Global AI Talent Shift Explained: From Origins to Now

Why it is suddenly everywhere

1. 01

   Carnegie China published AI Talent Tracker 3.0 on September 23. Its analysis of 10,280 complete career histories from NeurIPS 2025 authors puts 40.6% of elite AI talent at Chinese workplaces and 34.2% at US workplaces—a reversal of the 2022 order.

2. 02

   The underlying records come from OpenReview, the submission and profile system used by NeurIPS. Carnegie began with all 25,677 unique authors and used the 40% whose profiles contained undergraduate, graduate and workplace histories.

3. 03

   China is not simply importing a finished workforce: 57% of the study cohort received undergraduate education in China, up 11 percentage points from 2022, and 69% of that China-origin group now works there.

4. 04

   The counterintuitive part is that the United States still recorded the largest net brain gain, adding 2,145 researchers relative to undergraduate origin, while China recorded a net loss of 1,729. China became the largest workplace because its production base grew even faster than its remaining outflow.

The global AI talent shift is not a clean story of researchers abandoning one country for another. It is a change in where a rapidly expanding cohort is educated, retained and employed. China now appears first when workplaces are counted in Carnegie’s new NeurIPS sample. The United States remains the strongest importer, retains a larger share of its own graduates and still attracts a substantial stream from China.

That distinction explains why two apparently contradictory headlines can both be true: “China overtakes the US” and “America remains the global talent magnet.” One measures the final location of all researchers; the other measures how much talent a country gains beyond the people its own universities originally produced.

The timeline

1. 1998A Beijing research bridge opens

   Microsoft Research Asia is founded, training generations of Chinese computer scientists in an internationally connected lab without requiring them to leave the region.

2. 2017–2019The US-centered pipeline

   Top Chinese graduates commonly move to American graduate schools and labs. Earlier MacroPolo tracking finds only 11% of China-origin elite researchers working in China in 2019.

3. 2020Mobility meets national security

   US Presidential Proclamation 10043 restricts visas for some Chinese graduate students and researchers linked to institutions associated with military-civil fusion.

4. 2022China becomes the largest origin

   The previous talent tracker finds Chinese universities producing 47% of top-tier AI researchers, while US institutions remain the leading workplace at roughly 46% versus China at 27%.

5. 2023–2025Domestic labs become frontier workplaces

   Chinese model companies expand; DeepSeek recruits heavily from local universities, while senior researchers also return from Google, OpenAI and other overseas labs.

6. 2025Performance narrows the credibility gap

   DeepSeek and other open-weight models show that researchers can work on globally competitive systems inside China rather than treating a domestic job as a second choice.

7. 2026-09The workplace ranking flips

   Carnegie reports China at about 41% of elite AI workplaces and the United States at 34%, with Peking University replacing Google at the top of its institutional ranking.

First, understand what the study counts

“Elite AI talent” sounds like a census, but the tracker is a carefully defined sample. Carnegie starts with authors of accepted papers at NeurIPS, one of machine learning’s most selective conferences. The 2025 conference had 5,823 accepted papers and 25,677 unique authors across the included tracks. OpenReview profiles supplied three career stages: undergraduate institution, graduate education and current workplace. Complete records existed for 10,280 people, or 40% of all unique authors.

Origin is therefore a proxy: it means the country of a researcher’s undergraduate institution, not citizenship, ethnicity or birthplace. Workplace means the institution listed in the observation year, and a multinational employer may be assigned to its home country rather than the physical office. The sample also slightly overrepresents academia and underrepresents industry. Carnegie says confidence is highest in the aggregate China-US reversal and lower for individual locations; Europe’s direction changes under broader sample checks.

Those caveats do not erase the finding. They tell us exactly what changed: among visible authors publishing at one frontier research venue, Chinese institutions now account for the largest workplace share. The result should not be inflated into “41% of every AI engineer on Earth,” but it is a meaningful signal about the people extending the research frontier.

How the old one-way pipeline formed

For two decades, the standard route ran through three stations: undergraduate study in China, doctoral training in the United States, then employment at an American university or technology company. China supplied an enormous pool of mathematically trained students; US graduate programs, research funding, computing infrastructure and immigration pathways concentrated them in a small number of labs. Brookings summarized the 2022 version starkly: China produced 47% of the tracked top talent, versus 18% for the United States, yet America employed the largest share.

That arrangement was not simply “brain drain.” It was a transnational production line. Microsoft Research Asia, founded in Beijing in 1998, helped connect local universities to global research norms. Chinese undergraduates moved through American PhD programs and often collaborated across borders. US labs gained a workforce larger than their domestic education pipeline could supply, while Chinese universities gained alumni, networks and eventually returnees.

The first crack was visible before the newest numbers. In 2019, only 11% of tracked China-origin researchers worked in China; by 2022 the share was 28%. Carnegie’s reconstructed comparison now puts retention at 57% for the 2022 cohort and 69% for 2025 under its complete-career method. The levels differ because editions and samples changed, but the direction is consistent: the pipeline is no longer overwhelmingly one-way.

Staying became a frontier option

A researcher chooses more than a salary. The decisive question is whether a workplace offers important problems, compute, colleagues and enough authority to pursue an idea. China’s model boom changed that bundle. Rest of World documented how DeepSeek recruited recent graduates and interns from Tsinghua, Peking and other leading universities, offering young researchers consequential roles rather than making Silicon Valley the obligatory finishing school. Zizheng Pan chose the then-obscure DeepSeek after an Nvidia internship in 2023; less than two years later, his team helped produce a model that forced the global industry to pay attention.

That success mattered as evidence. Once a domestic lab could produce frontier work under chip constraints, staying home no longer necessarily meant stepping away from the most difficult problems — a constraint that runs deeper than accelerators, as the AI memory supercycle now decides who can even buy enough HBM to train at scale. The same ecosystem is producing the software that powers AI agent phones, the specialists building autonomous AI cyber agents and control teams able to turn the Unitree GD01 into a half-tonne balancing machine. Applications, models and research jobs reinforce one another: products create demand for researchers, and visible breakthroughs make the next cohort more willing to stay.

Returnees added another layer. ThinkChina recorded a cluster of senior moves: former Google DeepMind research vice-president Wu Yonghui joined ByteDance; former OpenAI agent-project leader Vinces Yao became Tencent’s chief AI scientist; and former Gemini reinforcement-learning head Zhou Hao joined Alibaba in early 2026. These are prominent examples rather than proof of mass migration, but senior hires can redirect teams, budgets and research agendas disproportionate to their number.

Push factors matter, but they are not the whole cause

US policy made the old route less predictable. Presidential Proclamation 10043, issued in 2020, restricted entry for certain Chinese graduate students and researchers associated with institutions deemed connected to military-civil fusion. Visa lotteries, security reviews and a broader climate of suspicion added friction even for people outside the proclamation’s scope. Stanford’s 2026 AI Index reports that the number of AI researchers and developers moving to the United States has fallen 89% since 2017, including an 80% decline in the latest year.

Yet attributing the reversal only to American barriers misses the larger arithmetic. The China-origin share in Carnegie’s cohort rose to 57%, and China retained 69% of that much larger base. Even if no established scientist had “returned,” more graduates taking their first serious job at home would shift the workplace map. This is as much a story of a domestic pipeline reaching scale as it is one of migration.

Nor has the United States stopped attracting talent. Its 89% retention rate exceeds China’s 69%, and its net gain of 2,145 is the largest in the study. The bilateral flow remains lopsided: about thirty China-origin researchers work in the United States for every US-origin researcher working in China, down from forty-six to one in 2022. America lost first place in total workplace share while preserving the stronger immigration machine.

Why the map changed now

Three trends converged. First, Chinese universities enlarged the upstream pool: the China-origin share reached 57%, and Peking University rose above Google in Carnegie’s institutional ranking. Second, domestic labs offered frontier projects with visible global impact, turning “stay” into a positive career choice rather than a visa fallback. Third, cross-border movement became harder just as opportunities at home became more credible.

The technical race also tightened. Stanford’s 2026 AI Index says US and Chinese models traded the lead several times after early 2025, while China led in publication volume and citations. Talent responds to that feedback loop. Researchers go where ambitious work seems possible; their arrival then makes more ambitious work possible.

But concentration has costs. A world in which researchers must “pick a side” loses the collisions of ideas that made the old pipeline productive. ThinkChina’s interviews also describe intense performance metrics, rigid boundaries between universities and companies, and returnees who struggle with institutional culture. China can become the largest workplace without becoming the easiest research environment. The United States can remain the largest brain-gain country without assuming that attraction is permanent.

The real change, then, is not a trophy moving from one flag to another. It is the end of a default assumption: that the best AI talent produced in China will eventually work in America. There are now two credible centers, plus rising hubs in South Korea and Singapore. The geography of AI research is becoming less like a funnel and more like a network—just as policy is trying to split that network apart.

Key facts

NeurIPS 2025 authors

25,677 unique

Complete histories

10,280 (40%)

Working in China

40.6%

Working in US

34.2%

China-origin share

57%

China retention

69%

What people are saying

Common questions

Does the study prove China has more AI workers than the United States?

No. It studies authors at NeurIPS with complete education and employment records, not every AI employee. It is best read as a measure of visible frontier-research talent within a consistent conference cohort.

How can China lead in workplaces but still have a brain drain?

Because China starts with a much larger origin pool. It retains 69% of China-origin researchers but still loses more people than it gains from abroad. The United States produces a smaller share itself but imports enough researchers to record the largest net gain.

Are researchers counted by nationality?

No. Undergraduate institution is used as the origin proxy. A Chinese citizen who studied for a bachelor degree in the United States is counted as US-origin, while citizenship and ethnicity are not the classification.

Why use NeurIPS?

Its accepted authors provide a repeatable proxy for researchers doing frontier machine-learning work, and OpenReview supplies structured career records. The trade-off is that one conference cannot represent applied engineers, closed industrial labs or the entire AI workforce.