What Adam Smith knew about AI dependence (that we keep forgetting)

A nation can wake up to find that a capability woven into its hospitals, its courts, and its code has simply been switched off. Not broken. Not priced out. Switched off — by a decision made somewhere else.

The capability still exists. It still works.

It’s just no longer available to you, your firms, your researchers, or even your own citizens working abroad.

For any country leaning on frontier AI, it didn’t build and can’t reproduce; this stopped being a thought experiment a while ago. It’s the defining strategic exposure of the decade.

And the clearest guide to thinking about it was written 250 years ago.

We remember The Wealth of Nations as a hymn to free markets. It’s actually something more useful: a careful map of when a nation should accept dependence on others, and when it absolutely must not.

That map is exactly what AI policy is missing right now.

The pin factory still wins

Smith opens his book in a pin factory for a reason. One worker doing every step makes a handful of pins a day. Specialised workers, each to a task, make tens of thousands.

The same logic governs nations. A country that insists on producing everything itself; its own chips, its own data, its own frontier models at every layer, forfeits the very gains that make a modern economy modern.

Autarky isn’t a strength. It’s a vow of poverty dressed up as prudence.

So the maximalist version of “sovereign AI” — we’ll build the entire stack, beholden to no one — is the pin-maker insisting on smelting his own steel. He’ll make worse pins, far fewer of them, at ruinous cost.

Specialise and trade by default. The gains are real, and pride is not a strategy.

But Smith drew one hard line — and named it

Here’s the part people forget.

Smith, the supposed apostle of free trade, defended one of the most protectionist laws of his age — the Navigation Acts — on a single explicit ground:

“Defence is of much more importance than opulence.”

Naval capacity wasn’t an ordinary commodity to buy from the cheapest supplier. It was the precondition of survival. On security goods — and only on security goods — Smith would pay the efficiency penalty of self-reliance with a clear conscience.

This is the strongest case for sovereign AI. Not industrial pride. Not economic nationalism. Just this: if a capability is genuinely defence-critical, dependence on a supplier who can withdraw it at will isn’t a tolerable inefficiency. It’s a vulnerability that nothing can excuse.

The imperative isn’t “build everything.” It’s harder than that:

Identify the specific capabilities whose sudden loss would be catastrophic — and secure those, and only those.

The catch Smith never resolved? The defence exception has no natural boundary. Almost anything can be called defence-critical. A technology that writes code, reads documents, and accelerates research touches everything — and “everything” can’t be the protected category without collapsing back into autarky.

The entire policy fight lives in where you draw that line. Smith won’t draw it for you. He only insists you draw it deliberately, and then stop.

Plurality, not a flag on the data centre

Smith sharpens a distinction that the current debate keeps blurring.

Wanting multiple, independent sources of a capability is not the same as wanting your own.

His hostility to monopoly was specific: the monopolist keeps supply short and “raises the price above the natural price,” because you have nowhere else to turn. A nation afraid that a single provider — or a single jurisdiction — could choke off what it depends on is making a textbook anti-monopoly argument.

But look at Smith’s actual remedy. It was never to nationalise the monopoly and crown a single state champion. That just relocates the choke point. Now you’re hostage to your own government’s single supplier instead of a foreign one.

His remedy was competition. Many suppliers, none able to dictate terms, because the buyer can always walk.

Resilience comes from plurality — not from a flag painted on the side of a data centre.

Listen to the merchant “with great precaution”

Smith saved his sharpest suspicion for one figure: the merchant who shows up with a proposal for the national good that happens, on inspection, to enrich him.

Any plan to restrain commerce, he warned, “comes from an order of men whose interest is never exactly the same with that of the public” — and “ought always to be listened to with great precaution.”

No line travels better into 2026.

Every champion of a domestic AI capability has an incentive to define the national interest as identical to its own balance sheet — to recast subsidy as sovereignty, and the exclusion of rivals as security.

A genuine security argument and a piece of rent-seeking sound exactly alike from the outside. Both speak the language of the flag.

So when someone proposes AI protection or promotion, ask first: who profits if this is adopted? The discomfort of the question is the whole point.

What only the sovereign can do

Smith was no minarchist. He gave the state three duties — defence, justice, and public works; no private actor will build because they can’t be made to pay.

That third duty is the legitimate home of state action in AI: safety research, independent evaluation and red-teaming, shared compute for those priced out, the basic science that benefits all and is owned by none.

Not because the state runs things better. Because some things otherwise don’t get built at all.

The five imperatives, plainly

Specialise and trade by default. The gains from exchange are real and large.
Reserve self-sufficiency for the genuinely defence-critical few — and force yourself to name them, before the category swells to cover everything.
Pursue supplier plurality over national monopoly. Competition delivers resilience; a domestic flag doesn’t.
Distrust every security argument that enriches whoever is making it.
Have the state build only what the market structurally won’t — the public goods of safety, evaluation, and access.

The argument is yet to be settled.

Serious thinkers reject Smith’s balance.

Friedrich List argued that free trade entrenches whoever already leads, and that a developing nation must shelter its “infant industries” long enough to catch up. Alexander Hamilton made the same case for a young United States. Both map directly onto any country watching a handful of foreign firms pull away in AI.

And the security hawk answers in Smith’s own defensive idiom: a capability dangerous enough to cause catastrophe may warrant control precisely because defence outranks opulence. The power to withdraw isn’t only a vulnerability — it can be a legitimate instrument.

Smith doesn’t settle this. What he offers is better than a settlement: a discipline.

Accept dependence where it buys you abundance and costs you nothing you can’t afford to lose. Refuse it — deliberately, narrowly — where its loss would be ruinous. Build plurality so no single hand, foreign or your own, can close around the things you’ve come to need.

The sovereignty worth having isn’t the sovereignty of building everything yourself. It’s the sovereignty of never being one decree — made elsewhere — away from having nothing.

The hard boundary every nation now has to draw is where “defence-critical” ends and “everything” begins. How that line is drawn will define sovereign AI strategy for the decade ahead.