GearFlow Lab

Part 1 of this series broke the semiconductor down to its smallest unit. To understand the core semiconductor valuation models that Wall Street uses, we must realize that the entire $772 billion chip market splits into two distinct tribes. One tribe stores data; the other tribe thinks with it.

Get this split wrong and every headline about “memory price rebounds” or “the AI chip war” blurs into noise. Get it right, and you understand why $MU can swing from a 36.8% gross margin to a 74.9% gross margin in twelve months while $NVDA holds steady at roughly 75% quarter after quarter. Same semiconductor industry. Two entirely different businesses.

The 2025 Map: Where $772 Billion in Silicon Actually Sits

When thinking about semiconductor valuation, look at the scoreboard. Before picking a side, look at the scoreboard. Based on 2025 market-share data, the chip industry splits like this:

Logic alone — the “thinking” category that houses CPUs, GPUs, and NPUs — is bigger than the entire memory market by a wide margin. That asymmetry matters once you understand that Wall Street doesn’t price these two tribes on the same multiple, the same cycle logic, or the same risk model. It prices them as two different asset classes that happen to share the word “semiconductor.” Mastering this 2-tribe framework is the ultimate shortcut to accurate semiconductor valuation.

Memory: Working Space vs. Cold Storage

Memory’s entire job is holding data, nothing more. Split it into two physical roles and the category stops being abstract.

DRAM: The Desk Next to the Brain

DRAM is the memory sitting directly beside the CPU, holding whatever you’re actively running — your spreadsheet, your browser tabs, your messaging app, all open at once. Proximity to the processor is the entire value proposition: the CPU needs data instantly, and DRAM delivers it at near-zero latency. The catch is that DRAM is volatile. Kill the power and everything on it vanishes, the same way a cleared desk loses whatever wasn’t filed away before you left.

NAND: The Filing Cabinet That Survives a Power Cut

Your photos, documents, and installed apps live somewhere that doesn’t erase itself when the device powers down. That’s NAND flash — non-volatile memory, slower than DRAM but durable. Every time a phone spec sheet advertises “256GB” or “512GB,” that number describes NAND capacity, not DRAM.

Here’s the part that actually matters for your portfolio: a 16GB DRAM module from Samsung and a 16GB DRAM module from SK hynix perform almost identically. There’s no meaningful technical differentiation a buyer would pay a premium for. That makes memory a commodity in the literal economic sense — closer to wheat or crude oil than to software. Competitive advantage comes down to a single lever: who can manufacture more bits at a lower cost per bit.

Demand spikes, prices rip higher, and the entire sector prints money simultaneously. Everyone over-builds capacity in response, supply floods the market, and prices collapse just as violently. This is precisely what’s playing out at $MU right now.

Micron’s own fiscal Q2 2026 earnings release shows GAAP gross margin more than doubling year-over-year, from 36.8% to 74.9%, and climbing again to roughly 84.9% non-GAAP by fiscal Q3 2026 as AI-driven DRAM and HBM demand outran available supply.

That’s not operational genius compounding quarter after quarter — that’s a commodity price cycle hitting a supply wall. Investors who forget that this swing can reverse just as fast are the ones who get hurt when the next capacity wave lands.

Logic: The Four Brains Running Modern Compute

Logic chips take stored data and do something with it — add, subtract, render, decide. The category splits by specialization:

Unlike memory, logic chips are custom. An Nvidia GPU and an Intel CPU are not interchangeable — they’re built around different instruction sets, different software stacks, and different customer workloads. That single fact rewrites the entire competitive game.

Instead of racing to manufacture the most units at the lowest cost, logic companies compete on irreplaceable design IP. A developer who builds an AI training pipeline around Nvidia’s CUDA software stack faces real switching costs to leave — re-engineering, re-validating, retraining teams.

Wall Street has a name for this: the lock-in effect, and it’s the single biggest reason logic valuations behave nothing like memory valuations.

Semiconductor Valuation Framework: Ingredients vs. Chefs

Cooking is the fastest way to internalize this split. Memory is the ingredient. Logic is the chef.
A Michelin-starred chef (logic) with an empty refrigerator (no memory) cannot cook.

A refrigerator stacked with prime wagyu beef (memory) with nobody to cook it (no logic) is just expensive, rotting inventory. Each tribe is worthless without the other — but they are not worth the same amount, and they are not paid the same way.

The core distinction an investor needs: Memory competes on cost per unit produced. Logic competes on irreplaceability of design. One is a manufacturing race. The other is an IP moat.
AI has only tightened the bond between the two.

Training and running large models means feeding a GPU staggering volumes of data with zero tolerance for delay — which is exactly why chipmakers now stack memory (specifically High Bandwidth Memory, or HBM) directly next to the GPU die itself, rather than routing data across a motherboard. The chef stopped walking to the fridge. The fridge moved next to the stove.

Why the Money Differs:
Commodity Economics vs. Monopoly Economics

This is the section that should change how you screen semiconductor names.

Memory Trades on the Cycle, Not the Story

Because DRAM and NAND are standardized products, pricing power lives entirely in the supply-demand balance, not in brand or feature differentiation. When you study $MU, you are not underwriting a story — you are underwriting a cycle.

The current upswing is dramatic: Micron’s fiscal Q2 2026 DRAM revenue hit a record $18.8 billion, up 207% year-over-year and representing 79% of total revenue, driven by what management has bluntly called a structural supply shortfall against AI server demand.

Operating margin followed gross margin into record territory, hitting roughly 69% for the quarter. Those are extraordinary numbers — and that’s exactly the warning sign. Commodity businesses don’t sustain extraordinary margins indefinitely.

Capacity gets built in response to price signals, and price signals always overshoot in both directions. The right question for a memory investor isn’t “how high can margins go this quarter” — it’s “where are we in the build-out cycle, and who’s adding wafer capacity right now that lands in eighteen months.”

This produces a specific trap in P/E ratio structures that catches retail investors constantly: trailing earnings multiples on cyclical commodity stocks are most deceptive exactly when they look most attractive. At the peak of a pricing cycle, earnings spike, which mechanically compresses the trailing P/E even though the stock hasn’t gotten cheaper in any real sense — the “E” is inflated and unsustainable.

At the trough, the opposite happens: earnings collapse, the trailing P/E balloons, and the stock screens as “expensive” right when it’s structurally cheapest. Anyone running a simple P/E screen on $MU without adjusting for cycle position is reading the instrument backward.

Logic Trades on the Moat, Not the Quarter

Now contrast that with $NVDA. Nvidia’s fiscal Q1 2027 results (reported May 2026) show revenue of $81.6 billion, up 85% year-over-year, with Data Center revenue alone hitting $75.2 billion, up 92% year-over-year. The number that should actually grab your attention isn’t the growth rate — it’s the consistency of the margin underneath it.

GAAP gross margin landed at 74.9%, essentially identical to the prior quarter’s 75.0%, and Nvidia has now held gross margin in the low-to-mid 70s for multiple consecutive quarters even as revenue nearly doubled year-over-year.

That stability, sitting on top of hypergrowth, is the signature of a monopoly-style moat rather than a commodity cycle. Operating margin tells the same story, running north of 60%, a level commodity producers essentially never sustain.

This is also where P/E structure flips. Because logic earnings are viewed as durable rather than cyclical, the market is willing to underwrite forward multiples with more confidence — but it also means the bar for “cheap” sits much higher in absolute terms.

Recent data actually shows Nvidia trading at a forward P/E in the low-20s, roughly in line with — even slightly below — the broader semiconductor industry average. That’s a notable signal on its own: even with a structurally defensible monopoly position, the market is already pricing in eventual growth normalization. Monopoly status earns margin durability. It does not earn an unlimited multiple.

The Forgotten Middle Layer: $TSM and $ASML

A complete framework needs the two companies that don’t fit neatly into “memory” or “logic” at all, because they sit underneath both.

$TSM doesn’t design chips — it manufactures other companies’ logic designs, including Nvidia’s and AMD’s, as the world’s dominant pure-play foundry. TSMC’s Q1 2026 results show a 66.2% gross margin and a 58.1% operating margin, on the back of roughly 72% global foundry share and over 90% share at advanced process nodes (7-nanometer and below).

That’s not quite Nvidia-level monopoly pricing power, but it’s far closer to a monopoly than a commodity — what you’d call a scaled oligopoly. The catch is the cost of staying there: TSMC’s 2026 capital expenditure guidance sits at $52–56 billion, an outlay that effectively prices smaller foundry competitors out of the advanced-node race entirely. The moat here is bought, continuously, in cash.

$ASML is the purest monopoly in the entire chip supply chain. It is the sole global supplier of EUV lithography systems — the machines that physically etch the smallest features onto both logic and memory wafers. Every advanced chip from $NVDA, $AMD, $TSM, $MU, Samsung, and SK hynix passes through an ASML machine at some stage.

ASML’s Q1 2026 gross margin came in at 53.0%, with a backlog of roughly €38.8 billion locked in heading into the year. Notice that 53% sits well below Nvidia’s 75% — a reminder that even an outright technological monopoly doesn’t automatically command software-like margins when the product itself is a billion-dollar capital good sold to a small, powerful customer base that negotiates hard.

Quick Reference: The Two Investor Playbooks

The Wall Is Already Cracking: HBM

Everything above describes a clean separation that’s currently dissolving in real time. High Bandwidth Memory — DRAM stacked vertically and welded directly onto the GPU package — no longer behaves like commodity memory at all.

Micron has confirmed volume production of HBM4 for Nvidia’s next-generation Vera Rubin platform, with next-gen HBM4e ramping into 2027 and custom HBM already slated for Nvidia’s Feynman GPU in 2028.

When the ingredient gets engineered into a long-term, co-designed contract with the chef rather than sold on a spot market, it starts picking up the pricing characteristics — and the margin profile — of logic.

That single shift is arguably the most important structural change happening in semiconductors right now, and it’s why management at Micron keeps describing memory as a “strategic asset” rather than a commodity. We’ll dig into exactly why HBM breaks the old rules, and what it means for the next leg of the cycle, in Part 3 and Part 4.

For now, the framework to carry forward is simple: storage trades on the cycle, thinking trades on the moat, and the foundry and equipment layer underneath both is starting to behave like a moat of its own.

Before any of that physically exists, though, it starts as raw material — and that’s where this series goes next. Time to step onto the sand and walk through how a silicon wafer, semiconductor’s true mother material, actually becomes a chip.