Lesson 25 · Chips & LLMs · Capstone finale (3 of 3)

Portfolio Synthesis

Twenty-four lessons gave you seven theses. This one gives you the method for holding them together — because the hardest mistake in this whole space isn't picking the wrong name, it's owning seven names that are secretly the same bet.

Builds on: the entire course + THESIS.md Skill: size a book by conviction × independence

You can now value each name from first principles (L24) and you've kept a falsifier for every one in THESIS.md. The finale is the step that turns a watch-list into a portfolio: deciding how much of each to hold. The naive answer — "size by conviction × upside" — is exactly the trap, because it ignores the one fact that dominates this sector: your seven names ride the same dollar of AI capex, so they don't fail independently. The missing term is independence, and you already have it — it's encoded in how unrelated each name's falsifier is from the others.

Core skill: Position size = conviction × falsifier-independence. Conviction you've built lesson by lesson. Independence is the new idea: two names whose falsifiers trip for the same reason are one position wearing two tickers; a name whose falsifier is genuinely orthogonal earns its own slot. The whole lesson is learning to read your THESIS.md falsifiers as a correlation map, then sizing off it.

Warm-up · Retrieve First
spacingL20·dollar-traceTHESIS
Before reading on: name the single underlying factor that TSMC, NVDA, AVGO, SK Hynix, ASML, AMD and Cadence/Synopsys all depend on. Then ask yourself — if that factor halved, how many of the seven would draw down at the same time?
The shared factor Hyperscaler AI capex — the ~$750B/yr that flows down the stack (your L20 "follow one dollar" trace). Every one of the seven is, at root, a differently-shaped claim on that same flow: ASML/EDA/TSMC tax it on the way in, NVDA/AVGO/AMD convert it into accelerators, SK Hynix feeds them memory. If AI capex air-pockets, all seven draw down together. That common factor is why holding all seven is not diversification — and why independence, not just conviction, has to drive sizing.
01 — The Sizing Identity

Why "Conviction × Upside" Is the Wrong Rule

Position size ∝ conviction × falsifier-independence conviction = how sure the thesis is right · independence = how unrelated its falsifier is to the rest of the book. The second term is the one investors who "understood the tech" still skip — and it's why their books blew up together.

Sizing by conviction × upside feels rigorous and quietly builds a time bomb: your highest-upside names (NVDA, AMD, SK Hynix) are also the ones most levered to the same capex factor, so a "diversified" book of your best ideas is really one giant position on AI spending continuing. The fix is to demand a second thing from every slot — that its falsifier is its own. A position only diversifies the book to the extent the thing that would break it is unrelated to what would break everything else.

The diversification illusion: seven tickers across "different" companies looks diversified and behaves like one bet. Correlation — not the count of names — determines how many independent bets you actually hold. You'll measure exactly this below.

02 — Correlation = the Independence Input

Read the Falsifiers as a Correlation Map

Independence isn't a vibe — you can derive it directly from the falsifier field you've maintained all along. Group the names by what would actually trip the falsifier. Names that trip on the same event are correlated; names that trip on unrelated events are the real diversifiers:

Capex-air-pocket cluster
NVDA · SK Hynix · TSMC
All trip if hyperscaler capex / GPU demand rolls over. Highly correlated — closest to pure capex beta.
Design-starts cluster
ASML · Cadence/Synopsys
Trip on a sustained drop in leading-edge bookings / design starts — same capex, but lagged & cushioned by installed-base service.
Share-shift cluster
AMD ↔ NVDA · AVGO
Partly internal: AMD's bull is NVDA's bear (ASIC/rack-scale share moves between them). A hedge within the leaders, not across the factor.
Orthogonal: power leg
Utilities · cooling · SMR
Trips on grid/energy economics, not capex mix. The one cluster that can hold up if the silicon thesis wobbles — your true diversifier.

Two insights fall out. First, the share-shift cluster (AMD vs NVDA) is mostly an internal hedge — sizing both is a bet on the segment, not new independence. Second, the only cluster whose falsifier is genuinely orthogonal to the silicon stack is the power leg from L20: it fails if grid interconnect, PPAs, or energy costs break — a different physical bottleneck. That's what makes it the portfolio's real diversifier rather than just another capex proxy.

03 — Tier-Specific Sizing Logic

Three Silicon Tiers + One Hedge, Each Sized Differently

TierNamesSize onCap the size because
toll boothASML, Cadence/Synopsys, TSMCDurability — highest, steadiest weight. Socket-agnostic; get paid whoever wins downstream.Shared capex-cycle risk (lagged); single-customer-base concentration.
leaderNVDA, AVGO, AMDConviction in share + margin-trend durability. The growth engine of the book.Competition falsifiers + the highest capex beta. Don't let "best idea" become the whole book.
cyclicalSK HynixCycle position, sized small. Own the HBM lever, not the memory cycle.The L24 double-count: peak earnings × peak multiple. Smallest, most cycle-timed slot.
power legUtilities / IPPs, cooling, SMRIndependence. Sized as the hedge — the slot that earns its keep by not being capex-beta.Its own risks (rate cases, project timelines), but uncorrelated to the silicon falsifiers.

Notice the logic flips per tier: toll booths are sized up for durability, leaders are sized to conviction but capped for correlated risk, the cyclical is sized small and cycle-aware, and the power leg is sized for the independence it adds — not its standalone upside. Same identity (conviction × independence), four different bindings.

04 — Build a Book, Then Stress It

Portfolio Allocator + Capex Stress Test

Set tier weights (auto-normalized to 100%). Watch how many independent bets you actually hold, then pull the AI-capex lever to see how the book draws down. Illustrative model, not advice.

Capex-factor loading
0.76
1.0 = fully exposed to the one factor
Effective independent bets
1.4
of 4 tiers held
Names you think you hold
7
vs. bets you actually hold →
Book drawdown on the shock: −23%

Model: effective bets = 1 ÷ (wᵀRw) with silicon-tier correlation ≈ 0.8 and power-leg correlation ≈ 0.2; drawdown = Σ(weight × tier-beta × shock), betas toll 0.5 / leader 1.1 / cyclical 1.6 / power 0.25. Numbers are stylized to teach the shape, not to forecast.

05 — A Worked Book

One Illustrative Allocation (Not Advice)

Teaching artifact only — not investment advice. The point isn't these specific weights; it's seeing the sizing identity produce a shape: durable toll booths anchor, leaders drive returns but are capped, the cyclical is a small cycle-timed lever, and the power leg buys real independence.

SlotTierIllustrative weightWhy this size
ASML / EDA / TSMCtoll~40% (split)Highest conviction × durability; socket-agnostic. Anchor of the book.
NVDA / AVGO / AMDleader~35% (split)Return engine, but capped — highest capex beta + live competition falsifiers. AMD/NVDA partly hedge each other.
SK Hynixcyclical~10%Own the HBM lever; kept small and cycle-aware to dodge the double-count.
Power legpower~15%The orthogonal hedge — the slot that lifts effective independent bets, sized for independence not upside.

Drop these weights into the calculator above (40 / 35 / 10 / 15) and you'll see the uncomfortable result: a book that looks like seven diversified names resolves to only ~1.4–1.6 effective bets, and the power leg is doing most of the diversifying work. That's the whole lesson in one number.

06 — The Quarterly Watch-List

One Falsifier Tripwire Per Name

A portfolio you can't monitor is a hope, not a thesis. The durable artifact this course produces is a quarterly watch-list: the single observable that would move each name toward its falsifier. These live in THESIS.md — the capstone just makes them operational. Check them every earnings season and ask: did anything move toward the falsifier?

NameTierThe ONE quarterly tripwire
TSMCtollCoWoS capacity stops being the binding constraint; a flagship leading-edge customer defects to Samsung/Intel.
ASMLtollNet new EUV + High-NA bookings roll over; a sustained down-cycle in the order book.
Cadence/SynopsystollRecurring-revenue % / renewals weaken; an open-source flow wins a production advanced-node tape-out.
NVIDIAleaderGross-margin downtrend (competition biting) or a qualified rack-scale NVSwitch rival adopted at a hyperscaler.
BroadcomleaderA major ASIC customer in-sources design; Ethernet scale-out share stalls vs InfiniBand/NVLink.
AMDleader2026 Helios/UALink rack qualification + ROCm production parity (bull-confirm) — or no inference share despite specs (bull-break).
SK HynixcyclicalSamsung wins full HBM4 NVIDIA qualification (3-way price war); HBM ASPs roll over; SSM/hybrid adoption cuts KV-cache HBM.
Power legpowerGrid-interconnect approvals + hyperscaler PPA volume — the leading indicator for whether ordered GPUs can actually be energized.
07 — Putting It Together

From Twenty-Four Lessons to One Book

Count bets, not names
Seven tickers can be one bet. Before adding a position, ask what would break it — if the answer rhymes with what would break the rest of the book, you're concentrating, not diversifying. The "effective independent bets" number is the honest scoreboard.
Independence is sized, not assumed
The power leg earns its weight by failing for a different reason (grid/energy), not by having the best standalone story. Pay up — in allocation — for orthogonal falsifiers; that's what actually cushions a capex air-pocket.
The falsifier is the position size
A name whose falsifier you can't articulate can't be sized — it's a belief, not a position. THESIS.md isn't a notebook beside the portfolio; its falsifier field is the independence input the sizing identity needs.
Monitor the tripwires, not the price
Conviction degrades silently. The watch-list converts each thesis into a quarterly yes/no: did the falsifier signal move? That's how durable judgment beats in-the-moment fluency — you change your mind on evidence, not on the tape.
Primary Source

Go Deeper

Read first: on holding correlated bets and why "number of positions" lies about diversification, the cleanest framing is the idea of the effective number of independent bets (the inverse-correlation intuition the calculator uses). Pair with: SemiAnalysis for the quarterly data behind each watch-list tripwire, and read the 10-Ks / earnings releases yourself (margins, bookings, ASPs) so your watch-list runs on primary signals, not headlines. See RESOURCES.md.

Comprehension Check

Quiz — 6 Questions

Select the best answer for each.

1. In the rule Position size = conviction × falsifier-independence, the second term captures:

How much upside the thesis offers if it works out
How unrelated this thesis's failure is to the others
How many analysts currently agree with the thesis
How large the company's total market value is now

2. Holding all seven semiconductor names is mostly one bet because they share:

The same chief executive officer across the whole group
Exposure to one factor: hyperscaler AI capital spending
Identical gross margins reported every single quarter
A legal requirement to move their stock prices together

3. The power leg (utilities / cooling / SMR) is the portfolio's real diversifier because:

Its falsifier trips on grid and energy, not on capex mix
It always delivers the highest standalone return of any slot
It has no falsifier and therefore cannot ever be wrong
It moves in perfect lockstep with the leader-tier names

4. "Effective independent bets" falling well below the number of tiers held means:

The book is more concentrated than the name count suggests
The book is safely diversified across unrelated outcomes now
The total expected return of the book has just increased
The positions are all equally weighted by definition here

5. Why is SK Hynix sized small and cycle-aware rather than to its upside?

Because memory companies are barred from large positions
To avoid paying peak multiple on peak-cycle memory earnings
Because it has no exposure to the AI capex factor at all
Since its falsifier is fully orthogonal to every other name

6. The quarterly watch-list turns each thesis into:

A precise forecast of next quarter's earnings-per-share value
An observable signal that says whether the falsifier moved
A fixed price target that should never be revised over time
A chart pattern used to time short-term entries and exits
From your instructor: The frame to keep — size = conviction × falsifier-independence; seven names on one capex factor is one bet; the power leg buys the only orthogonal falsifier; count effective bets, not tickers; and monitor each name's single tripwire quarterly. THESIS.md is now your portfolio: the falsifier field is the independence input, and the watch-list is the rebalancing trigger. That completes the original roadmap. Ask me anything — how to translate tier weights into actual position sizes, how to add a name outside the seven, or how I'd build the post-capstone retention checkpoint. And whenever you're ready, tell me which Checkpoint 2 & 3 cards felt shaky and I'll build the overdue spaced re-test.