Why Semiconductor Research Got Repriced (It Wasn't "AI Made Chips Sexy")

The story everyone tells is "AI made semiconductors sexy." That story is wrong about causation. The sector didn't get more glamorous. The people reading about it started making much bigger decisions.

Here's the answer up front, so you can decide whether to keep reading. Independent semiconductor research got repriced from a niche into a premium category because the marginal buyer changed, not because the topic changed. The words on the page are roughly the same as they were in 2019. What changed is the size of the decision the reader makes off those words. When the decision goes from thousands to eight figures, price follows. That's the mechanism. Everything below is just working it out one step at a time.

Editorial note: This analysis draws on a public conversation — Dylan Patel of SemiAnalysis speaking on the Sequoia Capital podcast, in the episode "Why Hardware-Software Co-Design Is AI's Real 100x." The phrases quoted below ("semis were not sexy in the West," "you went very long") are from that interview. LensUp.ai has no relationship with SemiAnalysis or Sequoia; we're just decomposing what was said.

The dumb question worth asking

Why would the same information be worth twenty times more than it was five years ago?

The information didn't get better that fast. So it has to be something about who's paying for it. And once you look there, the answer is obvious in retrospect.

In 2019, the person buying chip research was a long-only portfolio manager. His decision was position sizing on a few large-cap names. His error tolerance was wide — he could be a little wrong and it barely mattered. He paid two or three thousand dollars a seat, mostly as insurance against looking stupid in a meeting. That's what research was worth to him: the price of not being caught flat-footed.

The buyer swapped out from under the product

Now look at who reads it today.

One is a hyperscaler capacity planner deciding whether to commit billions of dollars of packaging allocation twelve to eighteen months forward. Another is a pod at a multi-manager fund sizing a half-billion-dollar single-name book, where one supply-chain data point moves the mark before the earnings call.

Same page. But the reader's decision is now eight figures wide. If a research note shaves even a little uncertainty off a $10B commitment, it's trivially worth six figures. So willingness to pay jumps twenty to a hundred times — for information that barely changed.

Put it as a formula you can carry: the price of research is the reader's decision size times how much the information moves that decision. AI didn't make chips important. It made the decisions built on chip data enormous. That's the causal engine. "Sexy" is just what people call a P&L consequence after the fact. In the Sequoia conversation, Patel frames the sector's old reputation plainly — semis "were not sexy in the West" — which is exactly the thing this formula explains away.

Demand alone doesn't explain who wins

If it were only about big decisions, every Substack writer with a chip newsletter would be rich. They're not. So there's a supply side, and it's where the real moat lives.

The defensible thing is a bottom-up cost, yield, and capacity model that a normal Wall Street analyst structurally can't rebuild. Here's the difference, and it matters.

A sell-side analyst models a chipmaker's revenue as average selling price times units, calibrated off company guidance. That model is derivative of what the company already told the street. By construction, it can't front-run the disclosure — it's downstream of it.

A real cost model is built from primitives instead:

• Mask-layer count and litho step count per process node
• Defect-density curves — how many killer defects per unit area at a given node
• Reticle size and how large a die you can print
• The gap between wafer yield and die yield

That last one is the whole game, so it's worth slowing down.

Wafer yield is not die yield

Wafer yield is the fraction of the silicon area that comes out good. Die yield is the fraction of individual chips that come out good. On a big chip — a reticle-limited GPU die — those two numbers pull apart hard. A single killer defect kills the whole big chip, so a small increase in defect density hits a large die much worse than a small one. A tiny shift in defect density can swing your cost per good die dramatically.

Stack packaging on top of that: how much advanced-packaging interposer capacity exists per quarter, what the stacking yield on high-bandwidth memory looks like, what packaging costs per GPU. Now you can estimate allocation and pricing before the earnings call confirms them.

That's the alpha. And that's what Patel means in the Sequoia interview when he says "you went very long." It isn't patience on a hot theme. It's having spent years assembling the supply-chain graph while it was unglamorous, cheap, and ignored — so that when the decisions got big, you already had the only model that answers them.

What the "$100M revenue" rumor actually means

There's a widely-repeated rumor that SemiAnalysis passed nine figures in revenue. Treat it as a rumor — it hasn't been confirmed. The interesting part isn't the number; it's that most people reading it assume it's subscription ARR.

It can't be. A newsletter at roughly a thousand dollars a year would need on the order of 100,000 paying subscribers to hit that — implausible in a niche this technical. So the number, if real, is composed differently:

• Newsletter tier (~$1–2k): a loss-leader at the top of the funnel. Its job is building the audience and the citation surface, not the revenue.
• Enterprise seats ($50k–500k+): for corporates and funds that need the model, not the opinion.
• Bespoke consulting and data licenses: teardown data, wafer-cost models, supply-chain scheduling feeds — sold per engagement. This is the bulk of the tail.
• Events, plus a rumored fund (also unconfirmed).

So the recurring subscription revenue is probably the minority; consulting, data, and events the majority. Which means the margin structure looks nothing like software. Anyone saying "$100M ARR" is wrong about the composition — and the composition is the whole point.

The citation flywheel

A model in a drawer is a private hedge-fund edge, not a business. The trick that turns it into one is distribution — specifically, becoming a citable primary source.

A proprietary data layer — physical teardowns, fab-equipment import scrapes, customs and shipping data — converts opinion into evidence someone else has to quote. Once your capacity estimate gets cited on an earnings call and picked up by journalists, you stop being a commentator and become infrastructure. And that loop compounds: citation builds credibility, credibility gets you cited more, and each cite is a top-of-funnel event that upsells into the enterprise and consulting tail.

So there are two halves, and neither works alone. The cost model is the supply half. The citation flywheel is the distribution half. A model with no distribution is a private edge; distribution with no model is just another newsletter.

Headcount is a tell, not a flex

The team is reportedly several dozen people. People read that as a brag. Read it as an admission instead.

Primary research is labor-in, not code-in. Every teardown, every import-data scrape, every yield estimate is human hours. There's no version where a script does it overnight. That headcount is simultaneously the barrier to entry — nobody spins this up in a weekend — and the margin ceiling. It's a consulting firm wearing a media brand's clothes. Model it as SaaS and you'll be wrong about everything downstream.

Cyclical revenue, structural moat — don't mix them up

Here's the honest tension. A lot of the current revenue is levered to one thing: the AI-capex growth window. When hyperscaler spending decelerates, the eight-figure decisions get smaller, and the consulting tail compresses with them.

But notice what doesn't compress. The cost and capacity model still works when capacity is loose, not just when it's tight — supply constraints exist in every phase of the cycle, up and down. What shrinks in a downturn is willingness to pay, not the moat.

So split it cleanly: the revenue is cyclical; the category position is structural. The organization built durable infrastructure inside a cyclical window. Don't conflate the two, or you'll misjudge both the risk and the durability.

The one line that outs an outsider

If you want a single tell for whether someone actually understands this sector: watch whether they quote transistor density or cost per transistor at yield.

Headline density numbers come from press-release slides. They're marketing. The number that matters is what a working transistor actually costs once you account for defect-density-adjusted die yield — and that number is never on the slide. Anyone reciting the roadmap density figures as if they were research is telling you they've never built the model.

FAQ

Why did semiconductors "become important again"?

They didn't stop being important. The decisions built on chip data got big enough that people would finally pay serious money to get them right. Attention followed the money, not the other way around.

Is SemiAnalysis really doing $100M in revenue?

It's an unconfirmed rumor. Even if true, it's almost certainly mostly consulting, data licenses, and events — not subscription ARR, which is the funnel.

Can someone just start a competing chip newsletter?

The writing, yes. The moat, no. The moat is a labor-intensive bottom-up cost model plus a citation flywheel that takes years to build. That's why headcount, not code, is the barrier.

Do these three things tonight

• Test one company's cost model. Pick a chipmaker whose stock you follow — say Nvidia or TSMC — and ask: is your view derived from company guidance (ASP × units), or from primitives (mask layers, defect density, die yield)? If it's the former, you're downstream of the disclosure, not ahead of it. Write down which one it is.
• Decompose the last "$100M" claim you read. Split any research-business revenue rumor into subscription vs. consulting vs. data vs. events. The exercise alone will change how you read every one after it.
• Check one number you rely on. Is it a headline figure or a yield-adjusted one? Transistor density vs. cost per transistor at yield is the fastest example. If you can't tell which you're looking at, that's the gap to close first — and it's usually one Google search away.

One open question worth sitting with: if the revenue is cyclical but the moat is structural, what does this business look like on the far side of an AI-capex slowdown — smaller, or just cheaper to buy from? What's your guess?