"Who's Ahead in the Global AI Talent Race?" is a data feature by Damien Ma and Binyi Yang of Carnegie China, the China-focused center of the Carnegie Endowment for International Peace, published on September 23, 2026. It is presented as AI Talent v3.0, a new edition of the Global AI Talent Tracker that MacroPolo ran until it ceased operations in 2025. Like its predecessors, it measures the balance of elite AI research talent by tracing the undergraduate origin, graduate training and current workplace of authors of papers accepted at the Conference on Neural Information Processing Systems (NeurIPS), comparing the 2025 cohort with the 2022 cohort. Its headline finding is that China overtook the United States as the leading work location for the 2025 cohort.
Framing and sample
The authors frame U.S.–China competition as most visible at the model layer, one of the five layers of AI (energy, chips, infrastructure, models, applications) articulated by Nvidia's Jensen Huang, and describe that layer as "essentially a proxy competition for talent." They treat NeurIPS authors as elite research talent and argue that the conference allows like-for-like comparison over time and that the series is "the only continuous longitudinal study on elite AI talent over the last six years."
| Unit | NeurIPS 2022 | NeurIPS 2025 |
|---|---|---|
| Accepted papers | 2,834 | 5,823 |
| Author total appearances | 13,191 | 34,206 |
| Unique authors | 9,617 | 25,677 |
| Complete three-stage career histories (sample used) | 3,153 (33%) | 10,280 (40%) |
The 2022 cohort covers the main and Datasets & Benchmarks tracks, excluding sixty-seven Transactions on Machine Learning Research articles presented at the conference; the 2025 cohort adds the Position Papers track, introduced that year. Current workplace is observed one year after each conference: 2023 for the 2022 cohort and 2026 for the 2025 cohort. Undergraduate institution is the proxy for country of origin. Hong Kong is grouped with China, the United Kingdom is separate from Europe, and Singapore is its own category. Net gains and losses compare undergraduate origin with current workplace and do not track individual migration.
Findings
Where top researchers work
| Work location | 2022 cohort | 2025 cohort |
|---|---|---|
| China | 27.1% | 40.6% |
| United States | 46.4% | 34.2% |
| Europe | 6.4% | 6.9% |
| South Korea | 4.1% | 4.7% |
| United Kingdom | 3.9% | 3.0% |
| Canada | 3.7% | 2.5% |
| Other | 8.3% | 8.0% |
Where top researchers come from
China is described as "the largest originating source of elite AI talent globally." Chinese-origin researchers, defined as those who received undergraduate degrees in China, rose by 11 percentage points to 57 percent of the sample, while the U.S. share fell to 13 percent. India's share fell by 2 percentage points.
| Undergraduate origin | 2022 cohort | 2025 cohort |
|---|---|---|
| China | 46.3% | 57.4% |
| United States | 19.8% | 13.3% |
| Europe | 7.2% | 6.4% |
| India | 6.2% | 4.0% |
| South Korea | 4.8% | 5.7% |
| United Kingdom | 2.8% | 2.1% |
| Other | 12.9% | 11.0% |
Flows between origin and workplace
The United States "remains a magnet for global AI talent, particularly Chinese talent." Among researchers working in the United States, the 2025 share holding undergraduate degrees from China rose by 4 percentage points. The ratio of Chinese-origin researchers working in the United States to U.S.-origin researchers working in China fell from 46:1 for the 2022 cohort to 30:1 for the 2025 cohort, which the authors describe as "still largely a one-way flow of talent."
Net flows for the 2025 cohort, comparing undergraduate origin with current workplace:
| Country/region | Left | Arrived | Net |
|---|---|---|---|
| United States | 155 | 2,300 | +2,145 |
| Singapore | 31 | 249 | +218 |
| United Kingdom | 114 | 210 | +96 |
| Canada | 122 | 178 | +56 |
| Europe | 264 | 308 | +44 |
| Australia | 31 | 65 | +34 |
| Russia | 54 | 2 | −52 |
| South Korea | 135 | 31 | −104 |
| Other | 503 | 162 | −341 |
| India | 370 | 3 | −367 |
| China | 1,847 | 118 | −1,729 |
The most common career paths in the 2025 cohort were China–China–China (3,619 people), U.S.–U.S.–U.S. (1,093) and China undergraduate, U.S. graduate school, U.S. work (1,071), followed by South Korea–South Korea–South Korea (407) and Europe–Europe–Europe (376). Paths such as India undergraduate with U.S. graduate school and work (246), China undergraduate with graduate school and work in Singapore (147), and China undergraduate with U.S. graduate school returning to work in China (132) also appear in the top eighteen.
Retention
The share of Chinese-origin researchers working in China rose from 56.6% for the 2022 cohort to 68.7% for the 2025 cohort, and the share working in the United States fell from 31.8% to 21.2%. The authors attribute the rise in retention to two factors they describe as likely: easier job-finding in China's AI industry, and greater difficulty coming to the United States under "tighter visa restrictions," particularly for Chinese graduate students in STEM fields, linking to the June 2020 presidential proclamation suspending entry of certain students and researchers from the People's Republic of China. Among U.S.-origin researchers, 90.5% of the 2022 cohort and 88.7% of the 2025 cohort work in the United States.
Retention rates for the 2025 cohort (share of each country's own undergraduate graduates working there):
| Country/region | Own graduates | Retention |
|---|---|---|
| United States | 1,369 | 88.7% |
| South Korea | 591 | 77.2% |
| China | 5,903 | 68.7% |
| Europe | 662 | 60.1% |
| Australia | 59 | 47.5% |
| United Kingdom | 216 | 47.2% |
| Singapore | 53 | 41.5% |
| Canada | 205 | 40.5% |
| Other | 742 | 32.2% |
| Russia | 69 | 21.7% |
| India | 411 | 10.0% |
Institutions
Institutions are ranked by paper credit: each accepted paper contributes one unit, divided equally among its listed authors and assigned to each author's employer in the observation year. Peking University ranks first (55.2) and Tsinghua University second (51.9) for 2025, against Google first in 2022. The Korea Advanced Institute of Science and Technology (KAIST) ranks third (40.3), having risen twelve places since 2022, followed by Shanghai Jiao Tong University (40.0), MIT (36.0), Stanford (35.4) and Google (33.3). Of the top thirty, fourteen are in the United States, twelve in China (including three Hong Kong universities), two in South Korea (KAIST and Seoul National University, twentieth) and two in Singapore: the National University of Singapore ranks thirteenth (twenty-fifth in 2022) and Nanyang Technological University sixteenth (twenty-ninth in 2022). Industry entrants are Google (7), Meta (17), Microsoft (21), ByteDance (25), Alibaba Group (26) and Nvidia (29). ETH Zurich and the University of Oxford dropped out of the ranking, and no Indian institution appeared in it in either year.
Regional readings
The authors describe South Korea and Singapore as making "notable gains," with South Korea holding the second-highest retention rate. They describe Europe as continuing to "falter": slightly more attractive as a place to work than in 2022 but weaker on other measures. They describe India's share of top talent as down 2 percentage points, with the share of Indian talent working in the United States also lower.
Methodology
Career data come from OpenReview, the submission and peer-review platform NeurIPS uses, whose author profiles are maintained by the researchers themselves. Records were retrieved with rule-based Python scripts; institutions were matched mainly by recorded web domain. The study used Claude Code and OpenAI Codex to help write, test and run the scripts, "with human intervention in the loop," and states that no LLM generated career histories. A human researcher set cohorts and definitions, handled missing data and judgment-call institution matches, and approved corrections. Back-testing on the 2022 cohort exposed a profile-linking error: in the preliminary 2025 build, 1,705 of 25,682 authors (6.6 percent) were wrongly treated as having no profile. Both cohorts were rerun after the fix. The 2022 cohort was re-run with the new method rather than reusing the previous edition's hand-coded random sample.
Checks reported: eighteen automated consistency checks per edition, which both passed; an independent rebuild that reproduced the core population counts; a separate automated review of the handover folders that reproduced the main figures and corrected three numerical errors in the supporting documentation; and a comparison with the hand-coded 2.0 edition, where the four reconstructed 2022 results differed by 1.7 to 3.2 percentage points, within the previous edition's estimated uncertainty of ±5 to 7 points.
Limitations stated by the authors
- The 40 percent subset with complete career histories "is not assumed to be random." Broader sample checks preserve the main U.S.–China findings, but Europe's direction of change is sample-sensitive: its share rises slightly in the complete-path sample and falls slightly in the broader sample.
- The sample slightly over-represents academia and under-represents industry, by less than 1 percentage point in each direction.
- Human verification was selective and did not re-code a random or stratified sample, so the study cannot report a record-level accuracy rate.
- Profiles are self-reported and cannot be independently verified. Country labels follow the employer's recorded country, which for multinationals may be the home country rather than the researcher's physical office.
- Confidence is described as highest for the aggregate counts and for China overtaking the United States as the leading work location, and lower for individual profiles and physical work locations.
- The study covers a single conference.
The captured page also contains a third 2022-versus-2025 comparison chart (China 31.7% → 35.7%, United States 38.6% → 34.5%, India 10.2% → 8.1%) whose title is not preserved in the captured text; its measure is not recorded here.
Relationships
- supports: US-China AI Competition: Different Races, Different Metrics — research-talent measure alongside compute, capability and deployment gaps
- instance-of: AI Talent Flows
- depends-on: MacroPolo — continues MacroPolo's Global AI Talent Tracker and its 2023 methodology
- related: Carnegie Endowment for International Peace, Damien Ma, Binyi Yang
- related: ASPI's two-decade Critical Technology Tracker: The rewards of long-term research investment (August 2024) — publication-based measure of national research leadership