China’s AI research is booming. Where is it headed?

As his graduation from a top Chinese university approached earlier this year, computer scientist Lu Yi faced a tough choice. Should he move to the United States to pursue a PhD in artificial intelligence, as Chinese high-fliers have so often done before? Or should he stay in China?

The United States pays AI scientists higher salaries, has the best hardware and hosts the world’s most advanced AI laboratories. But “the gap is getting closer and closer”, says Lu, whose name has been changed because he was concerned about speaking publicly on what he saw as a sensitive issue.

China’s extraordinary rise in science and how it will change the world

AI research in China is advancing rapidly. Many more AI researchers emerge from Chinese universities than from US institutions, and graduates are increasingly opting to stay put, or to return to China after spending time in the United States. China-based researchers dominate global AI-related publications, and firms such as DeepSeek and Moonshot AI make AI models with capabilities not far off those of their US rivals — and at a fraction of the price.

At the same time, the United States has become a less appealing destination for foreign researchers, with an increasingly restrictive visa system; Lu has seen friends’ applications rejected or delayed. “It’s a worry,” he says. Ultimately, Lu decided to stay in China.

The AI ecosystem in China differs in several ways from that of California’s Silicon Valley. In China, the line between academia and industry is more porous, and access to computing resources is much lower. This presents both opportunities and challenges. Chinese firms have been accused of copying US advances in their race to catch up, but they also have found techniques that make models more efficient.

“There are some that definitely think that the US is still the best place in the world to build a life and do this research, but more are questioning it,” says Scott Singer, co-director of the China AI Initiative at the Carnegie Endowment for International Peace in Washington DC.

Where AI talent grows — and goes

China trains more AI talent than any other country. It has a pool of 51,000 active AI researchers compared with 14,000 in the United States, according to data from Digital Science (a firm operated by the Holtzbrinck Publishing Group, which has a share in Nature’s publisher, Springer Nature). The data, which are updated from a study the firm released last year¹, also show that many of the researchers are at an early stage in their careers, with fewer older superstars than in the United States (see ‘Active AI researchers’).

As a proxy for studying elite researchers, many analysts have examined the origins and locations of authors at the prestigious NeurIPS conference. A study released in September by Carnegie China, a Singapore-based research centre under the umbrella of the Carnegie Endowment, reported that, on the basis of a large sample, 57% of authors at the 2025 conference had obtained their undergraduate degrees in China — up from 46% three years earlier — but that the United States was still a magnet for those studying at graduate level, pulling away around one-fifth of those undergraduates (see go.nature.com/4yfsl4v). The attraction of the United States seems to be weakening, however; Chinese scientists are increasingly staying at home, the study found. Of those researchers publishing at NeurIPS who had received undergraduate degrees in China, 69% were still there, up from 57% in 2022 (see ‘AI talent career paths’). This ability to keep top home-grown AI talent is higher than Europe’s (60%), but trails that of the United States (89%).

A related statistic is that although across all levels 63,000 Chinese students studied mathematics or computer science in the United States in 2024–25, this was 20% fewer than five years previously. One reason for this drop is that the effects of the US–China geopolitical rivalry are trickling down to research, says Grace Shao, an independent analyst covering AI and technology, based in Hong Kong. A lot of Chinese people now “feel less comfortable or a feeling of animosity amongst peers”, in the United States, she says. Every Chinese student who spoke to Nature cited visa troubles as a primary concern of going to the United States. Increasingly, visas are delayed, and the US administration has sought to limit their duration, although this is being challenged in court. Because new visa rules could affect re-entry to the United States for even existing students, some US-based Chinese nationals are too anxious to visit their home country.

Direction of travel

The data are not as clear on whether AI researchers already in the United States are flooding back to China. Media coverage tends to focus on high-profile returnees, such as Moonshot’s founder, Yang Zhilin, who created the Beijing-based company in 2023, having finished a PhD at Carnegie Mellon University in Pittsburgh, Pennsylvania, in 2019. But the true picture is more nuanced. A study by the Carnegie Endowment suggests that the United States is still able to retain high-profile AI researchers (see go.nature.com/3vxhdic). Of its sample of authors publishing at NeurIPS, 87% of researchers of Chinese origin — those who did their undergraduate degree in China — who had been in the United States in 2019 remained there in 2025. But the Digital Science study suggests that when looking at the changing affiliations of all AI researchers actively publishing, the net direction of flow is the US losing talent to China.

Does China’s latest AI model finally equal US rivals? What scientists think

Analyses of AI-related academic papers show China ahead of the United States in terms of sheer numbers and, increasingly, influence. The Digital Science study found that China dominates academic output — by citation as well as volume — and AI-related patents. Key universities and AI firms are spread across China, with hubs in Beijing, Shanghai, Hangzhou and Shenzhen, and with 156 research organizations each publishing more than 50 AI papers a year, the study noted. Research quality is improving, too: in 2025, for the first time, the NeurIPS conference had more lead authors based in China than in the United States or Europe, according to an analysis by The Economist (see go.nature.com/4i1ka6e).

But other statistics tell a different story. Overall, the United States produces higher-impact patents and more notable AI models than China does, according to the Stanford AI Index 2026². And publication data can offer a skewed portrait of AI advances, because Chinese firms are more likely to publish scientific papers than are US start-ups, according to a preprint published in July³. Chinese academics also report high pressure to publish papers to advance their careers.

Fundamentally, the United States still makes the world’s most advanced AI. Models from OpenAI and Anthropic, both based in San Francisco, California, top the leader boards for system capabilities. “China is probably somewhere between three and seven months behind the frontier in US models,” says Singer.

‘Robot Olympics’ reveal humanoids’ rapid progress — and flaws

One reason for the lag is that China has many fewer of the computing chips that are required to train and run cutting-edge models. Since 2022, the United States has restricted export of powerful chips known as graphics processing units (GPUs) to China, owing to national-security concerns. Chinese firms are racing to manufacture their own high-calibre chips, but none yet rivals those used by US firms.

Cash is also generally less abundant in China, with private AI investment in the United States in 2025 reaching more than 23 times that invested in China, according to the Stanford report. Chinese firms’ lack of these resources is their main bottleneck, says Shao.

Efficiency is everything

Lower access to chips and cash affects research strategies in several ways. A lack of computing resources, or ‘compute’, means, for instance, that Chinese companies are not devoting as high a volume of resources as US firms are to a process known as recursive self-improvement, says Lu. This seeks to use AI to create better models but does not guarantee success and sucks up GPU time. Shao, similarly, says that US firms have the freedom and budget to pursue more speculative research directions.

But compute constraints push researchers in China to “come up with more compute-efficient algorithms and training methods”, says Zilan Qian, who studies China’s AI development at the non-profit organization GovAI in London. For example, DeepSeek, based in Hangzhou, created its own ‘sparse attention’ mechanism, which reduces overall computation when training or using models by deciding which tokens matter most, and running calculations on the basis of those alone.

One much-debated question is to what extent Chinese companies are catching up by using a process called distillation on US models. Distillation involves either training a model on outputs from a ‘teacher’ AI, to capture a similar level of ability, or, in a more sophisticated approach that avoids direct emulation, using the teacher AI to grade and improve a smaller model’s responses. In September, the US government said that distillation campaigns formed “the core” of development by Chinese AI firms. “In the industry, the understanding is that there has definitely been distillation,” says Shao, although no one has openly acknowledged it. But in August, Zhang Yiming, founder of the Beijing-based tech firm ByteDance, which produces the Doubao chatbot, publicly rebuked the practice, telling staff to avoid distillation in favour of “long-term goals”.