[Translation] The bulls are divided. What phase have we entered?

Before we begin: This is an English translation of a piece written by someone working in sales at a local sell-side firm in China.

I found it to be the most valuable piece I’ve read this month. Since I don’t speak Chinese, I used an LLM to help with the translation, so there may be errors.

---

From mid to late September, hardware quickly rolled over into a wave B decline. China and Korea went first, and more recently Japan and Taiwan have started correcting too. SOXX was the last. Views in the market are sharply divided. I'm not going to talk about the rebound. It doesn't mean much. What matters more is the cause: what exactly is driving the decline. Is AI now entering the dreaded midcycle collapse? There's little disagreement that the industry is still moving forward. But not every technological revolution is completed in one go. So what is the market really worried about? And what will decide the next inflection point?

Over the past week, I interviewed several investors. In long conversations, each of them walked me through a range of views. I've sorted and organized how they think and how they build their arguments. I hope this offers some good questions and perspectives. Personally, I think all of the questions below are quite interesting. I hope this helps.

1. This time China was the first to correct, and its correction has also been the deepest and the longest. So are we the smartest ones? What exactly caused the drop?

Setting flows aside, crowded positioning already caused one selloff back in July. Of course, you could argue that positioning in July was so crowded that a single drop couldn't clear all of it. But that isn't the main reason for this crash. The cause people point to most right now is that US long rates are too high. Funding structures and cost burdens at companies like CoreWeave have deteriorated sharply. A shortage of money looks like a problem the whole world shares. The July crash also started with abnormal moves in CDS.

The biggest problems AI faces right now come down to a few things.

1. CAPEX has started to require external financing just as US long rates are rising fast. The two have begun competing for capital. PTJ voiced this concern in an interview too: capital is being consumed too quickly around the world, while assets keep multiplying. The massive buybacks and abundant cash flow that have supported the market are now being challenged as well.

2. If large AI models can't secure a monopoly position, the pace of ARR growth is only half the problem. Without a monopoly, prices inevitably fall. But CAPEX and costs can't be cut quickly. The supply chain is still short of all sorts of items. So how do you close this gap between price and cost?

So with this combination, having enough money in the market is a necessary condition. Previously we had the Fed and fiscal policy working together, a "fiscal and monetary brew" in which both were easing at once. As PTJ said earlier, he had never seen monetary and fiscal policy ease at the same time with inflation at these levels. But since Warsh took office, the direction has changed.

One investor said everyone already knows AI's ROIC won't show up in the short run. US total factor productivity has fallen over the past three years. In other words, this concern has been there all along. People are looking for use cases beyond coding, but there's been no breakthrough worth getting excited about yet.

An industry advancing and its stocks crashing once in the short run don't seem to contradict each other. Each of the last several technological revolutions went through a collapse. The argument is that it takes one collapse to drive costs down sharply, and only then can the second cycle begin. There has never been a case that succeeded on the first try.

Right now the whole world is in "i" mode, and excluding AI, the US is half in recession as well. Home sales are very weak, and retail demand is slowing after the rise in rates and inflation. Consumer indicators are underwhelming across the board.

So isn't this quite similar to Nvidia (NV) and the memory sector in 2024?

ROIC wasn't there back then either. During the May to July correction, the market worried about whether CSP CAPEX was sustainable. Then cloud revenue came through. During the October to December correction, first, US long rates went up, and second, the market believed training compute demand was already saturated. Nobody yet realized how big inference demand would turn out to be.

The worries this time are similar. Neither US long rates nor CDS look healthy. The CSPs have also started raising external financing. Token prices are falling, but the gap with supply chain costs isn't closing quickly. The internet had zero marginal cost. This time it doesn't. On top of that, after the "three giants," money in the market has become even scarcer.

To this day, nobody can prove that AI's ROIC doesn't work. But nobody can confirm that it does, either.

2. Then how do you explain the strong guidance from both ASML and TSMC this quarter, if we've really reached the end in the near term?

One investor described US AI this way: however nicely you package it, it doesn't change the fact that leverage is high. It's similar to Chinese real estate. There's one layer of leverage when you buy the land, and another layer in running the company. AI also has multiple stacked layers, and each layer carries its own leverage and its own inventory.

One seasoned investor compared Sunac (融创) and Vanke (万科) with Dehua TB (兔宝宝) and Skshu Paint (三棵树). Sunac's share price peaked in January 2020, and Vanke's in February 2021. Dehua TB, by contrast, peaked in 2022, and Skshu in June 2021.

The point is that today's equivalents are CoreWeave, Nvidia, ASML, and TSMC.

This doesn't mean things have to collapse right now. It's just that the bullwhip effect takes time to show up. People keep asking, "Why aren't Taiwan and Japan falling?" Flows aside, it's because their whip is longer. A company's own leverage and which layer of the industry value chain it sits in also matter. The extreme divergence among the neoclouds says a lot in itself.

So with money in the market getting tight and faith in ROIC starting to waver, if funding costs at the long end rise as well and the LLM companies' own finances deteriorate, volatility could get far bigger than people think.

3. There are also some interesting counterarguments. Is the cost of long term funding a coincident indicator or a leading one? Are we only now finding out about the doubts over ROIC, or that total factor productivity isn't improving fast enough? In the end it's about a shortage of money, but does this particular shortage have to be tied to rates?

One client said that in my interviews I love asking about extreme scenarios: either a super boom or everyone goes down together. Embarrassing, but that might be true. Born contrarian, I suppose...

My take is that the period when everyone cheered even as US long rates rose is probably over. But look back at the six months before the dotcom bubble burst in 2000: the Dow kept falling while the Nasdaq could keep rising. It's really hard to call exactly where the top is.

It's true that there's less money in the market this time. But whatever exists has its reasons. Treasury yields are where they are, yet that doesn't seem to be affecting Anthropic's IPO. That means the market is accepting current rate levels. Bridgewater also explained that this time the link between the economy and rates has started to weaken, because most of the contribution to GDP growth is coming from AI spending.

If anything, crashes have always come alongside rate cuts. So high rates are clearly not a good thing, but you can't say high rates necessarily lead to a collapse.

If I had to give my own view: on the facts, Treasury yields have risen and the ARR data hasn't been great. Add to that the fact that, ahead of the midterms, arguments against AI can win a certain share of votes. I think these three things together created the high volatility we're seeing now. But to say a collapse is about to start from here, you'd need more evidence.

All you can say is that AI's ROIC is hard to deliver in the short run. You can't prove right now that it's impossible. We already went through this once with Nvidia in 2024. ROIC is just a trigger, not the core cause that ultimately explains what's happening. After all, even after the shift from training to inference, ROIC was never confirmed. There were just more use cases.

The dollar is also starting to show signs of turning lower. Lately, crypto has actually fallen less than gold. That may also reflect this shift. As for US long rates, last week's sharp drop did reflect recession expectations, but either way they're turning. At the very least, they've started to move away from extreme levels.

Stocks are a market where human psychology gets traded. If AI is the only thing in the world that can solve inflation and make factors of production more efficient, then isn't the only option to stay on this path to the end, even at rates this high?

Put differently, breaking this enormous inertia would take a massive external shock, and I haven't yet seen what a shock of that size would be.

Besides, as I said earlier, after years of pumping out so much monetary and fiscal stimulus, telling everyone to just hold cash now is a very unrealistic demand. So until more evidence comes in, I lean toward this being a phase of high volatility rather than a collapse.

That said, I don't think the decline is over either. Right now everyone expects the market to start falling after the midterms, but my sense is the opposite. The market falls first, and once the election is actually over, the correction wraps up as well.

Keep an eye on the charts, and stay humble before the market.

4. Since the midterms came up, will the final result have a big impact on the market?

I don't think the impact will be big. Even if both the House and the Senate go to the Democrats, it may amount to little more than further constraints on some especially unrealistic AIDC projects. Other taxes can't really be cut by much either. That has played out countless times throughout history. Environmental regulation will matter even less.

So my sense, if anything, is that the market will start to find its footing again after the midterms.

US earnings season starts next week, and we need to check how several of the big tech companies are doing. There's a line going around the market right now: "The index is at new highs, but the number of stocks going up is at its lowest." In a way, that's also a coincident indicator of high rates and going all in on AI. But it's hard to say a collapse is inevitable just because breadth is narrow.

The recent break in Treasury yields above the top of their range probably isn't entirely about oil either. AI financing competing with other demand for funds may be the bigger factor. Or maybe Bessent, who intervened so actively before, now intends to let the market move on its own for a while. If so, volatility naturally rises.

5. A superstition of sorts? Every technological revolution goes through a collapse once CAPEX passes 1.5 trillion? Though railroads went far beyond that level.

This is something a senior who has followed tech stocks for years once said half jokingly over dinner. Looking across several cycles, from optical fiber to IDC, once spending broke through roughly 1.3 to 1.5 trillion, there was always at least one collapse.

Of course, given how much time has passed, comparing absolute amounts isn't fair. On CAPEX/GDP, AI this time is roughly 2%, above the dotcom bubble but still short of the railroads.

Again, what decides life or death is not the size of the investment but the financing structure. As I said earlier, you can keep investing while living with high rates. High rates don't necessarily mean a collapse.

6. How much real progress is there outside of coding?

In AIDD (AI drug discovery), the ESMO congress takes place in Spain next week, October 23 to 27. Let's see what new progress comes out of it.

On the SaaS side, I've also caught some progress from podcasts and the news. Palantir's Q2 2026 revenue grew 93%, with US commercial up 149%. monday.com's AI product ARR doubled in a single quarter. Airbnb's customer support cost per booking fell 16% year over year, and AI resolved about 45% of customer complaints on its own. The measurable value Celonis customers have realized cumulatively has reached $10 billion.

From what I hear, enterprise use cases fall into three broad categories.

• Saving work hours. Heineken is one example: integrated management and allocation of its supply chain. Enterprise SaaS is actually already in place, so once AI features are added, it can be deployed on the ground quickly. That's also why PLTR's sales have been going so well lately.

• Expanding usable capacity. In one case, a medical center used AI to put more than 50 additional beds to use every month, significantly raising turnover.

• Plugging money leaks. Personal assistants like MUSE fit here, and this is the area with the strongest real demand right now.

So this is also an important thing to watch in Q3 earnings. If enterprise AI adoption progresses smoothly, it could ease some of the ROIC concerns.

There's one thing to watch, though. If workflow improvements through SaaS work very effectively, these services charge by the project, not by the token. In that case, where the profits accrue could shift from hardware to these outsourced project vendors, and corporate ROE could start to rise significantly.

That's also the view some investors hold. Breadth is this narrow right now because the time simply hasn't come yet. Once it starts becoming clear which companies are benefiting, ROE improvement across sectors of the real economy will become a certainty.

7. Finally, let's talk about the Chinese market.

This drop was truly unexpected, but in a way it was understandable. To be honest, though, only two or three of the people I interviewed saw a "reverse 9/24" coming.

There's little disagreement that the market will bounce next week. But whether it's a bounce or a trend reversal is a different question. Sentiment is so bad right now that 90% think it's just a bounce, and not a quick one. The view is that we won't get a crash followed by a surge like in July, and that the rebound won't take the index above its August high.

Looking at the charts, people feel this year looks a lot like 2015. Opinions are split on whether, after the inverted V, we'll get another rally like the one at the end of 2015.

But if you also factor in the rotation from hardware to applications, hardware is most likely no longer the best choice in terms of sector selection. Or you could say the disagreement is over whether the ChiNext bull market ends here.

But if everyone piles into application names, it may be hard for the index to make new highs.

8. Over the weekend, I reread every sales note I wrote in November and December 2025.

It brought up a flood of mixed feelings. My conclusion: "I can never know even what I don't know..." What I think now and what reality looks like a year later are completely different. These days, a single year is more than enough to play out the stories of both a past life and this one.

Back then, people were talking about the Fed's inevitable rate cuts, how far dollar weakness could go, Google emerging as the unexpected powerhouse, the high certainty of supply chain price increases and whether that was the "fish tail" at the very end of the rally, and the flow of Chinese deposits into the stock market. The talk was that once the market broke above 4,100, that shift of money would likely accelerate. Another focus was whether the neglected corners of China that nobody would even look at could turn around. Precious metals were going through the roof too.

A year later, it's a completely different world. The geopolitical and macro variables are the exact mirror image of what they were then. The mania of the final leg of the rally went beyond anyone's imagination. After breaking above 4,100, the index has been heading down ever since. Only the Big Four banks' YTD returns, dividends included, stand as solid as Mount Tai.

I'm thinking back to the last piece I wrote last year.

2024 was a year when China's various indices performed very well, but investors felt terrible. The stretch after the "Great Miracle Day" produced 90% of the excess returns, and for the year as a whole, only a tiny handful of stocks actually went up.

In 2025, the market felt much better. Whether you stuck with dividend assets on one side of the barbell or invested in the dazzling range of growth stories (new consumption, innovative drugs, AI hardware), making money at least wasn't that hard, because you could catch one of them.

Last year I also asked myself: going forward, is it momentum or mean reversion? I was at a loss for words for a while too.

2026 is now down to its last two months. Looking forward to this year's final piece...