China Now Employs the Most Elite AI Researchers
For two decades the geography of elite artificial intelligence research was easy to describe. The most ambitious graduate students in the field, wherever they were born, tended to end up in the United States. They trained at a short list of American universities, published at NeurIPS and ICML, and then took jobs at Google, Meta, OpenAI, or a handful of West Coast labs. China produced a large and growing share of the raw talent. America captured it.
That map has flipped.
A study released this week by Carnegie China, titled “Who’s Ahead in the Global AI Talent Race,” finds that China employed about 41 percent of the world’s leading AI researchers last year, ahead of the United States at 34 percent. Three years earlier the United States still held a commanding lead: 46 percent against China’s 27 percent. The South China Morning Post, reporting the finding, put the change in a single sentence that would have sounded implausible in 2019: for China’s elite AI researchers, Silicon Valley no longer holds the unrivalled appeal it once did.
The more precise figures, circulated first by The Information and confirmed across secondary coverage, are 40.6 percent working in China in 2025 versus 34.2 percent in the United States, compared with 27.1 percent and 46.4 percent in 2022. The swing is not a rounding error. It is a structural reversal in where the people who write the field’s best papers actually sit.
That is the headline. The rest of the story is more interesting than the headline, because the same report shows that the United States is still a magnet, still a net importer of talent, and still the place where a large majority of already-arrived Chinese researchers choose to stay. The race has not been decided. The pipeline has been rerouted.
What the study is actually counting
Carnegie China’s update is the successor to the Global AI Talent Tracker built at MacroPolo, the now-shuttered Paulson Institute project that first made this method famous. The authors associated with the work include Matt Sheehan, a Carnegie fellow who worked on the original tracker, and research assistant Sophie Zhuang. The method is simple enough to describe and easy to misuse if you do not keep two distinctions in mind.
First: “elite AI researcher” here means an author on papers accepted at the field’s top conferences, principally NeurIPS, and in some versions of the tracker also ICML and ICLR. One detailed summary says the 2025 analysis covered 10,280 researchers, roughly 40 percent of authors on accepted NeurIPS 2025 papers. Conference acceptance is a noisy proxy. It privileges people who publish. It undercounts researchers locked inside closed frontier labs that do not send papers out. It may over-represent Chinese academics, who face stronger career incentives to place papers at prestigious venues. Those caveats matter, and they cut in more than one direction.
Second: the tracker separates origin from workplace. Origin is usually the country of the researcher’s undergraduate institution, a rough proxy for where they were formed. Workplace is current affiliation. A Tsinghua bachelor who does a PhD at Stanford and joins Anthropic is Chinese-origin and US-based. A Peking University graduate who never leaves Beijing is Chinese-origin and China-based. Confusing those two numbers is how this debate produces slogans instead of analysis.
On origin, China is not newly dominant. It has been the largest producer of this talent for years. The new tracker puts the undergraduate-origin share at about 57 percent in 2025, up 11 percentage points from the previous cycle. Only about 13 percent of the elite sample earned their first degree in the United States. Jensen Huang’s oft-repeated line that China produces half the world’s AI researchers was already close to true; the new figure is closer to three-fifths.
On workplace, the change is new. China’s share of current affiliations rose from just over a quarter to just over two-fifths in three years. The United States fell from nearly half to just over a third. That is the finding SCMP led with, and it is the one that should rearrange how governments and labs think about the next decade.
Retention, not a mass exodus from America
The mechanism is not a sudden flight of established stars from Menlo Park to Zhongguancun. It is a change in the marginal decision of younger researchers: stay, or leave.
Among researchers who completed their undergraduate degrees in China, 69 percent were working in China in 2025, up from 57 percent in 2022. The Economist, using a related NeurIPS sample earlier in 2026, found a similar climb: roughly a third of China-educated NeurIPS authors stayed in 2019, 58 percent in 2022, and about 68 percent in 2025. The return rate among those who did go abroad for graduate school also rose, from 12 percent in 2019 to 28 percent in 2025. Fewer people leave after the bachelor’s degree. More of those who leave come back.
The stock of people already in the United States looks different. Sheehan and Zhuang traced 100 Chinese-origin researchers who were at US institutions in the 2019 NeurIPS cohort. By 2025, 87 were still in the United States: 41 at private companies, 40 at universities, and six in doctoral or postdoctoral posts. Ten had moved to Chinese companies or universities. Three had gone elsewhere. Sheehan’s earlier public comment on that cohort still holds: long-standing institutions do not disappear overnight, and the American AI industry remains the largest beneficiary of Chinese-trained talent that already made the crossing.
The ten who left were not anonymous. Coverage of the report notes two startup founders, two tech executives, and five professors. One of them is Yang Zhilin, the Carnegie Mellon PhD who founded Moonshot AI in 2023, raised more than a billion dollars, and released open-source models that US startups now use. A handful of high-leverage returnees can matter more than the raw headcount suggests. But the 87-to-10 split is the number that should discipline anyone claiming an emptying of American labs.
The net-flow figures make the same point from another angle. Inside the 2025 conference sample, the United States had a net gain of 2,145 researchers. China had a net loss of 1,729. About half of those leaving China ended up in America. The United States is still the world’s largest importer of this talent. China is still a net exporter. What changed is the rate of export, and therefore the workplace share, as China’s own production of elite researchers exploded and a much larger fraction of that production stayed home.
There is even a wrinkle that cuts against a simple “America is losing Chinese talent” story. Sina’s write-up of the Carnegie numbers notes that the share of China-undergraduate researchers currently working in the United States rose by four percentage points. That can happen at the same time as China’s workplace share rises if the global pool is growing fast enough, and if the United States continues to skim a slice of a much larger pie. Volume and share are not the same variable.
Why home started to look better
SCMP’s frame is the right one: a booming domestic industry plus tighter US visa conditions. Both are real. Neither is sufficient on its own.
On the pull side, China’s commercial AI sector is no longer a provincial afterthought. ByteDance and Alibaba now appear among the tracker’s top institutions alongside Peking, Tsinghua, and Shanghai Jiao Tong. Of the top 30 institutions in the latest Carnegie cut, twelve are Chinese, double the count from three years earlier; three of the top five are Chinese universities. Compensation in the leading Chinese labs has moved much closer to Silicon Valley packages than it was in the mid-2010s, especially once housing support, compute access, and the option to work on products with a billion-user domestic market are included. The Qiming-style recruitment packages that The Economist described earlier this year — high salaries, grants, housing — are the visible edge of a much larger industrial policy.
DeepSeek is the emblem, and not only because of the models. A Hoover Institution and Stanford HAI reconstruction of the authors behind DeepSeek’s core papers found a research culture that can now produce frontier work with a workforce trained almost entirely inside China. The Economist noted that none of the core contributors to DeepSeek R1 held degrees from outside the country. That is a different fact from “China has more NeurIPS authors.” It says the domestic pipeline can close the loop from undergraduate training to systems that force American labs to respond.
On the push side, US visa policy, campus politics, and export-control spillover have raised the expected cost of the traditional path. Chinese STEM graduates who want to do a US PhD or take a US industry job face longer uncertainty, more scrutiny, and a non-trivial chance that a conference, a lab visit, or a job offer simply does not materialise. NeurIPS 2025 itself was split between San Diego and a Mexico City backup partly so researchers who could not get US visas could still present. Some American universities have rescinded offers. Some Chinese researchers now treat a US affiliation as a career risk rather than a prize. You do not need a formal ban to change a twenty-two-year-old’s decision. You only need to change the odds.
Family and language are underrated in Washington discussions of this problem. A researcher who can do serious work in Beijing, Shenzhen, or Hangzhou, see parents on weekends, and avoid a multi-year immigration lottery will need a very large professional premium to leave. As that premium shrinks, the default flips.
The American system is still built on imported talent
The uncomfortable implication for US policy is that American leadership in AI has never been a story about American undergraduates. Among authors affiliated with US institutions in The Economist’s 2025 NeurIPS sample, roughly 35 percent had a Chinese undergraduate degree — about the same share as had an American one. When Meta staffed its superintelligence effort, reporting at the time described about half the named researchers as being from China. An analysis of 483 contributors to OpenAI’s GPT-5 found that 15 percent had at least one degree from a Chinese institution. The US AI stack is a global labour market with an American headquarters.
That is a strength until the inflow slows. It is also a political vulnerability. The same years in which American labs became more dependent on Chinese-trained researchers were the years in which American politics became more hostile to that dependence. The Carnegie numbers do not show that hostility emptying the labs of people already inside. They show it changing who never comes.
Sheehan put the point bluntly in earlier work: the US AI industry is the biggest beneficiary of Chinese talent. Studies of top papers have shown authors originally from China contributing as much as, if not more than, authors from the United States to American output. If the policy goal is American capability, the rational response is to keep that inflow open and to grow the domestic pipeline. If the policy goal is decoupling for its own sake, the Carnegie workplace reversal is what success looks like — for Beijing.
How much should we trust a conference?
Every talent tracker in this field has the same original sin. NeurIPS, ICML, and ICLR are excellent at identifying people who do publishable machine-learning research. They are worse at identifying the people who matter most at the frontier in 2026.
American capability is increasingly concentrated in labs that treat publication as optional. OpenAI, Anthropic, and the most secretive teams at Google DeepMind and xAI do not generate author lines in proportion to their influence. A tracker that never sees those researchers will understate the US workplace share and overstate the academic share. Chinese labs, by contrast, still use top-conference papers as currency for promotions, funding, and prestige, and China’s open-source model culture pushes more work into the public record. The measurement system is not neutral.
There are other distortions. First-author conventions differ. Multi-affiliation papers can be coded in ways that inflate or shrink a country’s count. Undergraduate institution is a crude origin proxy for people who moved as children or trained in several countries. The 10,280-researcher sample is large, but it is not a census of “the world’s leading AI researchers” in any metaphysical sense. It is a census of a visible, prestigious, academic-adjacent slice.
Those limits do not make the trend fake. Independent work points the same way. The Economist’s own 2025 NeurIPS sample found more lead authors based in China than in America or Europe, nine of the top ten undergraduate institutions in China, and Tsinghua alone accounting for about 4 percent of authors against MIT’s 1 percent. Digital Science’s Dimensions analysis found China producing more AI papers in 2024 than the United States, the United Kingdom, and the European Union combined, and filing AI patents at a scale that dwarfs the rest of the leading pack. China also has a younger researcher cohort: nearly half students, against about 30 percent in the West. Pipeline plus youth plus retention is how workplace share moves this fast.
The honest statement is narrower than the headline and stronger than the sceptic’s dismissal. Among researchers visible in the top conference literature, China has become the largest workplace, because it now keeps most of the people it trains. Whether that literature is the right scoreboard for frontier capability is a separate argument. It is not an argument that the scoreboard is broken.
What the numbers do not settle
They do not settle who will build the most capable models. Talent is an input. Compute, data, capital, chip supply, energy, and the willingness to deploy at scale are others. The United States still leads on several of those. Export controls still bite. The most valuable closed models are still, as of this writing, associated with American labs. A country can lead on NeurIPS affiliations and lag on the systems that matter to militaries and markets.
They do not settle quality. Conference volume can rise because a system got better at producing papers, or because it got better at producing science. China’s citation impact, independence from foreign co-authors, and ability to train models under chip constraints have all improved. That is not the same as claiming every incremental Chinese affiliation is a DeepSeek-level researcher. Averages hide tails. The tail is what labs hire for.
They do not settle permanence. Labour markets in a hot field move. A US immigration thaw, a Chinese growth scare, a crash in domestic AI salaries, or a new generation of closed American labs that suddenly start publishing could all rewrite the 2028 snapshot. The Economist’s extrapolation — that China-based elite researchers could outnumber US-based ones two to one by 2028 if the last decade’s trend continues — is a warning, not a forecast. Trends this steep rarely run in a straight line.
They also do not settle the ethics of the contest. A world in which talented people can choose where to work is better than a world of exit bans and loyalty tests. Both governments have been willing to treat researchers as strategic assets. China’s recruitment drive and travel controls sit on one side of that line. America’s visa suspicion and entity-list spillover sit on the other. The Carnegie tracker is a map of those choices as they accumulate in CVs.
The policy fork
If you sit in Washington, the report is an argument for treating high-skill immigration as industrial policy rather than as a security afterthought. The United States still wins the people who arrive. It is losing the people who never book the flight. Expanding domestic STEM production would help, but it will not replace a 57 percent origin share in three years. The complementary move is obvious and politically hard: make the US the easiest rich-world place for a Tsinghua graduate to build a career, not the most uncertain.
If you sit in Beijing, the report is validation of a bet placed a decade ago: train at home, pay at home, give people compute and products at home, and the workplace share will follow the origin share. The remaining weakness is the net outflow. China is winning retention and still losing people at the margin. That is why the returnee stories — Moonshot, the professors, the executives — get so much attention. They are scarce, and they transfer tacit knowledge that papers do not.
If you sit in London, Singapore, Toronto, or Paris, the report is a reminder that the talent market is no longer a two-node graph. Singapore already shows up as a secondary gainer in the net-flow charts, partly as a third place for researchers who want distance from both Washington’s politics and Beijing’s. Europe’s problem is not that China overtook America. It is that neither giant is leaving much of the surplus for everyone else.
And if you sit in a lab, the operational lesson is simpler. Hiring as if the world’s best twenty-six-year-olds will keep appearing in California by default is now a strategy with a dated map. Some of them will. Most of the new ones will not have to.
The deeper shift
The SCMP story is easy to read as another round in a tiresome scoreboard war. Resist that reading. The interesting change is not that one country’s percentage ticked above another’s. It is that the old bargain of the AI age — China trains, America employs — is no longer the governing rule.
That bargain created the industry as we know it. American labs are full of people whose first serious linear algebra happened in a Chinese classroom. Chinese labs are now full of people who never needed the second half of that sentence. The Carnegie tracker is the first widely cited dataset in which the second population outnumbers the first.
Whether that produces a bipolar research world, a fragmented one, or a Chinese-centred one depends on decisions that have not been taken yet: visas, funding, publication norms, chip access, and whether the closed American labs remain closed. The study cannot answer those questions. It can only tell you that the people who will answer them are, increasingly, not working where the last generation assumed they would.
Silicon Valley did not stop being a great place to do AI. It stopped being the only one that the best people from the largest talent pool felt they had to reach. That is a smaller sentence than “China has won.” It is also the one the data supports, and it is large enough.
Sources for this piece include the South China Morning Post report by Chong Ming Lee (25 September 2026); Carnegie China’s “Who’s Ahead in the Global AI Talent Race,” as detailed in coverage by The Information, AI Weekly, Sina Finance, and Interconnected; earlier Carnegie and MacroPolo cohort tracking by Matt Sheehan and Sophie Zhuang; and The Economist’s March 2026 NeurIPS analysis.
