AI Talent: China Overtakes the US Despite the Chip Embargo

Pawtucket, 1789: the machine stays home, the mind leaves

In 1785 the British Parliament banned the export of textile machinery. The same legal framework prohibited the emigration of skilled workers. The mechanism was: block the object.

In 1789 Samuel Slater left England disguised as a farm labourer. He carried zero machines and a complete memory of Arkwright's frames. In 1793 he opened in Pawtucket, Rhode Island, the water-powered spinning mill that bears his name.

Capital embodied in people travels with whoever owns it. Licences govern containers, patents and machine tools; tacit knowledge crosses the border inside a suitcase.

The embargo held on the artefact. It gave way on the mind.

London repealed the emigration ban in 1825 and the machinery ban in 1843, by which point it was defending an asset already lost. The context of 2026 is different; the structure is identical.

This is the lens for reading the US-China technological decoupling on AI: the factor that can be sanctioned and the factor that escapes sanction.

The data point that overturns the consensus

On 25 September 2026 the South China Morning Post reported on a study by the think tank Carnegie China, titled «Who's ahead in the global AI talent race».

In 2025 China hosted 41 per cent of the top AI researchers tracked by the study, against 34 per cent for the United States^[1]. The same survey in 2022 put the Americans at 46 per cent and the Chinese at 27.

Nineteen points of American advantage become seven points of disadvantage. Twenty-six points of rotation in three years.

The same research was picked up by The Information^[2]. Two outlets, one primary document: the think tank's country-by-country count.

A stock of human capital moves slowly, because a career lasts decades. A rotation of this magnitude in thirty-six months therefore implies a sign change in the flow: new entrants stay home, and some of the veterans come back.

Three precedents are enough to call it a pattern

An isolated case is an anecdote.

Britain, 1785-1843: embargo on machines and artisans, American textile industry up and running in Pawtucket within eight years of the law's passage.

COCOM, 1949-1994: the Western allies and Japan coordinate the embargo on machine tools, computers and semiconductors bound for the Soviet Union. Moscow stays at the frontier of mathematics: Kolmogorov, Gelfand and the Landau school sign first-rate results for forty years. The gap opens in volume production, never in theory.

Germany, 7 April 1933: the civil service law expels Jewish faculty and political opponents from the universities. Enrico Fermi leaves Italy in December 1938, after the racial laws, and lands in New York on 2 January 1939. Within a decade the centre of world physics shifts from Göttingen and Rome towards Chicago and Princeton.

The third case shows the opposite face of the same mechanics: the power that turns its own institutions hostile exports the human capital it had accumulated.

The mechanism: the constraint raises the price of brains

On 7 October 2022 the US Bureau of Industry and Security extended controls on advanced accelerators and chipmaking equipment bound for China. Subsequent revisions lowered the thresholds.

The causal chain runs in this order.

1. Compute becomes scarce and expensive inside the sanctioned perimeter.
2. The binding constraint shifts from hardware to efficiency per unit of compute.
3. Efficiency is a research output, so the marginal value of the researcher rises where compute is missing.
4. Salaries, equity and state laboratories bid up that cohort.
5. Friction on American visas reduces the inbound flow.

The decisive step is the second. When compute is abundant, the advantage lies in scale, and therefore in the budget. When compute is scarce, the advantage lies in the mind of whoever writes the kernel.

The domestic career thus becomes competitive with the Bay Area, and returning home stops being an economic sacrifice.

Visa friction works on the other side, and reduces entry precisely as exit slows. Two derivatives pointing the same way.

This is a structural regime change, with a multi-decade horizon. The cohort that is twenty-eight today will publish for thirty years.

This desk's position, and what would disprove it

Export controls have shifted the competition from silicon to human capital, and human capital escapes any licence.

I have long argued that this contest is about access to semiconductors, never about model capability. That thesis needs correcting by half. The bottleneck remains the fab, because lithography and production capacity govern the physical outcome; the gap was finance, with Hong Kong capitalising Chinese laboratories. Now there are two gaps, and the second one walks.

The counter-argument deserves respect. The count measures the institution of affiliation, so it tells you where an author sits, never what their output is worth. And the chip thresholds continue to limit the largest training runs, the ones that produce the frontier model.

The reply is a question of duration. An accelerator loses its value in twenty-four months and can be sourced through intermediaries; a researcher composes for thirty years and stays where family, laboratory and options are.

I would change my reading in the face of two pieces of evidence: the next edition of the same dataset reporting the Chinese share below 38 per cent, or evidence that this cohort signs its own work using American compute rented outside the country.

Three implications for capital

1. Horizon 12-36 months. The premium on owning fabs and lithography holds, because the physics of production stays scarce. The premium paid to American laboratories as the exclusive home of talent compresses.

2. Horizon 24-36 months. Debt-financed AI capex presupposes a durable model advantage. An advantage that thins out shortens the economic life of the data centre, and that life is the only variable holding up the private credit covenant.

3. Horizon 36 months. Control will extend towards what travels with people: cloud access for foreign nationals, research collaborations, identity verification on remote compute. Whoever sells compute at a distance will inherit a new compliance obligation.

The market has priced in chip scarcity. Researcher scarcity remains outside the models.

Four tables, four questions

Family offices and sovereign funds: the sensible reallocation over the next thirty-six months rewards production capacity and the countries that host fabs, and treats the laboratory's location as a movable variable.

CEOs and boards: the geopolitical risk missing from strategic plans is loss of access to personnel, never to the end market. A product roadmap that assumes hiring the best cohort in San Francisco rests on a fragile premise.

Chief risk officers: the scenario outside the VAR models is a convergence of capability at lower compute, with the collapse of the residual value of compute assets. Historical variance is silent on this case, because it has never observed it.

CFOs and investor relations: the narrative of a structural American advantage in AI risks ringing false within eighteen months. Better to replace it with a thesis on costs, energy and access to production capacity.

The forecast

Expected event: the next edition of the Carnegie China survey on AI talent places the Chinese share of top researchers at or above 41 per cent, and the American share below 36 per cent.

Confidence: 70 per cent, Medium. Horizon: 31 December 2027. Verification: the country table in the next edition of the study.

Kill signal: the same source reports the Chinese share below 38 per cent, or the American share above 38 per cent, on the 2026 cohort.

What to watch:

• The affiliations declared in papers accepted at the major machine learning conferences of 2027.
• American data on H-1B and O-1 visas issued to computer science graduates.
• The recruitment packages offered by Chinese laboratories, public and private, for senior research positions.

The divergence between sanctioned hardware and free human capital always resolves. The question is how.