更新：DeepSeek AI和人才竞争

Amy Zegart andEmerson Johnston White PaperUpdate: DeepSeek AI andthe Great Talent Competition Authors Emerson Johnstonis a full-time research assistant at the Hoover Institution’sTechnology Policy Accelerator. She is a graduate of Stanford’s Ford Dorsey Master’sin International Policy and received her B.S. and B.A. from Northeastern University. Amy Zegartis the Morris Arnold and Nona Jean Cox Senior Fellow at the HooverInstitution, director of Hoover’s Technology Policy Accelerator, and a senior fellowand associate director at Stanford University’s Institute for Human-Centered AI. How to cite this paper Emerson Johnston and Amy Zegart, “Update: DeepSeek AI and the GreatTalent Competition” Technology Policy Accelerator, Hoover Institution andthe Stanford Institute for Human-Centered Artificial Intelligence, June 2026.https://doi.org/10.64576/0626 When DeepSeek AI released its R1paper in January 2025that revealed a “… a third of the most integralcontributors to a Chinesemodel performing comparablyto OpenAI’s o1 model nevertrained, published, or workedin an international researchenvironment.” model rivaling OpenAI’s o1 for far less compute, medianarratives were dominated by two misconceptions.First, that the Chinese company’s breakthroughs werethe result of Chinese researchers who could imitatebut not innovate; and second, that the research teamwas young, green, and generally not highly cited in thecomputer science field. Our 2025 report, “A Deep Peekinto DeepSeek AI’s Talent and Implications for USInnovation,” extensively examined the 223 researchersbehind the company’s five foundational papers andfound a very different picture. We found that DeepSeekdidinnovate in significantways, that its research team was far more experiencedthan most assumed, and that more than half of theteam had never left China for school or work. Together,the findings pointed to a domestic AI talent pipelinein China that is larger and more capable than theprevailing narrative allowed. They also suggested thatany U.S. strategy built on the assumption that Chinalacks the talent or know-how to surpass the U.S. on itsown was already out of step with the evidence. China’s domestic AI talent pipeline is stillgoing strong •The share of DeepSeek AI researchers affiliatedexclusively with Chinese institutions throughouttheir recorded careers stands at 53.5 percent. •Ten of the 31 Key Team members—researcherslisted on all seven DeepSeek AI papers publishedin the last two years—are also affiliated exclusivelywith Chinese institutions, meaning a third ofthe most integral contributors to a Chinesemodel performing comparably to OpenAI’s o1model never trained, published, or worked in aninternational research environment. We argued then that the United States needs tocompete much more aggressively to attract, welcome,and retain the world’s best and brightest while urgentlygrowing domestic capabilities by improving K–12 andhigher education in STEM fields at home. New and better data reveals longer U.S.training periods, especially for top talent This report asks how much has changed in the yearsince. We extended our analysis from five to sevenpapers—adding DeepSeek V3.2 (December 2025) andV4 (April 2026) to incorporate the company’s latestresearch. By expanding our original dataset, we find theauthor pool has grown from 223 to 356 and we wereable to build comprehensive profiles tracking wherethey studied, lived, and worked for 282 of them. A year ago, 63.3 percent of the U.S.-affiliated cohorthad recorded only a single year of U.S. experience. Thatfigure is now 35 percent, with nearly half in the 2- to4- year range. This shift reflects both new DeepSeektalent and improved data about the original cohortwe examined last year. Notably, the 141 net new contributors to DeepSeek Papers 6 and 7 in 2026 area more internationally mobile cohort than the original223 that contributed to papers in 2025. In addition, wefound that OpenAlex’s affiliation coverage has improvedacross the board, surfacing prior U.S. positions forresearchers we tracked last year that were not yet in thedatabase when we ran our original queries. simply land in the U.S. stay for many years andthen return to China. Instead, they made manyinternational transitions over many years. •The 80 DeepSeek AI researchers with any U.S.institutional affiliation are, on average, the mostacademically accomplished in the pool, and ameaningful share of them spent five or moreyears embedded in U.S. institutions beforereturning to China. The picture emerging from both the new researchers andour improved data is not of brief, exchange-style visits tothe U.S., but of meaningful embedding in U.S. researchenvironments followed by a return rate of 70.2 percentamong internationally mobile DeepSeek contributors. The implication is, at best, uncomfortable. Knowledgethat has already moved with these researchers cannot berecovered by tightening visa policy or export controlsgoing forward. Restricti