# In the AI Boom, Who Is Actually Getting Stronger and Who Just Caught the Wave

> Personal reflections after listening to the 2026-09-03 episode of the Caibaogou podcast on the SEMICON show floor. From co-development between equipment makers and their customers, to who owns the heat problem, to facility lead times and local materials sourcing — a method for reading who is actually getting stronger during a boom. Educational content, not investment advice; no stock picks or price targets.

Published: 2026-09-04
Locale: en
Tags: semiconductors, SEMICON, supply chain, AI infrastructure, investing mindset


![A long aisle inside a semiconductor trade show, booth lights receding into depth, people talking in front of tool enclosures](/covers/caibaogou-2026-09-03-561-semicon-cover.png)

> Draw on the wisdom of the many, and no task is beyond you; draw on the strength of the many, and no obstacle stands.
> —— *Huainanzi, "The Ruler's Craft"* (Western Han, translation mine)

## What This Episode Is About

Episode 561 of the Caibaogou podcast, "A Quick Look at SEMICON" (2026-09-03), was recorded on a weeknight during the show, with the guest walking in more or less straight from the exhibition hall. So it isn't a polished industry report. It's three people emptying out what they saw that day — which booths were crowded, whose tools looked strange, what everyone was talking about.

I like unprocessed records like this. A report gives you the conclusion; the fragments from the floor show you how the conclusion grew.

One line stopped me. The guest described how the relationship between equipment vendors and their customers has changed, in the tone of someone reporting something dull. Underneath that dull sentence is a moat forming in real time.

Here's my summary, plus where my own thinking went.

## Key Points

**One — equipment makers went from selling machines to betting alongside customers.** The old flow was: customer writes a spec, vendor builds it, ships it, done. What the episode describes now is specs changing in under a year, forcing joint development, where vendor and customer decide together which direction to take. The investment meaning isn't in revenue, it's in the relationship. Once you're inside a customer's development process, replacing you costs more than the price of the tool.

![On the left a straight process that ends when it reaches the finish; on the right a closed loop that locks customer and equipment maker together and keeps coming back around](/figures/equipment-vendor-relationship-shape-en.svg)

**Two — heat stopped being a technical question and became a question of ownership.** The die has its thermal budget, the cold plate has its own, the rack has another, and facility cooling is yet another layer. Each layer used to solve its own problem. Now that per-system power keeps climbing, no single layer can absorb it, so every layer has to give up a slice. The guest laughed at how odd this sounds — your laptop runs hot, so you sit down with the air-conditioning company? But when the heat genuinely can't be removed, the customer says "go work it out among yourselves," and that's how alliances get forced into existence.

![Heat travels up through four stacked layers, each catching only part of it, with three dashed seams left between the layers](/figures/thermal-responsibility-four-layers-en.svg)

**Three — digital twins became necessary because of a labor shortage, not because the technology matured.** This one is counterintuitive. Teams used to converge by running physical experiments. Now there aren't enough people, there isn't enough time, and revisions come too fast, so the experiments move into models. And for a model to be right, everyone up and down the chain has to agree on what data they each hand over. So the simulation push and the co-development push are two faces of the same thing.

**Four — facilities are the tightest link right now.** Cleanrooms, piping, gases, power and water — the parts of semiconductor manufacturing that barely register in most discussions now determine when orders can be recognized. The guest's phrasing: if your lead time works, you get the business. Technical quality ranks behind delivery at this stage.

**Five — Taiwanese machinery and materials firms are getting their ticket in.** This year's floor featured automation, machinery and optical-component makers who weren't in semis before. Same on the materials side: consumables, chemicals, gases, intermediate chemistries now have local suppliers appearing. The reason isn't romantic — the market is hot enough that the Japanese and German incumbents are stretched, so the opening fell through. The guest added something worth keeping: this kind of work is pass or fail. It works or it doesn't. There's no partial credit.

**Six — the bar for startups went up, not down.** Specs change every six months. A small team has to burn cash chasing the moving spec while also building the product. So even startups are clustering into alliances. A hot market doesn't mean new entrants have an easy time — the opposite of what intuition suggests.

**Seven — watch durability, and remember capex is a lagging indicator.** The guest put it plainly: by the time you see capex being cut, the stock has already fallen a long way. The leading clues have to come from the industry itself — whether new technical themes are appearing, how fast specs are turning over, whether facility lead times are loosening.

One detail from the floor: three people in cleanroom suits were dancing at a booth, and the guest's reaction on walking in was simply "what is this supposed to be." There's also a shared understanding about live demos — a live demo will fail, so speakers inoculate themselves in advance by warning the audience, and if nothing breaks you thank the heavens. None of this shows up in a report, and all of it is what a trade show actually looks like.

## Going Further

### "This sector is booming — am I about to be the last one in?"

Probably the most common thought after hearing a show-floor recap. I've had it myself.

The way I handle it now is to separate "making money today" from "will win later." One exchange in the episode draws that line cleanly. The guest brings up cluster effects, joint development, switching costs — and the host replies that right now everyone is doing well, so you can do well without a competitive advantage at all.

That's worth writing down. A moat is built for bad weather. When every order book is full and shipping on time is enough to earn, the moat doesn't show up in the financials, because nothing is testing it. Compare two companies' profits in a period like this and you're mostly comparing who expanded capacity faster, not whose position is harder to take.

![Two profit curves overlap exactly while order books are full, and only separate into a higher and a lower one once capacity catches up](/figures/moat-shows-only-in-bad-weather-en.svg)

So what do you look at during a boom? My answer is: watch how relationships change, not how numbers change. Two companies both post a great quarter. One got pulled into a customer's next-generation development program and now helps decide where the spec goes. The other happened to have a lead time two months shorter than everyone else. When the tide goes out, the first is still on the customer's list; the second disappears the day someone else's capacity comes online.

![A large circle on the left holds one supplier while three others queue outside on the right, their arrows reaching the boundary without crossing it](/figures/inside-the-loop-vs-at-the-door-en.svg)

Concretely, I look for a few things: whether the company shows up in a customer's joint development program, consortium, or test-site roster; whether its product is written into someone else's specification; whether it's building shared simulation or data interfaces with its suppliers and customers. None of that converts into this quarter's earnings. All of it is evidence they'll be called again next round.

I should be honest that I've gotten this wrong. I once took a jump in gross margin as proof a company had won pricing power, and later found it was a temporary supply squeeze that reverted the moment capacity caught up. Tight supply and real bargaining power are different things that look almost identical in the numbers.

### "I trade off capex revisions — why am I always late?"

That line about capex cuts arriving after the decline is the most useful sentence in the episode for an individual investor.

The trouble with capex is that it's the output of a decision, not the process. For a company to cut next year's capex, it has already seen demand soften, held the meetings, run the numbers, and been to the board. By the time the figure is published, months have passed, and the market has been pricing whatever leaked out during those months.

![The share price line above starts falling at the far left, while the capex announcement is only the fourth point on the timeline below, joined by a dashed line to the low part of the curve](/figures/capex-is-a-lagging-signal-en.svg)

Where does it leak? The episode points at several entrances without organizing them. Here are the three questions I check on a schedule:

First, is the spec revision cadence still fast? Fast revisions mean demand is still pushing everyone forward. When revisions slow, direction has settled — which usually also means the urgency to expand has dropped.

Second, has the tightest bottleneck loosened? Here that's facilities and engineering labor. When "the money is ready, we just can't find people" starts to ease, it's either supply catching up or demand losing urgency. To tell them apart, watch lead times and pricing together: shorter lead times with firm pricing means supply caught up; shorter lead times with softening pricing means demand backed off.

![One phenomenon runs down into a test circle, then forks into two conclusions pointing in completely opposite directions](/figures/shorter-lead-time-two-readings-en.svg)

Third, is there a new theme? The softest of the three, but the guest's method is practical — see what technical topics are getting dense discussion at the conference sessions. This year he pointed at glass substrates and silicon photonics. When there's nothing new left to talk about on stage, that industry's narrative fuel is running low.

All three lead capex. All three are also easier to misread. So I don't use them to buy or sell — I use them to decide how often to check a position. When a signal lights up, I raise the review frequency. That's more conservative than cutting, and more active than doing nothing.

![A small hollow dot at each end of an axis, with the enlarged solid dot in the middle marking the option chosen](/figures/dial-the-check-frequency-en.svg)

### "This all sounds like good news — where's the risk?"

Short of people, short of materials, short of space, lead time as a shield — it does read like a clean bull case. I got wary somewhere in the second half and went back to look for where the risk sits.

Two places, I think. The first is the one above: a boom hides differences in quality, and when the new capacity starts competing, those differences surface all at once, by which point the price has already moved. The second is subtler. The blurred ownership of the heat problem is itself a systemic risk. When a problem takes four layers of suppliers each absorbing a slice, nobody is accountable for the end result. That structure is efficient when things go well and slow when they don't — and the time spent working out whose fault it is tends to be exactly where delays and deferred orders come from.

I wouldn't call that a bear case; the episode doesn't support a conclusion that strong. I treat it as something to watch. If I later hear that integration broke down somewhere in the chain, I'll know it isn't one company's execution problem — it's the structure paying its bill.

## Worth a Look

- Caibaogou podcast, episode 561, "A Quick Look at SEMICON," 2026-09-03
- Episode 542 in the same series covers inspection equipment, referenced here when 3D packaging inspection comes up — a natural follow-on
- To understand why tight supply isn't the same as pricing power, look at gross margin, long-term contract share, and price-increase history together; any one alone is easy to misread
- Trade show session agendas are public information, and the organizer's site usually lists the talk titles — the distribution of topics tells you where that year's center of gravity sits

## One Thing to Take With You

During a boom, performance can't tell you who is strong. What can is whose process a company got wired into while the boom was on.

This holds well outside semiconductors. When business is good, projects are plentiful and everyone is busy, everyone's output looks fine. The difference is that some people are merely absorbing this wave, and some have become a precondition for somebody else's next plan.

Here's something I tried; it takes under ten minutes. Pick one thing you've been doing smoothly lately — the kind of request that comes to you and you just handle it. Then answer one question: if I weren't here next week, who would they go to instead?

If a name comes to mind, you're replaceable on that task, and your good run comes from volume rather than position. If no name comes, or the name you wrote would need two weeks just to understand how you do it, then you're genuinely wired in.

![On the left a bypass runs around you; on the right the only road goes through you, and it has two gaps left in it](/figures/detour-or-no-detour-en.svg)

The first time I ran this, I wrote down three names, and I wrote them fast. It wasn't comfortable. It was better than finding out at low tide.