The Talent War Nobody Tracks

Every year, the MacroPolo think tank publishes an analysis of where the authors of papers accepted to NeurIPS — the field’s most prestigious conference — were trained as undergraduates and where they work when the papers are published. The dataset is not comprehensive, but it is the best publicly available measure of where AI talent is produced and where it ends up.

The 2026 edition of that analysis contained a number that should have been more widely reported. For the first time since MacroPolo began tracking, the share of top AI researchers working at Chinese institutions exceeded the share working at American institutions in the highest-tier paper category — the roughly 5 percent of papers that receive oral presentations and awards. The margin was small. The direction was significant.

AI capability, at the frontier, is not primarily a function of hardware or capital. It is a function of researchers who can formulate the right questions, design the experiments that answer them, and iterate on failures with the speed that the field demands. The people who wrote the transformer architecture paper (Vaswani et al., 2017), the people who designed the RLHF training regime that made GPT models useful (Ziegler, Stiennon, and collaborators in 2019-2020), the people who developed mixture-of-experts scaling at Google — these are not fungible inputs. They are specific individuals whose specific insights changed the trajectory of the field.

The US has maintained its AI research lead through a combination of factors that are well understood but often discussed as if they were permanent rather than contingent. American universities attract the best students from everywhere, including China, because their research environments, funding levels, and career trajectories are uniquely appealing. American AI companies pay salaries that no other country’s private sector can match. American immigration policy, despite being genuinely dysfunctional in many respects, has been permissive enough that researchers who want to stay after their PhD usually can.

None of these advantages are guaranteed to persist.

The talent question became explicitly political in 2019, when the US government began implementing visa restrictions that affected Chinese graduate students in STEM fields, particularly those with affiliations to institutions that had relationships with the Chinese military. The restrictions were narrow but the signaling was broad: China-born researchers began calculating whether a US career path was reliably available to them, whether their work would be subject to export control restrictions, and whether the political environment made staying worth the uncertainty.

The calculations are individual and difficult to aggregate. But the pattern in the data is visible. The share of China-born AI researchers who complete US PhDs and then return to China has increased. The share who choose European universities for their graduate training, partly to reduce US visa uncertainty, has increased. The flow of researchers from US academic positions to Chinese AI labs — which was notable in the 2015-2020 period, when Chinese company salaries became competitive — has slowed, partly due to political pressure on US institutions to enforce technology transfer rules, but the underlying pull factors remain.

DeepSeek’s research team is a case study in what this looks like in practice. The company has recruited heavily from Tsinghua, Peking University, and Shanghai Jiao Tong University — China’s leading technical institutions — and from the diaspora of researchers who returned to China in the 2018-2022 period, when Chinese AI lab compensation reached parity with US tech companies. Many of the DeepSeek researchers have academic backgrounds that include US graduate programs, sometimes completed before the visa restriction environment changed. They brought with them not just technical knowledge but the research culture — the emphasis on publication, peer review, and methodological transparency — that American research institutions had cultivated.

The reciprocal dynamic is what US policymakers have been most focused on: the risk of technology transfer from US research institutions and companies to Chinese entities. The restrictions on Confucius Institutes, the export control rules that apply to certain dual-use research, the FBI’s focus on academic espionage — these are responses to a real phenomenon. China has made systematic efforts to accelerate its AI development by acquiring knowledge from US researchers through legal channels (collaboration, publication) and occasionally through less legal ones.

But the policy response has been calibrated to the espionage risk more than to the talent development risk. Restricting which researchers can work at US companies is a different problem from ensuring that US universities continue to attract the best global AI talent. These are not the same policy lever. The former is about preventing outflows. The latter is about ensuring inflows.

The inflow question is the one that the talent competition ultimately hinges on. For most of the past two decades, the US has attracted a disproportionate share of the world’s most talented AI researchers because the combination of academic prestige, industry salaries, and immigration accessibility was simply better than what any other country offered. That combination is now being actively degraded — academic prestige is stable but immigration accessibility has become uncertain, and Chinese industry salaries are now competitive for researchers who are willing to build their careers in China.

The calculation is different for different people. A Chinese researcher who wants to stay in the US for the research environment and career options available here will make that choice if the immigration environment permits it. A Chinese researcher who is uncertain about long-term visa security, or who has family considerations, or who believes that the AI opportunities in China’s current development moment are genuinely comparable to US opportunities, will make a different choice. Both choices are rational. The second choice is more common than it was in 2018.

The European dimension of the talent picture gets less attention than it deserves. Several European AI research centers — EPFL in Lausanne, the Max Planck Institute for Intelligent Systems in Tübingen, the Alan Turing Institute in London — have become more attractive options for researchers who want world-class research environments without the US-China political complications. This is a relatively recent development; European AI research in 2015 was genuinely weaker than US research, and the gap has narrowed substantially through deliberate investment and the recruitment of researchers who might previously have chosen US institutions.

Mistral AI, the Paris-based lab that has emerged as the most credible European AI company, has explicitly positioned itself as a destination for researchers who want European salary levels (lower than the US) in exchange for European regulatory certainty and a specific research culture that emphasizes efficiency and openness. Mistral’s research outputs have been disproportionate to its size, which suggests that the research culture premise is at least partially validated. Whether it can eventually compete for the researchers who choose between Mistral and Anthropic (rather than between Mistral and a European government lab) is the next test.

The export controls, whatever their effect on chip supply, cannot restrict the movement of published research. The transformer architecture is public knowledge. The RLHF training techniques are in published papers. The architectural innovations that DeepSeek and Mistral have developed are documented in technical reports that anyone can read. The knowledge is diffusing regardless of what hardware policy does.

This is simultaneously reassuring and alarming, depending on your perspective. It is reassuring because it means that research productivity — the ability to generate insights that advance the field — is not monopolized by any country. It is alarming because it means that the US cannot maintain AI leadership by restricting hardware access while allowing knowledge to flow freely.

The talent war — who attracts, retains, and develops the researchers who make the next architectural breakthrough — is the competition that determines long-run outcomes. The chip war gets the headlines. The talent war gets the results.