The prompt. What should countries which are not currently in the AI production chain (semis, energy, frontier models, robotics) do in order to not get totally sidestepped by transformative AI? If you’re the leader of India or Nigeria, what do you do right now?

I. Outside the Chain

Suppose transformative AI happens. Take a large developing economy outside the production chain (India, Nigeria, Indonesia). “Sidestepped” can mean three different things. Are people materially worse off than they would have been (absolute wealth)? Does the government keep meaningful control over its own affairs (sovereignty)? Does the country’s share of global wealth and power shrink (relative position)? The first one, it turns out, runs on a separate track from the other two.

II. Absolute wealth rises anyway.

In any fast-AI scenario worth taking seriously, AI lowers the cost of producing tradeable things. An AI-discovered drug is cheaper to manufacture. A robot-built car is cheaper to assemble. The model that wrote the textbook also translates it into Yoruba. Even a country whose firms adopt none of this is going to see its import-price index fall.

Two caveats. Tyler Cowen’s point: in poor countries the binding constraint on welfare is institutions, capital, and infrastructure, not intelligence. Cognitive labor on tap does not, by itself, get the borehole drilled. And AI deployment often lags behind richer economies. But the net effect will stay positive because they capture the consumer surplus from cheaper tradeables.

III. Sovereignty and relative position

Both depend on owning a meaningful piece of the AI economy. There are two ways to try: invest in someone else’s, or supply something the chain can’t supply itself.

Investment alone won’t secure either. Picture the ceiling case, which would obviously never clear parliament: the prime minister of India burns every bit of political capital, mobilizes the state pension and insurance pools, draws down FX reserves, issues new sovereign debt, and concentrates it all into a single-thesis AI fund. The binding constraint isn’t total wealth. It’s how much converts to dollars without the rupee breaking. So realistically $300 to $500 billion over five years is the ceiling. Pledging other assets doesn’t help, because collateral only works if a lender can seize it, and no court is handing over a port to a foreign bondholder.

Markets might be underpricing AI. Long-term real rates around 2.5% are hard to reconcile with transformative growth, and AI capex at 1.6% of US GDP sits below historical buildouts like the railroads at 6%. But even granting that, India’s $400B buys at most 3% of the AI-exposed US float. And equity is only a slice of total AI value. A Nvidia shareholder earns the profits Nvidia retains but not the agglomeration premium around it, or the security and procurement advantages flowing to the country where the technology was developed. Those channels are much larger than the equity channel, and none of them can be bought from abroad.

This strategy wouldn’t solve the sovereignty aspect either: CFIUS blocks the control rights that would turn equity into influence, and the assets themselves can be frozen in a crisis. See Russian reserves & Chinese ADRs.

Building is the other route. The chain (US and China for foundation models, Taiwan and South Korea for fabs, the Netherlands for lithography, Japan for chemicals and optics) has physical openings it can’t close at home on a fast-AI timeline.

IV. Substrate, permitting, trust

Three things determine which openings a country can fill: substrate (minerals, chemistry, medical infrastructure, large populations), permitting speed (political room to override environmental, patient-rights, and labor frictions), and US strategic trust (Washington’s willingness to host weights, share FDA data). Different niches stress different premises.

Substrate plus permitting opens critical-mineral processing. Indonesia ran this play starting in 2020: a raw-nickel-export ban paired with industrial parks pulled in tens of billions of Chinese capital, grew the sector’s share of GDP from 4.3% to 9.2% in two years, and took Indonesia to 58% of global nickel production. Chile, the DRC, and Kazakhstan can run versions. I rate this durable, high-impact, and crucially, US trust isn’t really required - China funded most of Indonesia’s smelters.

Substrate plus permitting plus very high trust opens gigawatt data-center hosting and Special Administrative Regions. US grid hookups now run three to five years, and $42B of projects were cancelled after local opposition in Q1 2026, so the chain needs foreign hosts. SARs are the regulatory analogue for autonomous robotics and agentic deployment that cannot run under Western product-safety law. The UAE and Saudi Arabia are doing both. The bar is high, because hosted weights are strategically sensitive. The UAE paid for its deal by stripping Huawei equipment out of G42 and handing Microsoft a board seat. This is of course very high impact.

Substrate plus medium trust opens wet labs and clinical trials. If AI designs molecules faster than Western pipelines can test them, the bottleneck just shifts downstream to throughput, and you still have to recruit patients and file paperwork, none of which gets faster because the molecule was designed in three weeks. Recruitment and administration already take up 30 to 50% of trial time in the US. India runs 20% of global trials at a third of the cost, has the largest CRO base outside the US, and its chemistry industry has been plugged into Western pharma for decades. I would rate this medium-impact but durable.

V. The honest Answer

Nigeria has none of the three. India is the marginal case: Substrate in depth, trust recoverable, permitting that won’t deliver a gigawatt in eighteen months. So wet labs and clinical trials, which is enough to keep India a credible second-tier supplier in pharma but not enough to make it anyone’s peer. India can’t play data centers or SARs at scale. To open those, you have to permit fast and override patient-rights, environmental, and labor frictions, and Indian democracy doesn’t let you do that. Democracies generally have this problem. The moves that pay involve rolling back protections people voted for.

Today’s chain countries didn’t plan for AI. Taiwan started state-led foundry investment in the 1970s. ASML spent forty years on lithography before anyone noticed it was a monopoly. These are historical moats rather than strategic ones. The opportunities above can’t make a country a chain peer, only a country the chain needs.

Frederik Inderst

Essays