industry) has only emphasized making money, and ignored innovation. Innovation is not solely driven by business, it also needs curiosity and a desire to create.
— Liang Wenfeng, to Chinese media, 2024Our principle is not to lose money, nor to make huge profits … our starting point is not to take advantage of the opportunity to make a fortune, but to be at the forefront of technology and promote the development of the entire ecosystem.
Maxims
- Take less, win more. In a market too big to monopolize, the player who wants the smallest share sets the terms for everyone else.
- Price from cost, not from fear or greed. A fixed payback rule turns every efficiency gain into a lower price and every lower price into goodwill.
- Pick up the sesame, but walk toward the watermelon. Collect the easy revenue in passing, and never let it set the direction.
- Buy the bottleneck while it still looks like a hobby. By the time the constraint has a narrative, it has a price.
- Write nothing down; do it every day. A vision people can see in your decisions holds them together better than one framed on a wall.
- Leave half the week unassigned. The architecture that changes your industry may come from someone nobody told to build it.
- Inexperience is a search strategy. People who don't know the standard answer keep looking until they find a better one.
- Your team is the only thing you can't afford to lose. Money and compute only cost you time. Departures can end the project.
- Open the recipe and keep the kitchen. Publish what others can copy, and compete on what they can't run as cheaply.
- Build the easy staircase. Order the work so each step produces the tool that makes the next one lighter.
