The Premise

Intelligence too cheap to meter means the marginal cost of analysis, code, design, synthesis, and coordination falls toward the cost of inference.

The first-order thesis is obvious: compute and models capture the first wave of attention. Obvious trades get crowded. The better question is where scarcity moves after cognition gets cheap.

My answer: scarcity moves from thinking to permissioned action. Models can generate work; institutions still need to verify it, route it through distribution, attach trust to it, and power the physical infrastructure behind it.

COGNITION PRICE analysis / code / synthesis falls toward inference NEW BOTTLENECKS verification distribution energy + atoms regulatory trust capital access when thinking gets cheap, ownership moves to the bottlenecks that remain scarce

The AGI trade is not one trade. It is a migration of scarcity.

The Framework

I would underwrite AGI exposure with three questions from the mental-models page: what is the terminal valuation, why is the market not there yet, and what closes the gap?

The market is eventually efficient, not immediately efficient. AGI compresses that delay in domains where public information can be digested by machines, but it leaves gaps where the edge is private, physical, regulated, or operational.

The mistake is treating AGI as a uniform software deflator. It is deflationary for generic cognitive labor and inflationary for the scarce things that let that labor become shipped, trusted work.

Predictions Worth Underwriting

  1. 01
    Software splits. Seat-tax SaaS loses pricing power when agents do the work that justified the seat. Workflow owners with distribution, proprietary data, embedded permissions, and trusted execution gain share.
  2. 02
    Verification becomes the moat. Generation commoditizes before judgment does. In underwriting, drug discovery, chip design, compliance, and aviation, the winner is the organization that knows when an agent is safe enough to ship.
  3. 03
    Capability without distribution commoditizes. Distribution without capability decays. The best application companies own both: a route to the customer and enough model leverage to collapse the cost of serving that customer.
  4. 04
    Disease prevention is underpriced. If biology becomes more simulatable, value migrates upstream from chronic treatment to diagnostics, prevention, trial design, lab automation, and data-rich platforms that can close the loop.
  5. 05
    Public-market alpha thins. As machine analysis gets universal, edge migrates to private data, operating control, regulatory position, power contracts, and being early to compute itself.

Rates, Commodities, Real Estate

The macro path is not simply "AI is deflationary." It is software deflation colliding with physical capex inflation.

Short rates can fall if AGI weakens labor demand and pushes central banks toward insurance cuts. Long rates can rise first if the capex cycle absorbs savings, stresses grids, increases energy demand, and raises the real-growth ceiling. The curve can steepen before it flattens.

now later relative pressure short rates: cuts after labor shock long rates: capex + term premium first power / copper / cooling real estate: office down, powered land up AGI can be disinflationary in software and inflationary in atoms

The non-consensus macro view is a split path, not a single deflation story.

Commodities become the shadow price of intelligence. Compute turns electricity, land, water, transformers, copper, gas, nuclear, and cooling into cognition. The "AI commodity trade" is not glamour; it is the bill of materials for agency.

Real estate bifurcates. Generic office loses the knowledge-work scarcity premium. Data centers, powered land, labs, advanced manufacturing, and places where trust still requires physical presence get bid.

The Application Winner

The naive application thesis is "AI app wrappers win." Most wrappers will be competed into margin compression because capabilities diffuse.

The stronger thesis is distribution plus capability plus verification. Own demand, collapse service cost, and certify output. That is why the most interesting winners may look boring: insurers, banks, labs, logistics networks, payroll systems, industrial software, and medical platforms with proprietary feedback loops.

distribution owned → model leverage + verification → cool demo competed away compounder demand + agency + trust outsourced feature thin margin legacy channel decays unless upgraded

The durable app company owns the customer and the checking layer.

The Counter-Argument

The cleanest objection is speed. If general evaluators become good enough quickly, the verification moat becomes a trade, not a fortress. If open models reach frontier capability and distribution is easy to rent, application margins compress faster than expected.

The second objection is timing. A true AGI buildout can be right and still be uninvestable if the market capitalizes the terminal state years before the cash flows arrive. The gap between eventual and immediate efficiency is where returns live; paying terminal multiples too early gives that gap away.

So I would avoid the AI label and underwrite the bottleneck. What remains scarce after intelligence gets cheap? Who owns it? What closes the gap? How much of the answer is already in the price?