China’s AI workforce is growing faster than the talent drain to the United States
As graduation season approached at a leading Chinese university, computer scientist Lu Yi had to choose between two very different research futures. A doctorate in artificial intelligence in the United States still carries prestige, higher salaries and access to the most advanced hardware. Staying in China, however, increasingly looks like a serious option rather than a compromise.
Lu chose to remain in China. His concern about speaking publicly reflects a wider sensitivity around the geopolitical stakes of AI, but his decision also captures the central shift now reshaping the field: the United States still holds the frontier, yet China is closing distance quickly and pulling more researchers into its own orbit.
The scale of that change is easiest to see in personnel. China now trains more AI talent than any other country and has a pool of 51,000 active AI researchers, compared with 14,000 in the United States, based on Digital Science data. Many of those researchers are early in their careers, which matters because a young workforce can expand fast, but it also means China still has fewer older, globally recognized figures than the US ecosystem.
One sentence captures the pressure beneath those numbers: China is producing more AI researchers than it loses.
Publication volume is rising, but access to compute still shapes the ceiling
Academic output places China ahead on both volume and, increasingly, influence. Digital Science found that China dominates AI-related publication counts, citation impact and patent activity, with major hubs clustered in Beijing, Shanghai, Hangzhou and Shenzhen. More than 150 research organizations in China now publish at least 50 AI papers annually, a sign of how broad the national effort has become.
Research quality is rising too. In 2025, NeurIPS — one of the field’s most influential conferences — had more lead authors based in China than in the United States or Europe, according to an analysis by The Economist. That kind of result suggests the country’s academic base is no longer limited to imitation or scale alone; it is beginning to generate work that can compete at the top tier of the discipline.
Still, publication data can overstate genuine technological leadership. Chinese firms are more likely to publish scientific papers than US start-ups, and Chinese academics face strong incentives to publish for promotion. That makes paper counts an incomplete proxy for frontier capability, especially in a field where many advances never appear in journals or conference proceedings at all.
What remains harder to copy is compute. China has far fewer of the graphics processing units needed to train and run leading models, and US export controls since 2022 have restricted access to the most powerful chips on national-security grounds. In practice, that constraint does not just slow training runs; it shapes model design, forcing Chinese groups to squeeze more performance out of less hardware.
The US still leads the frontier, but the gap is no longer the one researchers assumed
The most advanced AI models remain American. OpenAI and Anthropic, both based in San Francisco, continue to top capability leaderboards, and Scott Singer of the Carnegie Endowment for International Peace estimates that China is roughly three to seven months behind the US frontier. That is a meaningful gap in a field where a few months can separate one generation of model performance from the next.
Yet a different metric tells a more unsettling story for the United States: talent retention is shifting. A Carnegie China study of NeurIPS authors found that 57% of 2025 authors had completed undergraduate degrees in China, up from 46% three years earlier. Among researchers who studied in China at the undergraduate level, 69% were still based there in 2025, compared with 57% in 2022. The US still holds on to most Chinese-origin researchers who enter its system, but the broader direction of travel is changing.
The numbers matter because AI progress depends on concentration. Strong models emerge where talent, compute and commercial deployment meet. China is building the first two at scale, while the US still dominates the third and retains a hardware advantage. The result is not a simple race with a visible finish line; it is a layered competition in which each side leads in different parts of the stack.
Visa barriers and geopolitics are changing individual career decisions
Several Chinese students told Nature that visa uncertainty was the main obstacle to studying in the United States. Delays, rejections and efforts by the US administration to limit visa duration have made the process less predictable. Even current students worry that travel back to China could complicate re-entry. For some, the anxiety is enough to tilt the decision toward staying home from the start.
Those concerns are filtering into the career pipeline. China still sends large numbers of students abroad — 63,000 Chinese students studied mathematics or computer science in the United States in 2024–25 — but that total was 20% lower than five years earlier. Grace Shao, an independent analyst based in Hong Kong, linked the change to a widening sense of geopolitical friction, including the possibility of discomfort or hostility in US academic settings.
This is where the talent story becomes more than a head count. Research ecosystems rely on trust, mobility and repetition: the ability to study, return, collaborate and come back again. When visa systems become less stable, the effect is cumulative. Fewer students leave. Fewer researchers return. Fewer partnerships form across borders.
China’s AI system is becoming more self-sustaining, even without full parity
China’s AI ecosystem looks different from Silicon Valley’s in ways that could matter for years. The boundary between academia and industry is more porous, which can speed the movement of ideas from lab to product. Companies such as DeepSeek and Moonshot AI have developed models with capabilities not far from US rivals, and they have done so at a fraction of the cost. That efficiency is not a trivial detail; it is one of the few ways to compensate for a hardware deficit.
At the same time, accusations of copying US advances continue to shadow parts of China’s AI rise. Yet the broader pattern is more complicated than imitation. Chinese groups have also found ways to make models more efficient, and their academic base is large enough to keep generating new approaches. The country’s strength now lies not only in scale, but in the density of its institutions and the increasing tendency of its own graduates to stay.
The result is a field in which the US still leads at the top, China leads in volume and talent production, and the balance between them is becoming less stable every year. The unanswered question is not whether China will matter in AI research; it already does. The more consequential issue is how much of the global frontier can be shaped by a country that is still operating under tight compute constraints, even as its universities and companies keep filling the pipeline with researchers.