Summary

No bubble is in sight for at least 3 years. “When everyone is talking about a bubble, it definitely hasn’t arrived.” 朱啸虎 believes bearish sentiment is being “deliberately guided” by secondary-market investors hoping for a pullback, while the pullback over the past 1-2 months “has actually laid a very good foundation for another burst of growth next year”; the idea that GPUs depreciate in 2-3 years is “nonsense,” with 5-6-year-old and even 7-8-year-old cards still in use—the real bottlenecks are data centers and power.
A pullback would be healthier for Nvidia. But when asked whether he still held it, he softened: “There should be better names than Nvidia”; for a $500M stability-first portfolio, he would look at Google, Microsoft and Apple. OpenAI’s $300B financing is oversubscribed and many friends could not get allocations, but whether it can hold the consumer super-gateway “remains a very significant test”; he would pass on Anthropic because “API users have zero loyalty.”
OpenAI’s major strategic turn is from weekly active users to daily active users. Pulse, the browser and group chat are all about increasing usage frequency; “the group-chat move is extremely, extremely elegant,” and if OpenAI builds a social graph on the back end, “Meta is genuinely in danger”—that is his threshold for betting on OpenAI. “If OpenAI is going to become a trillion-dollar company, it has to move into social networking.”
Technologically, “we are basically at the end of the road for the transformer architecture,” and language models have peaked. Application-layer technology is already good enough and beginning to break out: token consumption is up 10x-plus this year and could rise another 10x-plus next year; in China, token costs are about RMB0.20 per DAU per day, and simple advertising can earn back RMB0.10. US AI valuations are 100x China’s—“one side is definitely wrong.”
His investing rule is to move 15 degrees away from consensus. He passed on foundation models and humanoid robots, instead backing underwater ship-cleaning robots, sales-oriented massage robots and AI companion toys, one of which should already be a clear category leader in less than 6 months. His risk test is simple: if consensus never converts into commercial reality, “then consensus is a bubble”; today’s foundation-model companies resemble the last cycle’s four VC darlings, “and are even worse than the four.”
The startup ecosystem’s room to maneuver has narrowed sharply. The odds of big tech continuing to make strategic mistakes are lower, while “second- and third-tier companies are all-in”; a $10T market-cap company “will definitely emerge, perhaps within 3-5 years, even within 3 years.” Founders should “move 3 blocks away from big tech,” pursue vertical optimization and private deployment, and grow quietly for 3 years; the $100B slot is nearly impossible to find, while “the Ele.me possibility” still exists at $10B.
His highest-conviction 10-year call is that China’s AI industry will decisively lead the US. “Catching up in algorithms and models is much easier than building data centers and generating power—especially when building a nuclear plant in the US is so difficult.” He expects China’s open-source ecosystem to lead in 3-5 years and “pull decisively ahead” over 5-10 years. His message to founders has evolved from “stay at the table” 2 years ago to “seize the opportunity and move at full speed.”

Deep dive

1. No Bubble in Sight for 3 Years: “When Everyone Is Talking About a Bubble, It Hasn’t Arrived”

朱啸虎’s core call is that there will be no bubble in sight for at least 3 years, and that bearish sentiment is being “deliberately guided by many secondary-market investors” whose goal is to “get a pullback.” The pullback over the past 1-2 months “has actually laid a very good foundation for another burst of growth next year.”
Bubble indicators should be read in reverse: when a real bubble arrives, nobody talks about a bubble and everyone piles in, as with “more than 90% of fiber-optic capacity sitting unused.” Conversely, when nobody dared to say they were investing in consumer last year, “that was the true bottom.” Today, even the stated scale of “几十兆 cards” and the available data centers are still far from sufficient; the market is nowhere near that point.
His view on Nvidia is nuanced: “A pullback would be healthier,” but when asked whether he still held it, he replied, “I think there should be better names than Nvidia.”

2. The Bubble Case Is “Nonsense”: Depreciation, Earnings Alignment and the US-China Valuation Divergence

GPUs depreciated over 2-3 years? “How could that be?” The cards available today are “nowhere near enough”; many 5-6-year-old cards, even 7-8-year-old cards, are still in use—if you have a card, you can use it. Token consumption has surged 10x-plus this year and could do so again next year, while data-center construction is “far behind.”
His framework: the 2000 internet bubble was a case of “price and earnings coming apart,” while AI in 2025 is a case of price and earnings moving together—revenues are genuinely rising. But there is one divergence: US AI valuations are 100x China’s. “One side is definitely wrong”—either US revenue is unsustainable or Chinese revenue is undervalued. “It is still unclear how this will evolve.”
The OpenAI-Oracle GPU flywheel “does look somewhat like the internet bubble back then,” but it is merely an accounting technique that puts depreciation off the balance sheet. “The core question is whether anyone is actually using the GPUs and whether there are enough of them.” Today there are not enough, so the mechanism does not undermine his bullish view.

3. OpenAI’s Major Strategic Turn: From Weekly Active Users to Daily Active Users, Then From Group Chat to Social

Sam Altman has “barely mentioned AGI” this year. Pulse, which he calls “the AI version of Toutiao,” the browser and group chat all reflect a renewed focus on moving from weekly active users to daily active users. “Weekly-active use cases are hard to defend against big-tech attacks, but the defensive moat around daily active use will deepen.” His verdict on Sam: “He says one thing and does another—he may talk about AGI, but in practice everything he is doing is on the application side.” 朱啸虎 also called OpenAI “a company built by investors”; the host replied that this was probably a good thing, since the company has obligations to investors, employees and everyone else.
“The group-chat move is extremely, extremely elegant.” Could OpenAI then build a social map and social relationships? It attacks areas where Google is weak and moves into Meta’s home territory. “If OpenAI is going to become a trillion-dollar company, it has to move into social networking.” That is also his threshold for betting: “If it can build a social network through group chat, the moat will be extremely deep, and Meta will genuinely be in danger.”
One product detail: when a company group is brought into ChatGPT to summarize a discussion, “we praised OpenAI, and it clicked a like on itself.” This kind of anthropomorphic behavior “didn’t exist before.”
An IPO “may be a little early.” Fundraising in the private market remains very easy, and the $300B valuation round was “far oversubscribed”; there is no reason to rush public. His end-state view is that once products like ChatGPT shift to daily active use, there will be “at most 2 or 3” major players, much like chat apps.

4. China’s Gateway Battle Is Just Beginning: ByteDance Leads, Alibaba Faces a Strategic Test

Who is China’s OpenAI? “Doubao and Qwen.” ByteDance is clearly ahead for now: “Doubao’s experience is genuinely very good,” with a clear traffic advantage. Alibaba is slightly late to the consumer push but has model capability; Qwen and Ant’s Lingguang both have “particularly strong highlights.” Tencent is following its usual pattern: in the past 20 years it has never been the company spending first to test the waters, waiting until the game is clear and then moving from fifth or sixth to first.
Alibaba’s instant-retail fight with Meituan is “mutually destructive,” and “the meaning of the investment may be limited.” The pressure of burning cash on both sides is intense. Whether Alibaba refocuses on consumer AI applications is “a strategic test,” and Qwen or Lingguang may matter more to Alibaba.
The big change is that 2025 “is no longer a foundation-model battle.” Nobody talks about AGI anymore. It is a battle for daily active users and time spent—a return to the competition for a high-engagement super-gateway that accumulates user data, much like WeChat.

5. Scaling Hits a Wall, but the Technology Is Already Good Enough: The Blind Spot in Token Consumption

Elon says pretraining data will run out and a new recipe will be needed. 朱啸虎 says “老坤” has been saying this all along; people used to think he was “an old geezer throwing cold water on the younger generation,” but now they believe it because Elon said it. “At least for the transformer architecture, we are basically at the end of the road.” Language models themselves have peaked, although multimodal models “may still have some room to take another couple of steps.”
That does not prevent an application boom: industry-wide token consumption is up 10x-plus this year, proving that today’s AI technology is already good enough. His position is that both things can be true: he is not optimistic about technological progress, but is extremely optimistic about AI’s continued breakout.
The scale is hiding in plain sight: “Our tiny company consumes tens of billions of tokens a day. I was stunned when I heard that.” A company can reach that level with 1M DAUs, and an app with 1M DAUs “is nothing special in mobile internet.” OpenAI says monthly consumption has reached trillions of tokens among large customers, and “there are many companies like that in China.”
The cost math is straightforward: Chinese token prices are very low, with most usage coming from cheap input tokens. Each DAU costs roughly RMB0.20 per day, and simple monetization through advertising can “earn back RMB0.10.” Big tech companies will discount or even give tokens away if they see strategic value in a use case. But he also points out the fluff: token consumption is “at least an indicator of AI content,” though 70%-80% of it is token padding and has “somewhat lower technical content.”

6. How to Allocate $500M: Google, Microsoft and Apple for Stability; Avoid the API Business

If he had to invest $500M in Silicon Valley AI giants, “for stability, Google and Microsoft, including Apple, are relatively safer.” OpenAI’s financing at a valuation of roughly $300B is “oversubscribed—I have many friends who did not get an allocation,” but whether it can secure a consumer super-gateway “remains a very significant test.” The strategy may be right, but execution is still unproven; defending a daily-active-user gateway requires a completely different DNA and operating track record.
Anthropic is out because of its business model: “API users have zero loyalty. If any model is cheaper and better, they will switch everything immediately.” OpenAI’s profitability has two sides: training the base model is extremely expensive, but the other side has controllable costs and revenue, so “it may not be particularly far from profitability.”
Where will AI revenue come from? “It is replacing your job.” There are roughly 1B white-collar workers globally, each costing about $40K; replacing 1%-2% would generate enormous revenue. But gross margin matters: Cursor is running at negative gross margin, effectively subsidizing programmers. In China, ByteDance, Alibaba and Tencent are “all definitely good names.”

7. DeepSeek Is Underestimated: Without It, Human AI Might Be Controlled by a Few Private Companies

“People are still underestimating how much it has changed humanity and history. Ten years from now, looking back, DeepSeek will be seen as a very important turning point in the AI process.” Without DeepSeek, China’s open-source ecosystem would not have committed so firmly to openness; “humanity’s AI could very well have been controlled by the models of a few private companies, and frankly that would be dangerous for all humanity.”
DeepSeek “never considered commercialization, and did not even consider an ecosystem. It was single-mindedly focused on seeing whether AGI could be reached.” He is looking forward to R2, but says both China and the US have shown how difficult the path to AGI is. Recent Chinese open-source models “may still be based on DeepSeek’s architecture,” and that alone is valuable.
The end state for APIs is clear: “It will definitely no longer be a good business,” increasingly resembling utilities such as electricity and water.

8. The Six Little Dragons Cannot See the Endgame: With Big Tech Tokens Nearly Free, How Can Startups Compete?

Base-model training capability “is still scarce,” and diffusion has been slower than he expected. But commercialization is the hardest part in China: big tech companies offer tokens free or at near-cost; Volcano Engine and Alibaba “give away tokens at an incredible level.” The models are good enough and integrated into cloud services. “What are you going to compete with?”
Whether Zhipu and MiniMax list first is beside the point: “Whether you list or not frankly does not matter. The key is how you ultimately commercialize.” He cannot answer that question. “It is not visible right now. It really is not visible.”

9. The Old-Timer Investor’s Cycle Rhythm: 6x the PC Internet Pace

“It is so similar, genuinely so similar.” From hardware and chips to infrastructure and then applications, “every cycle follows exactly the same rhythm.” The only difference is speed: the PC internet era moved at 6x the pace, and mobile internet at 3x. The 20-year retrospective confirms the pattern: the previous cycle was slow because applications were not ready at the start; this year applications have broken out and the cycle is visibly accelerating.
His 2026 forecast is for a further breakout in the application layer and another surge in token consumption, testing whether underlying infrastructure can keep up. “This is mainly a test for the US market; China should be able to solve the problem relatively more easily.” If closed-source models such as GPT stop advancing, China’s open-source ecosystem has significant room for another breakout. Many startups are already using 30B-class open-source small models for private deployments, with “very low costs and very rapid revenue growth.”
A second switch for market activity is the reopening of the Hong Kong IPO exit channel: “Investors are willing to invest again.” His listing thresholds are clear: consumer companies need at least $100M in profit; SaaS and AI applications need at least RMB1B in revenue. If they cannot reach those levels, “they might as well not raise money.”

10. Move 15 Degrees Away from Consensus: Ship-Cleaning Robots, Sales-Oriented Massage Robots and AI Companion Toys

The investing discipline is simple: “We no longer touch sectors where consensus is too concentrated; we move 15 degrees away.” He did not invest in foundation models or humanoid robots—“they are all very, very expensive,” and “we do not understand them, so we avoid them.” He invested in more than 10 companies this year and has been “particularly busy,” focusing mainly on AI applications and “workhorse robots.” The value-for-money proposition “improves immediately.”
One target is a robot that cleans ships at sea. Seawater is highly corrosive, the waves are large and “China’s seawater is very murky”; seeing the underside of a ship clearly and removing buildup presents a “very, very high barrier.” A massage robot can also conduct sales: “How’s the massage, sis? Want to top up?” The founder was “very grassroots,” had studied the category for years, and won the investment after a 30-minute conversation.
An AI companion toy was one of his bolder bets in this year’s crowded red ocean. He decided after 10-20 minutes of discussion; in less than 6 months it “should already be a clear category leader,” with extremely strong token-consumption metrics. The moat is “taking the user experience to the extreme—even getting the Wi-Fi connection right is not easy.” The product also does not require massive datasets to work well, so “the data-flywheel moat is actually quite small.”
His meeting rule is well known: he is famous for 10-minute meetings, and “if a meeting lasts 30 minutes, it basically means I am going to invest.” He dislikes founders who talk about grand trends: “The bigger the story they tell, the less willing I am to invest.” Founders who flatter or tailor their stories dishonestly do not get a second meeting.

11. The “Enormous” Advantage of a Complete Supply Chain: Every Competitor Is Within 20 Kilometers

A Vietnam example: a menswear sportswear company he backed was selling T-shirts well, so he asked it to make electric shavers. “It was simple, but they just could not do it. They had to come to the Greater Bay Area.” Every individual product looks simple, but “the system is what matters. Once you have 100 components, you become very powerful.”
拓竹 is doing very well, while Plug’s US market share is “extremely stable.” “Chinese founders are decisively ahead in every niche.” Competition is mostly among Chinese founders; “foreigners do not even have to be considered. Your competitors are all within 20 kilometers of you.” Smart wearables—glasses, earbuds, rings and watches—will inevitably become fully AI-enabled, but first-, second- and third-tier companies are all entering, so “competition will be extremely intense.”

12. Do Not Build Tools: Move 3 Blocks Away from Big Tech, Build Vertical Agents and Private Deployments

The mobile-internet lesson is clear. Flashlights, perpetual calendars, the AI-native email product Mailbox and Android optimization tools once had hundreds of millions of users worldwide. Where are they today? “They were not fake needs; they were real needs. But the key is that you cannot defend them.” Once the OS or underlying model has time, “it is very easy for it to build these things.” That is why he has always had reservations about the general-purpose Agent story around My纳斯 (original subtitles), while Lafite (original subtitles) is different: “Almost every image designer in China is on Liblib,” giving the designer ecosystem some defensive value.
The playbook has moved from being 1 block away from big tech in the mobile-internet era to being 3 blocks away in the AR era: vertical optimization, private deployment and doing your own sales. “These are things big tech does not want to do.” The growth curve for one vertical project was “tens of millions last year, more than RMB100M this year, and RMB200M-300M next year.”
The core of a vertical Agent is pushing accuracy to the limit, such as a US medical patient-doctor consultation Agent with “more than 95% accuracy.” The north-star metrics are the same as in the app era: growth rate and retention.

13. Big Tech Will Stop Making Mistakes: A $10T Company Will Appear Within 3 Years, and Startups Must Grow Quietly

The survival pattern from the last cycle: the US pure-play mobile internet produced only Uber, DoorDash and Airbnb, all of them “difficult, exhausting offline work that half of the offline incumbents did not want to do.” Douyin, Kuaishou and Xiaohongshu found room because “big tech’s repeated strategic mistakes created ecosystem slots.” What about this cycle? “You cannot find it in China or the US; you would not dare to bet on it.” Not only first-tier big tech, but “second- and third-tier companies are all-in.” Meta made some mistakes, but ChatGPT will move to fill the gap.
The conclusion is that big tech will get bigger: “A company with a $10T market cap will emerge within 3-5 years, and I even think within 3 years.” The $100B startup slot has “almost disappeared”; at $10B, “the Ele.me possibility” still exists, but it must be “an opportunity everyone looks down on on day one,” grown quietly for 3 years until outsiders cannot understand it. nano banana can “wipe out a bunch of startups with a single launch event.” Multimodality will not change the structure: “Big tech is still big tech.”
SOTA has lost its meaning: “In many scenarios, a 30B small model is the best—cost is low and accuracy is high. What is the point of spending hundreds of millions or billions of dollars training an SOTA model that leads for less than a month?” It makes no sense for startups; for big tech companies selling cloud services, “it may still have some meaning.”

14. VC Consensus Is Bubble Risk: Foundation-Model Companies Are “Even Worse Than the Four VC Darlings”

At a Hong Kong VC event, LPs complained that GP consensus had become too concentrated. Each deal involved 10-plus funds in a couple deal, leaving everyone with tiny ownership stakes. “GPs cannot make big money, LPs are unhappy, and investing with different VCs feels roughly the same.” Even the rounds had been bid away: “A-plus-plus-plus, with endless pluses.” Why is consensus so high? “The aesthetics have become too uniform,” and everyone benchmarks against the US, where commercialization works in these sectors—“except robotics.”
The key distinction is that internet-era consensus eventually converted into commercial reality. Attention economics, bike sharing and shared power banks were all mocked at the time, but eventually caught up. “Some consensus today may be very far from commercialization. If commercialization never materializes, then consensus is a bubble.” Today’s foundation-model companies resemble the previous cycle’s four VC darlings, “and are even worse than the four—the four at least did not have commercialization problems at the time. Today, commercialization may genuinely be difficult.”
His scar is EVs: he positioned “5 years ahead of the entire cycle” and lost a lot of money, then failed to invest when the real cycle arrived. Are foundation models and humanoid robots the same type of trade? “We do not know. Only time can tell us.” For early-stage funds, his approach is to control fund size and entry costs; the first check remains RMB20M-30M.

15. China’s AI Will Definitely Lead the US in 10 Years: Algorithms Are Easy to Catch Up With, Nuclear Plants Are Hard to Build

The US-China gap has stabilized at 3-6 months and is no longer widening; it may narrow over the next few years. Over a 5-10-year horizon, his highest-conviction call is that “China’s open-source AI ecosystem will definitely pull decisively ahead,” because “catching up in algorithms and models is much, much easier than building data centers and generating power—especially when building a nuclear plant in the US is so difficult.” China can build more than 10 nuclear plants at once, along with large amounts of offshore solar and solar-power capacity. He believes that is “absolutely the right thing to do”: the second half of the AI race is a race to build data centers and secure electricity. He also cited a public-account article: “China and the US have already brushed past each other.”
His message to founders has evolved from “stay at the table,” delivered at an investee-company meeting 2 years ago, to “seize the opportunity and move at full speed.” “There are genuinely too many opportunities. Do not think about the biggest opportunity; find a space to survive in the gaps.”

Verification Notes

The identities or standardized names of “老坤,” “My纳斯,” “Lafite,” “Plug” and “富盛” in the original subtitles could not be fully confirmed; the original names have been retained or annotated.
The original subtitle’s reference to “几十兆卡” is ambiguous in unit; the original wording has been retained.

GPUs, TPUs, & The Economics of AI Explained | Gavin Baker Interview China's Rise, GPT-5.2, Anthropic IPO & the Battle for AI Trust w/ Emad, Salim, Dave & AWG | EP #214