AI投资的大问题：不是泡沫破灭，但三年硬件牛市结束了

By丨Shen Yuan Huang Junjie

Edit|Huang Junjie

On October 26, we interviewed Liu Difan for the fourth time during market turbulence. That day, the Philadelphia Semiconductor Index had already fallen 16% from its October peak. Two weeks later, when the article was published, the decline had widened to 29%.

Over the past two years, we have discussed the perennial questions of cycles and investment research several times:

Over the past three years, AI hardware stocks have taken turns rising: from GPUs to HBM, from power to optical modules, then to hard drives, memory, PCBs, power semiconductors, MLCCs, indium phosphide… In this wave, the cyclical voice seems to have been constantly drowned out.

This downturn has been attributed by the market to Intel’s earnings missing expectations and the Democrats’ landslide in the midterm elections. But in his view, a deeper reason is that the accounting for compute infrastructure is getting tighter. A fund manager calculated that the $2 trillion of compute infrastructure already invested requires enterprises and consumers to generate $3.7 trillion in annual revenue just to sustain it.

Demand is hard to estimate, but the changes in supply are already visible.

He gave an example: indium phosphide suppliers have grown from 3 to 7‑8 companies. The speed of this supply release, in his opinion, is a typical characteristic of the later stage of a cycle. The listing of ChangXin, and Samsung and Hynix announcing more factories, are also classic topping signals.

British asset management firm Marathon wrote this principle in Capital Account: the higher the return of an industry, the more it attracts capital expansion, and the added supply will depress returns. Therefore they advise not to take the demand‑growth story told by corporate management and investment banks at face value, but to look at whether capital expenditures and capacity are growing too fast. Historically, at every cycle peak, people always say “this time is different.” — Whether this time truly is different remains to be seen.

Liu Difan founded CT Associates (Red Hare Capital) in 2020. To design overseas asset‑allocation solutions for LPs (limited partners), he continuously interviews fund managers worldwide and has met managers from over 600 hedge funds. His view of this cycle comes from the regularities of cycles and from these people’s position changes.

The following is the interview from October 26. The two follow‑up questions asked later are marked separately.

Later: What’s the difference in background between those who are bearish and those who are bullish?

Liu Difan: All three have gone through China’s new energy vehicle market. Those who haven’t traded NEVs can’t grasp that when profit margins are highest and demand is tight, a cycle actually starts to turn down.

The seven bullish ones are all U.S. fund managers. They can’t understand how freely Asian hardware manufacturers can expand production once they start making money. Everyone fills capacity and competes until others run out of cash. Whether it’s model factories or storage factories, once Chinese influence rises, nobody makes money.

Later: In the new energy vehicle industry, the “car manufacturers” can’t clearly articulate the real demand.

Liu Difan: I learned early on from an automotive analyst at a major bank not to ask a ToB boss about his needs. He has no idea where his product ends up, and any forecast he gives you is wrong.

Using electric vehicles as an analogy, there are three layers: consumers, car manufacturers, and parts suppliers. Cloud providers are like car manufacturers; they expect electric vehicle penetration to eventually reach 100% and place huge orders with parts suppliers, who then see extreme scarcity. But if fewer people actually buy cars in the end, will the car manufacturers cut those orders?

Previously, everyone assumed that a cloud provider’s order meant final demand, but now we also have to see whether anyone actually uses it. Backlogged orders might include OpenAI’s orders, and cloud providers also invest in OpenAI, creating a loop that no one can fully calculate.

Later: Why does capacity expansion always go overboard? What signals can be seen earliest?

Liú Dǐfán: There is human greed inside, and also the “impossible tasks” of hardware manufacturers’ capacity planning. Indium phosphide suppliers expanding from 3 to 7‑8 companies is basically it. Changxin, Changcun going public, SK Hynix issuing ADRs in the US, Samsung and Hynix announcing more factories because of strong demand, all are peak signals. (A week after this interview, there was a report that Changcun’s IPO has been suspended)

This year’s bottlenecks include HBM, power semiconductors, analog chips, MLCC, PCB, optical modules—could these become oversupplied and turn into short targets again?

People with cyclical experience, when they see everyone expanding capacity unchecked and projecting current profits into the future, know it’s a signal—not a judgment derived from fully crunching the numbers.

Later: At the beginning of the year even Apple had to pay a high premium to get memory chips, what about now?

Liu Difan: Storage is starting to loosen, prices have dropped, and vendors are no longer as strong in front of cloud providers. But the core remains domestic substitution; as long as domestic players unite, no one will touch it. China’s storage expansion is a significant part of the bearish argument.

Later: If Changxin and Changcun basically don’t sell overseas, will it affect the global market?

Liu Difan: The long‑term possibility can impact short‑term valuations. If that possibility exists, valuations may decline.

Later: ChatGPT and Claude have both driven demand and sustained bull markets with new products. What about personal assistant AIs like Meta Muse?

Liu Difan: No fundamental impact. According to Morgan Stanley’s report, even if Muse has 1 billion users, and each user only uses 25% of the free token quota per week, 3 GW of compute power would be sufficient, which further shows that that much capacity won’t be needed. Some segments (like Neocloud and certain hardware parts) have already shown signs of over‑building; it remains to be seen whether new developments will absorb them.

Later: In such a market, will investors focus more on short‑term trends?

Liu Difan: I’m also surprised that some fund managers who trained on A‑shares can be so “decisive.” They know to make as much money as possible when the market is booming and not to cling when it turns down. Those who haven’t trained on new‑energy vehicles don’t have that feeling.

Investing in the transformation of the liquor sector was previously quite difficult. If you only ever encounter the best business models at the start, you won’t feel the competition. Seeing poor products helps you recognize good ones, and you can exit quickly without attachment—being bold on both longs and shorts.

Later: Will the next period be a waste of time for investors?

Di‑fan Liu: To some extent, yes. But what I’ve learned over the past few years is that you must never exit the market. In my fund, 30% is allocated to TMT capital; whether I invest in semiconductors or applications, I can’t get out because the theme is still alive. I just have to let go of some of the excess gains.

The key is not to exit. I’ve seen Chinese fund managers who survived 2022 keep their teams intact and avoid drift—shifting from focusing on Chinese stocks to a global view, adding new skills, and scaling up assets under management. However, long‑term investing does not mean holding the same company forever.

Later: So much capital has been poured into compute—how much AI consumption is actually needed to support it?

Liú Dǐfán: Based on the $2 trillion already invested in infrastructure, the AI sector would need $37 trillion in revenue per year. (Compute depreciation is roughly $300 billion annually; factoring in a 12% pre‑tax return requirement on these investments, cloud providers would need $560 billion in pre‑tax profit each year. At a 50% profit margin, that translates to $1.12 trillion in revenue for cloud providers. Assuming cloud providers capture 30% of AI consumer revenue, the entire AI industry would need about $37 trillion in annual revenue.)

The above is the return requirement; now looking at the revenue side. Current forecasts for the 2030 AI market range from $4.5 trillion to $7.2 trillion. Using the low end, that’s only 1.2 times the $37 trillion needed, so the coverage is already very tight. If the cost of capital rises from 12% to 15%, the required revenue jumps to $42 trillion, which is almost a break‑even point.

Now the two leading model companies together generate only $160 billion in annual revenue, and the pace of compute capacity build‑out is out of sync with revenue growth. The key issue remains whether model revenue can sustain compute depreciation.

Later: Does the $800 billion‑plus compute commitment in Anthropic’s prospectus affect this assessment? (Added on Oct 28)

Liú Dǐfán: Anthropic’s cash and profit margin are better than imagined, and inference compute spending is lower than expected. At a $1.5 trillion valuation, it’s reasonable and even a bit cheap.

But this doesn’t seem to add any new information for either side. The current focus is whether to trust that cloud providers’ capex is a wise decision—it’s really a confidence issue.

When macro conditions are good, capital embraces risk. When conditions are poor, capital abandons risky assets. Fundamental numbers can only be explained when the macro environment is favorable. It’s like driving in strong wind—there’s resistance; in a tornado, any car will flip.

Later: Why do some fund managers avoid risk by buying cloud provider stocks?

Liú Dǐfán: Whether it’s OpenAI or our Kimi K3, they’re both middle layers, tenants. AWS is a mall; all the shelves, décor, HVAC, elevators are paid for by it. OpenAI is a popular flagship store in the mall—it can also procure some things itself, but it mainly relies on the mall.

If China can enter a field, the product moat disappears and the middle layer can’t make money. The profit will eventually flow to the cloud providers.

Later: The ability to monetize compute varies greatly among cloud providers.

Liu Difan: Credit quality is starting to stratify. The more concerning one is Neocloud; its financing cost is a major indicator.

Amazon’s H100 GPU can be rented for $7 to $12 per hour, while Neocloud can only rent for $4 to $5. Amazon’s customers are enterprises; Neocloud’s customers are some innovative companies, paying on demand with greater elasticity.

The core issue is financing cost. Amazon’s EBITDA margin is about 57%, so Amazon and Microsoft borrow at a cost of 5% to 7% (recently up to around 7%). CoreWeave’s margin is 59%, but it expands with high‑interest debt, and its free cash flow is negative $15 billion. Traditional cloud providers have an extra layer of software business, so when compute prices drop they still have software revenue to support them. Neocloud is a pure compute landlord; when compute prices fall, it’s exposed.

When selling compute power, credit quality and customer composition matter more than how many GW you have.

Later: Some Neocloud loans are backed by Nvidia—what impact does that have?

Liu Difan: Just look at its financing cost; that reflects the market’s conclusion about it. If the market is rising, Nvidia’s support could be why it’s going up; if the market is falling, that support matters less.

If the model layer consolidates into a few players, the surviving cloud providers, key hardware, security firms, and software with data moats can still thrive. The losers are the small Neoclouds that borrow money and leverage, and the lenders to those Neoclouds.

Later: Besides cloud providers, what are AI‑focused fund managers still buying?

Liú Dǐfán: Nowadays, whether in the US or China, anyone with even a modest long position has Intel as one of the core holdings. The logic is that the CPU‑to‑GPU ratio in data centers will rise in the future. CPU shortage is a consensus, but there’s disagreement on whether Intel is worth it. Bears think this is another high‑margin segment that will attract new supply; once supply comes in, the rally disappears.

Later: Intel fell almost 30% in one day—what does that indicate? (Updated November 9)

Liú Dǐfán: It shows how hard it is for buyers to track the numbers, because even management can’t forecast demand—this was mentioned earlier. This earnings “black hole” reflects two issues: first, management discovered that much of the PC demand was actually inventory replenishment; second, they can’t secure enough advanced substrates, so the server CPUs customers want can’t be shipped. Moreover, in this stage of scarce bullish ideas, many fund managers adopt a “lazy long” approach, buying high‑logic, high‑liquidity stocks and holding them, leading to crowded trades.

This 30% plunge, followed by another sharp drop after the Democrats’ big win, I guess halfreflects the fundamentals, and the other half is caused by a stampede. The company has given a terrible outlook, leaving many fund managers with no reason to hold on.

Later: Many data centers are being used as collateral for mortgages—are you worried a stock market crash could lead to a crisis like in 2008?

Liú Dìfán: Given the U.S. capital structure, a 20%-30% drop in semiconductors isn’t particularly surprising. 2008 was a banking problem, which is a bigger issue. The financing guarantees currently provided by banks haven’t reached a dangerous level yet.

But if this continues—everyone borrowing high‑interest debt at 7%-8% or even 10%—it will become a problem.

Later: What signal would indicate that the market is no longer just a bear market, but that the AI bubble is about to burst?

Liú Dǐfán: First, GPU rental prices are starting to decline.

Second, the spread between bonds issued by cloud providers and investment‑grade bonds is widening (indicating investors see higher default risk for borrowers). If the credit of cloud providers themselves starts to be doubted, that would be a major reversal.

Third, the growth of back‑logged orders for cloud providers stalls, or contracts are canceled or cut. Fourth, accounts receivable are rising. Nvidia’s days sales outstanding jumped from 45.4 days to 59.6 days, while its peers only moved 4–5 days. Does this mean Nvidia is granting customers longer payment terms? Its November earnings report is therefore crucial.

If these four indicators trigger simultaneously, it’s not a cyclical adjustment—it’s a reversal.

Later: Could it get any worse?

Liú Dǐfán: The core is that oil prices must not reach 150, and U.S. Treasury yields must not reach 6%; once they do, it’s a recession that no one can rescue. When there’s no recession, everyone thinks they’re right, but once a recession hits, demand disappears and a huge correction is inevitable.

Later: How does that compare with fiber optics in 2000?

Liú Dǐfán: In 2000, only 2% to 5% of fiber was lit, Lucent’s quarterly revenue fell 18%, and Corning shut down production lines. Today it’s the opposite: Micron can only meet 50% to 67% of demand, and equipment orders are full. During the 1996 storage crash, DRAM prices fell 78% in a year, and Micron’s gross margin dropped from 39.8% to 9.3%. The current situation is clearly the reverse.

Later: In April this year you told us that some fund managers in Hong Kong achieved a 50% return in half a year, “gradually developing a strong confidence.” In the July crash, did anyone manage to exit successfully?

Liu Difan: Many Asian fund managers were up more than 100% in the first half of the year, then fell 30% in July. Many of them had already liquidated by the end of June because the economics didn’t add up, but they couldn’t resist jumping back in in July.

Fund managers are professionals; their research is solid and backed by fundamentals, so they anchor on a valuation and the more they look at it, the more they feel they should buy.

Experienced and inexperienced people look the same in the first half; the losses come in the second half.

Later: What did they see when they sold in June? Why did they buy back later?

Liú Dǐfán: My partner went to Korea in June and saw that locals were already selling. I think this is the most direct; Koreans have experience with their own cycles and don’t need analysis.

At that time there were several signals: foreign investors had bought up all the channels that could purchase Samsung and Hynix, investment banks had exhausted their swap limits for Korean stocks. Hong Kong had single‑stock leveraged products for Hynix, and Korea had many as well, plus things like Situational Awareness, the white‑hair “stock god” on Twitter, all of which were worshipped to the extreme. All the profit‑making effects pushed the frenzy to the limit. Usually at this point a crash is only a matter of time. But at the time, there were also various reasons to believe a new era was beginning, and every dip was seen as a buying opportunity.

Buying back is actually unavoidable. They look at it, see that the valuation is low according to their own framework, and reflexively buy back. This year we’ve already made so much profit, the safety cushion is still thick.

A more seasoned fund manager can tell which segment of a stock’s rise is irrational, knowing that the money they’re making is actually non‑fundamental, sentiment‑driven, and not treating all of their gains as if fundamentals are at work.

Later: In AI, who sells at just the right time? Have you ever seen such a person?

Liu Difan: I think I have seen one, and he’s not even in AI. He trades cyclical stocks and also credit bonds, and reduced them in May. He has many investment options, doesn’t have to invest only in this.

First, regardless of fundamentals, AI occupies a fixed allocation in his fund—that’s discipline. Second, he sells when the target price is reached and doesn’t raise the target just because conditions improve.

Discipline sometimes goes against not only human nature but also your fundamental analysis: even if a company’s fundamentals improve, discipline forces you to sell regardless.

Later: After going through this round, what’s your biggest takeaway?

Liu Difan: I’m increasingly convinced that common sense is especially important.

This year’s first half achieving 100% returns is unsustainable. This is just a pattern I discovered as a allocator after reviewing many funds’ historical performance and life cycles. But from a fund manager’s perspective, perhaps the point of effort is to break history. My database shows that in the U.S., virtually no fund can achieve a 20% annualized return over ten years. When something deviates from the pattern, it will revert.

Later: In this environment, how can individual investors lose less?

Liu Difan: As far as I know, in May another batch of high‑net‑worth investors rushed in to buy storage stocks and added positions before the sharp drop. FOMO (fear of missing out) is a natural instinct.

Professional fund managers actually don’t study Buffett or Dalio; they study Druker‑Miller. These are core experiences repeatedly validated, and the focus isn’t deep fundamental digging. Druker’s greatest emphasis is on admitting mistakes and cutting losses—if the logic breaks, whether the loss is 1% or 20%, you must exit; don’t cling to predicting the endpoint, prioritize identifying turning points. He also stresses that the most dangerous phase in a bull market is when everyone thinks it’s easy and making money feels comfortable; comfort often means risk is accumulating.

Individual investors don’t fit the VC mantra of “believing in the power of belief,” especially in an era with no hype to chase. Be clear that relying on fundamental research to pick stocks has limits; someone who taught themselves storage for two months and then, after half a month of studying the Strait of Hormuz, jumps into oil trading is following the same logic.

Later: Actually, Buffett didn’t buy any compute‑power stocks either.

Liu Difan: That’s the crux. Many Chinese investors don’t buy what Buffett buys; they just borrow his slogan to give confidence to whatever they purchase.

Later: A lot of people will say this time is different.

Liu Difan: We still need to maintain reverence. Someone drew a chart saying Nvidia is the new Cisco; it sounds far‑fetched, but history repeats itself because it’s human nature. Back then, fiber optics were first scarce, then over‑built, and before the 2001 crash, no one thought it was a bubble.

What it teaches us is to train ourselves not to have “missing‑the‑boat” anxiety, and not to treat a guru’s single remark as gospel. You can hold See’s Candies for the long term because you understand its moat. Technology changes so fast—new models appear every day—that holding it long term is basically fooling yourself.

Human behavioral biases repeat. Cyclical stocks have their own characteristics; you can’t say that once you start studying them they stop being cyclical. When it comes to stock prices, you can’t discuss valuation only after the right‑hand side. You need to gradually learn to identify which portion of the money is irrational. If you don’t understand this, you may lose a lot.

Image source: Margin Call (2011)

– FIN –