Topics Pioneers Insight Method Author
Back

Urgent Bonus Episode: When STAR 50 Becomes STAR 25, Has the Global AI Bubble Actually Burst? – Huiketang X Mossfire

·苔藓之火

Rendering the long image…

Topics AI Venture Funding and Bubble Risks Trillion-Dollar AI Capex and Bubble Risks Macro Asset Allocation in High-Interest Regimes
Pioneers 董艺婷

Summary

  • The AI industry can be booming even as tech stocks suffer a sharp pullback. Raymond sees paid demand in the US driving growth across model companies, applications and suppliers; Dong Yiting is looking at a market where the Nasdaq is still making new highs while China’s STAR 50 Index has fallen about 50%. They start from the same premise: China and the US each contain four distinct “time zones” across public and private markets, so industry prospects alone cannot tell you what your position will return.
  • The US AI market is already splitting: cracks in the funding boom coexist with a shortage of compute. Raymond points to Thinking Machines Lab’s $12B first round and the difficulty it faced raising money at a rumored $50B valuation for its second, evidence that “the next OpenAI” will not automatically find backers. Meanwhile, cloud providers remain in favor because they control compute. Today’s supply-demand imbalance does not rule out a glut in the next year or two, or stop public markets from pricing in default risk ahead of time.
  • The debate over $1T in annual capex ultimately comes down to whether future revenue can absorb depreciation. Raymond’s rough math: investing about $1T a year and depreciating it over four or five years could eventually require several trillion dollars in annual revenue to support the spend. If revenue falls short, oversupply could ripple through the supply chain. But he leaves room for another possibility: technological breakthroughs could expand demand. “We shouldn’t use the scale of the old civilization to estimate what the new one will look like.”
  • Cross-market moves have real transmission mechanisms, but they are no clock for calling a top or rotating positions. Dong Yiting distinguishes three links: investment in the industry and earnings, shared liquidity, and short-term sentiment extrapolation. The last makes it especially tempting to treat the first market to open as the answer. Raymond worries that after losses, investors may chase from A-shares into US stocks, only to get caught by a reversal and miss the most important up days in both markets.
  • Spotting crowded trades and correctly calling a bubble are different from trading it well. Dong Yiting sees this A-share selloff as driven mainly by capital flows and risk appetite, not as proof that the technology outlook has been disproved. Raymond argues that difficulty forecasting markets is no reason to stop studying entry points. The discussion comes back to preparation: “You can use any selloff to prove your bubble call was right. That doesn’t mean you can trade it right.”
  • The value of asset allocation is that a mistake in a tech position need not wreck a family’s wealth. Dong Yiting illustrates the point with a portfolio that has 60% or 70% in tech: a 50% index decline could cost roughly 30% of total wealth. Diversification and rebalancing protect an investor’s ability to stay invested. They cannot call the top precisely, but may help investors “go further”; her opposition to leverage runs through the entire discussion.
  • Whether you can make money from the technology wave depends first on your investment horizon, position size and financial constraints. Industry participants may benefit from “industry leverage layered on human capital leverage,” while ordinary investors need a capital structure that can withstand volatility. For those already nursing losses, preparing to enter, or looking to concentrate their bets, the recurring questions are not when the next breakthrough will arrive, but whether they can bear the losses, follow their rules and actually “sleep at night.”

Deep dive

1. US private-market growth is being pulled by paying customers

  • Raymond spent nearly a month in the US, mostly meeting tech founders and investors. Public information may be much the same, he said, but conversations with people on the ground reveal a picture beyond stock charts and earnings reports. “The state of mind is quite different.”
  • In his view, revenue growth at ecosystem leaders such as Anthropic and OpenAI is driving demand for compute, data and a range of services. In China, Doubao and DeepSeek are widely used, but far fewer users actually pay, which limits the size of the supporting ecosystem.
  • He points to Cognition as an example of how the US market tolerates second- and third-place players. Products differ, and the paying market is large enough for companies further down the rankings to grow revenue quickly. Investors’ attitude is: “You’re number two. So what?”

2. New Labs are showing cracks, while robotics is at a different stage

  • Raymond sees the clearest signs of a US private-market bubble in New Labs: researchers leaving big tech with new algorithm ideas, backed by investors who were once unusually forgiving. He cites Thinking Machines Lab, which raised $12B in its first round but struggled to raise money when its second-round valuation was reportedly pitched at $50B. In his view, “there’s a big crack” in this market.
  • Robotics is on a different timeline. Raymond feels China has already been through a cycle of fundraising, leading companies going public, share prices falling and then recovering. Many US robotics, data and application labs, by contrast, have only emerged in the past 12 months.
  • He even speculates that the US could see robotics New Labs valued at tens of billions of dollars before their products take shape. But in his view, that kind of bubble is not yet pronounced in robotics. Even within one market, some sectors are just getting started while older private shares have lost their appeal.

3. The scramble for compute and public-market default concerns coexist

  • Raymond mentions CoreWeave and Nebius: compute is such a good business that people at dinner parties offer them extra courtesies, hoping to get “a little bit of compute.” He explicitly says he has only heard rumors about companies using compute allocations to secure equity stakes or other benefits.
  • Resource allocation is contentious even within Google DeepMind. He relays that some researchers believe they need 5,000 or 10,000 GPUs to test their ideas. Unwilling to spend the prime years of their research careers waiting, some have left or started New Labs.
  • Yet public markets are already discussing default risk through CoreWeave CDS. Raymond draws no firm conclusion: compute could be in oversupply next year or the year after, but today the scramble is still on. This is one example of the four “parallel universes” he sees across China and the US, public and private markets.

4. A bullish long-term view does not remove the risks of overpaying or concentrating positions

  • Dong Yiting recalls her September 2024 view: an unprecedented technology wave coincided with what she saw as a Chinese stock market that “couldn’t fall any further,” creating an opportunity to the upside. She also warned that it would become very dangerous once skeptics started praising that call.
  • She observed that in June this year, some investors who had stuck with dividend strategies or barbell portfolios switched entirely into tech, only to ride the decline from July onward. One listener described watching “the STAR 50 slowly become the STAR 40, then the STAR 30, and now it’s the STAR 25.”
  • From long-term industry bulls to risk monitors and leveraged momentum traders, listeners face the same underlying problem: the long-term bullish case has become consensus, but how should they allocate their money to assets with “high opportunity, high risk and high volatility”?

5. Investors who can make money from beginning to end need the right capital base

  • Dong Yiting first distinguishes industry participants: people able to use their own capabilities to secure resources from companies such as OpenAI already have industry exposure and future valuation upside. They may benefit from “industry leverage layered on human capital leverage.”
  • Those outside the industry who want to hold for the long term need ample capital and no pressure from financial obligations, leverage, losses or time. Those who entered early and have already taken out their principal after a rise may also do well with the remaining position.
  • She is more cautious about people who enter late and then chase other markets, and firmly opposed to leverage. Whether someone can hold on or exit “doesn’t depend entirely on what you know”; capital and investment style still set the limits.

6. Real technological breakthroughs do not mean valuations cannot be in a bubble

  • Dong Yiting turns the question around: “Because the tech industry has a future, it must have a bubble.” Raymond, meanwhile, believes there are real underlying technological changes today, so this is not directly comparable to 2021, when massive liquidity drove sentiment without an obvious technological breakthrough.
  • If 10 represented the mood and absolute valuation excesses of 2021, Raymond’s subjective read is that today is around 7 or 8. Dong Yiting adds that A-shares may have been at 7 or 8 in June and could now be back to 5.5 or lower. The Nasdaq is in a different state, and making new highs does not mean its path has been smooth.
  • US investors are also buoyed by how quickly gains can be realized. Raymond says a portfolio company might be acquired by Nvidia just 18 months later, giving a fund DPI quickly and allowing it to raise its next fund. “Fake money” becomes “real money” fast, and that positive feedback fuels the next round of investment.

7. The capital behind a valuation shapes what investors can believe

  • Dong Yiting argues that China’s links between research and capital, and the handoffs from angel funding through later rounds to exit, are not as tight as in the US. When investors have low risk tolerance and strict short-term performance targets, they struggle to underwrite long-term uncertainty. “Who gives you courage? The capital behind you does.”
  • That is why comparing a robotics company with OpenAI or Nvidia on a 50x, 60x or 80x PE multiple does not explain the valuation. Beyond growth expectations, you have to ask how long the buyers’ capital can wait and what risks it can bear.
  • She goes further, offering a judgment she acknowledges is subjective: “More than 60%” of pricing is determined by the capital behind it. Even founders cannot fully predict where technology will end up. Investors do not first see the future clearly and then set a price; they often look for information that fits their funding constraints.

8. M&A exits can create a virtuous cycle, while funding gaps push companies toward IPOs

  • Raymond cites $159B in acquisitions of US AI startups over the past 12 months, including $60B for Cursor. He says a16z had originally planned to lead a round at a valuation of about $50B over the same period. He cannot remember whether the funding amount was $5B or $10B.
  • His comparison is with Chinese companies that, after approaching early-stage dollar funds, state-owned capital and industry investors, have little choice but to push for an IPO. A Series C that might be funded by one US fund becomes a public-market offering in China. Many US companies, by contrast, are acquired before listing.
  • He believes the difference also comes down to buyers: first-generation business owners are more likely to believe they can build the business themselves, while professional managers focus more on “Time to Market.” Weak stock prices also make stock-financed acquisitions less attractive for public companies, further narrowing exit options.
  • But he rejects any reason to feel inferior. At one point, models seemed to be falling further behind; three months later, “K3 came out” and caught up again. He admires Chinese companies for producing similar results with 1% or even one-thousandth of the resources their US peers use, and says, “May there be bubbles everywhere,” expressing his hope for more capital to support the industry.

9. There are at least three distinct channels for cross-market moves

  • Dong Yiting first explains the real links among semiconductor companies in the US, South Korea and Taiwan: demand for compute, chips and wafers drives cross-investment; cash flows and profits become more connected; improving earnings then spur another round of investment.
  • A second channel is shared liquidity. When large investors manage portfolios, they may take some gains from winning positions and allocate them to other assets. A larger pool of capital can also create synchronized moves across markets.
  • Short-term follow-through may instead come down to “one person, one vote”: investors without their own view of fundamentals or capital flows extrapolate linearly from the market that opened first. Long-term earnings links and overnight sentiment spillovers are different signals.

10. Assess the risks of trillion-dollar spending by tracing depreciation and revenue

  • Raymond sees capex first as evidence of a real boom. In the US, data center operators struggle to hire even construction workers, while immigration restrictions add to the labor shortage. Strong demand is more than a figure in an earnings report.
  • His rough math: if companies spend about $1T a year and depreciate it over four or five years, annual depreciation could rise by about $200B. With investment sustained, annual depreciation could eventually reach $1T, requiring several trillion dollars in revenue to support it.
  • Investors’ concerns therefore have a basis: if revenue falls short and too many data centers are built, oversupply could drive prices down quickly and then ripple from the industry leaders to suppliers. But that is a debate about the future; it does not disprove today’s shortage.
  • He also leaves room for a technology breakthrough that changes the picture. Over the past three years, sentiment has often turned pessimistic after the third quarter, only for a miracle to arrive the following January. He points to DeepSeek in 2025 and Claude Code in 2026 as reminders that you cannot use the scale of the old civilization to estimate what the new one will look like.

11. A few investors going all in do not mean the capital system has high risk appetite

  • Dong Yiting sets the discussion against the macro backdrop: US inflation and Treasury yields are high, while China has low inflation and low yields. She sees the latter as evidence of low risk appetite across society; a few people around you concentrating their money in tech do not negate that.
  • Leveraged speculation can even push risk appetite lower. When high-risk investors are liquidated and exit, others watching are frightened too. Her test is whether US and Chinese Treasury yields reverse direction; without that, local all-in bets do not signal a shift in the broader capital system.
  • On the Nasdaq and Bitcoin, her view is clear: in an extreme environment, both would be hit by shared liquidity and risk appetite. “The correlation may not look very high now, but when the time comes, it will become very high.”

12. Crowding can be observed, but it does not prove a bubble

  • Dong Yiting defines a bubble as a valuation far above the profit growth expected over a given period, or a major fundamental problem that makes earnings unattainable at current expectations. The difficulty is that AI technology keeps resetting expectations, making it hard for outside investors to distinguish a genuine gap between expectations and reality from a lack of information.
  • That is why she thinks bubbles are “hard to identify, almost impossible to identify,” especially for people outside the industry. Crowding reflects investor composition and concentrated trading; it does not directly measure the gap between valuation and fundamental expectations.
  • She characterizes this A-share selloff as a contraction in capital and risk appetite, not proof that the technology outlook has been disproved. Identifying industry change requires understanding the technology roadmap and ecosystem. Public markets already lag substantially, and ordinary investors may not benefit from making bubble prediction a daily concern.

13. Correct predictions and successful trades are separated by preparation

  • Raymond cites Michael Burry, the inspiration for The Big Short. As he tells it, Burry’s fund was only a little over $100M, while his Substack had more than 300,000 followers and annual revenue approaching $100M. He also mentions that Burry publicly shorted Nvidia and Palantir last year or the year before, though he cannot recall exactly when. “Pessimists are right; optimists make money.”
  • Dong Yiting adds the example of Xie Guozhong, who began calling a downturn in China’s property market in 2004. If you put the eventual turning point in 2022, he was nearly 20 years early. A later selloff can vindicate the prediction without making it a successful trade at the time.
  • Raymond still argues for studying entry points in the context of your position. If you already hold the asset and think it is expensive, you can trim; if you have no position and want exposure, you need to work out where to buy. He even thinks that if you insist on guessing the top or bottom, you should at least try—otherwise, when panic really arrives, you will not know what to do.
  • Dong Yiting brings it back to preparation: it is more practical to plan how to protect capital and achieve the returns you want under different scenarios than to fixate on the date. Their emphasis differs: she relies more on rebalancing, while he stresses combining research with position sizing.

14. Asset allocation protects your ability to bear risk; it cannot promise a perfectly timed exit

  • Dong Yiting says diversification first protects family wealth, so a mistake in a highly uncertain asset does not put the rest of your life at risk. Second, investors can keep holding tech only if the rest of their portfolio remains safe.
  • She compares two scenarios. If 60% or 70% of wealth is concentrated in tech, a 50% index decline could cost roughly 30% of total wealth. If the starting allocation is 30% tech and 70% capital preservation, then the tech position doubles and the portfolio is rebalanced, losses in her simplified example might later account for only the low teens as a percentage of initial wealth.
  • That protection can help investors “go further.” Asset allocation cannot call the top precisely, and a barbell strategy is not a competing principle; it is one way to diversify and protect a portfolio.

15. Research can bring peace of mind, but ample savings may not remove loss aversion

  • Raymond manages a tech fund himself and believes that understanding a company’s business, finances and management makes a position easier to hold. If you understand Nvidia well, a sudden 20% drop might even look like an opportunity rather than a reason to panic.
  • Asked whether stable cash flow means you can put all your other assets into high-risk investments, Dong Yiting’s condition is that savings and income must be sufficient to support the family, and losing the invested portion must not affect either your life or your state of mind.
  • Raymond is skeptical: “Nobody can put money in and then stop looking.” He has seen very wealthy people still fixate on losses, because money is also a scoreboard; losing feels like getting two points taken off an exam.
  • Dong Yiting points to the other side: looking back 20 years from now, today’s 40% to 50% swings may seem like small bumps, and people may regret not holding on. She calls this “not a personal finance lesson, but a meditation lesson.” In the end, both return to choosing an approach that fits you.

16. Existing losses, a first entry and a market switch call for different rules

  • For a losing position, Dong Yiting offers a conditional rebalancing example: if it originally accounted for 20% of the portfolio and falls by half to about 10%, you could consider adding back to 20%. The prerequisite is that the rest of your assets are stable and the overall allocation is safe enough.
  • For investors who have never held the asset and are preparing to enter, she recommends buying in stages and setting a cap. New investors are not the only ones who lack experience; even seasoned investors cannot always make accurate calls. Rules should take on part of the decision-making burden.
  • On selling A-shares to switch into US equity ETFs, Raymond asks: what if A-shares suddenly rise from 3,000 to 6,000 after you switch? Citing the idea that Bitcoin may have only about 10 major up days a year, he worries that repeatedly switching markets could mean missing the key up days in both.

17. Knowing what you own is different from earning executable alpha

  • Dong Yiting distinguishes product structure from company fundamentals. For publicly disclosed ETFs and linked products, you can check the underlying assets in disclosures, contracts, with an account manager or through AI. Understanding a listed company’s business takes much more time. For private products that do not disclose holdings, you may need specialized attribution tools.
  • In her view, asset allocation protects the family, discipline enables “fuzzy timing,” and stock picking depends on whether the industry offers scope for individual stocks to outperform. You cannot rank one approach as the most important without considering the market phase.
  • Raymond warns that it may take only “one second” for information to be fully reflected in prices, and “a lot of so-called alpha is fake.” Tariffs, wars, oil prices and changes in US Treasuries also keep exceeding what past experience can explain.
  • Even the boundaries of tech-stock research have changed: from familiar names such as Tencent, WeChat, Alibaba and Xiaohongshu to optical modules, PCBs, semiconductor niches and energy. He never expected studying AI to lead him to gas turbines, but that is the new normal he has to learn.

18. Compute rental prices offer clues, but leading research is hard to do

  • Raymond suggests tracking H100 and even A100 rental prices. What stands out to him is that “a chip from 8 years ago” is still being used, with people still willing to pay roughly the same price. That alone is a demand signal.
  • If new compute capacity becomes oversupplied, customers should favor newer chips such as B300, and rental prices for older chips should theoretically fall sharply. Spot memory prices are also worth watching. But by the time the inventory chain starts dumping product, it may already be too late.
  • He gives the example of Citrini’s research team examining the $500B “Stargate” project: they use drones to identify construction equipment and liquid-cooling brands, then thermal imaging to check whether the machines are actually running.
  • Dong Yiting compares this with CTA firms using drones, satellites and remote sensing to study crops. Her conclusion is that if well-resourced institutions are already tracking leading indicators, prices may have reflected the change long ago. Deep industry research can satisfy your curiosity, but it does not automatically solve family wealth management.