Owning the Bottlenecks

August 14, 2026

The obvious trade of the decade is to buy AI. The problem with the obvious trade is that AI’s product is the fastest-deflating good in modern economic history. You do not earn a lasting return on the thing whose price is collapsing. You earn it on whatever is still scarce once the collapse is finished.

So the investment question is not “will AI work?” It is: what is the last thing standing between AI and the outcome, and can I own it?

Four Scarcities, Four Speeds

Sort economic activity not by industry but by what actually blocks its automation, and it separates into four kinds of scarcity, each running on its own clock.

Scarcity	What it is	Clock	Investment character
Cognition	Code, analysis, documents, support	Months	Abundant and deflating. Own it as capped upside, never as collateral
Physical capacity	Power, grids, cooling, transformers, construction	Years	Genuinely scarce now. Self-correcting later
Permission	Licences, sign-off, liability, trust	Decades — with one fast hand	Valuable, and legally reversible
Position	Located land, entitlement, attention, defaults	Never	The terminal sink. Also the future tax target

The money does not accrue to the fast one. It accrues to whatever the fast one runs into. (The economics behind this — why cheapening one input shifts spending toward everything it can’t replace — is the subject of a companion piece; this essay assumes it and goes to the portfolio.)

Rents Move Up the Stack

Run the same sorting on AI’s own industry and the answer is uncomfortable for the consensus trade. Models are the cognition layer: commoditizing, price collapsing, open weights nipping at the frontier. Compute and power are physical: gated by turbines, transformers, and interconnection queues that no amount of intelligence shortens. The trust wrapper is permission — enterprises do not buy model weights, they buy somebody to sue, and indemnified, compliance-grade deployment is a legal product, not a technical one. Distribution and defaults are position: attention is zero-sum and the customer relationship is owned.

The durable margins in AI are not in models. They are in electrons, indemnities, and defaults. Which means the layer most people mean when they say “AI exposure” is the layer with the weakest claim on the eventual rents.

Every Bottleneck Is a Countdown

Here is the discipline that keeps this from becoming a bottleneck shopping list. Each scarcity contains the mechanism that ends it, and the mechanisms are different in kind:

Scarcity	Why it pays now	What ends it	On what timescale
Cognition	Novelty	The next model	Months
Physical capacity	Building is slow	Everyone builds	Years — its own success
Permission	The signature is required	Insurers price the machine below the human	Sudden, once actuarial
Position	It cannot be manufactured	Taxation, zoning, politics	Whenever politics arrives

Two consequences follow, and they are the whole strategy in miniature.

Buying a bottleneck is always simultaneously holding its reversal. Every position should be entered with its kill mechanism written down. If you cannot name what destroys the scarcity, you have not underwritten the position; you have bought a story.

The diversification that matters is across kill mechanisms, not sectors. A book spread across grid equipment, data-centre landlords, and AI-adjacent industrials is not diversified — it is one bet on continued capex wearing three costumes, and all three die on the same day. A book spread across a technical reversal, a supply reversal, a legal reversal, and a political reversal actually is.

And note which scarcity dies last. Position can only be ended by politics — which is exactly why the most durable asset in the framework is also the eventual object of taxation. The sophisticated money and the future taxman are converging on the same coordinates from opposite directions, and neither is saying so out loud.

Finance Loses Both Ways

The first phase of the buildout was paid for out of hyperscaler cash flow: capability risk sitting on diversified, daily-marked equity — the most benign arrangement possible. That phase is over. The marginal dollar now arrives as debt, through special-purpose vehicles, private credit funds, and insurance balance sheets, secured substantially against chips whose useful life is itself disputed — five to six years by the operators’ own depreciation schedules, two to three according to sceptics. The gap between those numbers is the entire question, and around 2028–29 it gets asked directly: does inference revenue cover debt service plus perpetual hardware replacement?

Now look at both tails.

If AI disappoints, that financing chain breaks: cash flow, collateral, and refinancing all fail together.

If AI triumphs, something else breaks: offices empty, generic software and services firms get stranded, older chips are obsoleted by newer chips, and broadly falling prices make fixed debts heavier in real terms against government balance sheets that are already stretched.

The financial system is safe only in the orderly middle. And the two tails are held by different institutions, each structurally unable to hold the other’s risk. Private credit funds and insurers own the disappointment tail, largely without knowing it, because the paper is marked rarely. Sovereigns, landlords, and legacy service equity own the success tail, held under the labels “duration,” “real estate,” and “stable cash flows.” Each population is the other’s natural hedge and nothing connects them. The unhedged straddle is not a market view; it is a coordination failure wearing a market structure.

Three portfolio consequences: own the equity of this complex, never its credit; hold real cash, because the workout is where the returns are; and buy tail protection only when its price confirms the mispricing. That last clause is load-bearing. “The tails are unhedged” is a reason to go looking. It is not a reason to pay any price, and an empty hedge sleeve with full liquidity is a perfectly valid end state.

Your First Portfolio Is Your Life

Before a single weight is set: you are already a portfolio, and probably a badly concentrated one.

Write down which class pays your salary. Write down your employer stock, vested and unvested. Write down your housing — tenure, metro, and how much of the property value is land. Look through your pension and funds for AI-infrastructure credit and concentrated tech beta. Note your jurisdiction, your currency, and your fixed obligations.

Then build the financial account to complete that picture rather than to express views in a vacuum.

A software engineer with vesting stock, renting in a supply-constrained city, is already maximally long frontier cognition — through both career and equity — and short the land inflating around her. Her book should hold less frontier optionality and more position exposure: precisely the opposite of the aggressive AI portfolio she is most likely to want.
A licensed physician with a paid-off house is already long permission and long local position. Hers should underweight both.

The generic AI-heavy portfolio is the wrong book for almost everyone drawn to it, because the people drawn to it are usually already long the trade through the way they earn money.

Two Scoreboards

The single most confident price prediction available is not a direction but a divergence: cognition collapsing in price while housing, healthcare, education, insurance, and credentials keep climbing — with headline inflation averaging the two into a placid, uninformative number.

That breaks the usual scoreboard. A portfolio can beat inflation by three points a year and still lose ground against the things money is actually for. So score twice:

Against the market — a passive global equity/short-bond composite. Quarterly. Am I compounding acceptably?
Against your own liabilities — expected spending on the constrained-access basket (housing, healthcare, education, insurance) versus the income and assets that hedge it. Annually. Am I keeping power over scarce necessities?

The trap worth naming in advance: in the most likely world — the grind, where AI improves, robots stay scarce, and permission holds — equities drift pleasantly upward while unavoidable costs compound faster. The first scoreboard reports a win. The second reports the truth.

The Survival Line Does the Sizing

“I can hold through a 30% drawdown” is a feeling until it is arithmetic. Turn it into arithmetic and it stops being a tolerance and becomes a constraint that sizes the whole book.

Take the correlated cluster: everything that falls together when the AI trade breaks. Assume it draws down 40–50%. Assume hedges deliver 50–70% of their modelled value after basis and timing slippage. Then solve for what fits.

Configuration	Cluster loss	Hedge payoff	Net	Inside 30%?
Cluster 70%, unhedged	−31%	—	−31%	No
Cluster 70%, hedged	−31%	+4%	−28%	Barely
Cluster 65%, hedged	−29%	+4%	−26%	Yes
Cluster 65%, hedges fail entirely	−29%	0	−29%	Barely

Three outputs, none of them a matter of taste: the correlated cluster caps around 65% of the book; protective cash floors around 12%; and — the row that does the real work — the book must barely survive even if the hedges pay nothing at all. Hedges are allowed to help at the margin. They are not allowed to be the reason the plan is coherent.

Then the test that actually matters, which is not about stomach: in the worst month, are you forced to sell the assets you most want to own at the bottom? With no leverage, distant option expiries, and protective cash untouched, the answer is no. That is what an aggressive posture with a hard survival line buys you.

The Book

Illustrative midpoints for an aggressive mandate with a 30% hold-through line and public instruments only. These are portfolio jobs, not sector quotas. A candidate belongs in a sleeve only when a material share of its economics actually performs that job.

Sleeve	Weight	What it is for
Broad productive ownership + operational adopters	34%	The compounding engine; also pays if the framework is wrong
Power, grid, and buildout capacity	17%	Harvest physical scarcity without lending against obsolete hardware
Land and position	11%	Hedge scarce living costs and participate in the rent sink
Trust and governance plumbing	6%	Own the infrastructure around permission, liability, and oversight
Frontier AI	6%	Capped insurance against capability arriving faster than expected
Cash: survival + opportunity	17%	Prevent forced selling and fund purchases during a financing bust
Tail hedges and shorts	7% gross	Protect against both AI disappointment and disruptive success
Residual	2%	Flows, fees, and incomplete allocations

The hedge line is gross exposure, not necessarily seven cents of net capital for every portfolio dollar; collateral, premium, borrow costs, and net exposure must be reported separately. More importantly, none of these weights comes before the investor’s existing exposures. Employer stock, index holdings, pension assets, housing, and career income all count.

Every candidate, in every sleeve, must answer six questions:

Purity: Is a material share of its cash flow actually tied to the portfolio job?
Valuation: How much of the scarcity or growth is already in the price?
Balance sheet: Can it survive a refinancing shock without issuing capital at the bottom?
Reversal: What technical, supply, legal, or political mechanism destroys the advantage?
Portfolio fit: Does it duplicate exposure already held through another sleeve, a job, or a home?
Observability: Is there a measurable condition that would prove the thesis wrong—and enough liquidity to act on it?

The rule beneath all six: do not buy a sector; buy a specific economic mechanism at a price, with its reversal attached.

1. Broad productive ownership and operational adopters — 34%

This is the foundation: ordinary productive businesses plus a smaller overlay of firms measurably converting cheap intelligence into better operations.

The core provides broad participation if the framework is wrong-shaped—if cheap cognition creates much more demand than expected, permission barriers erode quickly, or the economy adapts more smoothly than the scenarios assume. The adopter overlay seeks something more specific: companies that use AI as an input without selling a product whose price is being driven toward zero.

Where to look

Hunting ground	What adoption should improve
Industrial distribution	Inventory placement, procurement, quotation speed, route density
Logistics and transportation	Cost per shipment, warehouse throughput, maintenance, empty miles
Manufacturing	First-pass yield, scrap, downtime, changeover time, design cycles
Banks and payments	Fraud losses, servicing cost, onboarding time, manual review
Insurance operations and brokerage	Claims cycle time, underwriting support, documentation, expense ratios
Healthcare administration	Scheduling, coding, billing, denials, capacity utilization
Retail and consumer businesses	Forecasting, assortment, markdowns, fulfillment, customer support
Workflow and data businesses	Automation embedded in proprietary records and customer processes

The strongest candidates possess proprietary context: internal transaction histories, regulated records, physical operations, customer relationships, or systems of record that a generic model cannot reproduce. They consume commodity intelligence while retaining control of the scarce asset around it.

What a candidate must prove

Volume and margin improve together. Margin expansion from layoffs alone is not productivity; the company should produce or process more at lower unit cost.
Gross profit per employee rises over several periods, not merely after one restructuring quarter.
Cycle times, errors, or inventory needs fall without service quality deteriorating.
The gain is retained. If every saving is competed through to customers immediately, consumers win but the equity does not earn an exceptional return.
Reorganization is internally funded, rather than financed through repeated debt or equity issuance.
Valuation does not already assume permanent margin expansion.

A press release about an AI partnership is not evidence. Neither is a reduction in headcount without a corresponding increase in output.

The same-industry test

Almost every industry contains both an adopter and a victim. The distinction is often what gets billed:

A firm billing by the hour for routine analysis sells cognition; a firm owning the customer’s system of record may use cognition to deepen workflow control.
A healthcare payer automating claims may be an adopter; a vendor selling manual coding hours may be exposed.
A distributor improving warehouse throughput may benefit; a generic software intermediary with no data or distribution moat may not.
A media owner with scarce distribution may benefit from cheaper production; an agency selling production hours may see its product commoditized.

If the company sells undifferentiated cognition, it probably does not belong here. If it uses cognition to deliver something constrained by physics, law, data, or relationship, it may.

Kill condition: if productivity dispersion between leaders and laggards fails to persist for roughly three years, remove the adopter overlay and retain the broad core.

2. Power, grid, and buildout capacity — 17%

This sleeve owns the physical system that intelligence cannot summon into existence: electrical equipment, generation, transmission, cooling, entitled sites, and construction capacity.

It is strategic but rotating. Physical scarcity is real now; it is also the scarcity most likely to be destroyed by its own success as high returns attract new factories and infrastructure.

Where to look

Sub-sector	What is scarce
Electrical equipment	Large transformers, switchgear, breakers, substations, turbines and generators
Grid components and materials	High-voltage cable, conductors, connectors, insulation, specialized electrical steel
Cooling and thermal management	Chillers, heat exchangers, liquid-cooling systems, power-density management
Transmission and interconnection	Grid access, scarce queue positions, difficult-to-permit network capacity
Contracted generation	Firm power under long-dated agreements with creditworthy customers
Engineering and electrical construction	Skilled execution capacity, especially where projects are cost-plus or reimbursable
Entitled sites and reusable shells	Land with secured power, zoning, water, fiber, and deliverable interconnection
Selective utilities	Franchise value and rate-base growth where regulation still leaves a return for equity

The preference order generally runs from the less legible bottlenecks—components, equipment, entitlement and land basis—toward the headline utility trade. The most visible expression is often the one whose price already assumes the thesis and whose rents attract the most political scrutiny.

What a candidate must prove

The conservative-demand test: the investment should still work if projected AI power demand is cut materially, supported instead by electrification, replacement demand, grid hardening, or reshoring.
Firm backlog rather than promotional backlog: deposits, cancellation protection, and improving backlog margins matter more than the headline order figure.
Repricing rights: inflation escalators or cost pass-through. A fixed-price backlog can turn scarcity into losses.
Aftermarket or installed-base revenue: service and replacement demand that survives the construction cycle.
Fixed-rate, distant debt maturities, preferably beyond the financing-test window.
Limited customer concentration and no dependence on one developer completing every announced project.
Separable asset value: land, power contracts, and shells should retain value if the current compute generation becomes obsolete.
Political survivability: regulation cannot be expected to confiscate the entire scarcity rent.

What does not qualify

A conglomerate deriving only a token share of revenue from grid equipment.
A speculative data-centre developer whose economics require every planned megawatt to be leased.
A utility priced as if ratepayers will absorb unlimited construction costs without backlash.
A fixed-price engineering contractor carrying the inflation risk its customer refused.
Hardware lessors treating rapidly depreciating accelerators as durable collateral.
Merchant power exposure with no contracted floor, unless it is explicitly sized as a cyclical commodity position.

The same sub-sector can sit on opposite sides of the book. An entitled site with deliverable power, diversified tenants, and fixed-rate debt may qualify here. A speculative developer with unleased capacity and a near-term maturity wall may belong in the disappointment-tail short book.

Trim when lead times normalize, competitors announce aggressive capacity additions, order growth outruns plausible demand, or valuations still imply permanent scarcity after supply begins responding. The buildout sleeve’s sell signal is often the thesis working.

3. Land and position — 11%

This sleeve owns claims whose value comes from location, entitlement, access, network position, or an irreplaceable right. It is not a generic “real assets” allocation. A building can be rebuilt; a particular entitled site connected to a constrained grid cannot.

It has two jobs: hedge the household’s exposure to scarce necessities and participate in the tendency of productivity gains to capitalize into rent.

Where to look

Ground leases: a relatively direct cash-flow claim on land rather than on the building above it.
Land-rich operating companies: timber, farmland, urban landbanks, or industrial acreage where owned land is material to enterprise value.
Supply-constrained property: housing, logistics, or industrial sites where geography and entitlement—not temporary construction delays—limit supply.
Entitlement and infrastructure-adjacent land: sites near ports, transmission, fiber, transport chokepoints, or population centers.
Royalty interests: claims on geological or other fixed scarcities without assuming the full operating cost of extraction. These are X-like rather than purely positional and should be underwritten as such.
Distribution and attention platforms: marketplaces, defaults, or customer relationships with genuine network position and high switching costs.

What a candidate must prove

A high positional share of value: land, entitlement, distribution, or royalty value should matter more than replaceable structures or ordinary operating assets.
Current income where possible. A rent, royalty, or toll is preferable to relying entirely on the next buyer’s enthusiasm.
Supply is genuinely difficult to create, for legal, geographic, geological, or network reasons.
Leverage does not consume the scarcity rent.
Management cannot easily dilute the asset through poor capital allocation or repeated issuance.
An explicit political haircut has already been applied for property taxes, rent regulation, zoning reform, antitrust, or windfall taxes.
The valuation does not require both permanent scarcity and permanently low discount rates.

This sleeve must be netted against personal housing. A homeowner in a supply-constrained city may already hold a large, leveraged position in local land. A renter in the same city has the opposite exposure.

What does not qualify

Generic office property in a weak location.
Homebuilders, which are physical-capacity businesses rather than claims on land scarcity.
Property in regions where new supply can expand easily.
Commodity producers whose value depends primarily on operating execution rather than ownership of the scarce claim.
A platform whose users would leave immediately for an identical service at half the price.

Higher interest rates can crash the price level of land without reversing the trend in land’s share of income. The former can create an entry point; the latter would kill the thesis.

Gold is excluded from the base sleeve. If held at all, it is a separate capped diversifier. Land carries an enforceable cash-flow claim; monetary metal does not.

Kill condition: a sustained reversal in the underlying rent-share trend. Serious redistribution or land-tax legislation is a trim signal, not necessarily a surprise—the political reversal is embedded in the asset from the beginning.

4. Trust and governance plumbing — 6%

This sleeve does not primarily own today’s expensive licensed signatures. Those are incumbent permission rents and can collapse when law or insurance changes. It seeks the infrastructure required to deploy, insure, monitor, challenge, and govern automated decisions.

It is one of the strongest ideas conceptually and one of the weakest in public-market expression, so purity matters more here than anywhere else.

Three sub-lines

Sub-line	Possible activities	Pays when
Permission to automate	Liability brokerage, compliance-grade deployment, model validation, testing, certification, audit trails, data provenance	Private deployment expands—or regulation makes compliance more demanding
Institutional inference	Identity, eligibility systems, fraud detection, records modernization, secure data exchange, case management	Governments and large institutions automate their own administration
Citizen-side counterforce	Privacy, explanation, appeals, evidence preservation, algorithmic audit, adversarial-agent tools	Courts, scandals, or legislation demand contestability and due process

The second and third lines hedge the same political cycle from opposite directions: institutions expand their ability to infer and decide; citizens eventually demand tools to inspect and contest those decisions.

What a candidate must prove

Thesis-linked revenue is material today, not merely part of a total-addressable-market presentation.
Recurring or mandate-driven fees dominate one-off consulting projects.
Vendor neutrality: the rail should work across several model providers.
Workflow, record, or regulatory lock-in creates switching costs that do not depend on frontier-model leadership.
Production deployment rather than pilot activity drives revenue.
Liability is contractually bounded where possible.
Cash generation scales faster than headcount.
The business benefits under more than one gate state. It should be possible to name the cash flow if permission opens and the cash flow if it stays constrained.

The most important distinction is fee-earning rails versus risk-bearing balance sheets. A broker, audit platform, certification franchise, or identity network can collect fees while passing risk elsewhere. An insurer may simultaneously write correlated AI liability and hold AI-infrastructure private credit. That can place both sides of the capability straddle on one opaque balance sheet.

Risk-bearing insurers are not categorically excluded, but they require unusually clear disclosure of both underwriting exposure and investment assets. Until then, fee earners are the cleaner expression.

What does not qualify

Generic cybersecurity or consulting businesses where the thesis explains only a small fraction of revenue.
Manual compliance labor marketed as scalable infrastructure.
Incumbent professional firms priced as if the legal wedge will last forever.
Black-box decision providers whose contracts depend on practices likely to be invalidated by courts.
A company selected solely because regulation is increasing; regulation can create costs without creating attractive shareholder economics.

Kill condition: gate economics never translate into rail cash flow, or licensed services become cheaper at scale without any corresponding change in law, insurance, or institutional practice. In the latter case, the permission barrier was overstated.

5. Frontier AI — 6%

This sleeve is insurance against capability arriving faster and compounding more powerfully than the rest of the portfolio assumes. It is intentionally capped: a participant in the upside tail, not the book’s center of gravity.

Where to look, in rough preference order

Sub-sector	Investment logic	Principal risk
Semiconductor equipment and production tools	Sells to many racers rather than betting on one winner	Capex cyclicality and geopolitical concentration
Advanced packaging, memory, networking and optics	Increasing share of system cost; benefits across architectures	Supply response and product cycles
Financially strong cloud platforms	Distribution, infrastructure and non-AI cash flow can fund the race	Overbuilding and weak returns on capital
Foundry and advanced logic capacity	Genuine physical scarcity	Capital intensity and jurisdictional risk
Accelerator designers	Most direct capability participation	Architecture risk and consensus valuation
Small R&D-loop basket	Optionality on demonstrable self-improving research	High probability of permanent loss

The ordering favors businesses that sell to every racer over those that must remain the winning racer.

What a candidate must prove

Equity, not long-duration credit. The hardware and model advantage may melt faster than the debt matures.
Capex is internally fundable through a multi-year utilization shock.
The balance sheet can survive a closed financing window.
Paid production usage exists, rather than subsidized experimentation alone.
Revenue remains plausible as model and inference prices fall.
Customer concentration and circular financing are limited.
Valuation does not require permanent frontier leadership.
The business retains an advantage against open weights, through manufacturing, distribution, data, workflow, or customer ownership.

Apply the cap after look-through. Broad indexes already contain frontier mega-caps. Buildout suppliers carry second-order AI exposure. Employer stock may be the largest frontier holding of all. A nominal 6% sleeve can become 20% economically if those exposures are ignored.

Per-name positions should remain small enough that total loss is survivable. Long-dated calls can substitute for some equity where pricing is sensible, but only as explicit optionality—not as hidden leverage.

Kill condition—and it inverts the thesis rather than ending it: if the model layer sustains pricing power for several years despite open-weight parity, then rents did not migrate as expected. Re-admit the model layer instead of defending the underweight.

6. Cash: survival and opportunity — 17%

Cash has two separate jobs. Record them separately, or the protective reserve will quietly become a trading account.

Tranche	Approximate weight	Job
Protective liquidity	12%	Keep the book inside the drawdown limit even if hedges fail
Opportunity liquidity	5%	Buy durable assets from forced sellers during a financing bust

Protective liquidity

Appropriate instruments are short sovereign bills, high-quality government money-market vehicles, and short maturity ladders matched to expected liabilities.

The criteria are deliberately boring:

minimal credit risk;
minimal duration;
immediate or near-immediate liquidity;
currency matched to near-term obligations;
no dependence on selling another risk asset;
no reach for yield through private credit, structured notes, or gated funds.

Long-duration government bonds are not cash. Short corporate credit is not necessarily cash. A fund promising quarterly liquidity against illiquid assets is emphatically not cash.

Opportunity liquidity

This tranche is committed to a written dislocation list before the dislocation occurs. Possible uses include:

power contracts and interconnection rights beneath failed financing structures;
entitled land and reusable shells sold below separable asset value;
senior claims in workouts, where publicly accessible and conservatively structured;
high-quality operational adopters caught in indiscriminate selling;
buildout assets previously rejected only because their valuations were too rich.

Deployment still requires the physical scarcity to remain intact, financing terms to be conservative, and the asset to pass the same admission test used before the crisis.

Protective cash is not spent merely because markets are down. Opportunity cash is not a general budget for whatever happens to be down most.

7. Tail hedges and shorts — 7% gross

This sleeve is optional. It enters only where a specific structural risk exists, current market pricing offers acceptable value, and the rest of the portfolio does not already offset that risk.

There are three sub-books.

The disappointment tail

This pays if the AI buildout fails financially even though the technology may remain useful.

Possible targets include:

levered data-centre developers or operators;
hardware lessors relying on optimistic residual values;
businesses with near-term refinancing walls;
suppliers whose forecasts require every announced project to be completed;
lenders or listed financial vehicles with concentrated infrastructure exposure;
broader technology or semiconductor indexes where single-name hedges are unsuitable.

Possible instruments include puts, put spreads, defined-risk shorts, and accessible credit-protection vehicles. Direct protection on private special-purpose vehicles is generally unavailable, so listed proxies carry basis risk and should be sized accordingly.

Look for the combination of:

inference revenue falling behind debt service and replacement needs;
declining secondary hardware values;
aggressive depreciation assumptions;
customer concentration;
covenant weakening;
dependence on continuous refinancing;
protection that remains reasonably priced before the stress becomes obvious.

The disruptive-success tail

This pays if AI works so well that legacy business models are stranded.

Possible hunting grounds include:

generic IT outsourcing and business-process outsourcing;
staffing firms dependent on junior cognitive placements;
undifferentiated consulting and advisory services;
seat-priced software with weak workflow lock-in;
commodity content or information providers;
office property dependent on traditional knowledge-work occupancy.

Technological vulnerability is not enough. A viable short also needs:

a fragile valuation or capital structure;
limited ability to reposition;
weak proprietary data or distribution;
a catalyst inside the investment horizon;
liquid instruments and manageable borrow costs;
no hidden hard assets or takeover value overwhelming the thesis.

These positions are not recession hedges. Office property, for example, can rally when rates fall during an AI financing bust. The success-tail book may lose precisely when the disappointment-tail book pays.

Rates and trend

There is no permanent “AI means bonds rise” or “AI means bonds fall” position. The sleeve may use modest trend-following or convex structures, but directional duration waits until the evidence distinguishes falling inflation expectations and monetary easing from rising real yields and fiscal term premiums.

Hedge discipline

For every hedge, record:

the exact scenario and portfolio loss it offsets;
annual premium, borrow, dividend, and roll costs;
expected payoff after basis and timing slippage;
expiry relative to the risk window;
liquidity under stress;
the monetization rule;
the event that stops further spending.

The premium budget is zero up to a ceiling, not a quota. If protection is expensive, hold more cash or less of the underlying exposure. An empty hedge sleeve with full liquidity is a valid portfolio.

8. Residual — 2%

This absorbs settlement timing, fees, tax flows, incomplete allocations, and ordinary implementation friction. It is not a hidden speculative sleeve.

If no qualifying investment exists for a strategic allocation, the unused amount returns to broad ownership or cash. It does not go into an approximate substitute merely because the table contains a target.

Count Each Exposure Once

Sleeve labels can conceal duplication:

A cloud platform may be both an adopter and frontier exposure.
A data-centre landlord may contain physical capacity, positional land, and tenant-credit risk.
An insurer may appear to be an adopter and a trust asset while also holding AI-infrastructure credit.
An industrial conglomerate may contain a genuine equipment bottleneck inside a much larger unrelated business.

Assign every holding one primary portfolio job, then report its full exposure separately by class, geography, factor, liquidity, and reversal mechanism. The same dollar cannot count twice merely because it supports two narratives.

That final look-through is what distinguishes a portfolio from a collection of themes. Four sleeves that fail through four different mechanisms—technical, supply, legal, and political—provide genuine diversification. Four sleeves that all require continued AI capital spending are one trade wearing four labels.

What Public Markets Cost You

An honest defect, stated rather than buried: the two highest-conviction ideas here — entitled land and trust/indemnity rails — are the worst-served by listed instruments. Land arrives filtered through corporate structure, leverage, and management. Trust plumbing exists mostly as cyber, identity, and audit franchises carrying heavy unrelated beta.

The rule when the instrument doesn’t exist: shrink the sleeve; do not buy the costume. A wrong-shaped position is not partial credit — it is a different bet wearing the right label, and it will fail in a way the plan never modelled. The visible consequence is that weight shifts toward the sleeve public markets serve well (operational adopters) and away from the two ideas the analysis likes most. That distortion is real, it is priced as basis risk, and it is better named than concealed.

The Boring Half Is Deliberate

A third of the book is broad ownership and cash, and it is there partly because the thesis may be wrong-shaped — if cheap cognition unlocks far more new demand than it destroys, if the permission barrier turns out to have been imaginary, if the whole class structure simply overpredicts turbulence.

There is a sharper reason. When the parent framework audited its own factual claims against primary sources, six of seven errors ran in the direction of a more dramatic story: faster robots, bigger investment numbers, broader contagion, a wider prize. A framework with a documented bias toward drama should hold assets that pay in calm worlds. That is not hedged conviction. It is what lets the aggressive sleeves be aggressive: they never have to be right for the household to survive.

Rules of Play

Entry price is the strategy. Being right and being early are compatible states; the entry price is the only buffer between them. Every strategic addition is gated on how much of the scarcity is already capitalized.
No leverage, and no credit exposure to melting collateral. Ever.
Hedges go on before stress, on a schedule and a budget — a ceiling of roughly 1% of the portfolio per year in premium, with unspent budget treated as a decision rather than a failure. Buying protection at −20% spends the drawdown tolerance on overpriced insurance.
No standing bet on interest rates. In a fast-transition world two forces pull opposite ways: broad disinflation plus central-bank easing (bonds rally) against productivity repricing plus government borrowing (bonds fall). Which one wins depends on an unobservable choice about which price index the central bank decides to look at. “AI means deflation” and “AI means higher real rates” are both half- theories. Stay short duration; take convexity, not direction.
One signal is not a trade. Every adjustment must clear three gates: did the evidence change, is the candidate attractive at today’s price, and does it improve the exposure map. Correlated signals count as one observation, not three — a real financing bust fires the credit indicators, flattens the capability readings, and spikes the political ones simultaneously.
Never rebalance mid-cascade. Adjustments happen at quarter-ends and after recoveries.

When to Sell

Exit because the thesis died, not because the price moved. Wrong, not early.

Sleeve	It dies when
Operational adopters	The productivity gap between leaders and laggards fails to persist ~3 years
Buildout capacity	It underperforms in confirmed grind conditions — the scarcity was never worth what you paid
Trust plumbing	Gate economics never turn into cash flow — or licensed services get cheaper with no change in law or insurance
Frontier AI	The model layer sustains pricing power for 3+ years against open-weight parity — in which case invert the tilt and re-admit it
Land and position	The trend in land’s share of income reverses. Price drawdowns from rate rises explicitly do not count
Success-tail shorts	Junior white-collar hiring and wages recover while capability keeps advancing
Hedges	Market pricing eliminates the value — then you simply stop spending

The rotation has a peculiar signature worth internalizing: for the buildout sleeve, the sell signal is the thesis working. Scarcity in transformers and turbines is resolved by everyone building transformers and turbines. When robot economics and actual shipments confirm, that is when you begin trimming physical capacity and drifting toward position — because position is what survives the buildout’s own success, right up until politics arrives for it. And around year seven the book should start planning its own wind-down: harvest the temporary scarcities, review every concentrated position against its status as a tax target, let hedges expire rather than roll from habit.

Six Dials

The monitoring set, in plain language, each attached to an action:

AI liability insurance pricing versus human malpractice premiums. The earliest signal the permission barrier is cracking. Sustained below parity: buy the successor plumbing, sell the incumbent rents.
Hiring and quit rates, plus employment of 22–25-year-olds in exposed occupations. The signature is not layoffs but a frozen market.
Robots actually shipped, and cost per hour against roughly twice the median wage. Production numbers, never demos.
AI revenue against debt service plus replacement capex — the 2028–29 test. Deterioration raises cash; no new risk for a cycle.
Land’s share of national income and of house prices. The scarcity thermometer, and the exit signal for the position sleeve.
Government AI procurement and algorithm registries. Whether the state is automating itself faster than it permits the economy to be automated — which decides whether protected-sector employment stays a shock absorber or quietly stops being one.

The Likeliest Way This Fails

Not by being wrong about AI. By the wrapper. The right thesis bought through an impure proxy. Three sleeves that turn out to be one exposure. Hedge instruments unavailable exactly when wanted, or borrow vanishing in the squeeze. Fees and taxes eating an edge that was real on paper.

And the quiet one, which is the most probable of all: the maintenance lapses. The quarterly review slips, the triggers go stale, and the plan keeps reading as current while its assumptions rot — a worse epistemic state than being visibly wrong. If the cadence cannot be sustained, the correct simplification is fewer clever sleeves and more foundation, not a complicated book with dead switches.

The Proposition

This is not a bullish AI strategy and it is not a bearish one. It is long productive adaptation, selectively long scarce complements, conditionally long the plumbing of institutional trust, and short the durability of both today’s cognition rents and the financing structures behind them. It holds enough real cash to survive the financing bust and to be a buyer inside it. It pays for insurance openly, sizes its aggression against a stated survival line, and treats a job, a house, and a pension as the first portfolio.

The edge is not knowing which future arrives. It is owning claims that survive several, noticing when a scarcity has already been priced, and moving when the last unrelieved constraint moves.

This investment strategy sits on a scenario framework, summarized here and in full here, whose probabilities are stated subjective priors rather than measurements and whose own error register is worth reading before leaning on any number.

Nothing above is investment advice. It is one mandate — an aggressive posture, a 30% hold-through line, public instruments only — worked through to weights so that the reasoning is auditable. Change the mandate and the weights change; the four scarcities, the reversal mechanisms, and the survival arithmetic are the parts built to travel.