At the end of last year, former OpenAI researcher Yao Shunyu made a high-profile move to Tencent, prompting the company to issue an urgent clarification: Yao's compensation package was not in the nine-figure range. In the AI wealth-creation wave, it's not just startup founders seeing their net worths soar—top-tier employees are also commanding unprecedented packages, and a new class of New Money is being minted in bulk.

As one of them, Yao Shunyu, born in 1998, has made four pivotal decisions in his career. The first was choosing his major: in 2015, he entered Tsinghua's prestigious Yao Class with a college entrance exam score of 704, studying computer science under the tutelage of senior Tsinghua alumnus and deep learning expert Wu Jiajun, where he began systematic exposure to general artificial intelligence. In 2017, the landmark Transformer architecture emerged, making natural language processing a hot field.

The second decision was selecting his advisor: in 2019, Yao went to Princeton for his PhD, initially focused on computer vision. But Yao believed NLP held more potential, so he pivoted to natural language processing and reinforcement learning, connecting with advisor Karthik Narasimhan. Narasimhan had just arrived at Princeton, having previously spent a year as a visiting researcher at OpenAI, where he was a second author on the seminal GPT-1 paper. In 2019, OpenAI secured a $1 billion investment from Microsoft, gaining prominence in Silicon Valley. Yao and Narasimhan developed a close mentor-mentee bond—when Yao got married, his advisor served as best man.

The third decision was his choice of company: during his Princeton years, Yao authored two first-author papers, ReAct and Tree of Thoughts. The former proposed enabling AI to call tools while reasoning, alternating between the two—now a fundamental paradigm for AI agents. After graduation, Yao joined OpenAI, where his mentor had once worked, leading the company's agent products and contributing to the Deep Research project. Many core R&D roles at major companies follow a similar trajectory: pick the right major, follow the right direction, join the right company. When new technologies first emerge, talent is always in short supply. Each correct choice effectively raises one's market value.

But unlike most people, Yao's value skyrocketed because of his fourth decision: at the height of the AI race, he encountered a company that lacked everything except money. Tencent had fallen behind in the large model competition, and as a chaser, it was more generous with salaries. From another angle, for a company like Tencent, handing significant authority to a 27-year-old is a move born of necessity—they had no other choice. Choose the right research direction, graduate as the industry explodes, join a frontier company, then jump to a deep-pocketed chaser. Yao Shunyu's life so far can be summarized in one sentence: four times leveraged long on artificial intelligence.

It's not unusual for tech industry incomes to be high, but capitalists being this lavish is probably a first in human history. Sam Altman complained on a podcast last year that Meta was willing to pay its researchers $100 million signing bonuses to poach them. Meta has stumbled in large model development but has always been generous in talent acquisition. Last July, Meta poached Apple's foundational models team lead Pang Ruoming with a $200 million annual package. To put that in perspective: that's more than Tim Cook's entire compensation, with over $100 million left over. That same month, Meta spent $14.3 billion to acquire 49% of Scale AI, whose founder Alexandr Wang joined Meta to lead its AI team. With Zuckerberg's blessing, Wang became Forbes' youngest billionaire of 2025. Netizens who use AI to write Xiaohongshu posts shouldn't call themselves "super individuals"—this is what a real super individual looks like.

According to Equilar's compensation database, OpenAI's L5 senior engineer total annual package ranges from $829,000 to $871,000, while Anthropic's company-wide median compensation sits between $420,000 and $540,000. Top AI researchers can earn over $10 million annually—a tenfold difference. That's the direct distinction between genius and talent. Meta CTO Andrew Bosworth couldn't help complaining to CNBC: "The market is setting a price on a certain tier of talent that is truly incredible and unprecedented in my 20-year career as a tech executive."

Handing out cold hard cash directly to AI researchers obviously won't work. Precious cash needs to go to Jensen Huang for GPUs; motivating talent primarily relies on equity. Bosworth debunked the "$100 million signing bonus" rumor internally at Meta: the bulk was stock, only unlockable upon meeting KPIs. Issuing stock incentives to employees is common, but this AI wave has a distinguishing feature: startup valuations are inflating at astonishing speed. AI coding platform Cursor was founded in 2022, valued at $2.6 billion by end of 2024, and acquired by Elon Musk's SpaceX for $60 billion in 2026—a more than twentyfold increase in under two years. Buffett would want to close his account, and Soros would want to call the police. When ChatGPT burst onto the scene in late 2022, OpenAI was valued at around $29 billion. In March of this year, OpenAI completed a $122 billion funding round at an $852 billion valuation—a thirtyfold increase in just over three years. During the same period, Meta's market cap quadrupled, Google's rose 3.6 times, Microsoft's doubled, the Nasdaq gained over 2x, and as for the Hang Seng Tech Index... well.

In court testimony from Musk's lawsuit against OpenAI, OpenAI president Greg Brockman's equity was valued at approximately $30 billion, while former chief scientist Ilya Sutskever held $7 billion in stock. As valuations surge, employee stock becomes increasingly valuable, driving rapid wealth accumulation. So when big tech wants to poach top researchers from OpenAI and Anthropic, they must not only match salary levels but also factor in expectations of option appreciation. Domestic Chinese tech giants looking to recruit cross-border must also account for exchange rates. OpenAI knows it can't compete with big companies on financial resources, so it's exceptionally generous with employee stock. According to Wall Street Journal analysis, OpenAI's equity compensation expense as a percentage of revenue was around 45%—far higher than Silicon Valley giants. In 2024, that figure hit 119%. In October 2025, OpenAI bought back employee stock, with 600 employees pocketing $6.6 billion total, and 75 people selling the maximum allowed $30 million each. By US Department of Labor metrics, OpenAI and Anthropic have nearly identical median salaries—the competitive edge lies entirely in options. Long-termists never look at their pay stubs, but without an active, even feverish capital market, AI luminaries' net worths couldn't have risen this fast. You could argue this is all paper wealth, same as your stock portfolio—but at least their account balances are heading in the right direction.

Another characteristic of this AI technology wave is that the New Money class is remarkably young, achieving financial freedom multiple times over at tender ages. In 2024, AMiner analyzed the core research teams of the world's ten largest large models at the time, finding that 69% were under 40—and for Chinese teams alone, that figure was 84%. According to Lei Jun, the average age of Xiaomi's MiMo large model core team is 25. Its leader, Luo Fuli, born in 1995, is three years older than Tencent's Yao Shunyu. Alexandr Wang, whom Zuckerberg personally catapulted onto the wealth list, was born in 1997 and dropped out of college at 19 to start his company. In traditional industries, "ten years of experience" is core capital. But AI industrialization has an extremely short history—the core large model technology stack only took shape around 2018, GPT-3 proved the direction in 2020, and ChatGPT made its name in late 2022. Many AI labs have existed for less than a decade. In other words, an AI researcher who started their PhD in 2018 is already considered a veteran and expert in the field.

Counterexamples do exist, though. Wu Yonghui, head of ByteDance's large model research division, is literally an old-timer. When Wu entered Nanjing University's computer science program in 1997, Yao Shunyu hadn't even been born. After earning his PhD in 2008, Wu joined Google and stayed for 17 years. In early 2025, he joined ByteDance, taking full responsibility for foundational large model research, reporting directly to CEO Liang Rubo. So age is only superficial—the inflation of technical geniuses' value is tied to another broader context: the incredibly rapid pace of AI technology iteration. There's a broad consensus in the industry that even compared to just a few years ago, current large models represent a generational technological leap. The training methods of this year belong to a different world than two years ago. TensorFlow was once the absolute king of deep learning frameworks; today, PyTorch holds a near-monopoly in both research and production. Model scale, architecture, and infrastructure iterate constantly—five-year-old experience in AI is practically an archaeological finding. When the half-life of experience is so short, seniority offers no moat. This creates a practical problem: only when a team chooses the right technical direction do an individual's accumulated knowledge and experience hold value.

Wu Yonghui initially worked on search algorithms at Google. In 2014, he joined the Google Brain team, pivoting toward deep learning. In 2023, he was promoted to "Google Fellow" and research VP at Google DeepMind. Though not young in age, Dr. Wu has been fighting on the front lines of Google's AI research. Despite Google's current struggles, when Wu moved to ByteDance in 2025, Google's Gemini-3 series models were crushing the competition across the board—plenty of Silicon Valley pundits were busy defending the company. The knowledge and experience Wu accumulated at Google Brain and DeepMind, all those granular details and know-how, were like rain in a drought for ByteDance. If Dr. Wu had been in Google's Android or YouTube divisions, he almost certainly wouldn't have gotten a ByteDance badge. Since Wu joined over a year ago, Seed has iterated through four versions, Doubao's daily token usage has surged to 180 trillion, and the Seedance video model has improved by leaps and bounds. Yao Shunyu hasn't been at Tencent long, but the Hunyuan large model has already been transformed. The same logic applies—the clusters Yao touched and models he tuned at OpenAI are invaluable to Tencent. Tencent certainly doesn't lack PhDs—everyone has a doctorate. Is the gap really in IQ?

An AI luminary's value is largely determined by their company's technical architecture and research direction. Two people with identical backgrounds and degrees—one goes to Company A, the other to Company B—could see wildly different career trajectories five years down the line. Imagine this scenario: your classmate A is at Anthropic researching AI infra optimization, classmate B is at OpenAI researching harness design, classmate C is at some startup frantically studying the 512,000 lines of Claude Code that leaked, and you're racking your brain over your weekly report, wondering: why didn't the five users from yesterday's AI red packet campaign convert?

NASA flight director Robert Frost once proposed an idea: 21st-century America struggles to return to the moon not because it lacks rocket blueprints and factories, but because it's missing the engineers, technicians, scientists, and flight controllers who understood the Apollo manufacturing process. An organization's success comes partly from papers, patents, and documents, but also from vast amounts of experience, engineering acumen, and management know-how. These tacit knowledge assets are like dark matter—difficult to observe but certainly present. Where does tacit knowledge in high-tech industries come from? Trial and error. Companies at the industry frontier inevitably conduct the most experimentation, and the tacit knowledge沉淀 through this process is the dream asset for those playing catch-up—and the core yardstick for pricing talent. During GPT-4 training, OpenAI used 25,000 A100 GPUs, yet model floating-point utilization was only 32%-36%. The reason: too many failures requiring frequent re-training from checkpoints. Each training run cost approximately $63 million in compute, plus additional costs for experimentation and data collection. The lessons bought with burning cash and intuition earned through losses all become tacit knowledge.

When Tencent poached Yao Shunyu from OpenAI and ByteDance hired Wu Yonghui from Google, their annual packages included a price for that tacit knowledge. If an expert can save $100 million in trial-and-error costs, paying them $50 million is still a bargain. In 2017, SMIC brought in Liang Mengsong as co-CEO. Who is Liang? The inventor of nearly 500 TSMC patents, who led Samsung to mass-produce the 14nm process first and snatched Apple's A9 chip foundry orders away from TSMC. SMIC couldn't use Samsung or TSMC patents, but Liang knew which directions were wrong and which paths were dead ends. That's why Liang had the confidence to push SMIC to skip generations in R&D—jumping from 28nm directly to developing 14nm, successfully achieving mass production. That's the power of an expert. If Michael Schumacher says he's in a hurry, you'd better hand over the driver's seat.

The hot capital market, rapid technological iteration, and frontier accumulation have all combined to inflate AI luminaries' value. But there's one more crucial element in this chain: compute power. AI labs are bleeding money, cloud providers are drowning in debt, yet NVIDIA alone is raking it in. On NVIDIA's Q2 earnings call, Jensen Huang gave a 70% revenue growth guidance for 2027, specifically noting this was the "supply-constrained" figure—if you measure by actual customer demand, growth would be over 100%, but the lithography machines can't run fast enough to meet it. How severe is the compute shortage? Zhipu went public in January, and within six months of listing launched a major fundraising round, raising HK$31.4 billion through a placement—a testament to how fast the money burns. Long-time AI bear Ed Zitron disclosed OpenAI financial documents showing that in 2025, OpenAI lost $38.5 billion, with R&D spending consuming $19.18 billion and $17.2 billion paid to Microsoft for compute rental.

Expensive compute raises the cost of trial and error for AI labs. Due to compute scarcity, many large model companies' cash reserves might only cover training two to three generations of models. Some forward-looking experiments can't secure sufficient compute, and exploratory research must yield to core projects. Fewer trial opportunities and higher trial costs consequently impact product competitiveness and talent development. The first step to value investing is having capital; the first step to pursuing AGI is having compute. Whether you can scale up model parameters and size depends entirely on whether you have enough compute in hand. In the widely circulated "Wenfeng Memo," Liang Wenfeng addressed this exact issue: "Talent isn't the bottleneck—resources are. Resources first affect talent development, because with less compute, we can run fewer experiments, so our talent overall lags behind the US." "The difference between us and the US, I believe, is entirely a resource difference. Every difference we see—talent, model capability, applications—can all be attributed to differences in compute resources."

Higher trial costs raise the price of tacit knowledge. If compute were cheap, every AI lab could access abundant resources, research directions would be liberated, R&D efficiency would improve, and talent cultivation would accelerate. But with compute expensive and trial costs high, tacit knowledge born at the industry frontier becomes even scarcer. A well-traveled mind not only points the way forward but also genuinely saves compute and trial costs. As long as compute remains scarce and gaps between models persist, stories like Yao Shunyu's will keep repeating. Can Yao's four choices be replicated? Absolutely—but there's no need. Some of Yao's Yao Class classmates are at OpenAI, others at Anthropic, and many stayed in top academic institutions for research—all firmly rooted in frontier fields like AI and quantum computing. If you can get into Tsinghua's Yao Class, you don't have a choice problem. Every step you take afterward creates new choices for the rest of us.

Creating geniuses requires the right environment. When Andrew Ng chose to collaborate with Google on the "Google Cat" project, it was because Google was the only place in the world with the data and compute to train such algorithms—itself built on Google's enormous profitability. Behind Google stood active venture capital and capital markets, formidable industrial moats, and academic supply. China doesn't lack geniuses—it has long lacked the Tencents that can afford to hire them. Tencent offers more than just compensation; its financial strength impacts the development of upstream and downstream supply chains, its investment affects the resources geniuses can command, and its ambition shapes what geniuses can achieve.

Ukrainian Valery Babich would surely understand this best. Who is Babich? A ship design expert, design department director at the Black Sea Shipyard, involved in the design and construction of every Soviet aircraft carrier, and chief designer of the Varyag. When the Soviet Union collapsed, Ukraine's shipbuilding industry lost its funding, the Varyag was halted, and Babich was laid off. In 2000, he pivoted to literature, unlocking his novelist skill tree, joining both the Ukrainian and Russian writers' unions, winning the Golden Pen Award in 2006 and the Arcas Prize in 2012. In 2014, Babich temporarily left the literary world to join the Qingdao China-Ukraine Special Ship Research and Design Institute, earning the Shandong Province "Qilu Friendship Award." Genius IQ is determined by genes; genius achievements are determined by industrial foundations. When Chinese companies eventually claim the AI throne, every ad you've scrolled past on Douyin, every skin you've bought in Honor of Kings, and every yuan you've lost in the A-share market—those are your medals of honor.