The Portfolio

TAM-LEV.02 · The Levers · The Approximate Mind

Drafted by Claude with Syam, 2 September 2026.

Every state holding a position on artificial intelligence is holding a portfolio, whether or not its ministries use the word. The holdings are policy levers: instruments the state can pull, each with a price, each with a yield, each with a decay curve. Portfolios are managed well when the holder knows which assets waste and which compound. The argument of this essay is that almost every state has the classification backward, and that the mis-pricing is systematic rather than local: the same error, in the same direction, in Washington and Brussels and Beijing and New Delhi.

Sort the levers first. There are five that matter at the frontier, and they fall into three classes.

The Fast-Wasting Gates

#

The first class gates mobile inputs, and decays fast, because gating a mobile input transfers it.

Talent is the first holding and the subject of The Talent Valve, summarized here only as a classified asset: a valve that reroutes rather than retains, decaying with each use as the redirected flow becomes foreign institutions, and decaying with time as the willingness to migrate erodes on its own. The full mechanism, the ratchet of the two valves in series, is stated there and not retold.

Export controls are the second holding, and the decay mechanism has a special cruelty: the control finances its own substitute. Restrict the sale of frontier chips to a rival, and the restriction does not remove the rival’s demand. It removes the rival’s option to satisfy demand through purchase, which converts the entire blocked revenue stream into the business case for domestic substitution. The chip controls of the mid-2020s produced, on schedule, the largest state-directed fabrication program in the target country’s history, along with something the controls’ designers priced at zero: the elimination of the target’s coordination problem. A thousand firms that would each have preferred to buy American silicon were handed a single national reason not to. The gated input here is not the chip, which is only an artifact, but the capacity to make chips, and that capacity is mobile on a five-to-ten-year clock, slower than talent, faster than the control regime assumed. Each year the control operates, the substitute program compounds. The lever’s yield is real, a genuine multi-year drag on rival capability, and the yield is paid for out of principal.

The claim goes on the record dated. As of 2 September 2026: by September 2030, domestically produced accelerators in the primary target country will be training models within one recognized capability generation of that country’s frontier laboratories’ needs, such that export controls no longer bind the pace of its frontier training. Miss condition: if by that date frontier-scale training there remains materially gated by inability to obtain or substitute controlled hardware, the financing-the-substitute mechanism is slower than this portfolio prices it, and the export lever belongs closer to the slow class. Entered in the register.

Intellectual property is the third holding, and it is overrated. The ruling deserves its reasons stated plainly, since IP is the lever most beloved by the lawyers who staff delegations. The moats at the frontier were never patents. The two moats were compute, which is a purchasing problem, and tacit knowledge, which walks on two legs through passport control. The published paper was upstream of every frontier laboratory; the architecture that defines the era was given away in its founding document; the weights that matter leak, distill, or are released outright, and the corpus has already recorded what the efficiency curve does to any strategy premised on their scarcity. A state pulling the IP lever is guarding the artifact while the capacity to produce the next artifact, the only asset that was ever strategic, sits entirely outside the instrument’s reach. IP protection matters at the deployment layer, where products live and copying is commercially real. As a frontier gate it is a lever connected to nothing.

The Slow-Wasting Gates

#

The second class gates inputs that are genuinely hard to move, and decays slowly, but on the same curve.

Materials and energy form the fourth holding: lithography monopolies, rare-earth processing shares, the gigawatts a training campus drinks, the industrial gases, the substrate chemistry. These bind harder than anything in the first class because atoms are not files. A processing monopoly built over thirty years does not relocate in five, and grid interconnection is measured in decades. States holding these positions are holding the best gates in the portfolio.

But the class decays by the same two mechanisms, geared down. Use transfers: every exercise of a materials chokepoint is a subsidy announcement for substitution, and the substitution programs are running. Time commoditizes: the energy appetite that makes gigawatts a gate is itself a temporary fact about one era’s training regime, and the efficiency curve is the subject of The Mainframe Verdict. The honest classification is not durable but durable-for-now, with the expiry date load-bearing. A gate that binds for fifteen years and is priced as permanent is still mis-priced, just fatally later.

The Compounding Holds

#

The third class contains no gates at all, which is why it is systematically misfiled as background.

The fifth holding is really a bundle: the policy ecosystem aimed at the deployment layer. Rule of law that makes contracts involving machine outputs enforceable and liability allocable. Capital friendliness that lets ten thousand small deployments find financing. Procurement, the state as the anchor buyer in health, education, benefits, defense. Certification and licensing regimes that let a regulated profession absorb a capability without dissolving. Administrative competence itself, the state’s ability to run its own registries and windows well enough to deploy anything into them.

Notice what these instruments have in common. None of them gates a mobile input, so none of them decays by transfer; there is no flow to reroute. And their yields compound, because each deployment they enable becomes infrastructure for the next: a working procurement pathway, a settled liability rule, a certified integration is a paved road that stays paved. The corpus’s central structural finding is the reason this class exists: capability concentrates because compute concentrates, but deployment cannot concentrate, because the deployment frontier is tacit, contextual, and held by people standing in specific places. Value settles at the layer that cannot concentrate. The compounding levers are the only instruments in the portfolio aimed at the layer where value settles.

The gates operate on the frontier that will commoditize. The compounding levers operate on the frontier that cannot. That asymmetry is the whole portfolio theory in one sentence.

The Mis-Allocation

#

Now read any national AI strategy against the sort. The pattern is uniform enough to be uncanny. The fast-wasting gates receive the summits, the legislation, the enforcement budgets, the prime-time announcements: visa crackdowns and export-control expansions are the visible muscle of AI statecraft. The slow-wasting gates receive the sovereign wealth: fabrication subsidies, energy megaprojects, stockpiles. And the compounding class receives language. Chapters titled adoption and skills and government modernization, unfunded, unowned, assigned to no minister who loses a job when they fail.

Almost every state is spending its wasting assets as if they were durable and treating its compounding assets as background conditions. The direction of the error is what makes it a theory rather than a complaint: mis-allocation this uniform has causes. Three are visible. Gates are legible, they produce announcements with named adversaries, and legislatures fund legibility. Gates fit the last century’s template, in which controlling inputs to strategic industry genuinely was leverage, and institutions reach for the template they have. And the compounding levers are slow, boring, and self-implicating: funding procurement reform means admitting the state’s own machinery is the constraint, which no government announces at a summit.

There is a fourth cause, less comfortable. The gates fail on a delay measured in years, and the officials who pull them are promoted on a cycle measured in quarters. A wasting asset is, for its operator, indistinguishable from a durable one, right up until it is not.

Could a portfolio this uniformly mis-weighted be rebalanced by anything short of the losses arriving? The record of the mainframe era, which The Mainframe Verdict reads as a verdict, suggests what usually rebalances it is the losses.

The sort is now on the table: two classes of gates, decaying at two speeds by the same two mechanisms, and one class of holds, compounding at the only layer where the corpus finds value able to settle. What remains is to test the sort against history, state the inversion it implies, and walk the three holdings of the compounding class one at a time. The portfolio does not need new assets. It needs to be read at current prices.