斯坦福报告_用海量实证承认中国本土_AI_人才体系成熟_DeepSeek人工智能与顶尖人才全球竞争（中英）_53页_1mb

Summary of "Update: DeepSeek AI and the Great Talent Competition"

Core Content

This report updates the analysis of DeepSeek AI's research and engineering contributors, focusing on the evolution of its talent pipeline and the implications for U.S.-China AI competition over the past year. The study examines the institutional affiliations, career trajectories, and academic achievements of 356 researchers involved in DeepSeek's seven most recent papers, revealing a more complex and robust picture of China's domestic AI talent development.

Main Findings

1. China's Domestic AI Talent Pipeline is Strong and Growing

• 53.5% of DeepSeek's researchers have been affiliated exclusively with Chinese institutions throughout their careers.
• Over 97% of the researchers in the dataset have held at least one confirmed affiliation with a Chinese institution.
• The number of researchers has increased by 57%, from 223 in 2025 to 356 in 2026, indicating significant growth in the company's research capabilities.
• The Key Team, consisting of researchers who contributed to all seven papers, has remained at 31 members but now represents only 8.7% of the total, down from 13.9% in 2025.
• The one-paper contributor group has grown dramatically, from 23 to 136, representing 38.2% of the total pool, up from 10.3%.

2. U.S. Training and Exposure Still Influence Chinese AI Researchers

• 29.5% of the current DeepSeek author pool (80 researchers) has had a U.S. institutional affiliation, up from 24.4% in 2025.
• The average U.S. experience duration for these researchers has increased significantly: 35% had only one year of U.S. experience, while 48.8% had 2–4 years and 16.3% had five or more years.
• 70.2% of internationally mobile DeepSeek contributors have returned to China, indicating that U.S. exposure continues to benefit the Chinese AI ecosystem.

3. Longer U.S. Stays Do Not Guarantee Retention

• 13 long-stay researchers spent a combined 119+ years in U.S. institutions, but most are now back in China.
• Only 5 out of 13 long-stay researchers remain in the U.S., and U.S. experience duration does not predict final destination.
• These researchers often had complex international career paths, including multiple transitions between countries, suggesting that U.S. institutions are part of a broader transnational research network.

4. U.S. Institutions Are Not Dominant in the Network

• While Microsoft leads with 7 affiliated researchers, no single U.S. institution dominates the network.
• Carnegie Mellon, Stanford, UC Berkeley, and University of Pennsylvania are among the top U.S. institutions with 4–5 affiliated researchers.
• The U.S. institutional affiliations are spread across 113 institutions, with notable growth in the Midwest and South regions.

5. Talent Mobility Patterns

• The most common mobility pattern is China → U.S. → China, with 38.8% of researchers following this trajectory.
• 23.8% of researchers started in the U.S. and ended up in China.
• Only 12.5% of researchers followed the China → U.S. → stayed in the U.S. path, and 6.2% started and ended in the U.S.
• International mobility is widespread, with researchers frequently moving between countries.

6. Academic Performance and Contribution

• The average citation count for DeepSeek researchers increased from 1,000 to 1,763, and the median h-index more than doubled, from 249 to 681.
• 141 new researchers were added to the dataset, with 26% of them having at least one U.S. affiliation.
• The domestic pipeline is now producing frontier-model contributors without significant U.S. or international exposure, suggesting a growing self-sufficiency in China's AI research capabilities.

Key Implications

• The U.S. strategy based on the assumption that China lacks the talent or know-how to surpass the U.S. is no longer valid.
• U.S. training and exposure still provide value to Chinese AI firms, but the trend shows that China is increasingly capable of developing its own leading researchers.
• Policy responses such as visa restrictions or export controls may not be effective in curbing talent flow, as knowledge and skills have already been transferred.
• The two-track workforce model at DeepSeek has strengthened, with new researchers filling the rotating cast while the committed core remains stable.

Methodology

• Researchers were identified from arXiv (2024–2025) and Hugging Face (April 2026).
• 282 out of 356 researchers had usable profiles in OpenAlex, which were queried for institutional affiliation history, publication records, and citation metrics.
• The analysis now excludes non-research contributors such as those in data annotation, business, and compliance roles, focusing more on the research and engineering workforce.

Conclusion

This report highlights the growing strength of China's domestic AI talent pipeline, which is now producing high-impact researchers independently of U.S. institutions. While U.S. exposure still plays a role in the development of some researchers, the data shows that China is becoming more self-reliant in AI innovation, with a significant portion of its leading contributors having minimal or no international experience. The U.S. must reconsider its approach to talent competition, as the current strategies may not be sufficient to address the dual challenges of losing skilled researchers and seeing domestic talent rise.