Lining Up One Week of Public Information: Upstream Is Now Pricing in Lead Times

There was a man of Song who pulled at his sprouts because he was worried they were not growing. He went home exhausted and said to his family, “I am worn out today — I have been helping the sprouts to grow.” His son ran out to look, and the sprouts had withered. Few in the world refrain from helping their sprouts grow. Those who think it useless and abandon it are the ones who do not weed; those who help it grow are the ones who pull at the sprouts — this is not merely useless, it does harm.

— Mencius, Book II, Part A

What this piece is about

Over the past seven days I laid out every public post from one independent researcher who covers the upstream of AI infrastructure — optical, memory, passive components, power equipment — on a single timeline. Then I deliberately ignored what he bought or sold, and asked one question: if you treat all of this scattered public information as a single body of material, what structure does it point to?

A boundary first. Every number cited here comes from published financial results, remarks made at public events, and publicly available industry reporting. All source material was drawn from freely readable public posts; nothing behind a paywall, a subscription, or a private channel was used. This article deliberately names no tickers, recommends nothing, and sets no price targets. It is about how an industry is structured, not about what to buy.

I wrote it this way because the compounding insight in this week’s public information isn’t which link in the supply chain is short. It’s that the way scarcity gets priced has changed.

Four threads from this week

1. Upstream switched its unit of account from price to lead time

Line up a few of this week’s data points and the shape is remarkably tidy.

Multi-layer ceramic capacitors (MLCC): one Japanese supplier’s lead times ran roughly 24 weeks in June, about 30 weeks in July, and some parts reached around 36 weeks in August. Distributor shipment data put certain high-capacitance parts from a Korean supplier at roughly 40 weeks. Earlier this year the broadly reported figure was around 20 weeks. In half a year, the wait for the same component doubled.

Compound semiconductor substrates: indium phosphide (InP) substrate prices have been rising since the fourth quarter of last year, and this week brought news of a fourth consecutive increase, expected to exceed 10%. Epitaxial wafers made from those substrates have already been repriced twice and are heading toward a third. A supplier’s remark was quoted bluntly: even if you have the money, you may still not get the material.

NAND flash: a controller company’s chief executive stated publicly that 2027 capacity constraints will be more severe than 2026’s, because upstream suppliers need as much as four years from plant construction and equipment investment to actual production.

Three sub-industries, three different physical processes, one sentence: demand moves in quarters, supply moves in years.

This matters because it changes what “shortage” means. A price increase is a signal demand can absorb — pay more, get the goods. A lengthening lead time is not. It means money is no longer the allocation mechanism. Allocation shifts to who you are, when you signed, and where you sit on someone else’s customer list. When a market slides from price rationing to quota rationing, the price number itself stops being informative — and the lead time number starts.

Which is why the most notable item of the week, to me, wasn’t any price headline. It was the line that new capacity expansion is being pushed from Q4 2026 into 2027. In a market where lead times already run 40 weeks, a one-quarter delay stacks more time onto the far end. That is the Mencius parable rendered as capital expenditure: demand can double overnight, fabs cannot, and pulling at them kills them.

2. Crowding out: not a shortage, a change in the opportunity cost of the same line

The second thread is harder to see and easier to misread as plain scarcity.

In passive components, the reported pattern is that high-end AI server parts are squeezing capacity for general-purpose and consumer-grade parts, and that the spillover has moved from automotive into consumer markets. Over the same period, Japanese and Korean suppliers posted record monthly shipments.

The causality is not “consumer demand improved.” It is that the same equipment, the same engineers, the same raw material got redirected to the higher-margin part. Low-end parts didn’t get tight because they sold better. They got tight because they gave way.

Memory played out this exact shape a few years ago: advanced processes and high-bandwidth products absorbed capacity, and mature-node memory turned first on price. So when someone says the current tightness in passives “resembles the early phase of the NAND price cycle,” the analogy deserves serious attention — but what deserves attention is the similarity of the mechanism, not of the magnitude.

I’d split it like this:

• The mechanism transfers. Both are “high-end crowds out low-end,” not “total demand suddenly doubled.” That gives you a predictable transmission path — high-end first, then automotive, then consumer, with a lag between each.
• The magnitude does not. Memory is an oligopoly with highly standardised products and brutal capex thresholds, so when price turns it turns vertically. Passives have more suppliers, more fragmented specifications, and more room for design-around. The same mechanism need not produce the same violent curve.

So the right reaction to “this looks like memory” is not “therefore it rises as much.” It is “therefore I know which layer tightens next, and roughly how long the lag is.” The first is analogy abused; the second is analogy used.

The failure condition for this thread is clean: if high-end demand growth decelerates, the crowding-out unwinds, and the low end unwinds first. That means tracking consumer-grade lead times is really tracking the second derivative of AI server demand. It’s a mirror — and it reflects faster than an earnings report.

3. Selling out is not the same as collecting rent

This is the most valuable and most counter-intuitive thread of the week.

Within the same seven days you can read two coexisting sets of facts.

One is demand visibility. A laser and optical module supplier said at a public industry summit that it is sold out at least through the second half of next year and beyond. An executive elsewhere said the entire optical supply chain is facing major shortages that will persist for years. One laser company’s chief executive framed the imbalance as a three-to-five-year condition. One company went as far as saying that even combining the capacity of the two largest players, meeting customer demand over the next three years remains very difficult.

The other set is financial behaviour. That same “sold out” company drew criticism this week for successive equity raises in the hundreds of millions of dollars — dilution arriving fast and often for existing holders.

Put the two together and you get an uncomfortable but essential question: a company being sold out and a company making money are two different things.

My own test has three parts.

First, who pays for the expansion. If customers lock capacity with long-term agreements, prepayments, or direct equipment investment, that is rent. If you have to raise equity and build the line yourself just to earn the right to bid, that is a loan you extended — you carry the capital risk without necessarily holding the pricing power.

Second, whether you dare sign the long-term agreement. One detail this week was telling: a supplier said it had several long-term agreements available but was deliberately not signing, because it did not want its capacity blocked by one large customer. That carries a lot of information. In a seller’s market, a long-term contract is not necessarily a gift; it can be a shackle. Choosing to keep spot capacity and optionality is itself a display of bargaining power. Conversely, a company rushing to sign long-term deals for certainty probably needs that certainty.

Third, whether gross margin moves with lead time. Lead times stretching while margin stays flat usually means costs rose, not prices. Lead times stretching with margin expanding is pricing power actually landing. The remark that margins could exceed 40% once co-packaged optics ramps matters not because of the number but because it ties a product generation to a margin structure — a claim an income statement can falsify.

A company can be both a genuine industry bottleneck and a poor steward of shareholder capital. These are not contradictory, and historically they arrive together rather often.

4. The boundary between memory and optical is dissolving

The last thread only surfaced this week, but it has the longest time horizon.

The clue: a major memory maker disclosed a roadmap touching co-packaged optics, involving a photonic interposer linking optics to memory. Around it sit a few cross-checkable fragments — the company’s own site has referenced optical-to-memory links, it works with a chip design firm on custom memory solutions, and that design firm happens to own relevant optical interconnect assets.

I’m not going to guess at the supply chain. I care about the structural implication.

We are used to cutting AI infrastructure bottlenecks into separate buckets: compute, memory, optical interconnect, power. Tracking them separately is useful because their cycles differ — you can look at one bucket while another is cold. But if optical starts growing into memory, those buckets begin sharing the same physical constraints: the same lasers, the same photonic IC capacity, the same advanced packaging houses.

Two consequences run in opposite directions, and both need holding at once.

The expansion. The addressable market widens again. Components that only served network interconnect gain a “connect to memory” use case; continuous-wave lasers, already in short supply, pick up another structural demand source. That is why some expect another memory-style shortage in 2027–2028 — not memory running short again, but optics being dragged into shortage by memory.

The contraction. Diversification falls. When two previously independent chains start sharing a bottleneck layer, their cycles synchronise. For anyone who diversifies by holding several themes, this is worth writing down: you may believe you own two themes while actually owning one set of fabs.

Going further

You understand shortages fine — you just always arrive in the third stage

The first reaction many readers will have: I follow all of this, and I still buy at the end of the move.

That’s worth unpacking, because it isn’t a comprehension problem. It’s a question of which stage of information diffusion you subscribe to.

The same shortage shows up in three forms, in order.

Stage one is the observable quantity: lead time in weeks, number of price increases, days of inventory, utilisation. These sit in distributor data, trade-press newsletters, supplier price notices. They are unexciting and nobody shares them, because they read like a table.

Stage two is first-person testimony: earnings calls, summit Q&A, the letter a supplier sends its customers. Here the human voice arrives, and the quotable lines — but it is already a consequence of stage one.

Stage three is headlines and social momentum. Easiest to read, and last.

If you always feel late, the likely cause isn’t slow reflexes — it’s that your inputs are all subscribed at stage three. The fix isn’t refreshing faster; it’s moving one stage upstream. Pick three to five observable quantities you genuinely understand, and turn them into a table you update monthly instead of a stream of headlines. This week’s lead-time figures are useful precisely because they form a series: 24 weeks, 30 weeks, 36 weeks. The series itself is an indicator you can maintain yourself.

Related: this is why I keep attaching a month to every figure here. An industry number without a timestamp carries almost no information, because you cannot tell which part of the trend it belongs to.

How to tell whether you’re reasoning or agreeing

The second, harder-to-admit problem: after reading someone persuasive, I can’t tell whether the conclusion is mine or borrowed.

I use a crude method: write the failure conditions before you write the conclusion.

Try it on this week’s material. “Upstream tightness persists for years” — what would break it?

• Lead times stop lengthening, or start to shrink. Most direct, and observable.
• Expansion plans shift from “delayed” back to “pulled forward.” Supply tells on itself.
• High-end demand growth decelerates, crowding-out unwinds, low-end loosens first — the mirror from thread two.
• A specification-level alternative emerges that routes demand around the bottleneck — a technical path gets designed out.

Write those four and you’re reasoning. Fail to, and remember only “he said it’ll be short for years,” and you’re agreeing. The difference isn’t whether the conclusion is right. It’s whether you’ll know when it stops being right.

The same test applies in reverse. This week offered a good example: the same person can hold a position in a company and publicly criticise how it finances itself, while explicitly saying that what he holds shouldn’t affect your decisions. Separating “I like this industry structure” from “I approve of how this company treats shareholders” is a discipline most people can’t manage — because once you separate them, you have to admit part of your book is uncomfortable.

One bad example and one good one, both from this week

A note on sources, because this week produced a neat pair.

The bad example. The researcher publicly stated that fabricated screenshots attributed to him are circulating in volume — some as engagement-bait jokes, others maliciously inventing specific industry news, such as a fictional order win. He also said he has no power to stop the images spreading, and asked people to check primary sources.

This matters more than it looks. Fabrication targets the format you most want to believe: a short, specific line with a company name, a dollar figure, the texture of inside information. Nobody fabricates “lead times went from 30 to 36 weeks,” because nobody shares that. The more shareable the format, the more it deserves suspicion; the more boring the data, the closer it usually sits to stage one.

The good example. In the same week, asked about a biotech theme, he answered that it isn’t his domain and he can’t comment — that it’s hard for him to build conviction in themes or companies he doesn’t fully understand, and that this shows up whenever there’s a correction.

That line is worth more than any industry number this week. The real function of a circle of competence isn’t earning you more on the way up; it decides whether you can hold on the way down. A position built from your own reasoning and one built from someone else’s screenshot look identical while rising, and reveal entirely different shapes at minus 30%.

One more small item from the same week deserves reading alongside it: several government-level policy fact sheets were used as a map of which sectors policy is focused on — space transportation, drone components, critical minerals, polysilicon supply chains, and more. I think that’s a smart use, provided you keep the purpose straight: a policy document tells you the direction resources will flow, not who receives them. Treat it as a map and you’re right; treat it as a shopping list and you’re wrong.

---

Every piece of source material here came from public posts. The original wording, the full context, and any subsequent corrections live on Serenity’s own account on X — go read the originals. All I did was rearrange seven days of information, and rearranging distorts; the original doesn’t. Given that impersonated, fabricated content genuinely circulated this week, treat any specific industry claim carrying his name as unverified until you’ve checked it against his account directly.

The one thing to take away

Pick one upstream link you actually understand, find its lead time in weeks, put it in a table, and update it once a month.

Just that. No price forecast, no bottom-calling, no guessing who wins the order. Because in an industry where supply is measured in years and demand in quarters, lead time is the one number you can observe continuously without inside information, and it leads the income statement by months. When it stretches, it tells you tightness is moving downstream. When it contracts, it will tell you the cycle is over earlier than anyone’s opinion will.

Raw Markdown

This article is an educational discussion of investment method. It is not advice to buy or sell any individual security, offers no target prices, and does not analyze any current holding. Investing carries risk; make your own decisions or consult a qualified professional.