Liang Wenfeng has no life, and Yang Zhilin has no way out.

Original author: Jia Liu

This week after the release of Kimi K3, the U.S. tech world has collectively expressed deep regret—why didn’t a young, excellent person like Yang Zhilin stay in the United States back then?

Kimi K3 surpasses all models in front-end capabilities

This topic has racked up more than 200k views on X. Before that, David Sacks, the former White House AI point person under Trump—Trump’s direct confidant—said he has already switched from Claude to Kimi for handling large volumes of work: “It’s so much more fun, because it just gets the job done instead of lecturing you.”

Silicon Valley veteran VC Vinod Khosla, however, balanced the books onto immigration policy, saying the U.S. is scaring off outstanding talent with its own hands.

As speculation keeps snowballing, Yang Zhilin’s PhD supervisor, Salakhutdinov, while publicly “showing off” his protégé on social media, also clarified why Yang Zhilin didn’t stay in the U.S.: “If you don’t even have the courage to try starting a business, then Yang Zhilin said he’ll regret it for the rest of his life.”

On the other side of the ocean, there’s another Guangdong name that is also widely known in China’s AI circle: Liang Wenfeng.

Liang Wenfeng is seven years older than Yang Zhilin. Born in Wuchuan, Zhanjiang, to parents who were primary school teachers in their town. He was a programmer with 15 years of quantitative trading experience, barely ever gave interviews and had no social media account. His only one-line personality description from coworkers was: “No hobbies other than programming.”

After DeepSeek R1 was released in January 2025, Nvidia’s market value evaporated by nearly $600 billion in a single day—what Silicon Valley calls a Sputnik moment. While the whole world was looking for him, he hid back in his hometown and kicked a ball around for a few days.

Two people from Guangdong—one born in Wuchuan, the other in Shantou—separated by a Leizhou Peninsula. Their companies are on the same track, carving out trajectories that are almost perfectly mirrored, and every branching point actually comes from their core character.

A radio and a band

Liang Wenfeng was born in 1985 in Mililing Village, Qinba Town. Both his parents were primary school teachers in the town. There wasn’t much in the way of toys at home. The most important object in his childhood was a Feiyue brand radio. He took it apart and reassembled it—over and over—until he lost count.

This quiet kid showed something different early on. His middle-school class teacher remembers he wasn’t a bookworm, and he wasn’t necessarily more diligent than others. But by middle school, he had already self-studied through high school math, then started flipping through college textbooks: “It’s as if he doesn’t need to spend much time to learn each subject well.”

In the 2002 Gaokao, he scored 806 points, taking the top spot in Zhanjiang City. The prize photo is still searchable today: a burgundy short-sleeve shirt, a large red flower pinned to his chest, a stiff expression—clearly pushed onto the stage by teachers. That autumn he entered Zhejiang University for electronic information engineering. But over the next two decades, there was never again a moment worth photographing.

Seven years later in Shantou, another Guangdong kid grew up. Yang Zhilin was born in 1992, ranked first in his class for four years in Tsinghua’s computer science department, and published more than twenty papers. These wouldn’t be the strangest part at Tsinghua; the rare thing was that besides these grades, he formed a rock band called Splay, named after a data structure. He even became the drummer himself.

Later he explained: “Back then it felt like there were many things I wanted to express, including the pressure from reality and the absurdity of the broader environment.” They wrote a song about a daytime dream of getting rich overnight after创业 success—“half out of empathy, half to remind myself not to become too utilitarian.”

Many years later, this band would return to the story in a way no one could have predicted: their bandmate Zhou Xinyu became a co-founder of Moonshot’s the Dark Side (Moonshot AI).

At the entrance of the offices of Moonshot’s the Dark Side, there was a white Yamaha electric piano, with Pink Floyd’s 1973 album The Dark Side of the Moon pressed under it. The company name came from that album.

The Dark Side of the Moon album

Disassembling a radio doesn’t require anyone to see it, while drumming does—because something has to be expressed. Every choice these two people made over the next twenty years basically grew out of this.

A rental room in Chengdu and the CMU corridor

After graduating from Zhejiang University, Liang Wenfeng didn’t go to a big tech company to collect a technical business card. He ran to Chengdu, hid in a cheap rental room and tried all kinds of algorithms, aiming to add AI to traditional industries—and he wiped out every attempt. Most founding teams of China’s top quant funds had gilded resumes from overseas hedge funds, but Liang Wenfeng figured it out himself in a rental room.

In 2015, he and his Zhejiang University classmate founded Fanhang (幻方). Next came a chain of moves that almost nobody understood at the time: in 2019, they poured nearly $200 million into building their own cluster—1,100 GPUs; in 2021, they added another $1 billion and stockpiled about 10,000 A100s.

You don’t need that many cards for quant work, and Liang Wenfeng himself admitted it—he only trades, and a small number of cards is enough. People who had interacted with him early on recalled that when they saw him hoarding GPUs to train models, they only felt it was a techie with a terrible hairstyle burning money.

But what he was doing was exactly the opposite of what everyone understood. He wasn’t using AI to make finance cheaper and more efficient—he was using finance to supply funding to AI research. Most China AI companies follow a sequence: first raise financing, then find a product, then find cash flow. But he flipped the order entirely: first build a cash-flow machine, and then use it to “buy out” the freedom to do research.

Yang Zhilin took a different route—one paved with flowers and mines.

After graduating from Tsinghua, he went to CMU (Carnegie Mellon University) for a PhD, during which he did research in AI labs at Google and Meta. In 2017, he focused all his effort on language models, and later said that it was “the only important problem.” During his PhD, he published two papers: one taught AI to remember longer context, and the other beat Google’s strongest model in 20 tests, with the combined citation count nearing 20,000.

Years later, when Kimi became popular for handling long texts via input, many people thought that was a differentiated selling point temporarily found in 2023. In fact, it was the direction he had already locked in during his PhD—just changed its form.

In 2016, while still in his PhD, he also helped start a loop intelligence company (循环智能), doing sales call analytics. Sequoia and 金沙江 were both investors. This startup let him see early on how rough the landing of technology can be—but it also buried a mine under his feet, which didn’t get triggered until eight years later.

In 2019, after finishing his PhD, his advisor arranged an Apple executive who could report to Cook to ask whether Yang Zhilin was willing to join Apple—even with the possibility of going to Apple’s China unit. But Yang Zhilin rejected Apple’s emails and the Silicon Valley offers and decided to return to China.

At that time, Liang Wenfeng was hoarding GPUs in Hangzhou, while Yang Zhilin was waiting for the wind in Beijing. Neither knew about the other’s existence, nor did they know that these two names would come to represent today’s China AI circle.

A one-month window and a catfish

On November 30, 2022, ChatGPT launched, and the Silicon Valley tech community collectively lost sleep. Yang Zhilin recalled that many friends around him were anxious, suffering FOMO, unable to sleep—and many people turned around to start businesses.

“Starting February 2023, we focused on the first round of fundraising. We delayed until April, and basically there wasn’t a chance anymore. But even if we had done it in December 2022 or January, there wouldn’t have been a chance either—at that time there was a pandemic, and people didn’t react. ” Yang Zhilin seized that one month. He didn’t take a day off.

In March 2023, the Dark Side of the Moon was founded. The co-founders were Tsinghua alumni Zhou Xinyu and Wu Yixin. It’s said that the initial fundraising was $60 million, and within three months they gathered about 40 AI researchers. Then came the steepest curve in China’s large-model fundraising history: Sequoia and Zhenge entered, Alibaba led with $1 billion, and Tencent, Meituan, Xiaohongshu followed—pushing it all the way to $3.3 billion. Kimi’s ability to handle 200k-word long texts turned into many Chinese people’s first high-frequency large-model product.

During that period, Yang Zhilin lived in a kind of dual state. Externally, he told the grand narrative: he estimated the probability that scaling laws wouldn’t work would be nearly zero; he compared entrepreneurship to driving toward a long stretch of snow mountains; he described the first year as building a prototype rocket and getting a hint of the fuel recipe. Internally, he had to obsess over the most trivial algorithms: when compute was tight, one machine was 260 today, 340 tomorrow, then a few days later it dropped again—buy or rent, which channel to use. He tracked and adjusted every day. Half the mindset of a scientist’s certainty, half the sharpness of a small boss—both were reflected in this 31-year-old.

But everything capital gave came with a price already tagged. To sustain the growth curve, in October 2024 Kimi deployed $220 million in a single month, and in November another $200 million—burning over two months more than the entire third quarter combined. That drummer who wrote a song satirizing overnight riches became the most aggressive user-acquisition founder across the industry. He didn’t change; it was the $3.3 billion valuation doing the decision-making for him.

Liang Wenfeng entered in Hangzhou in a way that was almost intentionally “opposite.”

In 2023, DeepSeek spun out from Fanhang without taking any external investment. The team was under 140 people—almost no returnees, mostly fresh graduates and young people who had graduated only a few years earlier from domestic universities. There was no KPI, no hierarchy. If you had ideas, you could directly allocate GPUs and people.

Former employees recalled to The Washington Post that Liang Wenfeng would drill into the details of training strategies, watch papers with the researchers, and write code with them. “He didn’t look like a boss at all—more like an extreme geek.” He explained why he hired fresh grads instead of snatching top talent in the industry: “Experienced people will tell you what to do without thinking. But people without experience will keep exploring.”

In May 2024, DeepSeek-V2 cut the API price to 1 yuan per million tokens. ByteDance, Alibaba, Baidu, and Tencent were forced to follow. The whole industry assumed it was a long-planned battle of war. His response was: “We’re not trying on purpose to become a catfish; we just accidentally became one.”

Pricing was just a slight profit above cost—“not losing money, and not making windfall profits.” Internet people in price wars talk about market share, entry points, and network effects; he talked about cost accounting. And precisely because this price cut came without emotions, it was the most deadly kind: it dragged the large-model API from a high-margin narrative straight into the pricing logic of infrastructure.

Somewhere along the line, people started comparing Yang Zhilin and Liang Wenfeng—two totally different business models.

Yang Zhilin, suspended

The mine that Yang Zhilin buried eight years ago exploded in November 2024.

His previous entrepreneurial project, Loop Intelligence (循环智能), with its five old shareholders, filed arbitration in Hong Kong, China, accusing him of starting a new company’s fundraising before he obtained all shareholders’ waivers.

On December 5, Zhu Xiaohu’s朋友圈 fired off. The firepower was focused on Zhang Yuting: the former partner of Sequoia China had received an initial 14% in 9 million shares free at the Dark Side of the Moon, exceeding the 9.5% that Loop Intelligence as the “parent” company distributed. Zhu Xiaohu’s “solution” was nearly insulting: apologize, step down from the equity, or sever ties between the company and Zhang Yuting.

On the evening of December 6 at 9:40, Yang Zhilin posted a 1,300-character long statement. He didn’t sever ties; he made it definitive instead: Zhang Yuting was a co-founder, and her shares were the consideration for the work done over many years to come. The procedures for leaving Loop Intelligence had been obtained with each board member’s signature. People close to the company relayed the internal stance: she and the Dark Side of the Moon were already “one entity,” and they couldn’t cut her off.

Zhu Xiaohu publicly said he couldn’t understand it at all. In a purely commercial coordinate system, there truly was no solution—cutting ties would be the only rational option. But in the coordinate system where Yang Zhilin made his decision, there were other considerations. Salakhutdinov’s later clarification provided a footnote: this was exactly the kind of person who says, “If I don’t try, I’ll regret it for a lifetime.” Once he’s acknowledged what he decided, he doesn’t look back—no matter that the cost is already laid out in public.

The real heavy hammer landed more than 40 days later.

On January 20, 2025, R1 was released. Free, open-source, and inference capability nearly matched OpenAI’s o1. A Hangzhou team of over 100 people flipped the global capital markets inside out within a week. A researcher at Carnegie, Matt Sheehan, said something quite interesting: DeepSeek wasn’t the company China had pre-selected in advance; its breakout came as a surprise even to China.

Meanwhile, Liang Wenfeng was celebrating the New Year in his hometown in Wuchuan. On the afternoon of January 27, he played a game of soccer with his middle-school classmates in the village. The entrance of the village was packed with tourists taking photos and checking in—the main characters were on the field.

For Yang Zhilin, it was a double kill. The arbitration hadn’t even finished when R1 also directly sentenced the “death penalty” for the route he had taken over the past year: the users bought through user acquisition had little value in front of a free and stronger opponent. Public opinion turned its fire elsewhere. An article’s title said: “Yang Zhilin, the suspended idealist of a 90s-born generation.” “Suspended” means his feet never touch the ground. He worried when writing songs about becoming utilitarian; now the world says he’s both utilitarian and failed.

At the beginning of 2025, in the Dark Side of the Moon, there was still money on the books, but very little say in the discourse.

A comeback against the pull of gravity

Over the next year—this would be a critical year for Yang Zhilin.

He almost fully denied the version of himself from the past year. Stop ad spend, cut redundant businesses, shrink to the base model, then shift to open-source. In a conversation in Geek Park, he said that organizational inertia is to do more and more things: “We need to fight against this pull of gravity.”

It sounds light when spoken, but to do it means admitting the route was wrong; firing the people he hired who helped create the wrong direction; bowing to the opponent who almost killed him. Most founders at age 33 can’t get past that. Yang Zhilin crossed it extraordinarily decisively—perhaps because open-source and long-termism were already part of his default settings from the factory. Buying user acquisition through a closed model was the costume that capital put on him; now he just took it off.

In July 2025, K2 with a trillion-parameter scale was open-sourced. In November, K2 Thinking put GPT-5 under its feet on several of the toughest Agent benchmarks. Hugging Face co-founder Thomas Wolf asked on X/Twitter: “Is this another DeepSeek moment?”

In the early hours after the release, Yang Zhilin brought Zhou Xinyu and Wu Yixin onto Reddit to host an AMA and answered 21 questions. He clarified that the $4.6 million training cost wasn’t an official number, and admitted that the number of GPUs wasn’t as many as U.S. peers—“but we squeezed the performance out of every single card to the extreme.” Someone asked what he thought about OpenAI burning money; Zhou Xinyu answered casually: “We don’t know either—only Sam knows. We have our own pace.”

Their own pace. The Dark Side of the Moon in 2024 couldn’t say those five words; at that time its pace was the pace of investors, the pace of ROI from ad spend. Returning those five words to Yang Zhilin—precisely was what Liang Wenfeng did. R1 proved that open-source plus algorithm efficiency can work in China, which is essentially a roadshow for Yang Zhilin to his own board. The people who had nearly been killed by DeepSeek ended up wanting to thank it for that.

The market’s response was equally direct. On the last day of 2025, the Dark Side of the Moon officially announced a $500 million Series C. Alibaba, Tencent, and Wang Huiwen all added more, pushing the valuation to $4.3 billion. Cash on the books exceeded $10 billion. Less than 20 days after the release of K2.5, revenue surpassed the entire 2025 calendar year. Individual subscription orders increased by more than 80x month-over-month, and surged into the top 10 globally on the Stripe ranking. Then came K3 this week.

Seven years ago, the PhD who said, “If I don’t try, I’ll regret it for a lifetime,” is now making the U.S. tech world start to doubt reality—and became the example used to smack the White House.

The hermit’s bill

But reality never just slaps one side of a face. After 2025, it was Liang Wenfeng’s turn to pay the bill.

He had no life, but his employees did. Ruo Fuli, Wang Bingxuan, Wei Haoran, and Ruan Chong—these names were well-known core backbones inside DeepSeek. Starting in 2025, many of them left, and quite a few moved elsewhere to directly become heads of business.

A painful phrase goes around: “If people with about the same rank can jump out and get that much, then why can’t I?” The research utopia with no hierarchy, no KPI, and no talk of money depended on members’ loyalty to the problem itself. But R1 raised everyone’s market value by 10x—loyalty got its first priced rival.

Money also started to become a problem. Fanhang’s management scale shrank from its peak to just a bit over 20 billion yuan (i.e., 200+ billion in RMB terms). The battery powering the “do it for passion” project was leaking. That’s why an unimaginable scene emerged: the person who once said “there are no plans to raise financing in the short term” started meeting investors.

The opening price was 5 billion yuan per deal (minimum); later it dropped to 1.5 billion yuan. But throughout the negotiations, what he repeatedly emphasized wasn’t valuation or equity ratio—it was the same condition: no poaching DeepSeek’s people, and no encouraging them to start businesses elsewhere. A fundraising deal turned into a non-poaching agreement. He could give up equity—but he wouldn’t give up the atmosphere of that lab.

By 2026, with the two lines running, an outcome nobody expected appeared.

Yang Zhilin, cutting ad spend, squeezing every bit of performance out of each H800, talking about his own pace—was increasingly starting to look like Liang Wenfeng. Meanwhile, Liang Wenfeng, meeting VCs, worrying about retention, and for the first time having to distract himself from the lab to handle the mundane problem of organizational management, had no choice but to step out of the laboratory.

One used to be a reservoir, the other a tide chaser. Now the reservoir keeper realizes the reservoir can leak too; the tide chaser finally waited for his own tide.

Two people from Guangdong rewrite China’s AI playbook

This is the hottest time in China’s AI entrepreneurship, when two names were pushed to the forefront: Liang Wenfeng and Yang Zhilin.

They both don’t resemble the founders that Chinese internet culture has grown familiar with over the past decade. No slogan poster quotes, no legends of dinner parties, and no intense desire to craft themselves into legendary entrepreneurs. But the positions they stand in are closer to the intersection where money, power, and the emotions of the era meet than most business leaders in the past.

Liang Wenfeng’s special part is that he appears to have almost “no life.” In public reporting, he’s more like someone swallowed by work: from quant investing, to building self-owned compute clusters, to DeepSeek. His narrative rarely includes family, consumption, hobbies, or social life—almost everything is occupied by models, architectures, compute, organization, and originality.

Yang Zhilin’s special part is that he appears to have “no retreat” left. After the Dark Side of the Moon was established, Kimi quickly became one of China’s most watched AI applications. Financing, valuation, user growth, model iterations, commercialization, arbitration involving old shareholders—everything pressed down on this company at the same time. The faster it runs, the less it’s able to stop and explain for too long.

This is where these two people are most similar, and most unlike.

Liang Wenfeng looks like a technical idealist who emerged from the depths of the capital markets. First, inside Fanhang’s quant operations, he proved that AI can directly change where money flows. Then he turned that capability toward large models. DeepSeek’s story didn’t start by being driven by financing; it started with the funds, compute, and engineering culture accumulated by Fanhang. What the outside world later remembered was R1, V3, low-cost training, and the open-source shock that shook the U.S. market. But more critical than all that was: Liang Wenfeng early on defined the gap as one between “originality and imitation,” not as “how to make money with China’s large model applications.”

That made him look anti-commercial—and yet extremely commercial.

Anti-commercial because he repeatedly downplayed short-term monetization in public interviews and didn’t want to frame DeepSeek as an internet story of grabbing users and market share. Extremely commercial because every step he took targeted the most fundamental cost structure in the AI industry: training efficiency, inference costs, model architecture, talent density, and chip constraints. If a model can reach top-tier capability at lower cost, it doesn’t change a product ranking list—it changes the industry’s imagination of capital expenditures.

That’s also why when DeepSeek went viral, the global market reaction was so intense. It wasn’t simply “another Chinese chatbot.” It was a reminder to investors that the most expensive logic chain in the U.S. AI narrative over the past two years might not be as solid as people assumed. Bigger models, more GPUs, higher capital expenditure—those don’t necessarily translate into an unbeatable moat.

Yang Zhilin faced a different fate.

From birth, the Dark Side of the Moon stood in the spotlight. The founders’ resumes were impressive enough: Tsinghua undergraduate, Carnegie Mellon PhD, involvement in key research such as Transformer-XL and XLNet. The company was founded in 2023, and Kimi used long context to build differentiation, quickly becoming an AI product that ordinary users could sense. Unlike DeepSeek, which was discovered by the tech community first as a research organization, the Dark Side of the Moon entered the mass product world, capital markets, and industrial narratives earlier.

This gave Yang Zhilin huge advantages—and also massive burdens.

The advantage: Kimi has product mindshare. Many people’s first serious interaction with domestic AI wasn’t because they understood a technical report, but because Kimi can read long articles, organize materials, and handle workflows. When AI moves from the lab to the office, it needs an entry point that ordinary people can remember. Kimi once grabbed that position.

The burden: once product mindshare is established, it must be continuously fed. Users wait for stronger models; investors wait for higher revenue; the team waits for larger option value; competitors wait for you to make mistakes. The more the Dark Side of the Moon raises, the higher the valuation—and the less likely Yang Zhilin can return to the calm position of a researcher.

By 2026, this pressure became clearer. Public reports showed that by May 2026, the Dark Side of the Moon had completed about $2 billion in new financing, with post-investment valuation surpassing $20 billion. The company’s official website also put Kimi K2 and code and Agent capabilities in the core position. What the capital markets gave it wasn’t applause but a bill: you have to prove you’re not only “the best large-model company at making products,” but also that you can maintain both technological and commercial leadership amid assaults from DeepSeek, Alibaba, ByteDance, MiniMax, Zhipu, and others.

What’s even more troublesome is the shadow of Yang Zhilin’s old entrepreneurial project that followed him. In 2024, media outlets such as the South China Morning Post reported that Moonshot’s the Dark Side founder Yang Zhilin and co-founder Zhang Yutong were sued for arbitration in Hong Kong, China by some former investors from Loop Intelligence (循环智能). Yang Zhilin’s side claimed that leaving Loop Intelligence and starting a business had already completed all necessary procedures, while the investors presented a different account. The core of the dispute wasn’t just personal problems between founders—it was a sharper issue in China’s AI startup boom: when a new company’s value skyrockets, where exactly are the boundaries among the old company, old shareholders, the old team, old intellectual property, and new financing?

Yang Zhilin can’t only tell the story that he’s a genius researcher. The scale of financing, commercialization revenue, product iteration, IPO expectations, and legal disputes will all demand that he becomes a more complete, colder CEO. A researcher can prove themselves through papers; a CEO must prove themselves through the organization. Papers can carry a name; an organization can’t just run on names.

The more low-key someone is, the more the era magnifies them. The more someone wants to move forward, the more the past catches up.

This round of China’s AI story may not ultimately be won by the best storyteller. It’s more likely to be won by those who can simultaneously endure three things: technological uncertainty, capital’s patience consumption, and the cost of founders being mythologized and also judged.

Whether Liang Wenfeng had a life, the outside world doesn’t really know. Whether Yang Zhilin has an escape route—there still isn’t a final answer.

But at least right now, they both don’t have much room to return to being ordinary people.