Summary

王煜全 lays out a firm timeline: he expects the AI bubble to emerge in 2029, and when the crisis peaks and turns back up, “gold will be everywhere.” The main drivers are a clear winner in robotaxis around 2029 and a surge in public expectations—fares falling to 1/4-1/5 of today’s levels and a market “worth at least $2T”; AI programming becoming another trillion-dollar market; crypto going mainstream alongside the tokenization of mainstream securities, amplifying volatility; and Trump leaving office in 2029 and inevitably facing a reckoning, starting with crypto. The discipline is to hold gains through the upswing, exit as the crisis approaches but “don’t go far”—the US knows how to throw money around and keep companies too big to fail, so recovery should take only 6 months to 1 year.
The endgame for autonomous driving is a subscription mobility service, with Tesla essentially on the brink of victory; in China, Huawei may be the only player that can still challenge it. FSD collects data daily from more than 1M vehicles, overwhelming Chinese automakers trapped in “data silos”; BYD uses Nvidia at the high end, Horizon Robotics in the mid-range and Black Sesame at the low end, exposing the mistake of thinking “the car is just the shell while forgetting that the soul is what matters.” His prescription: Chinese automakers should form a data alliance, while Didi’s Cheng Wei should go directly to Huawei, which does not want to take the front-facing role—“Can I be the frontman? We’ll use only your cars.”
The Magnificent Seven have made their positions clear: Tesla is “bad now, good later, a long-term winner,” while Google “will be very troublesome in the future.” Musk was hurt by Tesla’s board and got distracted playing with SpaceX; Cybercab, Cybertruck and the 1M-robot target are all strategic mistakes, but EVs plus autonomous driving have no rival in the US or Europe. More than 50% of Google’s revenue still comes from advertising, while foundation models deliver answers directly—“where is the ad inventory?”—so advertising spend will disappear within 10 years. Vision Pro has sold no more than roughly 400K units and “may be” the most embarrassing product in Apple’s history; but if Apple can make AI glasses work, “Meta’s fate is not in its own hands—it is in Apple’s hands.”
A $10T company is certain to emerge, potentially within 10-15 years; Tesla is competing for the title, while Nvidia could reach $6T, perhaps even $7T-$8T, before losing momentum, making $10T possible but a stretch. It could just as easily be a new entrant. OpenAI “doesn’t look likely right now”; he would rather bet on Anthropic because programming is a trillion-dollar market. Nvidia is building infrastructure while the infrastructure market continues to expand, making “at least the next 5 years of upside” a safe call.
“I oppose all of them”—humanoid robots, the low-altitude economy and commercial spaceflight—with the verdict that “humanoids are a false demand; pseudo-humanoids are the real demand.” A machine must first understand people, which requires cultural AI that has not even begun to be studied; a capable humanoid servant is “probably at least 20 years away.” Training legs is a big bubble, training arms a small bubble—Tesla talks about hands, while China talks about legs. Humanoid robotics is an entire industry whose supply chain is complete before the market exists; the real demand is closer to Yunji’s 4-wheeled “big lump” robots deployed across more than 30K hotels.
The most underestimated opportunity is mobile “behavioral intelligence” and user-side Agents: once foundation models enter the phone operating system, “all the gateways—or walls—of the super apps will be breached.” An Agent “stands on my side; it may even be me,” while Meituan and Alibaba remain oriented toward merchants or the platform itself and may be unable to turn around. Privacy requires local processing, which will drive up demand for on-device compute—one reason he calls Apple’s decision to discontinue Mac Pro shortsighted.
China’s turnaround play is to take open-source models and cheap chips overseas to build cloud infrastructure—a new “rural encircles the city” strategy, using the rest of the world to surround North America and Europe. After OpenClaw took off, the inference market expanded rapidly: “Anthropic’s Claude is too expensive; I might as well use Kimi.” Zhipu and MiniMax generated more overseas revenue in the first 2 months of this year than in all of last year. The key is to use the momentum to build cloud data centers in Southeast Asia and elsewhere, with hybrid clouds that train on Nvidia and infer on Chinese chips; governments should support this “highway.” In the US, Trump visited the UK with a delegation last October and pledged $150B to help Britain build AI infrastructure.

Deep dive

1. Methodology: The edge is being early—betting on day 27 of the lily pond

In 2017, 王煜全 predicted that Nvidia would overtake Intel, based not on financial statements but on the fact that cloud computing had clearly arrived, its biggest source of demand would inevitably be AI, and CUDA’s power was already evident—“it was almost impossible for anyone else to do this.” At the time, “everyone already thought you were crazy.”
His core distinction is that revenue and profit are lagging indicators. Once an industry war has already been decided, it still takes time for the public to see the value in financial statements—“usually at least 5-6 years.” Investing means finding the lead time, the foresight: knowing in advance who will win.
The lily-pond analogy: if lily pads double every day and cover the pond on day 30, day 27 has only 1/8 of the surface covered, with patches scattered across the water. “To ordinary people, you’re talking nonsense—7/8 of the water is still empty.” His forecasts are made on day 27.

2. The record of being wrong: a lidar bet, and Tesla “was exceptionally lucky”

He once believed lidar would be the mainstream autonomous-driving route: point clouds provide precise 3D distance measurements, while Moore’s law would halve hardware costs every 18 months, so “wouldn’t it become ubiquitous?” In hindsight, he had overlooked the cost structure of pure vision—once high-definition cameras are installed, the hardware may not need to be replaced for years; only the software gets updated.
The foundation-model revolution suddenly made pure vision viable and even capable of pulling ahead. “You can only say Tesla was exceptionally lucky—it caught the software revolution.” When 钦文 asked whether Musk had seen the foundation-model breakthrough coming, he flatly rejected the idea: “Not very likely.” Even the key people at OpenAI who used Transformer to ignite the revolution admitted they had not anticipated it; they had simply bet on the scaling law and kept spending to increase scale. “If even OpenAI didn’t expect the revolution to arrive at this point, how could Tesla have?”

3. The best call: 3 waves for EVs, with subscription mobility as the endgame

In 2018-2019, he laid out 3 waves: electrification first—although China was then pushing hybrids, he said the future would be pure EVs; intelligence next, as the 3 core EV systems converged and autonomous driving created differentiation; and mobility-as-a-service last. Once a car can drive better than a human, “there is no point owning a car that is completely outside your control.”
The endgame is analogous to mobile-phone pricing. You do not remember how many calls you make or what each one costs, but paying for every taxi ride separately “doesn’t make sense”; the future must be a monthly subscription. Autonomous taxis will become mainstream, leaving automakers as suppliers whose margins will not be the largest. The ultimate winner will be the mobility service provider.

4. Fleet management and Chinese automakers’ “small tragedy”

The term on Tesla’s 2022 Investor Day presentation—written down but not discussed—was fleet management. When owners are not using their cars, Tesla can take control and put them on the road to earn money. Musk estimated that a vehicle could generate more than $30K a year, against a car price of only $30K-$40K: “basically, it pays for itself in 1 year.” Chinese companies dismissed it as hype until Tesla abruptly began laying out the business last year, catching them off guard.
His diagnosis is that most Chinese companies “have not studied the future deeply enough.” They think with the trend: the car sells well today, so add another model tomorrow. Investment works in the opposite direction: “You won’t get to invest in today’s winners. You invest in tomorrow’s winners.” First determine who will hold the most power in the endgame.
During 1 month in the US, he saw 3 robotaxi contenders competing: Tesla, Amazon’s Zoox and Google’s Waymo. Waymo does not dare race Tesla’s expansion because FSD “really works everywhere”; Waymo cannot easily drive routes it does not know. Americans are already discussing when Tesla will cover the entire country—“it is basically on the verge of winning.”

5. Who in China can respond: data silos and Huawei

Faced with competition between Chinese automakers and Didi-type operators, he backs the automakers: the deciding factor is data scale. Chinese autonomous-driving companies use models similar to those in the US—end-to-end models and VLA—but differ in data volume. FSD has more than 1M vehicles collecting data on the road every day. BYD has a large installed base but cannot collect enough data because it lacks comparable intelligent-driving capabilities; Huawei has the capability but is not an automaker, and he suspects it is aggregating data through its partnerships with multiple manufacturers.
“Frankly, we believe that so far, Huawei may be the only player capable of taking on Tesla.” He calls on automakers to form an alliance and share a common dataset, but the current situation is “a very serious silo problem.” The less companies think about the future, the less they cooperate; the more they wait for the government to intervene, “the more vicious the cycle becomes.” He wants Tesla to serve as an effective catfish.

6. What happens to Didi and Uber—and BYD’s “soul and body”

His prescription for the operators is to bind themselves deeply to an automaker, even form a joint venture, and eventually become 1 company: “Didi and Uber cannot win this fight on their own.” Uber’s alliance with Waymo is “the stupidest strategy imaginable”; it should “surrender while it is still worth something,” bind directly with Tesla, or even discuss a merger. Uber could exchange its $200B market value for Tesla shares: “You go up 10x, I go up 10x. How great would that be?” Cheng Wei should approach Huawei: “Can I be the frontman? We’ll use only your cars.”
王传福 is “super smart,” yet he was still saying in 2024 that autonomous driving was nonsense before making an emergency pivot in 2025. “No matter how smart he is, his position determines his thinking.” The test is architecture: autonomous driving is the soul and the car is the body. Tesla’s Model S, Model X, Model 3 and Model Y all run on 1 system because they share the same soul; BYD uses Nvidia at the high end, Horizon Robotics in the mid-range and Black Sesame at the low end.
When Musk announced publicly, “We are not an automotive company; we are an AI and robotics company,” he was also speaking to Tesla’s entire workforce. FSD is the company’s most important business. Cars keep getting cheaper because “I want to use an inexpensive carrier so that my soul can be everywhere.”

7. Intermediate species and the essence of the intelligence revolution

Companies that disappear “each have their own misfortune”: some failed to keep up with the trend, some chose the wrong direction, and some were “intermediate species”—EV makers that won the transition from combustion engines but were not decisive enough about mobility-as-a-service. Future mobility leaders will buy in volume from only 2-3 suppliers; they do not need the dozens of Chinese automakers that exist today. Consolidation and brand concentration will continue.
The Industrial Revolution’s keyword was the assembly line: mass-produce goods and let users serve themselves. You cannot deliver an iced watermelon to every person, but you can give each person a refrigerator to chill one. The intelligence revolution mass-produces human services. After AlphaGo beat 李世石, “you no longer need 聂卫平 as a Go coach”—once AI performs a profession, everyone can enjoy that service.
His keyword for the current era is “intelligent services”: expert-level, personalized and accessible to all. The prerequisite is continuity. Continuous data collection produces deeper understanding: “You don’t even need to move your eyes—I know what you want to do and immediately hand you what you want.”

8. The underestimated opportunity: behavioral intelligence and user-side Agents

The underestimated opportunity is mobile “behavioral intelligence.” “It is hard to see mobile-related opportunities in the US and Europe because their phones are weak,” just as TikTok was a mobile application where China led. Phones record enormous amounts of behavior; AI can analyze it to predict and satisfy the next action. Privacy requires processing locally rather than uploading to the cloud, “which means demand for phone compute will rise.” After you order food delivery 10 times, an Agent can place the 11th order: “It knows exactly what I want, while also recommending things I had not thought of.”
When 钦文 asked whether Meituan and Alibaba would have an advantage, he reversed the premise. Platform applications stand on the opposite side of the user—on the merchant or platform side—whereas an Agent “stands on my side; it may even be me.” The difference in position means the incumbent platforms may be unable to turn around. The future may contain both general-purpose Agents and many specialized Agents, with a “chief steward” coordinating the division of labor among Agents.
He cited “technological feudalism,” a concept widely discussed in the US. Super apps use the stickiness of their core applications to enclose users and then monopolize everything around them. Once foundation models enter the phone operating system, “all the gateways—or walls—of the super apps will be breached.” They will not disappear, but the model of monetizing adjacent businesses will no longer work. “Is Taobao really good enough? Is WeChat really good enough? Because they already have enough of a monopoly to pull every other business inside, why would they improve?”

9. Why the bubble arrives in 2029: triggers and the crisis scenario

The framework comes from Carlota Perez’s Technological Revolutions and Financial Capital: every technological revolution generates excessive expectations, and when actual returns fail to keep up, a bubble forms. But technology is not the same as tulip bulbs. Bandwidth can fall from 10 to 1 while retaining 5 of real value—“I actually come out ahead.” Recovery then gives way to rapid growth.
The first trigger is robotaxis. By 2029, it should be clear that Tesla has won the competition; public expectations will be that fares fall to 1/4-1/5 of today’s levels and the market expands to more than 10x its current size—“at least a $2T market.” The danger comes when everyone thinks they can earn that $2T tomorrow. The second trigger is AI programming. OpenClaw is only 1 example, yet it is already this popular; major US IT companies are seeing revenue and profit surge while cutting jobs. That is another trillion-dollar market. Both sectors should be clear around 2029.
The third trigger is crypto going mainstream at the same time that mainstream securities are tokenized, so “any disturbance will be amplified by orders of magnitude.” Another factor is Trump leaving office in 2029 and inevitably facing a reckoning, beginning with crypto: “How much did his family take from crypto?”
The bright side is that the US knows how to handle crises: “throw money at it and keep the system too big to fail.” It will be a major crisis but a rapid recovery—6 months to 1 year, or 3-4 months if things move quickly. The trading discipline is to hold gains during the upswing, exit as the crisis approaches but “don’t go far.” When the crisis peaks and rebounds, “gold will be everywhere.” Seeing the future is not about sitting out; it is about learning to use volatility.

10. The Magnificent Seven (I): Tesla is “bad now, good later”; Google is good short term and inevitably bad long term

Nvidia is building infrastructure while the infrastructure market continues to expand, making “at least the next 5 years of upside” a safe call; growth may slow after that. Tesla is bad now but good later: Musk “was hurt by Tesla’s board” and shifted toward the more controllable SpaceX. “He is making a lot of strategic mistakes.” Cybercab has no steering wheel or pedals, so it cannot be sold to ordinary consumers; scaling it depends on long negotiations with local governments. Tesla will either produce too much and leave cars gathering dust in warehouses or fail to supply enough. Why not use Model 3 and Model Y?
When Cybertruck launched, the market was unanimous in its praise. “I was probably the only person who said this car would definitely fail.” Its sloped sides make loading difficult, and it is bulletproof: “If you drive a work truck, do you need it to be bulletproof?” It sold fewer than 10K units in its first year. The target of 1M robots amounts to “1M robots sitting in warehouses gathering dust—where is the demand?” The long-term view remains bullish because EVs plus autonomous driving have no rival in the US or Europe; Tesla is a long-term winner.
Google is the classic case of being good short term and inevitably bad long term. TPU, DeepMind and the Nobel Prize are all positives, but “its research has always been strong while its productization has not.” Advertising still accounts for more than 50% of revenue. “In the AI era, who will advertise?” Foundation models replace search by giving direct answers, leaving no ad inventory, and there will be more than 1 replacement. The future model should charge intermediary and transaction fees for analyzing and executing behavior, not sell ads. “I think advertising spend will disappear within 10 years, so Google will be in serious trouble.”

11. The Magnificent Seven (II): Apple’s potentially most embarrassing product, Microsoft’s question mark, Amazon’s floor

Apple has been “hit over the head with gold bricks” but has not built around its advantage. AI-era devices need sufficient compute, and Apple happens to have it. After OpenClaw popularized local compute, “every household will run one, and the compute requirements will keep rising.” He had someone buy a Mac mini for him at a premium; the trend points to every household needing a Mac Pro in 2 years. Yet Apple is reportedly considering discontinuing Mac Pro—“completely shortsighted.” “Since Jobs left, they have had no strategic vision.”
Vision Pro was priced at $3,500 based on an expectation of 600K units, so the product itself would not lose money. If Apple were willing to spend $5B from its hundreds of billions in cash reserves to subsidize 10M units by $500 each, it could cut the price and “immediately become No. 1 in the world in VR.” “The more willing you are to take risks, the easier it is to win.” Industry insiders estimate actual sales at no more than roughly 400K units, making it “possibly the most embarrassing product in Apple’s history.” If Apple launches genuinely strong AI glasses this year, it could take a large share of Meta’s market: “Meta’s fate is not in its own hands—it is in Apple’s hands.” But he thinks Apple may not be able to launch a good product first; even if it does, the price may be wrong.
Microsoft’s problem can be summarized in 1 sentence: “It has no foundation model of its own.” It is not tightly bound to users or suppliers. “It is not certain to fall, but it carries risk.” Amazon’s shift online is irreversible, and the cloud business has already generated substantial revenue. “If AI is not good enough, I can still survive; if it is good enough, I can capture all of e-commerce revenue.” Amazon will rise overall; the question is whether it rises quickly or slowly.
A $10T company “is certain to emerge.” Tesla is competing for it now and would have a real chance if robotaxis break out. Nvidia could reach $6T and perhaps $7T-$8T before losing momentum; $10T is possible but would be a stretch. A $10T company could emerge within 10-15 years and “is very likely to be a new entrant.” OpenAI “doesn’t look likely right now”; he would rather bet on Anthropic because programming is a trillion-dollar market.

12. Overhyped narratives: humanoids are a false demand, pseudo-humanoids are the real demand

“I oppose them all”—humanoid robots, the low-altitude economy and commercial spaceflight. A humanoid servant must first understand you. The right temporoparietal junction in the human brain intuitively infers another person’s thoughts and intentions; “robots do not have this system.” The old focus was automation, or movement capability; the future is autonomous. Physical AI solves only the problem of understanding the environment. Understanding people requires cultural AI, and “research in this field has not even begun.” A competent servant is “probably at least 20 years away.”
The test for separating real from fake demand is simple: “If you are training legs, it is a big bubble; if you are training arms, it is a small bubble.” “Humanoids are a false demand; pseudo-humanoids are the real demand.” Chinese humanoid companies talk almost entirely about legs; Tesla talks about hands. They use the humanoid robot as a cover while selling “those 2 hands.”
He warns that humanoid robotics is an “entire industry”: “The market does not even exist, yet the supply chain is already complete.” AgiBot has shifted its focus toward pseudo-humanoids with chassis—“there is a real market for making coffee.” Unitree still sells because of university-lab and performance markets, but “after people get used to watching monkey tricks, will they keep watching?” Factory automation is already 100 years old, and in China it has been the main substitution theme for the past 20 years. “I have never heard of a factory application that must be done by a humanoid robot.”
The positive example is the unfashionable Yunji: 4 wheels and 1 big lump. Hotels “may even find it strange” not to have one delivering meals. Its robots are already deployed in more than 30K hotels. The point is not that there is no market for robots, but that companies must seriously examine how people will live in the future rather than build something cool and flashy with no application. The low-altitude economy faces the same problem: 4-rotor electric aircraft are disadvantaged in flight. There will be a small niche market, but it cannot scale.

13. China’s card: take open-source models overseas, build cloud infrastructure, and use chips as the “highway”

He agrees that foundation models and chips are important markets, but the strategy has to be right. For model companies without their own cloud—Kimi, MiniMax and Zhipu—the opportunity is overseas. After OpenClaw took off, the inference market expanded rapidly. “Anthropic’s Claude is too expensive; I might as well use Kimi.” These companies generated more overseas revenue in the first 2 months of this year than in all of last year; Zhipu and MiniMax shares are rising. The key to establishing a foothold is to use the momentum to raise capital and build cloud data centers in Southeast Asia and elsewhere. “They are all open-source models—how different can they really be?” Operations will decide the winner.
Chinese chips “are not actually strong, but they are cheap.” The strategy is to build hybrid clouds overseas: train on Nvidia and infer on Chinese chips, using the opportunity to build compute centers around the world. “That should be the main theme of the future; I even think governments should strongly support it.” This is the “highway”: help countries build the roads, and you become deeply embedded in their economies.
The US has already moved, with Europe as the focus. Last October, Trump visited the UK with a delegation that included corporate executives and announced $150B of investment to help Britain build AI infrastructure. “Put simply, it is helping you pave the entire highway.” China’s approach is to use its geopolitical advantages and its many poorer partners to enter Southeast Asia and the developing world. “It is another rural-encircles-the-city strategy, but in the broad sense: using the entire world to surround North America and Europe.” “The US and Europe will not do this; China is very good at it.”

14. Strength lies in depth of use: science, technology and engineering—and where China is strong

He frequently cites economist Diego Comin’s conclusion: “The strength of an economy does not depend on how quickly it adopts advanced technology, but on how deeply it uses advanced technology.” In the US, advanced technology covers only “2 narrow strips along the East and West Coasts,” and many people still use cash. China did not invent QR codes, but it has the highest usage rate and high efficiency—people “benefit from technology.”
China often lumps “technology” together, but it is really 3 separate things: leadership in scientific research—where the US is clearly ahead—does not equal technological leadership, and technological leadership does not equal commercial leadership. To scale technology so everyone can use it requires engineering leadership; “China is No. 1 in engineering capability worldwide.” The future form of globalization is “mid-tech, high design.” Only a few Americans, including Musk, understand this: if they did, they would not try to blockade China, because a blockade cuts the final link in the US chain from science to engineering. “If you play it right, form an alliance with China: you be the majority shareholder, I’ll be the second-largest shareholder. That would be best.”

15. Jobs, education and “diligently lazy”: think bigger

钦文 asked whether 3 coders doing the work of 10 means the 7 people laid off are “the cost of evolution.” His answer is redeployment. During the Industrial Revolution, factory employment fell with automation while white-collar and sales jobs surged; in the US in the 1950s and 1960s, white-collar employment grew by more than 50%. “If factories produce that much, someone has to sell it.” The intelligent-services era will create interface jobs for continuous service and new interest communities. “The increase in service jobs will be enormous, just as the increase in sales jobs was back then,” and what makes someone exceptional is “love.”
The barrier to entrepreneurship is not resources but insight into an industry. His hierarchy is: “The ability to define a problem is greater than the ability to ask a question, which is greater than the ability to solve a problem.” Jobs did not conduct market research; “he designed for himself, and it happened that the whole world liked it.” The new generation of entrepreneurs is “diligently lazy”—diligent about thinking through how to make AI do the work. “Everyone wants to be lazy, so this is a huge business.” In a major revolution, “do not occupy a small hill”; education is the biggest market, healthcare is the biggest market, and everyone’s self-actualization is the biggest market.
Education must move from the standardization of the Industrial Revolution to personalization. Children need an intelligent system that stays with them over the long term, understands their moods, emotions and changing interests, and guides them toward exploration and a lifelong passion. “That passion will be their core competitive advantage in the future.”
He closed by quoting Mark Twain: “History does not repeat itself, but it rhymes.” Find the rhyme and you can know where things are headed. His message was: “Do not let this era down. Do not think you can only do small things. This is an era for doing big things… In the future, you will certainly thank yourself for thinking big enough today.”

The Citadel Alum Reshaping The World Of Trading $PRKS: SeaWorld, an 8% cash yield, and a possible 80% short squeeze | Hawkins Entrekin