If you allocate AI assets, or hold a pile of AI names and don’t know what ruler to measure them with — this folds the book’s judgment into an operating manual you could put in front of an IC. No company names (companies go stale; layers and mechanisms don’t). How to allocate, how to discount, where the alpha is, when to tear up the whole playbook.

First, compress the book into one memo-ready line:

Intelligence is deflating, revenue concentrates in the model layer, but profit settles at the two ends — upstream on scarcity, downstream on lock-in. So: own the scarce ends, rent the deflating middle; discount the “fluff” at each end; alpha comes from the market still pricing a new structure with an old ruler; time it by the dashboard, not the calendar.

Now the actions.

1. The Barbell: Own the Ends, Rent the Middle

Upstream end (buy scarcity, and the right kind). Buy “watts” over “silicon”: physical scarcity (power, transmission, transformers, high-voltage substations, interconnection capacity, siting rights) has no erosion mechanism and gives the longest duration; engineering scarcity (chip design) enjoys the present but carries ongoing erosion-monitoring cost; cyclical scarcity (HBM) is a swing trade — watch expansion announcements and exit. The most lag-priced corner is the best value: power and the physical layer are still priced at traditional-industrial P/E while carrying the hardest, longest-duration scarcity.

Downstream end (buy lock-in, and the right kind). Buy “owned lock-in” over “traffic” (attention depreciates in the agent era; routing power is the new aggregation point); buy “compliance-barrier moats” over “switching-cost moats” (the former is reinforced as regulation tightens, the latter faces zeroing by mandated interoperability); buy apps with multiple stacked thickness sources, and avoid any thin wrapper whose only edge is “better prompts.”

Middle (model layer: rent, don’t own). Value a model company by splitting it into “a downstream company + a cash-burning model division,” and pay only for the former. Never pay a duration premium for “model-capability leadership” itself — that’s an asset with a shelf life measured in months. The model layer is to use, not to hold.

2. Three Discounts: Marking Down the “Fluff” at Each End

Neither end of the barbell is a mindless buy — each has a “fluff” to strip out:

• Real-vs-echo discount (upstream): today’s upstream profit is laced with circular-financing echo demand. Split each upstream name’s revenue into “organic” and “loop” — the bluntest test: how much of its order book comes from customers it (directly or indirectly) invested in.
• Chokepoint-exposure discount (upstream): for the names most dependent on a single geopolitical chokepoint, apply a tail discount that’s unquantifiable but must be managed. Not computed by probability — managed by survivability: cap exposure, don’t bet odds.
• Interoperability discount (downstream): for “switching-cost” moats, assume mandated interoperability lands someday, ask “what’s left,” and value by what’s left — not by today’s switching cost.

Together, the three correct the barbell from “both ends are good” to “only the real part of each end is good.”

3. The Mispricing List: Where the Alpha Is

Passively tracking the barbell gets you beta. Alpha comes from a more specific source: the mispricings where the market still prices a new structure with an old framework. The book’s identified ones, as a long/short list:

1. Revenue map ≠ profit map: the market eyes the brightest revenue point (the model layer); profit actually sits with the sellers of chips, power, and subscriptions. The root mispricing — the rest are its variants.
2. SaaS multiple on agents: pricing agents that carry token cost and a reliability tax at zero-marginal-cost SaaS multiples, without discounting the dirty income statement.
3. Traditional-industrial P/E on power: the hardest, longest-duration scarcity priced like a legacy utility — but clear the valuation gate before going long (is AI narrative already priced in, do earnings accrue to shareholders).
4. Single global market on a cut-up upstream: controls have split compute into camps; the market still values it at one global number.
5. Echo demand as organic demand: proving demand by order volume is circular reasoning in a closed-loop market.
6. Capability leadership as a duration asset: paying a duration premium for a lead with a months-long shelf life is systematic overvaluation.
7. Fast-growing thin wrapper as high-growth: the fastest-growing, most beautifully demand-proving, thinnest-moat app is doing the giants’ free market research — re-label as high-risk.

The common structure: new structure + old ruler = mispricing = alpha.

4. Timing and Two Lines in the Sand

Time it by the dashboard, not the calendar. The earliest smoke alarm is the neocloud credit spread — the thermometer of the system’s most fragile link. The break has a fixed shape: edge defaults + core write-downs, infrastructure stays, wealth changes hands. That gives a clear move: in the panic, core assets (built on own cash flow, physically scarce) get thrown out with the bathwater — the highest-odds candidate buy of this cycle (provided: unlevered, real cash flow, scarcity intact, discount deep enough). The three discounts keep you from standing guard (overpaying for fluff); watching for the core mispricing keeps you from missing out.

Above the whole portfolio, hang two lines-in-the-sand gauges that tell you when to stop believing this book:

• Upstream line · the power supply curve: if “released” (nuclear restart, SMR commercialization, storage biting), upstream scarcity is diluted and the barbell’s upstream-end duration marks down wholesale.
• Downstream line · agent portability (tech + regulation, combined): only when all four sub-needles turn red — tool interoperability, context/memory portability, permission/workflow migration, regulatory mandate — and retention and pricing power fall accordingly, is lock-in punched through and the barbell’s downstream end marked down.

A responsible framework must carry its own “when am I wrong” trigger. These two gauges are it.

One last warning, aimed at the playbook itself: a smart framework, once public and widely adopted, has its excess return arbitraged away. If everyone runs the barbell, prices power at a premium, discounts thin wrappers, then both ends get bid up and the mispricings flatten. So the truly sustainable alpha isn’t in the static act of “running the barbell” (that quickly becomes beta) — it’s in seeing the next migration before consensus does: silicon→watts, subscription→rent, attention→routing, global→camps. This book gives you not a static portfolio but a migration map. Alpha belongs to whoever stood at the next position before each migration happened.

Your portfolio — is it allocated by the “revenue map” or the “profit map”? Comments open.

Sources (framework argument, June 2026): this chapter is the allocation form of the book’s judgments and adds no new data; the cited scarcity/lock-in/discount/dashboard mechanics appear across Chapters 1–14.

— From Chapter 15 of a book in progress, working title The Deflation Sandwich

#AI #Investing #DeflationSandwich