China has offered foreign AI scientists a dedicated visa and generous research grants for two years, and the traffic still runs overwhelmingly the other way. The United States, meanwhile, is leading on the strength of people who chose it long before its current restrictions. Neither country comes out of the picture looking secure.
The evidence is a study by the Carnegie Endowment for International Peace, a think tank, led by Damien Ma and reported by Zeyi Yang of The New York Times on 6 October 2026. Every figure below belongs to that study. It found that for each 30 Chinese researchers working on AI in the United States in 2025, one person made the reverse trip into China. It also found that the share of Chinese-origin researchers in the US climbed four percentage points compared with three years earlier. Ma said the rise came despite tensions between the two governments.
The method deserves a close look. The team took the authors of papers accepted at last year’s NeurIPS, one of the field’s largest annual machine-learning research conferences, and traced where each studied as an undergraduate, earned graduate degrees, and now works. That is a strong signal about who produces frontier research. It is not a census of global AI talent, because one conference’s author list cannot capture everyone doing the work.
Within that sample, China’s institutional weight has grown fast. In 2022, six of the 30 institutions with the most NeurIPS papers were Chinese. By 2025, twelve were. Peking University and Tsinghua University ranked above Stanford, Carnegie Mellon and MIT. Chinese researchers accounted for 41 percent of published authors last year, against 34 percent for Americans.
Retention, not recruitment, explains much of that rise. The study found Chinese-trained researchers choosing to stay home more often, while the US keeps its edge by employing people who emigrated from China, India, South Korea and other countries. Kevin Xu, who founded Interconnected Capital, a hedge fund investing in AI technologies, argues that employers such as DeepSeek and Moonshot have changed the calculation. Working at DeepSeek now carries roughly the prestige of a job at OpenAI, he says, where five years ago the choice ran between Google and Tencent. Xu is an investor reading the market, not an author of the study.
Xu also points to a less flattering reason. He says the Chinese government is unwilling to let some of these people leave. The Times reports directly that this year Beijing stopped the two founders of Manus, a start-up Meta was set to buy for $2 billion, from exiting China. A further claim, that travel by Chinese AI founders, researchers and their families has been curtailed more broadly, comes from a Bloomberg report, which the Times cites rather than confirming through its own reporting.
Foreign recruits face friction of their own. Xu says anyone outside China’s cultural and digital “walled garden” finds it “still very hard to break into.” The Times adds that a newcomer also lands in a country with a small immigrant population and an unfamiliar language. Beijing has landed some returnees. Last year Tencent made Yao Shunyu, who had worked as a researcher at OpenAI, its chief AI scientist.
The study carries a second limitation that cuts against the American half of the headline. It counts researchers who moved but does not record when. Many may have arrived before the visa restrictions and research funding cuts imposed last year. Yingyi Ma, who teaches sociology at Syracuse University and studies Asian students in America, says “There is a delay effect.” The danger for the US, she says, is “a slow erosion of its appeal to attract talent that compounds over decades.” By the time it shows in the data, she adds, it is hard to reverse. That is why the numbers above may flatter the present.
Read together, the two findings describe a pair of policies that are each underperforming. China’s recruitment drive is not pulling in foreigners, and the US advantage is a lagging indicator, a stock of talent that was assembled by decisions made years ago and is not being replenished at the same rate. A repeat of this analysis on the authors of this year’s NeurIPS papers would be an early test of whether the delay has started to run out.
Reported by Zeyi Yang for The New York Times, published 6 October 2026, drawing on a Carnegie Endowment for International Peace study.
