Carnegie Report: China Overtakes the US for the First Time as the Top Destination for Leading AI Talent, but Talent Still Nets Outflow

[[**The 2025 Carnegie report shows that China has for the first time overtaken the United States to become the top destination for the world's leading AI talent, at 41% versus 34%, but China still had a net outflow of 1,729 people while the United States had a net inflow of 2,145; China's talent is almost entirely self-cultivated, whereas the United States relies heavily on overseas talent supply.**

**Key points**

**1. China tops the ranking but has a net talent outflow; the United States remains a net inflow country**

China's share of the world's top AI talent rose from 27% three years ago to 40.6%, while the U.S. share fell from 46% to 34.2%. But over the same period China had a net outflow of 1,729 people and the United States a net inflow of 2,145, a pattern of "a production powerhouse that is also a net exporter."

**2. China "produces and consumes its own," the United States "absorbs globally"**

98% of the top AI researchers working in China did their undergraduate studies in China; only about 30% of those working in the United States did their undergraduate studies in the United States. For every 30 researchers of Chinese undergraduate origin working in the United States, there is only 1 researcher of U.S. undergraduate origin working in China.

**3. Peking University tops the institutional ranking; university "student armies" are the main force**

Peking University ranks first among institutions, Tsinghua second and Shanghai Jiao Tong fourth. Frontier labs such as Google publish less publicly because of commercial confidentiality, so the list skews toward academia; Chinese companies such as DeepSeek use open-source papers as recruiting advertisements, with doctoral students as the main research force.

**4. Talent competition is measured block by block, and urban clustering effects are pronounced**

Peking University and Tsinghua University are just one street apart; South Korea's KAIST relies on the Daedeok research special zone; Singapore's two universities enter the top 20 on the strength of the city-state's advantages. Within China, Hangzhou has no top university on the list, yet it has become a major talent winner thanks to DeepSeek and Alibaba. 2026-10-13 08:15 Carnegie report: China overtakes the United States for the first time to become the top destination for leading AI talent TOP Innovation District Research Institute quick overview

This article is from the WeChat public account: TOP Innovation District Research Institute, author: Talent Research Group

At the end of last year, NeurIPS, the world's top AI academic conference, was held simultaneously in two cities for the first time.

The main venue was in San Diego, USA, with a branch venue in Mexico City with about 500 seats.

Authors whose papers were accepted but who could not get to San Diego were given priority for Mexico City.

The list of authors at this conference happens to be the entire sample of the global AI talent report released by Carnegie China on September 23.

The number of accepted NeurIPS 2025 papers from China increased by 45% compared with 2024

The report tracked the full resumes of 10,280 of these authors, and its conclusion is headline-worthy: in 2025, 41% of the world's top AI researchers work in China and 34% in the United States.

Three years ago, these figures were 27% versus 46%.

Carnegie report "Who's Ahead in the Global AI Talent Race?"

In 2025, 40.6% of top AI talent works in China and 34.2% in the United States

The "production" measure looks at where top researchers did their undergraduate studies. 57% did so in China, up 11 percentage points in three years; the United States accounts for only 13%.

The "retention" measure looks at how many of a country's own trained people stay in that country. The United States retains 89%, South Korea 77%, China rose from 57% to 69%, and Europe 60%.

The "attraction" measure reconciles undergraduate origin with current workplace, which yields the net inflow and net outflow figures at the beginning.

Cross-referencing the numbers yields a comparison that the report's main text does not state directly but that we consider most critical—

Among the top AI researchers working in China, about 98% did their undergraduate studies in China.

China is like a giant factory: output is astonishing, most products are sold domestically, and the rest are exported.

The United States is like a global exchange: its own output is limited, but the world's best goods are all willing to list here.

Factories compete on capacity and yield rates; exchanges compete on credibility and listing thresholds. Before asking "who leads," one must first ask which business is being compared.

China is growing too fast

Seeing 41% versus 34%, many people's first reaction is that the United States is bleeding talent.

The data do not support that judgment.

NeurIPS has expanded dramatically over the past three years, with accepted papers rising from 2,834 to 5,823 and authors from fewer than 10,000 to more than 25,000. The pie has more than doubled, so shares must be read differently.

Roughly converted by total authors, the number of top researchers working in the United States roughly doubled in three years, while the number working in China quadrupled.

The U.S. team is still expanding; it was overtaken by an opponent growing twice as fast.

The report's second conclusion also contains a detail: among these researchers working in the United States, the share who did their undergraduate studies in China rose another 4 percentage points.

Flows between the two countries also remain highly one-directional.

For every 30 researchers of Chinese undergraduate origin working in the United States, there is only 1 person of U.S. undergraduate origin working in China. Three years ago it was 46 to 1.

In 2022 the top institution was Google; in 2025 it became Peking University, with Tsinghua second, South Korea's KAIST jumping 12 places to third, and Shanghai Jiao Tong fourth.

In April 2025, the Financial Times, citing several current and former DeepMind scientists, reported that the company had imposed a six-month publication ban on "strategic" papers in generative AI and that internal approval had become cumbersome. One serving researcher said that today he finds it hard to imagine the company still making the original Transformer paper public.

DeepMind imposes a six-month publication ban on strategic papers in generative AI.

Frontier labs like OpenAI and Anthropic already publish little in public.

So strictly speaking, this ranking measures "people who publish papers publicly at top conferences." Many of the most frontier people in the U.S., the ones most involved in trade secrets, are invisible in the statistics.

The report itself admits that

the sample is somewhat skewed toward academia and underestimates industry.

In addition, employees of multinational companies may be counted under the company's home country; researchers at Microsoft Research Asia in Beijing are very likely counted on the U.S. side.

Conversely, China's large-model companies are precisely the most willing to open-source and publish papers. Openness is both a technical route and a recruiting ad. A paper circulated among peers worldwide is a researcher's strongest résumé.

Research statistics show that open outputs significantly boost research visibility and citation counts.

Most first authors of top-conference papers are doctoral students still in school. Peking University, Tsinghua, and Shanghai Jiao Tong topping the list is backed by a huge "student army" of doctoral students doing frontier research.

DeepSeek is the most typical product of this assembly line. In a 2024 interview with "Anyong," Liang Wenfeng said that the V2 model team had no one who had returned from overseas; all were domestically trained. He acknowledged that the top 50 talents might not be in China, "but perhaps we can build such people ourselves." The MLA architecture that the industry later repeatedly dissected originated from the personal interest of a young researcher on the team.

The upstream end of the assembly line is still expanding capacity.

In 2018, the first batch of 35 universities was approved to offer undergraduate AI majors; from 2020 to 2024, more than 400 additional program sites were added nationwide. By timing, most of these students have not yet reached the stage of being top-conference authors. More program sites does not necessarily mean more top talent, but the factory's capacity curve is still climbing.

So a considerable portion of that 41% figure is "in-transit inventory" that has not yet left campus.

When this cohort of doctoral students graduates in a few years, whether they become professors and start companies in Haidian, Shanghai, and Hangzhou, or board flights to San Francisco, is the real test (or check).

More of China's people staying

The retention rate rose from 57% to 69%, and the report gives two reasons.

First, China's AI industry has boomed, so there are more good jobs. Second, the U.S. has tightened visas for Chinese STEM graduate students, making it harder to go abroad.

The number of AI researchers and developers flowing into the U.S. is 89% lower than in 2017, and fell 80% in the past year alone.

Everyone can see what happened in the past year.

In September 2025, the White House announced that newly filed H-1B visa applications would require an additional $100,000 payment. In June this year, a federal district court vacated the policy implementing this fee; in July, an appeals court denied the government's request for a stay, and the money has still not been collected. On September 18, the White House extended the policy to September 2027. And in August, the Department of Homeland Security separately proposed a new fee of $103,000 for lottery-based H-1B applications.

At this moment, this money cannot legally be collected.

But a 22-year-old student deciding where to do a PhD does the math differently. He is betting on the policy direction of the next five to seven years. Issued today, withdrawn tomorrow, extended the day after—this back-and-forth is itself a cost, more damaging than any definite fee.

You may not believe it, but even NeurIPS is paying for this cost. Those 500 seats in Mexico City were added precisely for this uncertainty.

NeurIPS explicitly stated that setting up satellite venues was to address the surge in visa difficulties in recent years.

If U.S. policy turns one day, should China continue to export a portion?

China launched the K visa for young scientific and technological talent starting in October 2025, but the 30-to-1 flow ratio shows that this factory has so far had almost no "import" business. It is good at making people itself; attracting people from all over the world is another matter.

Talent clusters by neighborhood

Beyond China and the U.S., this report also has a map worth a closer look by cities and industrial parks.

Global No. 1 and No. 2, Peking University and Tsinghua, are separated only by Zhongguancun North Street, a 15-minute walk apart.

No. 3, KAIST, is in the Daedeok Research and Development Special Zone in Daejeon, South Korea, a science city planned by national will. South Korea also ranks second globally with a 77% retention rate.

The National University of Singapore rose from 25th to 13th, and Nanyang Technological University rose from 29th to 16th. A city-state squeezed into the global top 20 with two universities. Companies like Manus that moved their headquarters to Singapore make it look more and more like a buffer zone between China and the U.S.

At the other end, ETH Zurich and Oxford both fell off the list, and many people trained in Europe go to work elsewhere. India is even more extreme: its share of top talent fell another 2 percentage points, and in neither of the two counts did a single institution make the list.

This may be the most intelligent 3 square kilometers in all of China...

The domestic landscape is the same.

Of the three Chinese universities in the global top four, two are in Haidian, Beijing, and one is in Shanghai. The competition among cities for AI talent begins in the first round on the college application form. The second round happens at the moment of doctoral graduation. Hangzhou has no school in the top four, yet because of DeepSeek and Alibaba, it has become one of the biggest winners of this round.

Back to those 500 seats in Mexico City.

A conference can be held in two cities at once. A researcher's career can ultimately land in only one city.

For decades, the United States won this competition by making the world's smartest young people willing to settle there. Today its ranks are still expanding, but it increasingly relies on a door that opens and closes.

China is taking another path: train its own, retain its own, and try to put the top 50 into its own assembly line as well.

Neither path has been completed.

Next year's NeurIPS author list will tell us whether those 500 seats in Mexico City were a temporary addition or will keep growing. AI Venture Capital Daily Channel: Frontier Technology. This content comes from the internet and the original link; the views represent only the author and do not represent Huxiu's position. For copyright issues, please contact hezuo@huxiu.com, and we will verify and handle it promptly. Everyone who is changing and wants to change the world is on the Huxiu APP[ ]